UV Hyperspectral Imaging

Development of a UV Hyperspectral Imaging
Prototype for Industrial Applications
Dissertation
der Mathematisch-Naturwissenschaftlichen Fakultät
der Eberhard Karls Universität Tübingen
zur Erlangung des Grades eines
Doktors der Naturwissenschaften
(Dr. rer. nat.)
vorgelegt von
M.Sc. Mohammad Mahmoud Ahmad Al Ktash
aus Al Ramtha/Jordanien
Tübingen
2023
Gedruckt mit Genehmigung der Mathematisch-Naturwissenschaftlichen Fakultät der
Eberhard Karls Universität Tübingen.
Tag der mündlichen Qualifikation: 16.05.2024

Dekan: Prof. Dr. Thilo Stehle
1. Berichterstatter/-in: Prof. Dr. Marc Brecht
2. Berichterstatter/-in: Prof. Dr. Alfred Meixner
Kurzfassung
Baumwollfasern sind aufgrund ihrer Weichheit, Haltbarkeit und Saugfähigkeit für die Textilin-
dustrie unverzichtbar. Allerdings wirken sich Verunreinigungen negativ auf die Baumwollquali-
tät aus und es ist notwendig die Stärke der Kontamination zu bestimmen. Eine häufig auftretende
Verunreinigung ist Honigtau, welcher zu klebriger Baumwolle führt, und erhebliche Probleme
für die Textilindustrie verursacht. Verschiedene Methoden im sichtbaren (Vis) und nahen infra-
roten (NIR) Spektralbereich werden für Qualitätskontrollen und Sortierverfahren eingesetzt, wäh-
rend der ultraviolette (UV) Bereich bisher kaum genutzt wird.
In den letzten Jahren haben hyperspektrale Bildgebungssysteme aufgrund ihrer Multimodalität,

ihre räumliche Auflösung und ihrer Fähigkeit zur quantitativen Analyse im Vergleich zu her-
kömmlichen Verfahren zunehmend an Aufmerksamkeit gewonnen. Diese Vorteile haben sie für
verschiedene Anwendungen, z. B. in der Textilindustrie, sehr attraktiv gemacht.
Das Hauptaugenmerk der vorliegenden Arbeit liegt auf der Erkennung von Honigtaukontamina-
tionen und der Entwicklung eines hyperspektralen Bildgebungssystems im UV-Bereich. Der Auf-
bau basiert auf einem Spektrographen, der mit einer CCD-Kamera verbunden ist. Die Proben
werden auf ein Förderband gelegt, welches die Probe unter der hyperspektralen Kamera bewegt.
Dieses Verfahren wird als Pushbroom Imaging bezeichnet. Je nach Anwendung wurden Xenon-
oder Deuteriumlampen zur Beleuchtung verwendet, wobei Deuteriumlampen eine höhere Be-
leuchtungsstärke im UV-C-Bereich im Vergleich zur Xenon-Bogenlampe bieten. Zur Validierung
dieser neuartigen Bildgebungseinrichtung wurde eine Reihe von bekannten Substanzen wie ak-
tive pharmazeutische Wirkstoffe (APIs) und Schmerzmittel verwendet. Diese Proben waren Ibu-
profen, Acetylsalicylsäure und Paracetamol. Die Ergebnisse wurden mit lokalaufgenommenen
Einzelspektren verglichen und mittels multivariater Datenanalyse ausgewertet. Es wurde gezeigt,
dass die hyperspektrale Bildgebung im UV-Bereich zuverlässige Ergebnisse erzielt und eine ana-
lytische Methode wurde entwickelt, um kommerzielle Schmerzmitteltabletten mit dem neuen
Prototyp zu identifizieren. Anschließend wurde eine separate Probenreihe, einschließlich direct
bonded copper (DBC) Substrate, für eine sekundäre Bewertung getestet. Der entwickelte Prototyp
ist in der Lage wenige Nanometer dicke Oxidschichten, zu erkennen. Dabei können verschiedene
Oxidationszustände unterschieden werden, sogar nachdem die Proben vorgesehene Reinigungs-
verfahren durchlaufen haben. Im nächsten Schritt wurden Baumwollproben aus verschiedenen
Ländern verwendet und mittels Vis/NIR hyperspektraler Bildgebung untersucht. Die gewonnenen
I
Kurzfassung
Daten wurden mit lokal aufgenommen Einzelspektren verglichen und mittels multivariater Da-
tenanalyse analysiert. Die Ergebnisse zeigen, dass es möglich ist, anhand einiger ausgewählter
Wellenlängenbereiche zwischen verschiedenen Baumwollsorten zu unterscheiden. In einem letz-
ten Schritt wurde die Quantifizierung von Honigtaukontaminationen auf Baumwolle durchge-
führt. Hierfür wurde ein Verfahren zur Kalibrierung des Prototyps für hyperspektrales Imaging
im UV-Bereich entwickelt und mit realen Proben getestet. Mechanisch gereinigte Baumwollpro-
ben wurden in eine Lösung getaucht, die Zucker und Eiweiß in bekannten Konzentrationen ent-
hielt, um mit Honigtau verunreinigte Baumwolle zu imitieren. Diese Proben wurden nach 44
Stunden und nach einem Monat untersucht. Anhand dieser Proben wurde der Prototyp erweitert
und optimiert. Die gewonnenen Daten wurden chemometrisch analysiert, um ortsaufgelöst die
Honigtaumengen in Baumwollproben mit unterschiedlichen Mengen der Substanz erfolgreich
vorherzusagen. Zusammenfassend lässt sich sagen, dass die Menge von Honigtau auf Baumwolle
lateral aufgelöst quantifiziert wird. Dafür wurde ein Prototyp für hyperspektrale Bildgebung im
UV-Bereich entwickelt, der für industrielle Anwendungen geeignet ist.
Die Ergebnisse zeigten, dass die hyperspektrale Bildgebung mehrere Vorteile gegenüber etablier-
ten Bildgebungsverfahren oder der klassischen ortsaufgelösten Spektroskopie bietet, z. B. die la-
terale Auflösung, die Fähigkeit, Proben zerstörungsfrei zu analysieren, Stoffe in sehr geringen
Konzentrationen zu erkennen und Stoffe selbst dann zu identifizieren, wenn sie mit anderen Stof-
fen vermischt oder durch diese überlagert sind. Außerdem ist sie sehr empfindlich und kann
kleinste Veränderungen in der chemischen Zusammensetzung von Materialien in Abhängigkeit
der Zeit erkennen.
II
Abstract
Cotton fiber is essential for the textile industry due to its softness, durability, and absorbency.
Therefore, the assessment of the cotton quality is needed, which is determined by the degree of
contamination. The predominant contaminants in raw cotton come from insects that excrete sug-
ars called honeydew during feeding. Cotton contaminated by sugar causes significant problems
for textile equipment. Honeydew is the most common source of sticky cotton. However, various
methods in visible (Vis) and near-infrared (NIR) spectral ranges are regularly used for quality
control and sorting procedures, while the ultraviolet (UV) range has not been widely used. In
recent years, hyperspectral imaging systems have gained increased attention over traditional tech-
niques due to their multi-modality, spatial resolution, and ability for quantitative analysis. These
advantages have made them highly attractive for various applications, such as in the textile in-
dustry.
The main goal of this work is to develop a method to detect honeydew contamination in the UV
range. For this purpose, a UV hyperspectral imaging system based on a spectrograph connected
to a CCD camera was constructed. The samples were placed on a conveyor belt, which moved
them underneath the hyperspectral imaging camera. This technique is called pushbroom imaging.
Depending on the application, either Xenon or Deuterium lamps were used for illumination since
Deuterium lamps provide a higher illumination strength in the UV-C region compared to the
xenon-arc lamp. In order to validate this novel imaging setup, a set of well-known substances,
such as active pharmaceutical ingredients (APIs) and painkillers, was used. These sample are
ibuprofen, acetylsalicylic acid, and paracetamol. The results were compared with single-point
spectroscopy and analyzed using chemometric data analysis. It was shown that the hyperspectral
imaging achieved reliable results, and an analytical method was developed to identify commercial
painkiller tablets with the new prototype. Subsequently, a separate sample set, including direct
bonded copper (DBC) sheets, was tested for a secondary evaluation. The developed prototype is
able to detect very thin oxide layers, as thin as a few nanometers. It can also distinguish between
various oxidation states via a cleaning procedure for DBC samples. Consequently, cotton samples
from different countries were investigated using Vis/NIR hyperspectral imaging. The data ob-
tained were compared to that obtained from single-point spectroscopy and analyzed using multi-
variate data analysis. The results indicate that it is possible to distinguish between different cotton
III
Abstract
types based on specific wavelength ranges. In the last step, the quantification of honeydew con-
tamination on cotton was determined. A calibration procedure was developed using mechanically
cleaned cotton samples. These samples were immersed in different concentrations of sugar and
protein to mimic cotton contaminated with honeydew. Consequently, they were analyzed after 44
hours and one month. Further improvements were made to the UV hyperspectral imaging setup
in the later measurement. The data obtained were analyzed using chemometrics to predict the
local quantities of honeydew on cotton samples successfully. In conclusion, the present work aims
to quantify the spatial amount of honeydew contaminated on cotton by developing a hyperspectral
imaging prototype in the UV region that is advantageous for industrial applications.
The results showed that hyperspectral imaging has several advantages over established analytical
techniques, such as lateral resolution, the ability to analyze samples non-destructively, detect ma-
terials at very low concentrations, and identify materials even when mixed or obscured by other
materials. It is also highly sensitive and can detect subtle changes in the chemical composition of
materials over time.
IV
Table of Contents
Kurzfassung .....................................................................................................................I
Abstract .........................................................................................................................III
Table of Contents .......................................................................................................... V
1 Introduction ............................................................................................................... 1
1.1 Spectroscopy ....................................................................................................... 1
1.2 Hyperspectral imaging ........................................................................................ 3
1.3 Chemometrics ................................................................................................... 10
1.3.1 Principal component analysis (PCA).................................................... 10
1.3.2 Partial least squares regression (PLS-R) .............................................. 13
1.4 Model systems .................................................................................................. 17
1.4.1 Pharmaceutical tablets .......................................................................... 17
1.4.2 Direct bonded copper ........................................................................... 18
1.4.3 Cotton fiber .......................................................................................... 19
1.4.4 Honeydew............................................................................................. 21
2 Objective .................................................................................................................. 24
3 Paper I: Characterization of Pharmaceutical Tablets Using UV Hyperspectral
Imaging as a Rapid In-Line Analysis Tool ................................................................ 25
3.1 Abstract ............................................................................................................. 26
3.2 Introduction....................................................................................................... 26
3.3 Materials and Methods ..................................................................................... 28
3.3.1 Samples ................................................................................................ 28
3.3.2 API’s in Solution .................................................................................. 29
3.3.3 UV Spectroscopy.................................................................................. 30
3.3.4 UV Hyperspectral Imaging .................................................................. 30
3.3.5 Data Collection and Preprocessing....................................................... 31
3.3.6 Data Handling and Software ................................................................ 32
3.4 Results and Discussion ..................................................................................... 33
3.4.1 UV Spectroscopy.................................................................................. 33
3.4.2 UV Hyperspectral Imaging .................................................................. 35
3.5 Conclusions....................................................................................................... 40
4 Paper II: UV Hyperspectral Imaging as Process Analytical Tool for the
Characterization of Oxide Layers and Copper States on Direct Bonded Copper . 43
4.1 Abstract ............................................................................................................. 44
4.2 Introduction....................................................................................................... 44
4.3 Materials and Methods ..................................................................................... 46
4.3.1 2.1. Samples ......................................................................................... 46
4.3.2 Oxide Layer Thickness Measurement .................................................. 47
4.3.3 UV Spectroscopy.................................................................................. 47
V
Table of Contents
4.3.4 Data Collection and Preprocessing ....................................................... 47

4.3.5 Multivariate Data Analysis and Data Handling .................................... 49
4.4.1 UV Spectroscopy .................................................................................. 50
4.4.2 UV Hyperspectral Imaging ................................................................... 51
4.4.3 PLS-R ................................................................................................... 54
4.5 Conclusions ....................................................................................................... 57
5 Paper III: UV-Vis/NIR Spectroscopy and Hyperspectral Imaging to Study the
Different Types of Raw Cotton ................................................................................... 59
5.1 Abstract ............................................................................................................. 60
5.2 Introduction ....................................................................................................... 60
5.3 Materials and Methods ...................................................................................... 62
5.3.1 UV-Vis/NIR spectroscopy .................................................................... 62
5.3.2 NIR hyperspectral pushbroom online imaging system ......................... 62
5.3.3 Samples................................................................................................. 63
5.3.4 Data collection and preprocessing of hyperspectral data...................... 64
5.3.5 Data handling and software .................................................................. 66
5.4 Results and discussion ...................................................................................... 67
5.4.1 UV-Vis/NIR spectroscopy .................................................................... 67
5.4.1 NIR Hyperspectral Imaging .................................................................. 70
5.5 Conclusions ....................................................................................................... 73
6 Paper IV: Prediction of Honeydew Contaminations on Cotton Samples by In-
Line UV Hyperspectral Imaging ................................................................................. 75
6.1 Abstract ............................................................................................................. 76
6.2 Introduction ....................................................................................................... 76
6.3.1 Chemicals and preparation of solutions ................................................ 78
6.3.2 Sample set and sample preparation....................................................... 79
6.3.3 UV hyperspectral imaging setup .......................................................... 81
6.3.4 Data collection and preprocessing ........................................................ 82
6.3.5 Multivariate data analysis and model building ..................................... 83
6.4.1 Cotton samples impregnated with sugar ............................................... 84
6.4.2 Predicting the amount of sugar and honeydew based on the sugar PLS-
R model ........................................................................................................ 86
6.5 Conclusions ....................................................................................................... 88
7 Paper V: Applying UV Hyperspectral Imaging for Quantification of Honeydew
Content on Raw Cotton via PCA and PLS-R Models ............................................... 91
7.1 Abstract ............................................................................................................. 92
7.2 Introduction ....................................................................................................... 92
7.3.1 Chemicals and preparation of solutions and samples ........................... 93
7.3.2 UV hyperspectral imaging setup and data processing .......................... 94
VI
Table of Contents

7.5 Conclusions....................................................................................................... 97
8 Conclusion and Summary ...................................................................................... 99
9 Bibliography ............................................................................................................ IX
10 Appendix ............................................................................................................... XXI
10.1 Supplementary Information for Paper II: UV Hyperspectral Imaging as Process
Analytical Tool for the Characterization of Oxide Layers and Copper States on
Direct Bonded Copper ................................................................................... XXI
10.2 Supplementary Information for Paper IV: Prediction of Honeydew
Contaminations on Cotton Samples by In-Line UV Hyperspectral ImagingXXII
10.2.1 Cotton sample preparation................................................................ XXII
10.2.2 Additional figures of the principal component analysis of the sugar
cotton samples ....................................................................................... XXIII
10.2.3 Pure dried protein spectrum ........................................................... XXIV
10.2.4 X-loadings weights and x-loadings of the PLS-R model .................XXV
10.3 Supplementary Information for Paper V: Applying UV Hyperspectral Imaging
for Quantification of Honeydew Content on Raw Cotton via PCA and PLS-R
Models ....................................................................................................... XXVII
List of Abbreviations and Symbols ....................................................................... XXIX
List of Figures ......................................................................................................... XXXI
List of Tables ......................................................................................................XXXVII
List of Publications.............................................................................................. XXXIX
Declaration of Contribution ..................................................................................... XLI
Acknowledgments .................................................................................................. XLIII
VII
1 Introduction
1.1 Spectroscopy
Spectroscopy is a method of studying the interaction between light and objects. It helps us under-
stand the properties of different materials based on the absorption or emission over a specific
wavelength or energy range [1,2]. Different spectroscopic methods can be used to solve a wide
range of analytical problems. These methods differ based on the specific substances or species
that need to be analyzed, such as atomic or molecular spectroscopy [3]. It utilizes the ability of
molecules and atoms to absorb, emit, and scatter a type of energy called electromagnetic (EM)
radiation. EM radiation is a form of energy that covers a wide range of wavelengths and frequen-
cies, ranging from cosmic radiation at 10-14 m to infrasonic radiation at 1010 m. Spectroscopy can
be applied to different frequency ranges, such as Ultraviolet-Visible/Near-infrared (UV-Vis/NIR)
and nuclear magnetic resonance (NMR) spectroscopy. UV-Vis/NIR spectroscopy is one of the
most commonly used spectroscopic techniques, ranging from 100 nm to 2500 nm [4,5]. The ab-
sorption or emission of various forms of EM radiation is associated with different types of tran-
sitions. UV-Vis and NIR are associated with electronic transition and molecular vibration [6].
The absorption process is defined as the transfer of EM energy to atoms or molecules. Electrons
in the atoms are excited from a lower to a higher energy state, which exists on discrete levels.
Absorption occurs when the energy of an exciting photon matches the energy difference between
the ground and excited state. These energy differences are unique, providing a means of charac-
terizing a compound or material [7,8]. The following formula gives the energy E, which is related
to the absorption bands:
E = Eelectronic + Evibrational + Erotational (1.1)
Where the molecules' electronic energy is given by Eelectronic, the vibrational and rotational energy
is described as Evibrational and Erotational, respectively (see Figure 1.1).
1
1 Introduction
Energy
e"4
e"2
e"3
e"2
Rotational
e"1 energy levels
E2 e"0
e"1
eʹ4
eʹ3
Elctronic energy levels
eʹ2 Vibrational
eʹ1 energy levels
E1 eʹ0
e4
e3
e2
e1
E0 e0
Figure 1.1: Energy level diagram illustrates electronic, vibrational, and rotational energy.
Absorption occurs, for example, when light passes through a solution. Light absorption is defined
by Beer-Lambert law and is commonly used in spectroscopy [9]. Beer-lambert’s law can be used
for pure samples or samples that do not significantly scatter light. It defines the amount of the
energy absorbed A or transmitted from the solution as proportional to the molar absorptivity co-
efficient ε and the concentration of the solute 𝒸 the optical path length in cm 𝓁. The following
formula gives the relationship [10]:
I
A = ε𝓁𝒸 = − log10 (I ) (1.2)
0
Although it can also be written in terms of intensities, where I is the light intensity passing through
the sample cell, and the initial light intensity is defined by I0 .
Often, samples have varying sizes and shapes and can be either transparent or in-transparent in
specific wavelength regions due to absorption characteristics and refractive index in the case of
transparent samples or scattering in the case of in-transparent samples. Therefore, many spectral
systems acquire the reflected light from the samples. Reflection of light can be classified into two
types: classical reflection, such as mirrors, and diffuse reflection from rough surfaces. Diffuse
2
1.2 Hyperspectral imaging
reflection is described by Lambert's law and is observed in materials such as milk and Spectra-
lon®. Figure 1.2 illustrates both classical and diffuse reflected light [11].
(a) (b)
Figure 1.2: (a) Classical and (b) diffuse reflected light.

Optical spectroscopy is a widely used analytical technique in various fields, such as the internal
quality analysis of biological samples, due to its rapid, non-destructive, and environmentally
friendly nature [12,13]. It often requires minimal sample preparation and handling. These optical
methods can be applied as on-line or in-line techniques [14-16]. Over the years, optical spectros-
copy has become an essential traditional method in laboratory work and is considered a reference
method across a wide range of wavelength regions, from the far ultraviolet to the infrared, includ-
ing spectral imaging techniques [17,18]. Various instrumental techniques have been developed
based on these fundamental principles of optical spectroscopy, such as hyperspectral imaging
[19].
Hyperspectral imaging is a technique that involves capturing and analyzing images of an object
across a wide wavelength range, typically in the Vis and NIR and, recently, in UV regions
[17,19,20]. It allows for identifying materials and compounds within an image by analyzing those
materials' unique "spectral fingerprints". The combination of spectroscopic techniques with im-
aging is, therefore, a fast and cheap option with the ability to cover the whole production, which
has not been used so far. This method can operate as on-line or in-line technique [15,21].
Hyperspectral imaging has the potential to meet the requirements for spatial resolution and large-
area detection. It enables the rapid spatially resolved spectral analysis of surfaces. This technique
is well described in the literature for applications in medicine, food, mineralogy, agriculture, and
environmental monitoring [22-27].
3
1 Introduction
The imaging system can be designed in different modes of imaging. The most common are mul-
tispectral and hyperspectral imaging. The main difference between these imaging modes is in the
number of color channels, and the resolution see Figure 1.3 [28,29]. Multispectral imaging has 4
to 20 color channels and collects non-continued images from specific bands. In contrast, hyper-
spectral imaging continuously captures tens to hundreds of spectral bands [30].
Multispectral imaging
Hyperspectral imaging
400 500 600 700

Wavelength / nm
Figure 1.3: Schematic shows the difference between multispectral and hyperspectral.
Hyperspectral imaging is based on one of four different technologies to acquire complete spectral
information [30]. Figure 1.4a shows a point-by-point technique known as a whiskbroom or spot-
light sensor. It works like a broom sweeping across an area by moving a sensor back and forth to
collect data. The sensor collects the data from one pixel in the image at a time, which can obtain
a high spectral resolution. Therefore, discrimination between molecules can be acquired. Origi-
nally, this technique was used in satellite technology for scanning the earth by the Earth Resources
Technology Satellites, and the Airborne Imaging Spectrometer later adopted it. Figure 1.4b shows
a new strategy called snapshot. It captures a 3D area (x, y and λ) and spectral information into one
single measurement by using an image mapper and prism array. Unlike the other techniques, it
uses one exposure to record spatial and spectral information. Therefore, there is no need to scan
at all. This technique is suitable for capturing instantaneous images of a scene or object and is
often used in applications such as digital photography. Figure 1.4c shows an image-by-image or
area-scanning system, also called staring imaging. The staring technique captures the 2D area (x,
y) while changing a filter in front of the camera. The light passes through the optics, then it is
filtered, which produces a narrowband of the spectrum [18]. It is used in many applications, in-
cluding surveillance and astronomy, for capturing detailed and steady images of specific targets
or regions of interest. Figure 1.4d shows the line-by-line technique called pushbroom. Pushbroom,
named along-track scanner based on an on-line scanning system, acquires complete spectral in-
formation for each pixel in the line. These data result in a three-dimensional (3D) data matrix with
4
dimensions x, y, and λ, often referred to as a hypercube. The data of all these technique provides
high spectral resolution and is highly stable, making it useful for industrial quality control [31].
(a) Whiskbroom (b) Snapshot (c) Staring (d) Pushbroom
λ
y
x
Prism array
Prism Image mapper Filter Prism
Figure 1.4: Visualization of the different imaging technologies: (a) Whiskbroom imaging (single point scanning) (b)
Snapshot imaging (c) Staring imaging (2D scanning) (d) Pushbroom imaging (line scanning).
Such inspection systems require a minimum of sample preparation and can scan several samples
swiftly with high spectral resolution [23]. Hyperspectral imaging systems based on a pushbroom
scanner and a conveyor belt are commonly used in industrial applications [27,32-34]. The data is
collected with the camera placed perpendicular to the conveyor belt. As the conveyor belt moves,
the pushbroom scanner scans each line of the scene, and these lines are captured and combined
to form a complete image. Figure 1.5 illustrates the scheme and general setup of a pushbroom
scanner with a conveyor belt.
5
1 Introduction
Pushbroom imager
Halogen lamp
Halogen lamp
Figure 1.5: Scheme of Vis/NIR hyperspectral imaging (pushbroom).
Such systems have been primarily developed in the Vis/NIR region [23,35,36], but there is further
information on quality control in the UV region. Therefore, in this dissertation, we will address
the development of a new hyperspectral imaging prototype in the UV region and some applica-
tions combined with chemometrics.
A hyperspectral imager was developed in the UV region to quantify and classify different cotton
types with different amounts of honeydew contamination. Figure 1.6 shows a scheme for UV
hyperspectral imaging. Figure 1.6a illustrates the principle of continuous line-by-line spectral data
collection. The data results in a lateral resolved (x, y) 2D image, as shown in Figure 1.6b, c,
whereas each location contains a further spectroscopic dimension (λ), as shown in Figure 1.6d.
Thus, a 3D data matrix (hypercube) was recorded [19-21,37-39].
6
Pushbroom imager
(a)
Spectrograph Camera
Tunnel
PTFE
Objective
Xenon lamp
Power supply Xenon lamp
Power supply
(b) (c) (d)

2.0
1.5
 - log (R/R0) 1.0
0.5
x 0.0
200 225 250 275 300 325 350 375
Wavelength / nm
y
Moving direction Hyperspectral image Spectrum at one single pixel
Figure 1.6: (a) Schematic shows the concept of hyperspectral imaging based on the pushbroom (the tunnel in the scheme
was cut to show the inside). (b) Pushbroom Imager scanning principle. (c) Hyperspectral image produced
immediately during sample scanning. (d) UV spectrum for one single pixel extracted from the image in
(c).
Selecting an appropriate light source for UV irradiation is a significant challenge in various fields.
Several sources are available in the UV region, but only a limited number are suitable for our
purpose. Among others, there are synchrotron radiation [40], light-emitting diodes (LED), laser-
induced plasma, deuterium, xenon-arc (XBO), and mercury-arc (HBO) lamps.
Synchrotron radiation is a form of electromagnetic radiation emitted by a particle accelerator

when the velocity of the electrons is approaching the speed of light. These electrons are forced to
travel in curved paths by a magnetic field, which causes them to emit radiation when they are
moved within the synchrotron. This intense radiation covers a wide range of wavelengths, includ-
ing X-rays, UV-Vis, and IR light. These properties make synchrotron radiation a valuable tool for
scientists studying the nature and structure of molecules and materials. Having a synchrotron ra-
diance covering the entire UV range is challenging, but a spectrum within the 225 – 325 nm range
was obtained [40]. However, synchrotrons perform off-line measurements due to their transport-
ability limitations, making them unsuitable for industrial quality. Moreover, it is considered one
7
1 Introduction
of the most expensive radiation techniques, with approximately 70 synchrotrons worldwide.

Therefore, researchers need a cheaper and more available light source [41-43].
LEDs are diodes with several advantages over traditional incandescent and fluorescent lamps,
including their small size and ease of integration into various designs. LEDs have high quantum
efficiency, which refers to the ratio of the number of photons emitted by the LED to the number
of electrons that pass through it. They have a long lifetime of over 100,000 hours of operation.
However, one limitation of UV LED lamps is that they typically emit radiation within a narrow
wavelength band, making it challenging to generate a continuous spectrum of radiation with UV
LEDs [44,45].
Laser-induced plasma lamps, also known as plasma globes or spheres, are a light source contain-
ing a partially ionized gas in a glass sphere. It utilizes laser-driven technology, which is consid-
ered a leading choice for the UV-Vis/NIR region. These lamps have high spectral radiance inten-
sity across the aforementioned wavelength range and a longer lifetime than traditional lamps. Due
to their energy consumption and heat generation, plasma lamps require water cooling to prevent
overheating. Additionally, plasma lamps are relatively expensive compared to other lighting op-
tions [46].
Deuterium lamps are considered one of the most stable lamps in the UV region. This is due to the
ceramic electrode structure within the lamp, which results in low fluctuation levels in peak inten-
sity. One of the key advantages of deuterium lamps is their ability to produce a continuous spec-
trum of UV light from approximately 115 – 400 nm. This makes these lamps suitable for experi-
ments that require UV radiation. However, deuterium lamps have a few limitations, such as
restricted spectral range and low spectral radiance intensity in the UV region. These lamps are
commonly used in various scientific and analytical applications, such as the pharmaceutical in-
dustry and for atomic absorption spectroscopy [47,48].
High-pressure mercury vapor lamps, also known as mercury-arc lamps (HBO), are a type of light
source that can produce intense, bright illumination. These sources are stable and have a high flux
density, making them widely used in fluorescence microscopy. However, they have a relatively
low intensity in the UV region compared to xenon-arc (XBO) lamps. XBO lamps have a contin-
uous spectrum and produce high-intensity spectra within 240 nm to 400 nm. Despite the low
stability, the broad spectrum range and the long-life feature are added advantages of XBO lamps.
On the other hand, XBO lamps produce ozone and heat. In this dissertation, a xenon lamp was
used because it is inexpensive, easy to replace, and the aperture of the illumination is adjustable
8
[49-52]. One of the critical factors to consider when evaluating these sources is the distinct emis-
sion spectra they produce, see Figure 1.7. It is important to ensure that the chosen light source
emits the desired wavelengths to meet the application's specific requirements. This often involves
comparing the emission spectra of the different sources and selecting the one that best matches
the desired wavelength range. These lamps are available for use in our experiments, and the pro-
totype tested five different light sources: LED, plasma EQ-77, deuterium, XBO, and HBO. Figure
1.7 shows the different light source spectra used pushbroom hyperspectral imaging setup. The
pushbroom imager contains a back-illuminated CCD camera (Apogee Alta F47: Compact, inno-
spec GmbH, Nürnberg, Germany) and is connected to a spectrograph (RS 50-1938, inno-spec
GmbH, Nürnberg, Germany) [19,23,53-55].
(a) (b) (c)

1.0 1.0 1.0
P66 LED Plasma_EQ77
0.8 0.8 0.8
Normalize intensity
Normalize intensity
Normalize intensity
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0.0 0.0 0.0

225 250 275 300 325 225 250 275 300 325 350 375 400 225 250 275 300 325 350 375 400
Wavelength / nm Wavelength / nm Wavelength / nm
(d) (e) (f)

1.0 1.0 1.0
Deuterium Xenon-arc Mercury-arc
0.8 0.8 0.8
Normalize intensity
Normalize intensity
Normalize intensity
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0.0 0.0 0.0

225 250 275 300 325 350 375 400 225 250 275 300 325 350 375 400 225 250 275 300 325 350 375 400
Wavelength / nm Wavelength / nm Wavelength / nm
Figure 1.7: Spectral radiance for a different light source in the UV region. (a) Synchrotron radiation of P66 beamline
(b) LED radiance (Roithner LaserTechnik GmbH, Wien, Germany) (c) Plasma radiance (EQ-77, Energetiq
Technology LDLSTM, Wilmington, MA, USA) (d) Deuterium radiance (SL 3, StellarNet Inc, 24 V, 65.04
W, Tampa, Florida, USA) (e) Xenon-arc radiance (XBO, 14 V, 75 W, Osram, München, Germany) (f)
Mercury-arc radiance (HBO, 14 V, 75 W, OSRAM, München, Germany). All light sources were tested by
UV hyperspectral imaging except Synchrotron radiation of P66 taken from reference [40].
As a result, the synchrotron was excluded from this study because it did not serve our purpose; it
was deemed unsuitable for on-line measurements due to lack of portability and expense. LED and
mercury lamps were ruled out due to their narrow wavelength bands, see Figure 1.7b and f. Using
plasma was limited due to cost considerations and the requirement for cooling. In contrast, Deu-
terium and XBO-arc lamps were selected, offering a continuous spectrum (Figures 1.7d and e)
and low cost.
9
1 Introduction
1.3 Chemometrics
Chemometrics can be derived from “chemo” related to chemistry, and “metric” means measure-
ment. It is a sum of statistical methods that have introduced new algorithms capable of handling
the massive amount of chemical data using multivariate data analysis (MVA) [56,57]. MVA can
be a powerful tool for data reduction to find a small number of variables capable of explaining all
the variations from the data. Typical MVA techniques are principal component analysis (PCA),
Discriminant analysis (DA), and partial least squares regression (PLS-R) [58].
1.3.1 Principal component analysis (PCA)

This section is partially based on the work by Rebner K. [59] and has been further modified.
Principal Component Analysis (PCA) is a statistical method used to identify similarities and dif-
ferences among data sets. It reduces the dimensionality of the variables to their most essential
features without losing important information [60-62]. Data reduction and features are extracted
from a data matrix with p objects (rows) and t variables (columns). The first summary index, so-
called latent variables, the principal components (PCs), are calculated from the original variables
via a principal axis transformation. Mathematically, these factors are linear combinations of the
original variables. For example, an object in a two-dimensional space described by variables x1
and x2 can be transformed into a new vector space represented in PCs. The transformation is given
as an orthogonal matrix formed from the eigenvectors of the covariance matrix. The coordinate
system is rotated so that it points in the direction of maximum variance in the data, thus describing
the information content of the data (see Figure 1.8).
x2 PC1
PC2
x1
Figure 1.8: Graphical representation of the principal components. The original data in the original data space (x1, x2)
are transformed into new principal axes (PC1, PC2).
10
1.3 Chemometrics
In data analysis, the first principal component explains the most variance. The second principal
component explains the second highest variance in what remained after the effect of the first
component was removed [63]. Each new principal component describes the maximum variance
that was not captured by the previous components. The number of variables – 1 (maximum num-
ber of PCs) is mathematically determined until the data explain a certain percentage of variance.
The PCs are orthogonal to each other and thus are independent, which means the data is de-cor-
related. This also means that the principal axis transformation can look different depending on
the problem, and a separate transformation matrix must be calculated for each data set.
The principal components are determined by decomposing the data matrix X into a weight matrix
T, a transpose factor matrix PT, and an additional residual matrix E.
X = TP T + E (1.3)
TPT describes the new data structure, and the residual matrix E consists of noise and unexplained
data. Figure 1.9 shows how a matrix with PCA is decomposed. Where the data matrix X is de-
composed to PC1 (scores t1 and loading p1) and residual E1, PC2 will be calculated from residual
matrix E1, which contains information not explained by the first PC.
X = t1p1 + t2p2 + … + tapa + E (1.4)
PC1
p1
X = t1 + E1
PC2
p2
E1 = t2 + E2
Figure 1.9: Schematic description of a decomposition of a matrix X with PCA using two PCs.
In context of data analysis, after performing PCA to reduce the dimensionality of the data, another
statistical method called discriminant analysis (DA) can be used to validate the PCA model. DA
is a supervised classification method that assigns an unknown pattern to a group of similar objects.
Also, it is a separation technique that optimally divides a training set of objects into two or more
groups using a border, which can take the form of linear discriminant analysis (LDA), quadratic
discriminant analysis (QDA), or mahalanobis discriminant analysis (MDA) distance-based sepa-
rators see Figure 1.10. DA is a powerful classification method that simultaneously minimizes the
11
1 Introduction
variance within each group and maximizes the distance between the groups. This gives better
separation of the groups compared to PCA, which only looks at the variance of the data. The
optimized border may allow for minor group overlaps, and the choice of border shape depends
on the nature of the data. For example, quadratic borders help separate groups with differently-
oriented main variances, while Mahalanobis distance is suitable for measuring the distance
between objects and class centers using ellipses as distance calculators [64-66].
(a) (b) (c) PC3

PC3 PC3
x x x x x x
PC1 PC1 PC1
x x x
PC2 PC2 PC2
Figure 1.10: Schematic shows different discriminant analysis (DA). DA function creates a border of variable shape that
optimally separates a training data set into multiple groups. PCA can be used to reduce the dimension of
the training set before creating the discriminant function. The three most common types of DA separators
are linear, quadratic, and mahalanobis distance-based separators. Linear separators are represented by
straight orange lines (a), quadratic separators by orange curves (b), and mahalanobis distance-based sepa-
rators by ellipses (c). This figure is taken and modified from reference [66].
To create a discriminant function, the object groups must exhibit significant differences in their
variables, and the variables with the maximum variance are identified to achieve the highest pos-
sible separation between the groups. The performance of the discriminant functions can be eval-
uated using a classification matrix or confusion matrix. The confusion matrix, also known as the
error matrix, provides a table used to evaluate the performance of a model by comparing the pre-
dicted and actual values. It consists of four values: true positives (TP), false positives (FP), true
negatives (TN), and false negatives (FN) see Table 1.1. This table is used to evaluate the perfor-
mance of a binary classification model by comparing the predicted and actual values. It consists
of four values: true positives (TP), false positives (FP), true negatives (TN), and false negatives
(FN). Various performance measurements can be derived from the confusion matrix, such as pre-
cision, specificity, and accuracy, which provide insights into the model's effectiveness and poten-
tial limitations (see Chapter 3.4, Chapter 4.4) [67].
12
1.3 Chemometrics
Table 1.1: Confusion Matrix.
Predicted value
Positive (1) Negative (0)
Actual value
Positive (1) TP FP
Negative (0) FN TN
1.3.2 Partial least squares regression (PLS-R)

This section is partially based on the work by Rebner K. [59] and has been further modified.
PLS is considered one of the most important data analysis tools for regression and classification.
This analysis tool decomposes the variance between independent X variables (predictor values)
and dependent Y variables (response values) to calculate the variance and correlation between X
and Y to estimate PLS-R components [68,69]. The straight-line equation gives the relationship for
linear regression:
y = b0 + b1x (1.5)
Where b0 is the y-axis intercept and b1 is the slope of the straight line. This function is fitted to
the data x and y to calculate the variance and maximize the correlation between the data repre-
senting the first PLS. When using MVA, whose structure may be complex or even error-prone,
using a multivariate regression method is preferable to increase predictive accuracy. The PLS
method is related to PCA and already uses the structure of the Y data to find PCs. This has the
advantage that only a few PCs are often sufficient for the complete data description and can be
interpreted easily.
For each object, a target value yi is measured, which forms the vector y and, in the case of multiple
objects, the target value matrix Y. The target values can be chemical or physical parameters. As
in PCA, the data matrix is decomposed into a T and P matrix. As an intermediate step of the PLS,
an additional matrix is needed that calculates weighted loadings W and provides a link to the Y
variables. In this case, not the P loadings but the W loadings are orthogonal to the T scores.
Therefore, PLS components are used instead of PCs, as in PCA. The target matrix Y is decom-
posed into a factor matrix Q and a score matrix U, as well as an additional residual matrix F.
X = TP T + E (1.6)
Y = UQ T + F (1.7)
13
1 Introduction
Two data sets X and Y are now contrasted with their respective score vectors T and U, respec-
tively. Figure 1.11 shows a schematic description for a PLS-R model. The direction of each PLS
latent variable of the X matrix is now changed such that the covariance between it and the vectors
of the Y matrix is maximized. The exact intermediate mathematical steps are described in detail
in the literature [61].
Figure 1.11: Partial least square regression method. The Y variables influence the X variables.
However, unlike PCA, the PLS components do not contain the largest differences in variance
between spectra but the most relevant differences concerning the reference data. In the regression
approach between the x- and y-variables, the coefficients are calculated as follows:
B = W(P T) -1Q T (1.8)
and
b0 = ȳ - 𝐱̅TB (1.9)
Where B is the regression matrix W represents the loadings matrix for the predictor (X) variables,
Q represents the loadings matrix for the response (Y) variables, and P represents the score matrix
for the predictor variables. b0 is the y-axis intercept, and ȳ is the mean of the response variable (Y).
Finally, the target quantity yi can be given by
yi = b0 + 𝐱𝐢T B (1.10)
by using the measured values xT for each object. The equation represents the predicted value of
the response variable yi for a given set of predictor variables xi in a linear regression model. xiT
represents the transpose of the predictor variable vector xi. If the PLS regression is performed
with one Y variable, it is called PLS1. In contrast, PLS2 calculates a model for several Y variables
14
1.3 Chemometrics
simultaneously. The advantage is that all X and Y data are included in a common model; thus, all
correlations are considered. Mathematically, instead of the vectors, the matrices Y, F, and Q are
used and supplemented by the score matrix U of the Y values.
In addition to calibration, regression includes steps of validation and prediction. PLS-R is utilized
to unite so-called “factors” that describe the relationship between X and Y variables. The optimal
number of factors used to describe the model is essential. Too few factors (under fitting) or too
many factors (over fitting) lead to additional prediction errors. For an optimal model, this error is
minimal, but a difference must be made between a calibration error due to under fitting and an
estimation error due to overfitting. The calibration error and the residual variance will decrease
with the first PLS components as the relevant information increases to a certain degree. The op-
timum is exceeded if random changes in the form of noise are modeled with the addition of further
components. In this case, the estimation error increases, and the prediction error of unknown data
becomes larger than the calculated calibration error once a suitable model can be applied to un-
known data [61,62,68].
In chemometrics, a distinction is made between external and internal validation. Internal valida-
tion is often used for smaller data sets because the same data set can be used for calibration and
validation. Especially for method developments, cross-validation offers a very efficient method
when a large number of samples is not yet available. For this purpose, some objects are omitted
from the calibration data during the calculation, and a calibration model is created. Afterward, the
calibration model is used to predict the omitted objects and to determine the residuals [62]. This
process is repeated several times until all objects have been omitted once and predicted with the
model. The quality of calibration or validation can be indicated with different values. The most
important ones are briefly explained here.
1.3.2.1 Coefficient of determination

The correlation between the reference value y and the predicted value ŷ is often given for the
regression coefficients, and the coefficient of determination R2 is calculated. This expresses how
much of the variance of the dependent variable y can be explained by the independent variable x
and is calculated by the following formula [70]:
n
̂i ) 2
∑i=1(yi −y
R2 = 1 − n (1.11)
̂) 2
∑i=1(yi −y
15
1 Introduction
1.3.2.2 Root mean square error (RMSE)

Root mean squared error (RSME) is called the root mean square deviation (RSMD). It is consid-
ered a common function and evaluation matrix used in a regression model. RMSE is mainly used
to calculate the difference between values predicted by a model and observed values. It is given
with the suffix C for "Calibration" in the case of calibration RMSEC, the suffix CV for "Cross-
Validation" in case of internal validation RMSECV or the suffix P for "Prediction" in the case of
external validation RMSEP. The mean error is calculated by [71,72]:
n
∑ ̂i ) 2
(yi −y
RMSE = √ i=1 n
(1.12)
Where n is the number of samples.
1.3.2.3 Standard deviation

The standard deviation is a measure that shows how much data is scattered around the true value.
For example, how different the answers of your respondents are. It is summarized into two forms:
standard error of calibration (SEC) and standard error of prediction (SEP). The standard error
(SE) is the residuals' standard deviation [73]. Mathematically, the SE systematic error between
prediction and reference value (BIAS) must be determined in advance by the following formulas
[62,74].
n
̂ i )2
(yi −y
BIAS = ∑ n
(1.13)
i=1
n
̂i −BIAS)2
∑i=1(yi −y
SE = √ (1.14)
n−1
1.3.2.4 Prediction
PLS regression is also used for predicting dependent Y values from independent X values. De-
compose the following equation gives X value used for building up and predicting Y values [62]:
y = b0 + b1x (1.15)
Where b0 is the y-axis intercept and b1 is the slope of the straight line.
16
1.4 Model systems
1.4 Model systems

Quality control is an essential feature of products in industrial environments. Several studies ver-
ified UV hyperspectral imaging to achieve the quality aspect. This dissertation used various ap-
plications, such as pharmaceuticals, metals, thin layers, and natural products.
1.4.1 Pharmaceutical tablets

Chemical drugs have been increasingly used globally for treating and preventing infections and
pain [75]. Therefore, countless medicines, tablets, liquids, gels, and powders are made daily. Pain-
killers were used as a model system; painkillers may also consist of different liquids or powder
mixtures [76]. In this case, qualitative and quantitative analyses are needed to identify active com-
pounds, determine the content of active compounds, and measure the significant impurities of the
medicine [77]. Thus, these parameters are required rapidly and non-destructively during the man-
ufacturing process.
In the pharmaceutical industry, separation techniques are used for qualitative and quantitative
analysis. These analyses are mainly achieved by high-performance liquid chromatography
(HPLC) [63]. Snakar et al. [78] described the application of HPLC in pharmaceutical analysis.
This technique has many disadvantages, such as being expensive, time-consuming, needing sam-
ple preparation, and destructive. Using UV spectroscopy, Saeed et al. [79] estimated the quantity
of active ingredients in different tablets of some commercial dosages, such as ibuprofen, aspirin,
and paracetamol.
In the last 20 years, the food and drug administration (FDA) in the United States started the pro-
cess of analytical technology (PAT) to control medicine production processes [19,20]. For this
purpose, hyperspectral imaging is an excellent PAT tool for ensuring product quality. One of the
aims of this dissertation is to develop a laboratory prototype for hyperspectral imaging in the UV
region based on the pushbroom technique in combination with multivariate data analysis.
Active pharmaceutical ingredients (API) 100% and commercial painkiller tablets were used for
general testing of UV hyperspectral imaging [19]. The APIs were ibuprofen (IBU), acetylsalicylic
acid (ASA) and paracetamol (PAR). Ibuprofen tablets from two companies (ratio pharm IBUratio,
and beta pharm IBUbeTa), aspirin (Bayer GmbH ASPBAYER), paracetamol (ratio pharm PARratio),
and thomapyrin were used (see Figure 1.12). The overview of the APIs and painkiller tablets is
given in Table 3.1.
17
1 Introduction
4 g of ASApure and IBUpure were pressed at 10 tons for 2 min by a hydraulic press into the depicted
disc shape. Then, 4 g PARpure powder were dried in a vacuum oven for 1 h at 120 °C, and pressed
at 10 tons for 20 min (see Table 3.1). A mixture of 2 g ASApure and 2 g PARpure was prepared by
using a speed mixer and pressed at 10 tons for 2 min [19].
Figure 1.12: Drug samples. Reference API samples and painkiller tablets. This figure is taken from reference [19].
1.4.2 Direct bonded copper

Direct Bonded Copper (DBC) refers to a process in which copper and ceramic material are di-
rectly bonded. DBC substrates have been proven to be an excellent solution for electrical insula-
tion and thermal management of high-power semiconductor modules [80]. Therefore, they are
considered the most significant conductors compared to other materials, such as aluminum. The
advantages of DBC substrates are high electrical and thermal conductivity due to thick copper
metallization and thermal expansion close to the silicon copper surface due to the high adhesion
strength of copper to ceramic [81]. However, the copper surface interacts with oxygen to produce
copper (I) oxide (Cu2O) and copper (II) oxide (CuO). Therefore, the efficiency of the conductivity
becomes poor [81,82]. Optical techniques such as Auger electron and X-ray photoelectron spec-
troscopy were applied to solve this problem [83,84]. Such systems are time-consuming, destruc-
tive, and expensive.
In the past, progress in research led to the development of sensor technology [32,35,77,83,85,86].
Stiedl et al. [26,87] investigated the copper oxide layer thicknesses and copper state on a metallic
DBC using UV-Vis spectroscopy and visible hyperspectral imaging. In this thesis, we verified
the capability of the new UV hyperspectral imaging prototype to study the changes in the copper
state and thickness of the copper oxide on DBC (see Figure 1.13).
18
1.4 Model systems
1 2 3 4 5 6
Figure 1.13: Direct bonded copper Curamik®Power substrates. (1), (2), (3), (4), (5) and (6) are different samples with
different types of copper state and thickness of the copper oxide on DBC. This figure is taken and modi-
fied from reference [53].
1.4.3 Cotton fiber

Cotton is the most important natural raw material used in producing fabrics. It has been used
extensively for clothing people worldwide. Cotton lint (see Figure 1.14) is considered an essential
product that provides a source of high-quality fiber for the textile industry [88].
Figure 1.14: Cotton fibers.
Over 34 million hectares of land are used to grow cotton, and around 100 million households
worldwide are engaged in cotton production [89]. Cotton is an essential resource in the textile
industry, accounting for approximately 30% of all fibers utilized [90]. Almost all parts of the
cotton plant have a range of uses. Cottonseeds are an important oil source, which is a byproduct
of the plant. Additionally, cottonseeds have a high protein concentration, making them useful in
animal feed. The waste material left over from the ginning process, which separates the cotton
fibers from the seeds, can be used as fertilizer. The stalk of the cotton plant is also a potential
resource, as it contains cellulose, which can be used to make paper and cardboard [91,92].
Cotton is the most widely produced natural textile fiber and a significant global commodity, as it
is the most imported and exported raw material [89]. Cotton processing, through spinning or knit-
ting, plays an important role in the economies of many countries. Cotton fabric production begins
with the preparation of yarn, which is achieved by removing the seeds. This yarn is then woven
19
1 Introduction
or knit into fabric. The fabric undergoes a series of additional processing stages, including dyeing,
designing, and sewing, to produce a final product that is soft, clean, and ready for use [93]. Figure
1.15 presents the steps of the cotton manufacturing process.
Figure 1.15: Manufacture process for raw cotton; this photo is taken from reference [94].
In textile research, cotton plays a dominant role among textiles since cotton is the most important
naturally occurring raw material for fabric production. It is important to guarantee the quality of
the fibers in textile processing, as fiber properties highly impact the properties of finished yarns
and fabrics and how easy it is to manufacture. For example, fiber strength determines yarn
strength, and fiber maturity affects the dye uptake of fabrics. Raw cotton with poor fiber quality
causes problems in the textile mill [91,92,95].
Cotton fibers are soft, relaxed, and breathable and have high absorbency. They are natural hollow
fibers, capable of absorbing liquids such as water up to 24-27 times their weight. The elongation
of the fibers begins shortly after anthesis, which is the flowering stage, and continues for three to
four weeks, primarily increasing in length during this phase. Two weeks after the anthesis, depo-
sition of cellulose fibrils with varying orientations in the secondary fiber walls commences. The
growth process of cotton fibers begins shortly after anthesis, the flowering stage. The fibers in-
crease in length during this stage and elongate for three to four weeks. Two weeks after the an-
thesis, the deposition of cellulose fibrils with varying orientations in the secondary fiber walls
commences. This process reduces the inner space's size, known as the lumen, within the fiber (see
Figure 1.16) [96-98].
20
1.4 Model systems
Waxes
Primary walls
Secondary walls
Lumen
Figure 1.16: Schematic illustration of the structure of cotton fiber, showing its different layers.
Fiber growth and development primarily depend on plant photosynthesis and carbohydrate pro-
duction, which are impacted by various factors. Temperature and water status are two important
factors that can affect fiber growth, length, and quality. Studies have shown that temperature and
plant water status impact fiber growth and length. Nighttime temperatures below 22.0 °C inhibit
cellulose synthesis and deposition rate in the cotton fiber walls, leading to decreased productivity
and inferior fiber quality. Low temperatures during fiber maturation can also result in sucrose
accumulation, leading to "sticky cotton," a severe issue that negatively affects cotton quality and
value [99].
Under normal conditions, cotton consists of approximately 95% cellulose, while the remaining
5% consists of various substances, including sugar, wax, proteins, organic acids, and pectin [100].
One major quality issue of raw cotton is the impurity content after harvesting. The impurities
cause a significant economic loss because low-quality cotton is rejected during quality control.
The most relevant impurities in raw cotton arise from insects and are summarized under the um-
brella term “Honeydew” [101].
1.4.4 Honeydew
Cotton contaminated by sugar causes significant problems for textile equipment. These sugars are
produced by the cotton plant (physiological sugars) or insects (entomological sugars). Entomo-
logical sugars, also known as honeydew, are the most common source of sticky cotton (see Figure
1.17 a,b) [102,103].
21
1 Introduction
(a) (b) (c)
(d)
Figure 1.17: (a) and (b) Cotton fiber contaminated by sugar cause of (c) aphid and (d) Whiteflies insects these photos
are taken from [104,105].
Honeydew is excreted by aphids and whiteflies (Figure 1.17 c,d) and contains mainly trehalulose,
trehalose, melezitose, sucrose, fructose, and glucose (see Figure 1.18). These sugars vary in stick-
iness, such as sucrose, melezitose, and trehalulose are significantly stickier when deposited on
fiber than glucose or fructose. Furthermore, fibers contaminated with trehalulose are more sticky
than fibers contaminated with an equivalent amount of melezitose [101,106,107].
(a) Trehalulose (b) Trehalose (c) Melezitose
(d) Sucrose (e) Fructose (f) Glucose

Figure 1.18: Honeydew chemical structure contents (a) Trehalulose (b) Trehalose (c) Melezitose (d) Sucrose (e) Fruc-
tose (f) Glucose.
Cotton contaminated with a high amount of honeydew becomes sticky. Transferring sticky cotton
to a spinning machine causes severe problems by contaminating all mechanical components. This
can cause damage to the machines, and the final yarn is of lower quality [101,106]. In Figure 1.19,
the result of sticky cotton on a draw frame roll can be seen. Consequently, yarn from sticky raw
cotton reaches a lower quality and achieves a low price on the market [108].
22
1.4 Model systems
Figure 1.19: sticky cotton residue on a draw frame roll; this photo is taken from reference [102].
The stickiness of raw cotton depends not only on the amount of honeydew on the fiber but also
on the ambient conditions such as humidity and machine temperature. The stickiness can be sig-
nificantly reduced if the moisture content is increased [101]. One obvious idea is washing the raw
cotton to reduce the sugar content and thereby reduce the stickiness. It was not a suitable solution
because it consumes much water and requires an additional drying step. It makes the whole pro-
cess even more difficult [109-111]. The best and most economical solution seems to be blending
cotton with different honeydew contents to obtain an optimized blend [112]. Since the sugar com-
position determines cotton's stickiness, identifying the sugars is mandatory to optimize the blend-
ing process. If the sugar composition is known in detail, the blending process can be adapted to
the climate conditions (temperature, humidity) where the following production steps will occur
[101]. To reach this goal, fast on-line analytics is required to monitor the cotton quality with high
resolution in honeydew content and time.
In recent years, the increase in quality and processing requirements has led to the introduction of
modern techniques for processing and quality control [15,113-115]. Currently, the classification of
cotton is done by the United States Department of Agriculture (USDA) system [116]. It classifies
the quality by applying the High Volume Instrument (HVI). HVI is the standard instrument for
measuring cotton quality in the USA and has been adopted by other countries as it provides a
range of standards for measuring cotton quality, including micronaire, strength, length, color,
foreign cotton, and material [117-120]. HVI cannot measure the quality of single fibers and short
fibers. Therefore, another system, the Advanced Fiber Information System (AFIS), can be used.
Although both HVI and AFIS measure a set of quality characteristics, they are not able to quantify
and distinguish between cotton species, which is expensive and time-consuming [121]. To solve
this problem, spectroscopic methods have been developed and applied to identify and classify
different cotton fiber varieties in Vis/NIR region.
23
2 Objective
2 Objective
The present work aims to classify different types of cotton and determine the honeydew contents
on real cotton samples by developing hyperspectral imaging in the UV region. Traditional meth-
ods, such as fiber quality index (FQI), the spinning consistency index (SCI) and the premium-
discount index (PDI) used for this determination are not useful for large amounts of cotton bales.
In contrast, optical spectroscopy, UV-Vis/NIR spectroscopy, and hyperspectral imaging, espe-
cially UV hyperspectral imaging, are able to distinguish between different cotton types and hon-
eydew contaminated on cotton. The data evaluation, where a correlation between the spectro-
scopic and sensory data has to be established, is a special challenge. However, this can be
overcome by multivariate data analysis, for example, to evaluate the spectra in connection with
the results of the amount of honeydew contaminated on cotton sample measurement. Subse-
quently, the obtained information is linked by PCA and PLS-R. This is necessary because the
information is usually superimposed and thus cannot be derived directly from the respective spec-
tra.
The analysis method described above is already widely used for scientific purposes. However,
developing a method suitable for in-line analysis and monitoring industrial manufacturing pro-
cesses in the UV region has not been possible. The integration of the analysis method into the
manufacturing processes offers an enormous optimization potential of the existing value chain
since the time-consuming sample taking and preparation for an off-line measurement is no longer
necessary due to the in-line measurement. Therefore, developing UV hyperspectral imaging is
mandatory. The investigation of the effectiveness of such a prototype is tested by using standard
samples such as active pharmaceutical ingredients (APIs) as well as direct bonded copper (DBC).
To achieve this purpose, a calibration model is first developed using PLS-R. This model is then
used to build a prediction model for estimating the APIs content and oxide layer thickness on
DBC samples from the UV hyperspectral image. These experiments have demonstrated that UV
hyperspectral imaging can effectively identify and categorize pharmaceutical and DBC samples.
Based on this information, it is now possible to focus on the actual question. The question if
different concentrations of honeydew contaminated on cotton can be determined by UV hyper-
spectral imaging, with the necessary accuracy and safety, on-line or in-line during manufacturing
processes, should be answered. Therefore, PCA and PLS-R models are developed for honeydew
content. Finally, hyperspectral imaging is compared to conventional UV-Vis spectroscopy in
terms of prediction accuracy.
24
3 Paper I: Characterization of
Pharmaceutical Tablets Using UV
Hyperspectral Imaging as a
Rapid In-Line Analysis Tool
Mohammad Al Ktash1,2, Mona Stefanakis1,2, Barbara Boldrini1, Edwin Ostertag1 and Marc
Brecht1,2,*
1
Process Analysis and Technology PA & T, Reutlingen University, Alteburgstraße 150,
72762 Reutlingen, Germany
2
Institute of Physical and Theoretical Chemistry, Eberhard Karls University Tübingen,
Auf der Morgenstelle 182, 72076 Tübingen, Germany
*
Correspondence: [email protected]
This is originally published in sensors (https://doi.org/10.3390/s21134436) as
“Al Ktash, M.; Stefanakis, M.; Boldrini, B.; Ostertag, E.; Brecht, M. Characterization of Phar-
maceutical Tablets Using UV Hyperspectral Imaging as a Rapid In-Line Analysis Tool. Sen-
sors 2021, 21, 4436. https://doi.org/10.3390/s21134436”
25
3 Paper I: Characterization of Pharmaceutical Tablets Using UV Hyperspectral Imaging as a Rapid In-Line
Analysis Tool
3.1 Abstract
A laboratory prototype for hyperspectral imaging in ultra-violet (UV) region from 225 to 400
nm was developed and used to rapidly characterize active pharmaceutical ingredients (API)
in tablets. The APIs are ibuprofen (IBU), acetylsalicylic acid (ASA) and paracetamol (PAR).
Two sample sets were used for a comparison purpose. Sample set one comprises tablets of
100% API and sample set two consists of commercially available painkiller tablets. Reference
measurements were performed on the pure APIs in liquid solutions (transmission) and in solid
phase (reflection) using a commercial UV spectrometer. The spectroscopic part of the proto-
type is based on a pushbroom imager that contains a spectrograph and charge-coupled device
(CCD) camera. The tablets were scanned on a conveyor belt that is positioned inside a tunnel
made of polytetrafluoroethylene (PTFE) in order to increase the homogeneity of illumination
at the sample position. Principal component analysis (PCA) was used to differentiate the hy-
perspectral data of the drug samples. The first two PCs are sufficient to completely separate
all samples. The rugged design of the prototype opens new possibilities for further develop-
ment of this technique towards real large-scale application.
3.2 Introduction
A large number of remote sensing applications have been developed over the last decade
[122]. This also led to establish non-destructive imaging systems that are able to quickly iden-
tify quality problems within the scanned area [22,23]. Spectral imaging involves both spectral
and spatial information of any particular sample or region within an area of interest, thus each
pixel represents spectral and spatial information. Imaging systems can be realized in the
modes of hyperspectral and multispectral imaging. The difference between these modes is the
number and width of the recorded spectral bands. In multispectral imaging 3-10 bands are
used [123]. In hyperspectral imaging hundreds or thousands of correspondingly more narrow
bands are employed [18,29,86,124]. Therefore, hyperspectral imaging is also known as imag-
ing spectroscopy, a technique that combines conventional imaging with spectroscopy [86].
Hyperspectral imaging setups produce a 3D data matrix often referred to as hypercube. Two
of the dimensions are reserved for the spatial information (x, y coordinate) while the third
dimension represents the spectroscopic information (λ coordinate) [21,125,126].
Hyperspectral imaging is not restricted to the visible range, nowadays high performance sys-
tems are also available for the near infrared range (NIR) [35,86]. Hyperspectral imaging is a
26
3.2 Introduction
rapid and non-destructive method which analyzes samples without changing their physical
shape. This robust technique in combination with real-time chemometric analysis can be eas-
ily integrated into an industrial production environment. This enabled chemical sensing sys-
tems for very different applications in the fields of food quality monitoring, textile classifica-
tion, agriculture, detection target of military, astronomy, life science, medicine and
pharmaceutical drugs [22,23,25,26,127,128]. Traditional methods such as UV-Vis spectros-
copy, high performance liquid chromatography (HPLC) or mass spectrometry (MS) are, in
contrast, time consuming, expensive and require sample preparation and destruction [22,86].
Very recently, Tschannerl et al. reported an interesting application of hyperspectral imaging
in UV range. They were able to precisely discriminate between phenolic flavor concentrations
in melted barley by using hyperspectral imaging in UV and NIR regions [86].
In 2004, the food and drug administration (FDA) in the US started the Process Analytical
Technology (PAT) initiative to control manufacturing processes [22]. Hyperspectral imaging
is an attractive PAT tool for the quality assurance of final products. Hyperspectral imaging,
as expected, will be increasingly used as a PAT tool in the industry; it has been already applied
in the industrial manufacturing of pharmaceutical drugs and quality control of pharmaceutical
products [17,129]. Most drugs appear colorless to the eye, meaning that they do not absorb
light in the visible region but they may absorb in the UV region according to the chemical
structure [130]. Such drugs as ibuprofen (IBU), acetylsalicylic acid (ASA) and paracetamol
(PAR) show certain absorbance in UV region [131]. Up to now, a variety of drug studies in
the UV-Vis region have been performed; Saeed et al. investigated the active pharmaceutical
ingredients paracetamol, aspirin, ibuprofen, codeine and caffeine in different formulations by
UV-Vis spectroscopy [132]. Rote et al. developed a method to simultaneously quantify para-
cetamol and nabumetone by area under curve in bulk and tablet dosage form [133].
Hyperspectral imaging setups acquire thousands of spectra in short time resulting in a massive
amount of data. Therefore, techniques for data evaluation like the principle component anal-
ysis (PCA) methods are required. PCA is one of the most common statistical methods. This
technique is used for data evaluation/reduction but simultaneously minimizing information
loss in spectroscopy [38,134]. In addition, it is capable of visualizing common features in the
data set to detect possible groups and their heterogeneity within samples [135]. PCA combined
with hyperspectral imaging data can highlight the relative distributions of different compo-
nents in mixtures and reveal the spectral features in the spectroscopic data [67,127].
27
Analysis Tool
The aim of this study is to develop a hyperspectral imaging system in the UV wavelength
range for the in-line characterization of pharmaceutical tablets. The results show that hyper-
spectral imaging in the UV range is a suitable technique for in-line measurements with the
aim of a real-time classification at short time and low costs.
3.3 Materials and Methods
3.3.1 Samples
Two groups of samples were analyzed. Figure 3.1 shows photographs of all tablets used.
Further information is listed in Table 3.1. In the following, these samples are referred to as
IBUpure, ASApure, PARpure, IBUratio, IBUbeTa, ASPBAYER, PARratio and THO. For hyperspectral
imaging measurements, the coating of the commercially available painkiller tablets was re-
moved by sandpaper manually (grain size 320, Emil Lux GmbH & Co. KG, Wermelskirchen,
Germany). For each removal step a new stripe was brushed over it twice. The painkiller tab-
lets were measured at different depths. A layer of approximately 500 μm ± 50 μm thickness
was removed from the samples after each measurement. Three samples of each type were
collected (painkiller samples) and created (pure API samples) for the study.
Figure 3.1: Drug samples. Reference API samples and painkiller table.
28
Table 3.1: Types of drug samples

Abbre- CAS
Samples Descriptions Manufacturer
viation Number
Caesar & Loretz
Ibuprofen, 15687-
Ibuprofen IBUpure GmbH, Hilden, Ger-
>98%, API 27-1
many
Acetylsa- Acetylsalicylic acid, 99%, Acros organics, New
ASApure 50-78-2
licylic acid API Jersey, US
Hebei Jiheng (Group)
Paraceta-
Paracetamol, 99%, API PARpure Pharmaceutical Co., 103-90-2
mol
Ltd.
Ratiopharm GmbH,
Ibuprofen Ibuprofen (400 mg) IBUratio -
Ulm, Germany
Betapharm, Arzneimit-
Ibuprofen Ibuprofen (400 mg) IBUbeTa tel GmbH, Augsburg, -
Germany
Acetylsalicylic acid (500 AS- Bayer Vital GmbH, Le-
Aspirin -
mg) PBAYER verkusen, Germany
Paraceta- Ratiopharm GmbH,
Paracetamol (500 mg) PARratio -
mol Ulm, Germany
Thomapyrin (250 mg acetyl-
Thomapy- Sanofi-Aventis GmbH,
salicylic acid/ paracetamol, THO -
rin Frankfurt, Germany
50 mg coffin)
In total, 4 g of ASApure and IBUpure were pressed at 10 tons for 2 min by a hydraulic press
(PerkinElmer, Inc., Waltham, MA, USA) into the depicted disc shape. Then, 4 g PARpure
powder were dried in a vacuum oven (VACUTHERM, Thermo Scientific, Waltham, MA,
USA) for 1 h at 120 °C, and pressed at 10 tons for 20 min (see Table 3.1). A mixture of 2 g
ASApure and 2 g PARpure was prepared by using a SpeedMixerTM (DAC 150.1 CM41,
Hauschild GmbH & Co KG, Hamm, Germany), and pressed at 10 tons for 2 min.
3.3.2 API’s in Solution

A solution of ASApure (100 µg mL−1) was prepared by dissolving 50 mg ASApure in 500 mL
of 0.1 M HCl:methanol (1:1) in 500 mL volumetric flask with strong shaking.
For PARpure and IBUpure solutions, 10 mg of each API were dissolved in 15 mL methanol by
shaking. Then, 85 mL water was added to adjust the volume up to 100 mL (resulting to 100
ppm). From that, 5 mL were taken, and volume was adjusted up to 50 mL with diluent [132].
29
Analysis Tool
3.3.3 UV Spectroscopy
Total (specular and diffuse) reflectance spectra of all samples (pure API and painkiller tablets)
were recorded in the range of 200–380 nm using a commercial UV spectrometer (Lambda
1050+, PerkinElmer, Inc., Waltham, MA, USA). Both sides of the pure API samples were
measured. The UV-Vis/NIR spectrometer was equipped with a 150 mm Spectralon® integrat-
ing sphere to acquire data in reflection mode with an R6872-Photomultiplier (PMT). A deu-
terium lamp was used as light source in the spectrometer. The samples were placed at the
reflectance port of the integrating sphere with a diffused scattering Spectralon ® disk placed
behind the samples. The port measuring area is approximately 4.9 cm².
Absorbance spectra were measured using the aforementioned spectrometer in the range of
200–320 nm connected to the transmittance accessory. The liquid samples were measured at
2 nm spectral resolution. A 1 mm quartz SUPRASIL® cuvette (106-1-K-40, Hellma, Müll-
heim, Germany) was used for measuring the API’s in solution.
Fluorescence excitation spectra were recorded by using a commercial setup (Fluorolog–3,

HORIBA, Kyoto, Japan). The system includes a double grating monochromator in the exci-
tation (λEx = 270 nm) and emission (λEm = 280 nm–380 nm) paths in an “L” configuration. A
10 mm quartz SUPRASIL® cuvette (111-10-K-40, Hellma, Müllheim, Germany) was used
for measuring the samples.
3.3.4 UV Hyperspectral Imaging

Figure 3.2a shows a scheme of the hyperspectral imaging setup. The setup is based on a spec-
trograph (RS 50-1938, inno-spec GmbH, Nürnberg, Germany) connected to a CCD camera
(Apogee Alta F47: Compact, inno-spec GmbH, Nürnberg, Germany) with 300 ms integration
time. The samples were placed on a conveyor belt moving with speed 0.3 cm/s, which was
positioned completely in a tunnel made of PTFE. The purpose of the tunnel design is to have
an easily accessible system, which also ensures diffuse illumination of the samples and main-
tains a reasonable illumination strength and homogeneity. This minimizes an influence of the
sample shape and roughness on the spectra. The illumination is provided by a Xenon lamp
(XBO, 14 V, 75 W, OSRAM, München, Germany). Figure 3.2b–d illustrates the principle
and workflow of the data acquisition. The continuous line by line collection of spectral infor-
mation enables a lateral (x, y) 2D image as shown in Figure 3.2c, whereas each pixel contains
30
a further spectroscopic dimension (λ) as shown in Figure 3.2d. Thus, a 3D data matrix (hy-
percube) is recorded.
Pushbroom imager
(a)
Spectrograph Apogee camera
Cooling system
Tunnel
PTFE
Xenon lamp
Objective
(b) (c) (d)

0.6
- log (R/R0)
 0.4
0.2
x 250 300 350 400

Wavelength / nm
y
Moving direction Hyperspectral image Spectra at one single pixel
Figure 3.2: (a) Setup of a hyperspectral imaging system based on the pushbroom concept (the tunnel in the scheme
was cut to show the inside). (b) Pushbroom Imager scanning principle. (c) Hyperspectral image generated
immediately from the scanning of a sample. (d) UV spectrum for one single pixel extracted from the image
given in (c).
3.3.5 Data Collection and Preprocessing

Figure 3.3 shows the original images of the drug samples before and after background subtraction.
The UV hyperspectral images are captured by moving the drug samples at constant speed. For the
collection of UV hyperspectral imaging data set one sample of each type was chosen randomly.
A distinction between the respective spectral characteristics was made first to differentiate signal
and background. For this purpose, the regions assigned to the drug samples were manually se-
lected to eliminate the signals from background. The remaining hypercube was used as input for
the subsequent PCA classification.
31
Analysis Tool
Figure 3.3: Hyperspectral raw image of nine drug samples on the left. Images after subtraction of the background on
the right.
3.3.6 Data Handling and Software

The UV spectra were recorded with the Lambda 1050 UV WinLab software from PerkinElmer.
The UV hyperspectral imaging data were analyzed by the SI-Cap-GB version V3.3.x.0 software
(inno-spec GmbH, Nürnberg, Germany). Hyperspectral data matrices were analyzed by Pred-
iktera Evince version 2.7.11. PLS_Toolbox (PLS Toolbox 8.5.1, Eigenvector Research, Inc., WA,
USA) and MATLAB (MATLAB 9.2.0, Mathworks, MA, USA) were used for the data processing
and analysis. An initial baseline correction was followed by a Savitzky-Golay 1st derivative (15
points, 2nd polynomial order). PCA models were calculated with cross validation (venetian
blinds, 10 splits, 1 sample per split) and mean centering. A PCA combined with a quadratic dis-
criminant analysis (QDA, 2 PCs) was calculated by using the software Unscrambler X 10.5
(Camo Analytics AS, Oslo, Norway) including the same spectral preprocessing.
Lighting conditions may vary between the samples and even within the samples across the scan
line. A regular way to reduce this effect is to convert measured raw spectra to reflectance spectra
by the following formula [35,86]:
Isample − Idark
Reflectance = -log R/R0 = (3.1)
Ireference − Idark
where R and R0 represent the reflected intensity by the sample and a specific reference material
with high reflectance capability. Isample is the intensity of the original image data, Idark is the inten-
sity of the dark current image data and Ireference is the intensity of the white reflectance image [23].
For a better comparison of the reflectance spectra to the extinction spectra in solution (absorb-
ance) the negative decadic logarithm is calculated as -log R/R0.
32
3.4 Results and Discussion
There are numerous references for the APIs in solution [101,131,132,136] in the UV range, but for
solid API drug samples suitable references were not found. For this reason, first the liquid solu-
tions of the APIs were measured and then compared to the results found in the literature. In a
second step, samples in the solid phase, i.e., the pure API reference samples and the commercial
painkillers, were investigated.
3.4.1.1 APIs in Liquid Phase, Transmission Spectroscopy

The absorbance of IBUpure, ASApure, PARpure as well as a mixture of ASApure with PARpure in liquid
solution were analyzed in the UV range. Figure 3.4 shows their absorption spectra in the UV
region (200–320 nm). The smaller features of IBUpure and ASApure in the range of 240–300 nm
are shown in the inset in Figure 3.4. All samples show a strongly increasing absorbance below
310 nm. IBUpure presents one prominent maximum at 223 nm and three weaker maxima located
at approximately 258 (sh), 265 and 273 nm. ASApure exhibits a broad maximum at approximately
228 nm and a further, more pronounced but less intense maximum at around 277 nm. PARpure
shows a distinct band with a maximum at 244 nm and a weak shoulder at 284 nm. The mixture
of ASApure and PARpure presents a band maximum at 240 nm and a shoulder at 282 nm. These
findings are listed in Table 3.2 [131,132,136]. The determined band positions are consistent with
those reported by Saeed et al. (2016) and Lawson et al. (2017) [132,136].
0.8
PARpure
0.015
IBUpure
0.6
Absorbance
0.010
ASApure
Absorbance
0.005
IBUpure 0.000
0.4 -0.005
(ASA+PAR)pure 240 260 280 300
Wavelength / nm
0.2
ASApure
0.0
200 220 240 260 280 300 320
Wavelength / nm
Figure 3.4: UV absorbance spectra of APIs ibuprofen (IBU), acetylsalicylic acid (ASA), paracetamol (PAR) and a
mixture of acetylsalicylic acid and paracetamol (ASA+PAR) in liquid phase.
33
Analysis Tool
Table 3.2: UV band maxima positions of liquid and solid phase samples [131,132,137].
Drug Type
API Liquid Phase API Solid Phase Painkiller Tablets Solid Phase
223 nm
258 nm (sh) 240 nm 238 nm (IBUratio, IBUbeTa)
IBU
265 nm 275 nm 275 nm (IBUratio, IBUbeTa)
273 nm
228 nm (ASPBAYER)
230 nm
228 nm 280 nm (ASPBAYER)
ASA/ASP 277–310 nm
277 nm 294 nm (ASPBAYER)
328 nm (sh)
329 nm (ASPBAYER)
244 nm 232 nm 233 nm (PARratio)
PAR
284 nm 305 nm 300 nm (PARratio)
238 nm
THO - -
331 nm
240 nm 235 nm
ASA+PAR (mixture) -
282 nm 277–332 nm
3.4.1.2 API and Painkiller Tablets, Total Hemispherical Reflectance Spectroscopy

Two sample sets of tablets were used to study the total hemispherical reflectance in the solid
phase (see Figure 3.1 and Table 3.1). The first set consisted of pure APIs: IBUpure, ASApure and
PARpure and a mixture of ARApure and PARpure. The second set consisted of commercial painkiller
tablets. Three samples from each API were prepared and analyzed. Figure 3.5 shows the prepro-
cessed reflectance spectra of solid samples in the UV region (200–380 nm). Spectra were recorded
from each side of the samples (Figure 3.5a).
(a) (b)
1.0 PARpure Fluorescence 1.0 PARratio
(ASA+PAR)pure 1.2x108
THO
Intensity [cps / μA]
0.8 ASApure 0.8 ASPBAYER

8.0x107
- log (R/R0)
- log (R/R0)
IBUpure IBUbeTa
0.6 4.0x107
0.6 IBUratio
0.0
280 300 320 340 360 380
0.4
Wavelength / nm 0.4
0.2
0.2
0.0
0.0
200 220 240 260 280 300 320 340 360 380 200 220 240 260 280 300 320 340 360 380
Wavelength / nm Wavelength / nm
Figure 3.5: UV total hemispherical reflectance spectra of drug samples in the solid phase in the wavelength range 200–
380 nm. (a) API drugs IBUpure, ASApure, PARpure and a mixture of ASApure with PARpure. Upper right:
Fluorescence emission of IBU sample with excitation at 270 nm. (b) Painkiller tablets IBUratio, IBUbeTa,
ASPratio, PARratio and THO.
The most striking feature is the negative reflectance of IBUpure in the wavelength range of 288–
340 nm, which is due to fluorescence emission (inset in Figure 3a). All spectra show several
contributions, which are listed in Table 4.2.
34
The painkiller tablets were measured at different depth levels, i.e., one layer of approximately 500
µm was removed from the samples after each measurement; the resulting spectra are shown in
Figure 3b. The similarity of the spectra at all depth levels inside the tablets indicates an almost
regular distribution of ingredients. Although IBUratio and IBUbeta were manufactured from differ-
ent companies, they show similar spectral characteristics. The most prominent contributions are
also listed in Table 3.2.
The comparison of the spectra from the APIs and commercial painkiller tablets indicates that the
overall spectral characterizations are comparable. Nevertheless, several deviations are observed.
The spectral features are more pronounced in the API samples, also the negative absorbance ob-
served in the IBU sample is absent in the commercial painkillers. The reason for these differences
is mainly that the commercial tablets do not have 100% API content. For example, the IBUratio
tablets contains additionally pregelatinized corn starch, hypromellose, croscarmellose sodium,
stearic acid, highly dispersed silicon dioxide, macrogol 8000, titanium dioxide. Some of these
substances show some absorption in selected spectral range i.e., titanium dioxide shows a pro-
nounce absorbance [138]. Since the exact percentage of the composition is not known, a final
statement on the influence of these substances on the spectra cannot be made.

3.4.2.1 API Tablets, Hyperspectral Imaging
Figure 3.6 shows the results of UV hyperspectral imaging in the range from 225 to 400 nm. Figure
4.6a shows the raw image before (left) and after subtraction of the background (right). Figure 3.6b
shows a spectrum of an arbitrary but representative pixel for each API sample. The most dominant
contribution for IBUpure is observed around 275 nm and for PARpure at around 305 nm. For
ASApure, two strong contributions at 300 and 330 nm are observed. The mixture of
(ASA+PAR)pure shows—as expected—a combination of the spectral properties of ASApure and
PARpure. In the range 255–270 nm, all API preparations show a small peak in their reflectance at
around 265 nm. Towards lower wavelengths, the spectra show no additional features.
In the next step, a PCA model with cross validation (venetian blinds, 10 splits, 1 sample per split)
was calculated for the spectra of all preparations. The first two PCs ex-plain 98.9% of the total
variance. Figure 3.6c shows the scores plot of the PC1 and PC2. The scores plot shows that PC1
and PC2 are sufficient to separate all samples clearly from one another. PC1 yields a clear sepa-
ration of IBUpure from the other APIs, whereas the remaining APIs are separated with PC2. The
mixture (ASA+PAR)pure is found almost in the middle between ASApure and PARpure.
35
Analysis Tool
The loadings plot for PC1 and PC2 is given in Figure 3.6d. The loading of PC1 is dominated by
an overall positive contribution in the range between 280 and 350 nm, PC2 shows one more nar-
row negative contribution at 311 nm and one positive at 330 nm.
(a) (b)
0.4
IBUpure ASApure
0.3 (ASP+PAR)pure
- log (R/R)0
PARpure
ASApure 0.2
IBUpure
PARpure 0.1
(ASP+PAR)pure 0.0
225 250 275 300 325 350 375 400
Wavelength / nm
(c) (d)
0.10 PC1[90.67%]
1.5 ASApure PC2[8.22%]
0.08
1.0
0.06
PC2 [8.2%]
0.5 0.04
IBUpure (ASA+PAR)pure
0.0 0.02
PC1
0.00
-0.5
PARpure -0.02 PC2
-1.0 -0.04
-1.5 -0.06
-4 -2 0 2 4 225 250 275 300 325 350 375 400

PC1 [90.7%] Wavelength / nm
Figure 3.6: (a) Raw hyperspectral image for all API drug samples before and after subtracting the background. (b)
Spectrum recorded for a single pixel of each of pure API samples in the UV range 225–400 nm. (c,d)
Scores and corresponding loadings plot.
The comparison between the shape of spectra shown in Figure 3.5a or Figure 3.6b shows similar-
ities as well as some clear deviations. The shape of the spectra of all APIs is quite well reproduced
in the range above 275 nm. Most striking in this range is an intensity deviation of the different
spectral bands, i.e., for the ASApure the shoulder at 330 nm is much more pronounced in the hy-
perspectral imaging spectra. The same is valid also for the PARpure sample. In the range below
275 nm clear deviations are observed. The shape of the spectra shown in Figure 3.6b is charac-
terized by a continuously decreasing intensity, whereas the spectra of the APIs in Figure 3.5a
show clear variations in their shape in this range (see also below).
Figure 3.5 presents UV spectra with a good signal-to-noise-ratio recorded with a research grade
UV desktop spectrometer. Here, the measured area for one spectrum consisted of 12 × 5 mm.
Figures 3.6 and 4.7 show UV spectra with a less good signal-to-noise-ratio recorded with the UV
36
hyperspectral imager. These spectra result from one single pixel of the detector, representing a
much smaller area of the tablet which considered of 13 × 13 µm. A further reason for the low
signal-to-noise-ratio is weak irradiation intensity in the hyperspectral imaging setup by an XBO
lamp.
3.4.2.2 Commercial Painkiller Tablets, Hyperspectral Imaging

Figure 3.7 shows results from the hyperspectral imaging of the second sample set, consisting of
commercial painkiller tablets, in the range from 225 to 400 nm. Figure 3.7a shows the raw image
before (left) and after (right) subtraction of the background. Figure 3.7b shows a spectrum of an
arbitrary but representative pixel for each painkiller samples. For IBUratio and IBUbeTa, the most
dominant contribution is observed around 270 nm, and a further contribution with much lower
intensity is observed at around 315 and 333 nm. The spectra of IBUratio and IBUbeTa are quite
similar; it seems that the contributions from the further added chemical ingredients are spectro-
scopically comparable. The spectrum for ASPBAYER is dominated by a broad intensity distribution
at around 304 nm. Here, two contributions of different intensity are specifiable, a more intense
with maximum at 304 nm and a weaker one at 333 nm. For PARratio, only one strong contribution
with maximum at 310 nm is observed; whereas THO shows two contribution of different inten-
sity, a more in-tense with maximum at 304 nm and a weaker one at 333 nm. In the range 225–
275 nm, all painkiller samples show a minor peak in their reflectance at around 270 nm. Towards
lower wavelength, the spectra show no additional features.
The shape of the spectra of the commercial painkiller match those of the APIs (Figure 3.6b) quite
well. Slight deviations are most likely due to additional ingredients in the commercial samples.
37
Analysis Tool
(a) (b)
0.6
IBUratio ASPBAYER
0.5
PARratio
IBUbeTa
- log (R/R)0
0.4
THO
0.3 IBUratio
ASPBAYER IBUbeTa
0.2
PARratio
0.1
THO
225 250 275 300 325 350 375 400
Wavelength / nm
(c) (d)
1.5 0.08 PC1[94.1%]
ASPBAYER PC2[5.3%]
1.0 0.06
0.04
PC2 [5.3%]
0.5 IBUratio
THO 0.02 PC2
0.0
0.00
-0.5 IBUbeTa -0.02 PC1
PARratio
-1.0 -0.04
-0.06
-1.5
-6 -4 -2 0 2 4 6 225 250 275 300 325 350 375 400
Figure 3.7: (a) Raw hyperspectral image for all commercial painkiller tablets before and after subtracting the back-
ground. (b) Spectrum recorded for a single pixel of each painkiller tablet in the UV range 200–400 nm.
(c,d) Scores and corresponding loadings plot.
Figure 3.7d shows the loadings plot for PC1 and PC2. PCA model was calculated by cross vali-
dation (venetian blinds, 10 splits, 1 sample per split). The loading of PC1 is dominated by an
overall positive contribution in the range between 275 and 350 nm, whereas PC2 shows one more
narrow negative contribution at 308 nm and one positive at 327 nm. The distribution of the clus-
ters in the scores plot in Figure 3.7c shows a comparable variability with the scores plot in Figure
3.6c. Only for the THO sample an increased spreading is observed along PC2. In general, such
type of variability in the shape of the cluster can arise for several reasons; a change in the sample’s
properties on the scale of the resolution actually achieved, or changes due to shape effects of the
samples or positioning within the hyperspectral imaging setup. A general reason for deviations
between the hyperspectral imaging and UV-Vis spectroscopy (see Figure 3.5 vs. Figure 3.6b or
Figure 3.7b) are the different geometries used for illumination and detection in the setups. In the
UV spectrometer the light is collected in an almost perfectly reflecting integrating sphere, while
in case of the UV hyperspectral imaging, a tunnel made of PTFE is used for illumination and
collecting as shown in Figure 3.2a. As a consequence, a clear differentiation between specular
38
and diffuse reflection is not possible in the hyperspectral imaging setup, therefore, a mixture of
both contributions will be detected here.
Comparing the hyperspectral imaging spectra in Figure 3.6b or Figure 3.7b with the spectra given
in Figure 3.5 it is clear that the hyperspectral imaging data provide almost no useful spectroscopic
information in the region < 275 nm. The low performance in this range is due to the efficiency of
detector and the illumination in the hyperspectral imaging setup. A further consequence of this is
that contributions at higher wavelengths appear more dominant as they actually are. The tendency
of increasing sensitivity exists for the entire wavelength range. This is also why the shoulders
observed in the spectra of ASA/ASP and IBU at > 25 nm (in both sample sets) appear much more
enhanced compared to the spectra in Figure 3.5. As a consequence, the actual hyperspectral im-
aging setup yields valuable results for all samples, but reliable spectroscopic information is only
accessible in the range above 275 nm, and there, attention must be paid to the relative intensities.
Despite the spectroscopy weaknesses, the combination of UV hyperspectral imaging and chemo-
metric modeling enables a complete separation of all samples in both sample sets. The loadings
plots (Figure 3.6c or Figure 3.7c) indicate that a differentiation of all samples is possible consid-
ering only a few spectral channels, so that rapid classification is easily possible.
In order to validate the pure API PCA model (see Figure 3.6), the scores of PC1 and PC2 were
used to calculate a quadratic discriminant analysis (QDA). The confusion matrix resulted from
this model is listed in Table 3.3. A confusion matrix describes the performance of the classifica-
tion model based on QDA. An overall accuracy for the pure API tablets of 99.8% is reached,
which means the model can correctly classify approximately all spectra of the pure API tablets.
The highlighted diagonal describes how many spectra were predicted by the model as true. Only
19 spectra of (ASA+PAR)pure were predicted as PARpure and two spectra of PARpure as
(ASA+PAR)pure. This is because (ASA+PAR)pure contains both API components (ASApure, PAR-
pure).
Table 3.3: The confusion matrix of the pure API spectra.

Predicted
API Samples IBUpureASA purePARpure(ASA+PAR)pure
IBUpure 2365 0 0 0
Actual
ASA pure 0 2428 0 0

PARpure 0 0 2574 2
(ASA+PAR)pure 0 0 19 2586
This QDA model was used to classify all spectra of the painkiller tablets. Even 99.8% of the
spectra were predicted correctly (see Table 3.4). This means approximately all painkiller tablets
were predicted correctly in true API classes. Only two spectra of ASP BAYER were assigned as
39
Analysis Tool
(ASA+PAR)pure. This is because (ASA+PAR)pure contains both API components (ASApure, PAR-
pure).
Table 3.4: Classification of the painkiller tablets based on the pure API model.
Predicted
Samples IBUpureASP purePARpure(ASA+PAR)pure
IBUratio/IBUbeTa 469 0 0 0
Actual
ASP BAYER 0 283 0 2

PARratio 0 0 209 0
THO 0 0 0 394
The UV region is often preferred in process control and quality assurance, but hyperspectral im-
aging in this region is rarely reported. The aim of this study was to develop a simple UV hyper-
spectral imaging setup capable of distinguishing between different drug samples as an example
for a possible industrial application. With the prototype, a painkiller table can be measured at 4 s.
This speed is adequate for scientific purposes, but too low for industrial applications. The limiting
factor towards a setup for a production environment is the intensity of the illumination and the
quantum yield of the pushbroom imager. With an appropriate light source and imager then this
setup is capable for in-line data acquisition, process control, in-line classification/sorting, and thus
real-time release testing.
3.5 Conclusions
UV hyperspectral imaging was used to characterize active pharmaceutical ingredients in tablets.
Two sample sets were analyzed; sample set one consisted of tablets with 100% API content and
sample set two consisted of commercially available painkiller tablets. Reference measurements
were performed on the pure APIs in liquid solutions and in solid phase using a commercial UV
spectrometer.
Hyperspectral imaging in combination with PCA is a promising approach for the detection and
differentiation of all drug samples studied. The PCA model was able to separate all drug types
with the first two principle components. The advantage of the home-built setup is a high spa-
tial/spectral resolution and a data acquisition speed completely sufficient for scientific studies.
Based on the design and the data shown, a setup fulfilling the requirements of a real industrial
process can be easily realized.
40
3.5 Conclusions
Author Contributions: Conceptualization, E.O. and M.B.; methodology, M.A.K., M.S., B.B.
and E.O.; software, M.A.K, M.S. and B.B.; validation, M.A.K., M.S., B.B. and E.O.; formal anal-
ysis, M.A.K. and M.S.; investigation, M.A.K. and M.S.; resources, M.B.; data curation, M.A.K.,
M.S. and B.B.; writing—original draft preparation, M.A.K. and M.B.; writing—review and edit-
ing, M.A.K., M.S., B.B., E.O. and M.B. ; visualization, M.A.K and B.B.; supervision, M.B.; pro-
ject administration, M.B.; funding acquisition, M.A.K., E.O. and M.B.
Funding: Mohammad Al Ktash acknowledges the support of Katholischer Akademischer

Ausländer-Dienst (KAAD).
Data Availability Statement: The data presented in this study are available on request from the
corresponding author. The data are not publicly available due to privacy restrictions.
Acknowledgments: The authors thank Karsten Rebner, Tim Bäuerle and Tobias Drieschner for
valuable discussions.
Conflicts of Interest: The authors declare no conflict of interest.
41
4 Paper II: UV Hyperspectral
Imaging as Process Analytical
Tool for the Characterization of
Oxide Layers and Copper States
on Direct Bonded Copper
Mohammad Al Ktash1,2,†, Mona Stefanakis1,2,†, Tim Englert3,4, Maryam S. L. Drechsel1, Jan
Stiedl3, Simon Green3, Timo Jacob4, Barbara Boldrini1, Edwin Ostertag1, Karsten Rebner1 and
Marc Brecht1,2,*
1
Process Analysis and Technology PA & T, Reutlingen University, Alteburgstraße 150,
72762 Reutlingen
2
Institute of Physical and Theoretical Chemistry, Eberhard Karls University Tübingen,
Auf der Morgenstelle 18, 72076 Tübingen, Germany
3
Robert Bosch GmbH, Automotive Electronics, Postfach 1342, 72703 Reutlingen, Germany
4Institute of Electrochemistry, Ulm University, Albert-Einstein-Allee 47, 89081 Ulm, Ger-
many
*
†
These authors contributed equally to the work.
“Al Ktash, M.; Stefanakis, M.; Englert, T.; Drechsel, M.S.L.; Stiedl, J.; Green, S.; Jacob, T.; Bol-
drini, B.; Ostertag, E.; Rebner, K.; Brecht, M. UV Hyperspectral Imaging as Process Analytical
Tool for the Characterization of Oxide Layers and Copper States on Direct Bonded Copper. Sen-
sors 2021, 21, 7332. https://doi.org/10.3390/s21217332”
43
4 Paper II: UV Hyperspectral Imaging as Process Analytical Tool for the Characterization of Oxide Layers and
Copper States on Direct Bonded Copper
4.1 Abstract
Hyperspectral imaging and reflectance spectroscopy in the range from 200–380 nm were used to
rapidly detect and characterize copper oxidation states and their layer thicknesses on direct
bonded copper in a non-destructive way. Single-point UV reflectance spectroscopy, as a well-
established method, was utilized to compare the quality of the hyperspectral imaging results. For
the laterally resolved measurements of the copper surfaces an UV hyperspectral imaging setup
based on a pushbroom imager was used. Six different types of direct bonded copper were studied.
Each type had a different oxide layer thickness and was analyzed by depth profiling using X-ray
photoelectron spectroscopy. In total, 28 samples were measured to develop multivariate models
to characterize and predict the oxide layer thicknesses. The principal component analysis models
(PCA) enabled a general differentiation between the sample types on the first two PCs with
100.0% and 96% explained variance for UV spectroscopy and hyperspectral imaging, respec-
tively. Partial least squares regression (PLS-R) models showed reliable performance with R2c =
0.94 and 0.94 and RMSEC = 1.64 nm and 1.76 nm, respectively. The developed in-line prototype
system combined with multivariate data modeling shows high potential for further development
of this technique towards real large-scale processes.
4.2 Introduction
Copper is considered as one of the most important conductors for integrated circuit (IC) packaging
and wire bonding. It has significant advantages in comparison to other materials (e.g., aluminum)
and is thus a good alternative for smaller structures. Copper as a metal has a high mechanical
stability and excellent electrical and thermal conductivities at low cost [139]. However, copper
contact surfaces contaminate and interact with oxygen to copper (I) oxide (Cu2O) and copper (II)
oxide (CuO) layers. This process is considered a problem as it influences the conductivity effi-
ciency. Science and engineering progress has driven the development of sensor technology in the
past years [122,140]. This led to novel optical sensors, such as hyperspectral imagers, to identify
quality problems [20,141].
Hyperspectral imaging is a technique that integrates a conventional spectroscopic system with

imaging in order to acquire spectral and spatial information from the area of interest [21,39,86,142].
Therefore, hyperspectral imaging enables quantitative analysis with improved levels of accuracy
[19,23,143]. It is considered as a rapid, non-destructive and robust method. Combining spectral
44
4.2 Introduction
imaging with chemometric algorithms opens up new industrial applications, including manufac-
turing process control [144]. Such spectral imaging systems are used in different fields, such as
food, pharmaceutical and textile production, as well as agriculture, military, astronomy, life sci-
ences and medicine [20,25,26,128,141,145].
Hyperspectral imaging is able to capture images in different spectral bands, such as in the visible
(Vis), infrared (NIR) and ultraviolet (UV) range. In contrast, traditional methods, such as Auger
electron and X-ray photoelectron spectroscopy (XPS), which are used to analyze copper samples,
are time consuming, expensive and require sample preparation and destruction [87,146]. The in-
dustry demands a high lateral resolution, which cannot be fulfilled by single-point UV-Vis spec-
troscopy [19,147]. Several detection methods have been developed to classify and identify the
copper state and copper oxide layers. In the past, UV-Vis/NIR spectroscopic applications as well
as Vis/NIR hyperspectral imaging have been preferred in the industrial environment, especially
for copper and other metal conductors [26,87,146,148-150]. The detection and characterization of
oxide layers on metallic copper samples was studied by Stiedl et al. using visible hyperspectral
imaging and UV-Vis spectroscopy. They were able to detect the thickness of the oxide layers on
the technical copper [26,87].
Recently, Tschannerl et al. have shown the application of hyperspectral imaging in the UV range
to discriminate between phenolic flavor concentrations in melted barley [86]. In another recently
published study, Al Ktash et al. have developed this technology in the direction of real applica-
tions. The authors were able to precisely classify between different active pharmaceutical ingre-
dients (API) and painkiller tablets by using an UV hyperspectral imaging prototype [19].
Hyperspectral imaging collects information in three dimensions (x, y, λ), resulting in a massive
number of variables. Therefore, data reduction algorithms, such as principal component analysis
(PCA) and partial least squares regression (PLS-R), are required. PCA combined with hyperspec-
tral imaging data enables the detection of spectral features in the spectroscopic data along with
identifying the relative distribution of the components in mixtures [38,151]. The PLS-R is an em-
pirical data-driven modelling approach that relies on representative model building data for two
variable blocks (X and Y). It is used to search for a correlation between a simple and easily ac-
quirable data set (X) and a labor- as well as cost-intensive second set of measurements (Y) by
calculating a certain number of factors. In the present study, the X data contains the UV spectra,
and the Y data the oxide layer thickness of the direct bonded copper sheets. Consequently, quan-
titative descriptions and calibrations are possible [152].
45
Despite several studies having focused on the characterization of copper oxide films, sample ho-
mogeneity remains a big challenge in the estimation of their thicknesses over the complete sur-
face. We address this topic in the present contribution using a hyperspectral imaging system in
the UV wavelength range for the in-line characterization of copper states and oxide layers thick-
nesses on direct bonded copper. The data were evaluated by PCA and PLS-R. The results show
that hyperspectral imaging in the UV range has the potential to predict oxide layer thicknesses
and copper states in a rapid and non-destructive manner.
4.3.1 2.1. Samples

In total, 28 direct bonded copper Curamik® Power substrates (Rogers Corporation, Chandler, AZ,
USA) with dimensions of 21.0 mm × 21.0 mm × 1.1 mm were used for sample preparations. The
samples were first ultrasonically cleaned at 50 °C for 5 min with Vigon A 200 (Zestron, Ingol-
stadt, Germany) as cleaning medium and then rinsed with deionized water for 3 min. The copper
sheets were oxidized at five different preparation protocols (see Table 4.1). Sample type 1 was
left in its initial condition. Figure 4.1 shows an example of each copper sheet type.
Table 4.1: Sample preparation protocol for the direct bonded copper substrates
Sample Type 1 2 3 4 5 6
Number of measured samples 5* 4 5* 5* 5* 4
Temperature/°C - 110.0 142.5 142.5 175.0 175.0
110.0 - 2 11 20 11 20
Time/min 0 4.0 6.0 8.3 14.0 21.1
Mean oxide layer thickness/nm 0 5.9 3.0 4.5 7.0 8.2
* One of each sample set was used for PLS-R prediction.
Figure 4.1: Direct bonded copper Curamik®Power substrates. (1) is an example of sample type 1, (2) sample type 2,
(3) sample type 3, (4) sample type 4, (5) sample type 5 and (6) sample type 6.
46
4.3.2 Oxide Layer Thickness Measurement

The thicknesses of the oxide layers were determined by depth profiling using X-ray photoelectron
spectroscopy (XPS). The measurements were conducted under a system base pressure of 4.0 ×
10−10 mbar. A monochromatic Al Kα radiation was used and the anode tube operated at 12.5 kV
with 20 mA. The take-off angle for the electrons was 0° with respect to the surface normal. The
XPS core level spectra were measured with a standard X-ray source SPECS XR50 (SPECS Sur-
face Nano Analysis GmbH, Berlin, Germany) and a concentric hemispherical analyzer Phoibos
100, SPECS (SPECS Surface Nano Analysis GmbH, Berlin, Germany). The pass energy of the
concentric hemispherical analyzer was 50 eV for the survey and 20 eV for the high-resolution
spectra. The data acquisition was performed with 0.5 eV; 0.1 eV per step, respectively.
Total (specular and diffuse) reflectance spectra were recorded in the range of 200–380 nm using
a UV spectrometer (Lambda 1050+, PerkinElmer, Inc., Waltham, MA, USA). The 150 mm inte-
grating sphere module functioned as a detection unit and was deployed in reflectance with a
R6872-Photomultiplier (PMT). A deuterium lamp was used as light source in the spectrometer.
The samples were placed at the reflectance port of the inte-grating sphere with a diffused scatter-
ing Spectralon® disk placed behind the samples. The port measuring area is approximately 0.42
cm2. Three spectra were recorded for each direct bonded copper type while the sample was rotated
in different angles (see Figure 4.2). The UV spectra were recorded with the Lambda 1050 UV
WinLab software from PerkinElmer.
(a) (b) (c)
Figure 4.2: An example of a direct bonded copper sheet rotated according to the three different measurement angles
(a) 0°, (b) 45° and (c) 90°.
4.3.4 Data Collection and Preprocessing

The hyperspectral imaging setup was optimized compared to our previous work [19]. The
pushbroom imager is a BlueEye Tec (inno-spec GmbH, Nürnberg, Germany), consisting of a
47
spectrograph (RS 50–1938, inno-spec GmbH, Nürnberg, Germany), with a slit width of 80 µm,
connected to a back-illuminated CMOS camera with total size 2048 × 2048 pixel (spatial × spec-
tral) and pixel size of 6.5 µm × 6.5 µm. Additionally, the dispersion is approximately 0.1 nm/px
[153]. The quantum efficiency of the CMOS camera is between 30 and 50% [154]. The optimal
integration time was 10 ms. The samples were placed on a black conveyor belt (700 mm × 215
mm × 60 mm, Dobot Magician, Shenzhen Yuejiang Technology Co., Ltd., Shenzhen, China)
moving with a constant speed of 0.15 mm/s, which was positioned completely in a tunnel made
of PTFE. The illumination was provided by two ozone producing Xenon lamps (XBO, 14 V, 75
W, OSRAM, München, Germany). The ozone was eliminated by a laboratory vacuum system
(AirTracker, TEKA Absaug- und Entsorgungstechnologie GmbH, Coesfeld, Germany). Another
xenon lamp was added to the setup to increase the intensity and optimize the integration time. In
combination with the black conveyor belt and a state-of-the-art UV pushbroom imager a more
industrial-like prototype was created.
The principal and workflow of the data acquisition remained [19]. The UV hyperspectral imaging
data were acquired by the FluxRecorder version 4.2.1.17 (inno-spec GmbH, Nürnberg, Germany).
The reflectance was calculated by the FluxRecorder automatically according to the radiometric
calibration [19,21,39,155]. PTFE was used as white reference. For collecting the dark reference,
the objective was closed by its cover and the illumination was turned off.
Figure 4.3 shows the original images of the direct bonded copper samples before and after back-
ground subtraction. Hyperspectral data matrices were analyzed by Evince version 2.7.11 (Pred-
iktera AB, Umeå, Sweden). While importing the raw data in Evince, a data reduction was per-
formed by binning four columns and rows (x,y) and six channels (λ).
The background was removed by calculating a PCA and selecting the corresponding background
scores. Therefore, some edges and borders of the samples were also eliminated, resulting in dif-
ferent sample shapes (Figure 4.3). The reduced hypercube was then used as input for the subse-
quent PCA and PLS-R. In the end, approximately 2.0 million spectra remained from the initially
obtained 4.0 million spectra.
48
Samples for model building Samples for model building
Samples for model prediction

(a) (b)
Figure 4.3: Hyperspectral raw images of 28 direct bonded copper samples on the left (a). Images after subtraction of
the background on the right (b). In total, 24 samples were used for building the PLS-R model and four
samples were used for prediction.
4.3.5 Multivariate Data Analysis and Data Handling

Multivariate data analysis (MVA) was performed with “The Unscrambler X 10.5” (Camo Ana-
lytics AS, Oslo, Norway). All spectra recorded by UV hyperspectral imaging and commercial
spectroscopy were preprocessed in the same way: Gaussian smoothing with 15 points reduction
in the range from 200 nm to 380 nm. The spectral resolution of the hyperspectral imaging data
was further reduced to 1 nm by averaging to ensure comparability to the UV spectra of the single-
point spectrometer. The principal component analysis (PCA) was calculated with mean centering,
cross-validation and the NIPALS algorithm to distinguish between the direct bonded copper sam-
ple types.
Partial least square regression (PLS-R) models for the oxide layer thickness prediction were cre-
ated with mean centering, full cross-validation and the Kernel algorithm. Four direct bonded cop-
per sheets of each preparation type were used to develop the PLS-R model. Additionally, the
remaining samples of copper type 1, 3, 4 and 5 were used as prediction samples to test the final
PLS-R model. The predicted values were compared to the determined oxide layer thicknesses by
XPS. Finally, the oxide layer thickness of each pixel of the remaining samples was predicted by
the hyperspectral imaging PLS-R model. The distribution map thus generated was visualized by
MATLAB (R2020b 9.9.0, Mathworks, Natick, MA, USA). The samples were binned by factor 5
in the x and y direction due to the large amount of data and noise.
49
Direct bonded copper substrates were investigated using diffuse reflectance spectroscopy in the
UV region (200–380 nm). In total, 28 samples were measured. Generally, the thickness of the
oxide layers increases with the oxidation time and temperature. During the oxidation process,
copper is oxidized first to copper (I) oxide (CuO2) and then to copper (II) oxide (CuO). Figure
4.4a shows the preprocessed reflectance spectra. Based on the shape of the spectra, the different
steps of the oxidation process can be observed. Sample type 1 is representing copper in its initial
condition. The other samples have undergone an oxidation process, as detailed in Table 4.1. A
band minimum is detected approximately at 220 nm. A pronounced band maximum for all copper
samples occurs in the wavelength range from 315 to 320 nm. Weak shoulders at 243 nm (sh) and
266 nm (sh) are observed. Sample types 1, 2 and 3 present one prominent maximum at 295 nm.
Sample types 4, 5 and 6 show a distinct band with maximum at 378 nm. The band at 220 nm
could be ascribed to Cu2O. Increasing Cu2O pronounces the minimum. The band at 295 nm is as-
signed to the copper material (see Appendix Figure 10.1). This band started to fade away due to
the increase in the maximum band at 320 nm. This band is absent in sample types 4, 5 and 6. For
these sample types a band at 378 nm appears. These spectral differences were due to different
oxide layer thicknesses and copper states (Cu0, Cu2O and CuO) on the copper sheets. The remain-
ing small differences among the spectra were attributed to the roughness, measuring angles and
sample positions.
(a) (b) (c)
0.6
0.8 6 0.10
PC1[96.0%]
PC2[4.0%]
5 0.4 6 3 PC2
0.7 2 0.05
- log (R/R0)
0.2
4
PC2 [4.0%]
0.6
0.0
4 1 0.00
0.5
0.4 3 -0.2
-0.05 PC1
0.3
2 -0.4 5
1 -0.6 -0.10
0.2
200 220 240 260 280 300 320 340 360 380 -3 -2 -1 0 1 2 3
200 220 240 260 280 300 320 340 360 380
Wavelength / nm PC1 [96.0%]
Wavelength / nm
Figure 4.4: (a) UV reflectance spectra of copper sheets. Copper with initial condition type 1 (green), 2 (red), 3 (blue),
4 (light blue), 5 (pink) and 6 (yellow) represent the oxidation layer thicknesses 0 nm, 4 nm, 8.3 nm, 14 nm
and 21.1 nm, respectively. (b) PCA with scores and (c) the corresponding loadings plot.
Figure 4.4b shows the scores plot of the first two principal components (PC). The first two PCs
explain nearly 100.0% of the total variance. The scores of different sample types are clearly dis-
tinguished. Every copper sample type with a corresponding copper state and oxide layer thickness
50
appear as a distinct group. PC1 yields a clear separation of copper in the initial condition (type 1)
from the other copper types. The groups move below the average in PC1, with increasing oxide
layer thickness and conversion of copper states. Copper types 4 and 5 are slightly overlapped as
their oxide layer thicknesses are almost comparable (see Table 4.1). The variance in each cluster
results from the different samples for each type. The differences between the samples could be
due to temperature profiles in the oven while preparing the samples, roughness variation, or sam-
ple positioning during the measurements.
The loadings plot for PC1 and PC2 is given in Figure 4.4c. The shape of PC1 resembles the Cu0
spectrum (see Appendix Figure 10.1 and Table 10.1). This indicates that an increasing amount of
Cu0 on a sample results in a more positive sample arrangement on PC1. Vice versa, the less Cu 0
is present in the samples because of the growing oxide layer thickness, the more the samples are
shifted in the negative range of PC1. The influence of the oxidation state (Cu 2O, CuO) is ex-
pressed by PC2 (see Appendix Figure 10.1 and Table 10.1); these results are comparable with
previous studies [26].

All samples were analyzed by a UV hyperspectral imaging prototype, as described in Materials
and Methods. In order to make the data more comparable to the UV spectroscopy, the average
spectra were calculated to reduce the number of spectra. A total of 25 spectra was determined
from the hyperspectral imaging data for each of the 28 samples. Figure 4.5a shows the results of
the UV hyperspectral imaging in the range from 200 to 380 nm.
The comparison between the shapes of the spectra is given in Figures 4.4a and 4.5a, showing
similarities as well as a small deviation. They are due to the type of the illumination source and
the design of the experimental setups. For reflectance spectroscopy, a deuterium lamp was used,
while for hyperspectral imaging, two xenon lamps were available. Deuterium lamps have higher
spectral irradiances in the deep UV range compared to xenon lamps [22]. However, the xenon
illumination was sufficient for the characterization of direct bonded copper sheets. Therefore, the
interferences < 270 nm are more pronounced compared to the higher wavelengths. As a result,
the spectra shown in Figure 4.5a provide almost no clearly recognizable spectroscopic infor-
mation in the region <270 nm. The detector’s efficiency and illumination provide low perfor-
mance in this wavelength range. Therefore, the easily accessible tunnel design for hyperspectral
imaging was developed to ensure a diffuse illumination of the samples. As a result, a reasonable
illumination strength and homogeneity were reached.
51
As discussed before, the spectra were influenced by the copper states and thicknesses of the oxide
layers on the copper sheets. Copper in the initial condition is represented in the spectra originating
from sample type 1 (see Table 4.1, Figure 4.4a). The most dominant contributions for all copper
sample types are observed in the wavelength range 324–328 nm and at 241 nm. Copper types 1,
2 and 3 present a weak shoulder at 292 nm.
(a) 0.8
(b) (c)
PC1[77.0%]
0.9 6 0.6 2
0.10
PC2[19.0%]
PC1
5 0.4
3
0.8
4 0.05
PC2 [19.0%]
- log (R/R0)
3 0.2
0.7
2 0.0 1 4 0.00
0.6 1 -0.2 5
PC2
0.5
-0.4
6 -0.05
-0.6
-0.10
0.4 -0.8
200 220 240 260 280 300 320 340 360 380 -1.6 -1.2 -0.8 -0.4 0.0 0.4 0.8 1.2 1.6 200 220 240 260 280 300 320 340 360 380
Wavelength / nm PC1 [77.0%] Wavelength / nm
Figure 4.5: (a) Average UV hyperspectral imaging spectra of copper sheets. Copper with initial condition type 1
(green), 2 (red), 3 (blue), 4 (light blue), 5 (pink) and 6 (yellow) represent the oxidation layer thicknesses 0
nm, 4 nm, 8.3 nm, 14 nm and 21.1 nm, respectively. (b) PCA with scores and (c) the corresponding load-
ings.
In the next step, a PCA model with a cross-validation was calculated for the average spectra of
all samples. Figure 4.5b shows the scores plot of PC1 and PC2. The first two PCs explain nearly
96.0% of the total variance. The scores of different sample types are clearly distinguished. Every
copper sample type with a corresponding copper state and oxide layer thickness appears as a
distinct group. PC1 yields a clear separation of copper with initial condition (type 1) from the
other copper types. A discrimination of the copper state and oxide layer thickness is observed on
PC2. Beginning from the positive to the negative scores on PC2, the samples are arranged in the
order copper type 2, 3, 4, 5 and 6, respectively. Again, copper type 2 and 3 (positive scores) can
be separated from the other samples 4, 5 and 6 (negative scores).
The loadings plot for PC1 and PC2 is given in Figure 4.5c. PC1 shows the differences between
Cu0 and the oxidation states (Cu2O, CuO). The most dominant contribution is observed in the
range from 260 to 280 nm and the increasing shape > 280 nm. The loadings plot of PC2 mainly
shows increasing oxide layer thickness. The most prominent contribution is observed in the range
from 250 to 280 nm and the decreasing shape > 280 nm. Compared to PC1, PC2 has a positive
maximum at 263 nm. The minimum on PC1 is located at 273 nm. This region could include the
information about the copper state. The influence of the oxidation state (Cu2O, CuO) and oxide
layer thickness is observed by PC2.
52
Figure 4.4a presents UV spectra with a good signal-to-noise-ratio recorded by a UV spectrometer,

which collected one single spectrum over an area of 0.42 cm2. Figure 4a shows the UV spectra
recorded by the hyperspectral imaging setup. The spectra were averaged over an area size com-
parable to the UV spectrometer. The UV hyperspectral imager recorded raw spectra with a less
good signal-to-noise-ratio. These spectra result from one single pixel of the detector, representing
a much smaller area of the direct bonded copper, which is estimated to be 6.5 µm × 6.5 µm.
Additional reasons for the low signal-to-noise-ratio are the weak irradiation intensity by the xenon
illumination and the quantum efficiency of the camera in this spectral range of approximately 30–
50% [154]. Furthermore, ozone-producing xenon lamps were used. With the help of a vacuum
system, the influence of the ozone absorption at 250 nm was minimized.
The benefit of hyperspectral imaging is lateral information in real time. To get a visual impression
of the inhomogeneity of the copper states and oxide layer thickness, the thickness for every pixel
from the first two PCs was plotted as a distribution map, shown in Figure 4.6. A sample with high
absorbance has a high proportion of blue in the score image (e.g., Cu 0), while one with low ab-
sorbance shows a higher proportion of red (e.g., Cu(II)). Clear differences between the samples
are observed according to the oxidation time and temperature. As discussed before, PC1 yields a
clear separation of copper in the initial condition from the other copper types. A discrimination
of the copper state and oxide layer thickness can be observed on PC2. The regular distribution of
the pattern in PC2 indicates a common origin; this could be the variability of the temperature
inside the oven among each sample. Additionally, in the distribution maps, it is possible to clearly
identify oxidation hotspots on the direct bonded copper.
Score value
Samples for model building Samples for model building range
1 1 +2
2 2
3 3
0
4 4
5 5
6 6 -2
(a) PC1 (b) PC2
Figure 4.6: Distribution maps of the oxide layer PC1 (a) and PC2 (b). Each rectangle represents a single copper sheet.
The sample type for each row corresponds to Table 1. The samples are divided into two sets: model build-
ing and model prediction for PLS-R. The colored pixels (the score value range) represent the oxide content,
from low (blue) to high (red).
53
4.4.3 PLS-R
A PCA structures data sets according to their maximum variance, whereas PLS-R searches for
the optimal correlation between spectral characteristics and an external target value. PLS-R mod-
els of the direct bonded copper for each method have been established and compared, by using
the spectra of the UV reflectance spectroscopy and UV hyperspectral imaging. In this study, spec-
tral features were extracted from the spectral datasets and correlated to determine the oxide layer
thickness via XPS.
Gaussian smoothing with 15 points was performed to minimize the noise. A PLS-R model was
developed with a calibration set of n = 24 samples (see Figure 4.6), three factors, the Kernel
algorithm and full cross-validation. A prediction sample set was used to test the PLS-R model
performance with an external validation to assess the predictive ability. The prediction sample set
consisted of four samples with mean oxide layer thicknesses of 0 nm, 6 nm, 8.3 nm and 14 nm
(see Figure 4.6). Table 4.2 summarizes the overall chemometric model results for both the UV
spectroscopy and hyperspectral imaging.
The number of factors for each PLS-R model was optimized according to a high coefficient of
determination (R2) and a low root mean square error of calibration (RMSEC) and cross-validation
(RMSECV). This approach was applied to both the calibration (R2c) and cross-validation (R2cv)
model for each method (Table 4.2).
Table 4.2: Model statistics for the calibration and full cross-validation models for oxide layer thickness on the direct
bonded copper.
Method Number of Parameters Calibra- Parameters Valida-
Factors tion tion
2 2
Rc RMSEC/nm R cv RMSECV/nm
UV spectroscopy 3 0.94 1.64 0.93 1.74
UV hyperspectral imaging 3 0.94 1.76 0.93 1.88
The variances explained by the UV reflectance model for the X and Y variables were 99.0% and
95.0%, respectively, by using three factors. The variances of the X and Y variables were 98% and
94% for the UV hyperspectral imaging model, by using three factors as well. This indicated that
three PLS components (factors) were sufficient to describe most of the variance in the data ac-
cording to the spectral information.
The results show that the PLS-R models are very effective in correlating the oxide layer thickness
with both spectroscopic data sets. This is indicated by a high R2c and a low RMSEC and a high
R2cv with a low RMSECV (see Table 4.2). Figure 4.7a,b show the correlation between the refer-
54
ence and predicted values of the UV reflectance spectra and UV hyperspectral imaging, respec-
tively. The deviation and the variance within a sample type are increasing according to the oxide
layer growth on the direct bonded copper. Copper sample types 1, 2 and 3 have a smaller variance
within the sample type. In contrast, sample types 4, 5 and 6 have more variance in the UV reflec-
tance spectra model. For UV hyperspectral imaging all samples have nearly the same variance.
This variance is probably due to the efficiency of the detector and the illumination in both setups.
(a) UV spectroscopy (b) UV hyperspectral imaging

24 24
21 21
18 18
Predicted, Factor 3
Predicted, Factor 3
15 15
12 12
9 9
6 6
3 3
0 0
0 3 6 9 12 15 18 21 0 3 6 9 12 15 18 21
Reference, Factor 3 Reference, Factor 3
(c) UV spectroscopy (d) UV hyperspectral imaging

4
4
3
Regression coefficients
Regression coefficients
2
2
1 0
0 -2
-1 -4
-2
-6
-3
-8
200 220 240 260 280 300 320 340 360 380 200 220 240 260 280 300 320 340 360 380
Wavelength / nm, Factor 3 Wavelength / nm, Factor 3
Figure 4.7: Three-factor PLS-R models for the oxide layer thicknesses of direct bonded copper in the UV region (200–
380 nm). (a) Predicted vs. reference of UV spectra. (b) Predicted vs. reference of UV hyperspectral imag-
ing. (c) Regression coefficients of the UV spectra. (d) Regression coefficients of the UV hyperspectral
imaging.
The regression coefficients of the three-factor UV spectroscopy PLS-R model are shown in Figure
4.7c. Again, absorbance bands around 210 nm, 245 nm, 293 nm and 330 nm emerge, as displayed
in the spectra. Above 360 nm, an increasing baseline in the regression coefficient plot is regis-
tered. In Figure 4.7d, the corresponding regression coefficients of the UV hyperspectral imaging
PLS-R model are displayed. They have a comparable shape, but more details can be detected. For
example, in the range <260 nm and from 310 to 340 nm, more spectral features are pronounced.
At 370 nm, a defined band appears for the UV hyperspectral imaging model, while an increase
>360 nm in the UV spectroscopy model is registered.
55
Correlated to the bands at <260 nm, 320 nm, 335 nm and >360 nm, the oxide layer thickness
increases in the UV spectra, which is also comparable to UV hyperspectral imaging in the range
230–265 nm and 306–340 nm for increasing the oxide layer thickness. At 293 nm, the oxide layer
thickness decreases for both setups. Differences in the beginning and ending of the regression
coefficients between both methods could be due to the detectors limits with the UV hyperspectral
imaging setup, as already discussed in the literature [19,26].
In order to evaluate the PLS-R models, four samples of type 1, 3, 4 and 5 were used to test the
model’s performance by predicting the oxide layer thickness. The sample set contains four sam-
ples with mean oxide layer thicknesses of 0 nm, 6 nm, 8.3 nm and 14 nm (see Figure 4.6).
In addition, the results indicated that the PLS-R was very effective in predicting the oxide layer
thickness with three factors, R2p = 0.90 with RMSEP = 1.62 nm and bias = 0.51 for UV spectra,
and R2p = 0.85 with RMSEP = 1.98 nm and bias = 0.61 for UV hyperspectral imaging.
In Table 4.3, the results for the mean value of the predictions and deviations are given. The pre-
dicted values are matched well with the references.
Table 4.3: Prediction of the oxide layer thicknesses for direct bonded copper from PLS-R models.
Method Sample Type Reference/nm Predicted/nm Deviation/nm
1 0 1.59 0.93
3 6 6.00 1.02
UV spectroscopy
4 8.3 7.86 1.44
5 14 15.25 1.53
1 0 -0.87 1.49
UV hyperspectral 3 6 5.51 2.08
imaging 4 8.3 11.74 1.91
5 14 14.35 1.79
Mean values with a high standard deviation were measured by XPS (see Table 4.1), as reference
values for the PLS-R models. This average of one sample type is comparable to the UV spectra
recording. However, hyperspectral imaging enables to recognize different oxide layer thicknesses
among the samples. Therefore, the oxide layer thickness in nm of each pixel was calculated. The
UV hyperspectral imaging PLS-R model was applied to the four prediction samples of type 1, 3,
4 and 5. In Figure 4.8, the resulting distribution map is shown. The pixels represent the oxide
layer thicknesses in nm, from low (blue) to high (red). Sample type 1 displays the initial direct
bonded copper sheet without induced oxidation, while the other samples show an increasing oxide
layer thickness.
Although, PLS-R is a robust model to describe the majority of the variance of the data according
to the spectral information. Compared to the results of the Vis hyperspectral imaging [26], the UV
56
4.5 Conclusions
hyperspectral imaging models seems to be more robust. This is indicated by the fact that less
factors are necessary to achieve a model with better statistic parameters (higher R2, lower RMSE)
by using a new UV prototype.
(a) (b) (c) (d)

35 nm
30 nm
25 nm
20 nm
15 nm
10 nm
05 nm
00 nm
1 3 4 5
Figure 4.8: Distribution map predicted from the three-factor PLS-R model of the UV hyperspectral imaging data. The
oxide layer thicknesses for each pixel of samples (a) sample type 1, (b) sample type 3, (c) sample type 4
and (d) sample type 5 were calculated for model prediction.
Hyperspectral imaging in the UV range is rarely reported, although it is often chosen for process
control and quality assurance [19,86]. The aim of this study was to characterize the copper states
and oxide layer thicknesses by using a single-point UV spectrometer and a UV hyperspectral
imaging setup that can serve as an example for a possible real-time industrial application. With
our hyperspectral imaging prototype, a whole direct bonded copper sheet can be measured and
processed within 10 s. With the implemented pushbroom imager, hardware binning is also possi-
ble, and can decrease the measuring and processing time. The scan speed for the determination of
the oxide layer thicknesses on direct bonded copper can be optimized by selecting a few relevant
variables instead of the complete UV spectrum. The intensity and type of the illumination are the
limiting factors towards a setup for a production environment. This study opens a novel possibility
for further development of this method capable of rapid in-line data acquisition, process control
and in-line classification/sorting, which meets the requirements of a real-time process with indus-
trial standard and precision.
4.5 Conclusions
UV hyperspectral imaging and UV reflectance spectroscopy (200–380 nm) were used to charac-
terize 28 direct bonded copper samples. UV reflectance spectroscopy, as a well-known method,
was utilized to compare the quality of the UV hyperspectral imaging results.
Hyperspectral imaging in combination with PCA and PLS-R is a promising approach for the lat-
erally resolved detection and differentiation of copper states and the determination of oxide layer
57
thickness in the UV region. The PCA models were able to separate all direct bonded copper types
according to the copper states and oxide layer thicknesses, using only the first two principal com-
ponents. PLS-R models with three factors provided a high R2 and low RMSE for calibration,
validation (ncv = 24) and prediction (np = 4). To the best of our knowledge, this is the first work
reporting the identification and quantification of copper oxide thin films by UV hyperspectral
imaging. The advantage of the home-built setup is the high spatial and spectral resolution and a
relatively high data acquisition speed under laboratory conditions. Starting from the presented
design and data given in this contribution a setup fulfilling the requirements of a real industrial
process can be easily realized.
Supplementary Materials: The following are available online at

https://doi.org/10.3390/s21217332/s1, Appendix Figure 10.1: Reference spectra for the copper
Cu0, Cu2O and CuO by using UV spectrometer, Appendix Table 10.1: Description of the direct
bonded copper substrates and their sample preparation.
Author Contributions: Conceptualization, M.A.K., M.S. and K.R.; methodology, M.A.K., M.S.,
T.E., M.S.L.D., J.S. and B.B.; software, M.A.K. and M.S.; validation, M.A.K., M.S., M.S.L.D.
and E.O.; formal analysis, M.A.K. and M.S.; investigation, M.A.K., M.S. and T.E.; resources,
T.J., S.G., E.O., K.R. and M.B.; data curation, M.A.K., M.S., T.E. and B.B.; writing—original
draft preparation, M.A.K., M.S. and T.E.; writing—review and editing, M.A.K., M.S., T.E.,
M.S.L.D., J.S., S.G., T.J., B.B., E.O., K.R. and M.B.; visualization, M.A.K. and M.S.; supervi-
sion, S.G., T.J., K.R. and M.B.; project administration, K.R., M.B.; All authors have read and
agreed to the published version of the manuscript.
Data Availability Statement: The raw/processed data required to reproduce these findings can-
not be shared at this time as the data also forms part of an ongoing Ph.D. thesis.
Acknowledgments: We especially thank the company inno-spec GmbH (Nürnberg, Germany)

for the possibility to test the BlueEye Tec hyperspectral imaging system. Within these test meas-
urements the UV data were generated. We would like to thank Frank Wackenhut for his guidance
in programming and presenting the data. The authors also thank Bayan Ayyad for helping in the
data handling.
58
5 Paper III: UV-Vis/NIR
Spectroscopy and Hyperspectral
Imaging to Study the Different
Types of Raw Cotton
Mohammad Al Ktash1,2, Otto Hauler1, Edwin Ostertag1 and Marc Brecht1,2,*
1
Lehr- und Forschungszentrum Process Analysis and Technology (PA&T) der Hochschule Reut-
lingen, Alteburgstraße 150, 72762 Reutlingen, Germany
2
IPTC and LISA+ center, University of Eberhard Karls Tübingen, Auf der Morgenstelle 18,
72076 Tübingen, Germany
*
This is originally published in the Journal of Spectral Imaging

(https://doi.org/10.1255/jsi.2020.a18) as
“M. Al Ktash, O. Hauler, E. Ostertag and M. Brecht, “Ultraviolet-visible/ near infrared spectros-
copy and hyperspectral imaging to study the different types of raw cotton”, J. Spectral Imaging 9,
a18 (2020). https://doi.org/10.1255/jsi.2020.a18 © 2020”
59
5 Paper III: UV-Vis/NIR Spectroscopy and Hyperspectral Imaging to Study the Different Types of Raw Cotton
5.1 Abstract
Different types of raw cotton were investigated by a commercial ultraviolet-visible/near-infrared
(UV-Vis/NIR) spectrometer (210 nm - 2200 nm) as well as on a home-built setup for NIR hyper-
spectral imaging (NIR-hyperspectral imaging) in the range 1100 nm - 2200 nm. UV-Vis/NIR
reflection spectroscopy illustrates a dominant role of proteins, hydrocarbons and hydroxyl groups.
A similar result was revealed with NIR- hyperspectral imaging. Experimentally obtained data in
combination with principle component analysis (PCA) provides a general differentiation of dif-
ferent cotton types. For UV-Vis/NIR spectroscopy, the first two principal components (PC) rep-
resent 82 % and 78 % of the total data variance for UV-Vis and NIR regions respectively.
Whereas, for NIR- hyperspectral imaging due to the large amount of data acquired, two method-
ologies for data processing were applied in low and high lateral resolution. In the first method,
the average of the spectra from one sample was calculated and in the second method the spectra
of each pixel were used. Both methods are able to explain ≥ 90 % of total variance by the first
two PCs. The results show that it is possible to distinguish between different cotton types based
on a few selected wavelength ranges. The combination of hyperspectral imaging and multivariate
data analysis has a strong potential in industrial applications due to its short acquisition time and
low cost development. This study opens a novel possibility for a further development of this tech-
nique towards real large-scale processes.
5.2 Introduction
Hyperspectral imaging is an imaging technology that combines spatial information with spectros-
copy. It is a fast and non-destructive method, which has evolved into a powerful analysis tool for
product inspection. Thereby, spatial images with very detailed spectral information for each pixel
of an object are collected simultaneously [21,24,26]. In the past, spectroscopical applications as
well as hyperspectral imaging in the UV-Vis and NIR range are more frequently found in the
textile research and in industrial applications [156]. In textile research, cotton plays a dominant
role among textiles, since cotton is the most important naturally occurring raw material for the
production of fabrics [157,158]. More than 34 million hectares of land are used to grow cotton,
and around 100 million households worldwide are engaged in cotton production [159]. Cotton is
considered as a key resource in the textile industry and accounts for about 30 % of all fibers used
in this sector [160]. In recent years, the increase in quality and processing requirements has led to
the introduction of modern techniques for processing and quality control [15,113-115]. Neverthe-
less, distinguishing between different cotton species is still a demanding task.
60
5.2 Introduction
Several detection methods have been developed and applied to identify and classify different cot-
ton varieties [158,161]. Most of them are off-line techniques such as thermogravimetric analysis
and optical spectroscopy [162-164]. Only little information is expected in the visible range, since
most raw cotton and residuals are reflective (or transparent) [165,166]. Valuable information can
be expected in the NIR region from characteristic molecular vibration, e.g. CHn and OH groups
of cotton which are omnipresent [167]. Unfortunately, the overall sensitivity for small variations
of the sample as well as for small amounts of contaminations in the NIR range is low and they
are hard to detect [36,168]. Therefore, numerous studies in the NIR region used a combination of
spectroscopy and chemometric modeling [167,169-175].
With NIR- hyperspectral imaging system, a complete optical spectrum with innumerable spectra
are collected at all image pixels. This is in contrast to multispectral systems such as RGB cameras
where only a limited number of wavebands are collected [86,176].
Most of the hyperspectral imaging applications were focused on remote sensing systems such as
satellites or aircrafts to gather information for agricultural, geological inspections and military
purposes. Nowadays, hyperspectral imaging is evolving into a standard for in-line and on-line
inspection in process analytics and quality control. Prominent technical applications can be found
in quality control for medicine, food and agricultural products [86,177,178].
In industrial applications, a hyperspectral imaging system is based on a combination of a

pushbroom scanner and a conveyor belt. Pushbroom scanner is fixed over the conveyor belt as
shown in Figure 5.1. Such inspection systems require a minimum of sample preparation and are
able to scan several samples swiftly with high spectral resolution[179]. Here, the pushbroom scan-
ner captures the complete spectral information line by line. The data is collected with the camera
placed perpendicular to the conveyor belt. As the conveyor belt moves, images are continuously
captured by the pushbroom scanner, resulting in a three dimensional (3D) data matrix with di-
mensions x,y and  and is often referred to as hypercube [28].
For cotton research, hyperspectral imaging was used in the UV-Vis range to detect foreign matter
with differentiation and classification of lint in cotton samples [161]. The results showed great
potential using a hyperspectral imaging system for the classification of foreign matter
[15,115,156].
In this study, we used optical reflection spectroscopy in the UV-Vis/NIR range as well as hyper-
spectral imaging in the NIR range for the differentiation of cotton sample sets. For both methods,
61
a chemometric model was developed that is based on PCA. Using this model, we were able to
distinguish between the different cotton types of our sample sets.
5.3.1 UV-Vis/NIR spectroscopy

Reflectance spectra of the samples were recorded in the range from 210 nm to 2200 nm using a
UV-Vis/NIR spectrometer (Lambda 1050, Perkin Elmer Ltd). It was used to compare the data
from the NIR- hyperspectral imaging and validate to another device. The UV-Vis/NIR spectrom-
eter was equipped with an Ulbricht sphere covered by polytetrafluorethylen (PTFE) to acquire
data in diffusion reflection mode with two detectors: one is an indium gallium arsenide (InGaAs)
detector and the second one is photomultiplier inside the sphere. The samples were placed on this
rear of the sphere, and a diffused scattering PTFE as a white reference disc was placed behind the
sample. The complete measuring aperture area is approximately 4.9 cm². From every cotton sam-
ple disc, a spectrum was acquired on each side. In total, three discs were measured for each sample
and, thus, for each cotton sample disc, six spectra were recorded.
5.3.2 NIR hyperspectral pushbroom online imaging system

Figure 5.1a shows the setup of the used hyperspectral imaging system. The hyperspectral system
is based on a pushbroom imager connected to a Xencis, Xeva 2.5 – 320 camera equipped with a
mercury cadmium telluride (HgCdTe) detector of 8 nm of spectral resolution, and has a 30 µm
slit width. Two halogen lamps illuminate the sample area. PTFE is used as a white reference while
the dark reference is acquired by imaging without any light exposure to the sensor. Figure 5.1 (b-
d) illustrates the principle and workflow for hyperspectral imaging. Figure 5.1b shows complete
spectroscopic information acquired for each line. Thus, a continuous line by line collection of
spectral information forms a two dimensional (2D) image as shown Figure 5.1c. It is also possible
to extract a single spectrum from a given pixel or point in the 2D image as shown in Figure 5.1d.
62
(a)
Camera
Pushbroom Spectrograph
imager
Objective
Halogen lamp Halogen lamp
Assembly line
0.16
(b) (c) (d)
- log (R/R0)
0.12
0.08
0.04
1200 1350 1500 1650 1800 1950 2100
y Wavelength / nm
Moving direction line Hyperspectral imaging Spectra at one point
Figure 5.1: (a) Setup of a hyperspectral imaging system based on the pushbroom concept. (b) hyperspectral imaging
scanning principle. (c) hyperspectral imaging generated immediately from the scanning of a cotton sample
disc. (d) NIR-Spectrum for one single pixel extracted from the image.
5.3.3 Samples
Figure 5.2 show 5 types of raw cotton and one hemp sample which were investigated. The samples
are organic raw material cotton (RoB), hemp plant from China (HC), recycled cotton (RcO),
standard raw material cotton (RoSt), recycled organic bright cotton (RcBH) and mechanically
cleaned cotton sample (CLN). Three samples of the aforementioned cotton types were collected
from the bulk, amounting to 0.75 g form each sample. The samples were pressed at 10 tons for 2
minutes to have same physical properties by a hydraulic press into a disc shape. The hydraulic
press was cleaned after each sample to reduce the chance of any impurities.
63
Figure 5.2: Raw cotton sample discs.
5.3.4 Data collection and preprocessing of hyperspectral

data
The following two methods for data pre-processing are described resulting in low and high
lateral resolution. Matlab (MATLAB 9.2.0, Mathworks, MA, USA) scripts were written for
per-processing of the hyperspectral data cube.
Figure 5.3 shows the workflow for calculating the mean spectrum of each sample. The hy-
perspectral image is collected by moving any cotton sample disc at a constant speed, approx-
imately 50 spectra were collected manually within the indicated area of interest, as shown in
Figure 5.3a (marked as dash line) and plotted as shown in Figure 5.3b. The average of these
spectra is calculated and shown in Figure 5.3c.
(a) Hyperspectral imaging (b) Extracted spectra (c) Average Spectra

Figure 5.3: (a) hyperspectral imaging of a cotton sample disc with area of interest (dash line) with a diameter of 2.5
cm. (b) Spectra extracted from the selected area. (c) Average spectrum of all spectra shown in (b).
Figure 5.4 shows the workflow for the second preprocessing method. The hyperspectral im-
age is captured by moving the 18 cotton sample discs at a constant speed. To differentiate
64
between signal and background, a distinction is first made between the respective spectral
characteristics.
For this purpose, two parallel planes are fitted into each spectral channel, one for the back-
ground and one for the samples. The distance between these planes is then selected as the
parameter for the spectral difference between the sample and the background. The color chan-
nel with the highest value is used as mask for all other color channels. Half of this difference
is set as threshold value. All lateral points of the color channel whose intensity value is above
this threshold value are classified as background and removed. This clipping mask is applied
to the entire hyperspectral data cube. The remaining data corresponds to the spectral contri-
butions from the samples. These are converted from the 3D hyperspectral data set into a 2D
format by joining the lateral points of the X and Y dimensions. This creates a matrix in which
each row corresponds to a pixel with a complete spectrum. This matrix is used as input for
the PCA.
Figure 5.4a shows the image obtained through the hyperspectral camera. The color channel
with the highest differential value is displayed in Figure 5.4b. Figure 5.4c shows a single
color channel of the cotton sample discs’ hyperspectral data cube after removing all lateral
components associated with the background, the removal of outliers like dead pixels or cos-
mic events, and the application of a PCA filter, which removes all contributions of higher
PCA components. The PCA filter works as follows: the first three of the resulting PCs explain
about 88 % of the variance. The 4th and higher components, while contributing less than 7 %
to the overall variance, contain mainly noise and were therefore discarded for further analysis.
The remaining 5% of the total variance is found within the residuals, and do not contribute
significant information. Figure 5.4d shows an image, where the red-green-blue (RGB) value
corresponds to scores of the first (R), second (G) and third (B) components.
In the next step, all score values that are 90 % similar to another score in all the main compo-
nents considered are removed from the data set. From the remaining score values a reduced
data set with the load values of the considered main components was generated. The reduced
data set is then converted back into a 3D hyperspectral data cube by separating the combined
lateral information. Figure 5.4c shows the reduced data as lateral information for one spectral
channel. The principal component analysis of this data again shows a significant grouping of
the different types of cotton. In the end, approximately 120,000 spectra remain from initially
obtained 1.7 Million spectra.
65
(a) (b) (c) (d)
Figure 5.4: (a) Hyperspectral raw imaging of 18 cotton sample discs with a diameter of 3.1 cm. (b) Image of the color
channel with the highest variance between cotton disks and background. (c) Images after subtraction of
the background, removal of outliers, and application of filters. (d) Image of RGB value corresponds to
scores of the first (R), second (G) and third (B) components.
5.3.5 Data handling and software

The UV-Vis/NIR spectra are recorded with the Lambda 1050 UV WinLab software from
PerkinElmer. The NIR hyperspectral pushbroom images are analyzed by the Prediktera soft-
ware from Evince 2.7.9. PLS Toolbox 8.5.1 (Eigenvector Research, Inc., USA) which is used
for the data processing and analysis. Lighting conditions may vary between the samples and
even within the samples across the scan line. A regular way to calculate this effect is to con-
vert measured raw spectra to reflectance spectra by the following formula [35,86]:
Isample − Idark
Where R and R0 represent the transmitted and incident intensity. Isample is the intensity of the
original image data, Idark is the intensity of the dark current image data and Ireference is the
intensity of the white reflectance image. Pre-processing of the mean center, smoothing (Sav-
Gol) with filter width 15 and polynomial order one, and generalized least squares (GLS) are
applied to the data. GLS is used to achieve an efficiency by transforming variance covariance
matrix into a homoscedastic one36. It works as a filter that calculate the differences between
the samples. The differences are considered as interference or clutter and GLS aims to reduce
these interferences [180-182].
66
5.4 Results and discussion
5.4.1 UV-Vis/NIR spectroscopy

Figure 5.5a shows UV-Vis/NIR spectra (210 nm – 2200 nm) from all samples. Six spectra
were recorded for each cotton sample type, three on each side. As expected, the spectra show
a high similarity. All spectra show the strongest reflectance at 280 nm which can be attributed
to proteins on the samples, see Table 5.1 [165]. In the visible range from 400 nm – 750 nm,
the spectra do not show any distinct features since most of the raw cotton is reflective. In the
NIR region, several spectral features can be observed. Dominant contributions are found at
1500 nm, 1933 nm and 2100 nm corresponding to the functional groups CH, ROH and OH,
respectively.
Due to the high similarity of the spectra, a differentiation of the samples is demanding. As a
consequence, PCA is used to further differentiate the samples and was applied for processed
spectra.
The processing of spectra is described in the Materials and Methods section. Figure 5.5b
shows the scores plot of the first two principal components PC1 and PC2 for UV-Vis region
(210 nm – 1100 nm). The PCA model explains 70.1 % and 82.3 % of the spectral information
with the first two PCs respectively. The scores plot shows that PC1 and PC2 are sufficient to
separate all samples. In this representation, the hemp (HC) sample shows the most distinct
separation from the cotton group, as expected. Figure 5.5c shows the corresponding loadings
plot for PC1 and PC2. The most significant differences between those loadings are found in
the regions from 210 nm – 350 nm, 450 nm – 700 nm. In the UV range (210 nm – 350 nm),
the strongest influence on PC1 is found at 280 nm, 300 nm and for PC2 at 290 nm. They can
be assigned to proteins and amino acids (see Table 5.1)[165]. The contributions in the visible
range (450 nm – 700 nm) show a maximum/minimum at 680 nm, it can be assigned to the
color of the RcO samples (see also the inset in Figure 5.5a).
Figure 5.5d shows the scores plot of the first two principal components PC1 and PC2 for NIR
region (1100nm – 2200 nm). The PCA model explains 63.5 % and 78.0 % of the spectral
information with the first two PCs respectively. The scores plot shows that the first two PCs
are sufficient to separate all samples clearly from one another in NIR range. In the scores plot,
the hemp (HC) and CLN sample shows the most distinct separation from the cotton group.
67
Figure 5.5e shows the loadings plot for PC1 and PC2. The most significant differences be-
tween those loadings are found in the regions from 1100 nm – 1200 nm, 1350 nm – 1500 nm,
1600 nm – 1700 nm and 1850 nm – 2100 nm. In the NIR region (1100 nm – 2200 nm), several
spectral features are variable which are assigned to the hydrocarbons and hydroxides oscilla-
tion (see Table 5.1).
With UV-Vis/NIR, a separation of the analyzed cotton sample discs has been successfully
demonstrated. However, the large deviations between PC1 and PC2 are mainly found in the
UV-Vis and NIR region. Therefore, the application of an online method for characterization
is the most suitable for these spectral regions.
Table 5.1: UV-Vis/NIR reflectance maxima [165],[85,183].

Reflectance (nm) Functional groups
1240 nm CH
1525 nm ROH
1790 nm CH3, CH2
1955 nm OH
2117 nm ROH
2342 nm CH, CH2, CH3
68
(a)
1.2
RoB
HC
1.0 RcO
RcBH
RoSt
0.15
CLN
0.8
- log (R/R0)
- log (R/R0)
0.10
0.6 0.05
0.00
525 600 675 750 825
0.4
Wavelength / nm
0.2
0.0
250 500 750 1000 1250 1500 1750 2000
Wavelength / nm
(b) (c)
0.09 0.03
RoB RcBH
0.06 0.02 HC
0.03 CLN
PC2 [11.6%]
0.01
PC2 [14.4%]
RcO RcO CLN

HC RoSt RoB
0.00 0.00
RcBH RoSt
-0.03 -0.01
-0.06 -0.02
-0.09 -0.03
-0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 -0.050 -0.025 0.000 0.025 0.050
PC1 [70.7%] PC1 [63.6%]
(d) (e)
0.2
PC1[63.6%] PC1[63.6%]
0.3 PC2[14.4%] PC2[14.4%]
0.2
0.1
PC2
0.1 PC2
0.0
0.0
PC1 PC1
-0.1
-0.1
300 400 500 600 700 800 900 1000 1100 1200 1400 1600 1800 2000 2200
Figure 5.5: (a) UV-Vis/NIR spectra of cotton sample discs including one HC sample in the wavelength range 210 nm
– 2200 nm. Upper left: Image of a cotton sample disc where the region of integration for determining the
average spectra is indicated by a black area with a diameter of 2.5 cm. (b) Scores plot for the processed
spectra in the UV-Vis. The 2D projection of the 95 % confidence ellipse of the data collected from each
type of cotton is included to facilitate visualization of the obtained results. (c) Loadings plot for the UV-
Vis. (d) Scores plot for the NIR. (e) Loadings plot for the NIR.
69
5.4.1 NIR Hyperspectral Imaging

Two data processing techniques were applied to the NIR hyperspectral images to calculate
PCA models. As before, three samples of each raw fibers were analyzed. The setup for hy-
perspectral imaging as well as for determination of the spectra from the hyperspectral data
matrix is described in the Materials and Methods section.
In the first method, the mean value of the spectra was calculated for each cotton sample disc.
A total of six spectra are determined from the hyperspectral imaging data for each cotton
sample type.
Figure 5.6a shows hyperspectral NIR spectra in the range from 1100 nm – 2200 nm. The most
dominant contributions are observed around 1525 nm, they can be attributed to the presence
of OH groups. Four weaker peaks are observed around 1340 nm, 1790 nm, 1955 nm and
2117 nm, their assignment is given in Table 5.2[36].
The PCA of these spectra explains 93.7 %, 97.0 % or 98.3 % of the spectral information with
the first two, three or four PCs respectively. Figure 5.6b shows the results for the first three
PCs. In the scores plot it can be seen that the first three PCs are sufficient to separate all
samples clearly from one another.
(a) (b) (c)

0.20
RoB 0.2
RoB PC1[66.2%]
Average RoSt PC2[24.6%]
HC Spectra HC PC3[3.6%]
0.16 RcO RcBH 0.1
PC3 [3.6%]
PC3
- log (R/R0)
RcBH
CLN
0.12 RoSt
CLN 0.0
PC1
0.08 RcO
-0.1
0.04 PC2
1200 1350 1500 1650 1800 1950 2100 1200 1400 1600 1800 2000 2200
Figure 5.6: (a) Spectra recorded by hyperspectral imaging of cotton sample discs including one HC sample in the NIR
range from 1100nm - 2200nm. Upper right: Image of a cotton sample disc where the region of integration
for determining the average spectra for each sample is indicated by a black circle with a diameter of 2.5
cm. (b) Scores plot for the processed spectra in NIR- hyperspectral imaging. The 2D projection of the 70
% confidence ellipse of the data collected from each type of cotton is included to facilitate visualization of
the obtained results. (c) Loadings plot for the NIR- hyperspectral imaging.
Figure 5.6c shows the loadings plot for the first three PC’s. In the range from 1340 nm – 1663
nm the reflectance around 1508 nm can be assigned to the presence of ROH (see Table 5.2).
70
The reflectance in the range from 1789 nm - 2100 nm, 1973 nm can be assigned to the OH
group. The contribution approximately 2270 nm is due to CH [85,183].
Table 5.2: NIR Hyperspectral imaging reflectance maxima [85,183].

Reflectance (nm) Functional groups
1240 nm CH
1525 nm ROH
1790 nm CH3, CH2
1955 nm OH
2117 nm ROH
2342 nm CH, CH2, CH3
In the second method, several thousand spectra from every cotton sample disc were used to
calculate the PCA model. The preprocessing and workflow of the spectra from the hyperspec-
tral data matrix is described in the section Materials and Methods. Figure 5.7a presents ex-
amples of hyperspectral NIR spectra from a single pixel of each of the six cotton types in the
range 1100 nm – 2200 nm.
(a) (b) (c)

0.15
0.20
RoB RoB PC1[82.2%]
PC2[3.9%]
HC HC 0.10 PC3[2.1%]
PC3 [2.1%]
0.16 RcO RcO PC1

0.05
RcBH
- log (R/R0)
RcBH
0.12 RoSt RoSt 0.00 PC2
CLN CLN
-0.05 PC3
0.08
-0.10
0.04
-0.15
1200 1350 1500 1650 1800 1950 2100 1200 1400 1600 1800 2000 2200
Figure 5.7: Hyperspectral of cotton sample discs including one HC sample in the NIR range from 1100nm - 2200nm.
(a) Six example spectra recorded at individual pixels. (b) Scores plot calculated for the whole data set
including several thousand processed spectra. (c) Loadings plot for the NIR-Hyperspectral imaging.
The PCA of these spectra explains 86.0 %, 88.2 % or 89.0 % of the spectral information with
the first two, three or four PCs respectively. Figure 5.7b shows the results for the first three
PCs, the first three PCs are sufficient to separate all samples from one another. A clear sepa-
ration is observed for RoB and CLN, while the HC, RoSt, RcO and RcBH are slightly over-
lapping. Nevertheless, these samples can be separated only if one pair e.g. HC and RoSt is
included in a separate model (data not shown). Figure 5.7c shows the loadings plot of the first
71
three PCs. Overall, the loadings are comparable with the loadings shown in Figure 5.6c, ex-
cept a change of the sign. In the range from 1350 nm – 1700 nm, the reflectance around 1550
nm can be assigned to the presence of ROH (see Table 5.2). The reflectance in the range from
1800 nm - 1990 nm can be assigned to OH groups. The signal around 2302 nm is due to CH
[85,183].
The first three PCs explain a significant amount of the NIR hyperspectral data for both pre-
processing methods. Calculating the PCA model at each pixel or deriving it from the mean
spectra does not significantly change the data behavior of the model (Figure 5.6 and Figure
5.7). The advantage of using average spectra instead of the complete data set is fast data
processing. However, this method is limited to recognize or spectrally separate background
from the samples automatically. Therefore, a certain time is required to select the samples
manually and calculate the average spectra for each cotton sample disc. On the other hand,
when applying a filter (see Figure 5.4) the separation of the sample from the background
works automatically, but here the quantity of data hampers a fast processing. The scattering
in the scores plot in Figure 5.7 shows the huge variability of the properties of the samples,
theses only become visible if the spectra are taken with hyperspectral imaging. Compared
with the scattering where the spectral information is averaged over a larger area (Figure 5.5
and Figure 5.6) this is remarkably reduced. The large variability of the score values from the
hyperspectral imaging indicates a change of the samples properties on the scale of the reso-
lution actually achieved. For the hyperspectral imaging setup this is about 13µm. The high
lateral resolution achieved here shows that sample properties on this scale vary and are there-
fore relevant, as new insights into the heterogeneity of fiber samples can be gained. As a
consequence, the data show the high potential for hyperspectral imaging which is beyond the
differentiation of fiber types.
In the next step, a filter is required that combines the advantages of both methods to speed up
the data handling. Together with this, a simplified model can be developed that meets the
requirements of real online applications.
72
5.5 Conclusions
5.5 Conclusions
UV-Vis/NIR reflection spectroscopy and hyperspectral imaging in combination with PCA is
a promising approach for the detection and differentiation of raw cotton types. The most rel-
evant information for the differentiation of cotton types was found in both the UV and NIR
range (see Figure 5.5c).
The results obtained with UV-Vis/NIR spectroscopy revealed that the contribution in the UV
can be assigned to the presence of protein at 280 nm. The most dominant contribution to
absorbance in the NIR range can be assigned to CH3 for the most prominent band at 1775 nm
and to ROH vibrations at 1500 nm. The spectral data were analyzed with PCA in order to
achieve a differentiation of different cotton types. The PCA model was able to classify all
types with the first two PCs explaining the maximum variance of the data.
NIR- hyperspectral imaging results reveal the most dominant absorbance assigned to CH3 and
ROH at 2270 nm and 1525 nm respectively. Two methods were used for processing the large
amount of data. Both approaches resulted in a differentiation of all types. The advantages of
the rugged online home-built setup is a high spatial/spectral resolution and a rapid data ac-
quisition. With this method, several samples can be measured in a short time and at low cost.
Based on the data shown it is reasonable to develop a simplified chemometric model, which
meets the requirements of a real process with industrial standards and precision.
73
6 Paper IV: Prediction of
Honeydew Contaminations on
Cotton Samples by In-Line UV
Hyperspectral Imaging
Mohammad Al Ktash1,2,†, Mona Stefanakis1,2,†, Frank Wackenhut1, Volker Jehle3, Edwin Oster-
tag1, Karsten Rebner1 and Marc Brecht1,2,*
1
Center of Process Analysis and Technology (PA&T), School of Life Sciences, Reutlingen Uni-
versity, Alteburgstraße 150, 72762 Reutlingen, Germany
2
Institute of Physical and Theoretical Chemistry, Eberhard Karls University Tübingen, Auf der
Morgenstelle 18, 72076 Tübingen, Germany
3
Texoversum Faculty Textile, Reutlingen University, Alteburgstraße 150, 72762 Reutlingen, Ger-
many
*
Correspondence: [email protected].
†
“Al Ktash, M.; Stefanakis, M.; Wackenhut, F.; Jehle, V.; Ostertag, E.; Rebner, K.; Brecht, M.
Prediction of Honeydew Contaminations on Cotton Samples by In-Line UV Hyperspectral Imag-
ing. Sensors 2023, 23, 319. https://doi.org/10.3390/s23010319”
75
6 Paper IV: Prediction of Honeydew Contaminations on Cotton Samples by In-Line UV Hyperspectral Imaging
6.1 Abstract
UV hyperspectral imaging (225 nm – 410 nm) was used to identify and quantify the honeydew
content on real cotton samples. This study presents the implementation and application of UV
hyperspectral imaging as a non-destructive, high-resolution, and fast imaging modality. For this
novel approach a reference sample set, which consists of sugar and protein solutions that were
adapted to honeydew, was set up. In total, 21 samples with different amount of added sug-
ars/porteins were measured to calculate multivariate models to predict and classify the amount of
sugar and honeydew at each pixel of a hyperspectral image. The principal component analysis
models (PCA) enabled a general differentiation between different concentrations for sugar and
honeydew, respectively. A partial least squares regression (PLS-R) model was built based on the
cotton samples soaked in different sugar and protein concentrations. The result shows a reliable
performance with R2cv = 0.84 and low RMSECV = 0.009 g for the validation. The PLS-R refer-
ence model was able to predict the honeydew content laterally resolved in gram on real cotton
samples for each pixel with light, strong and very strong honeydew contaminations. Therefore,
in-line UV hyperspectral imaging combined with chemometric models can be a future effective
tool for the quality control of industrial processing of cotton fibers.
6.2 Introduction
Hyperspectral imaging is an imaging technology that combines video image analysis with spec-
troscopy [141,184]. Precisely, it is a series of images acquired by moving the object or the imager.
It is a fast and non-destructive technique, which has developed into a robust analysis tool for
product screening. Such systems are able to capture spectral and spatial information with high
resolution. As a result, a spectrum of an object can be obtained for each hyperspectral image pixel
simultaneously [20,185].
Hyperspectral imaging and spectroscopic applications are widely used in industrial environments
[19,53,186-189]. The importance of hyperspectral imaging is steadily growing in the textile indus-
try. For example, the visible (Vis) and near infrared (NIR) range are often applied for quality
control and sorting processes [190,191]. Where the UV range is rarely used so far.
In textile research, cotton is considered as one of the most important natural fibers for fabric
production [192,193]. It provides approximately 50 % of the world’s textile fibers [194]. Cotton
consists of approximately 95 % cellulose and 5 % sugar, wax, proteins, organic acids and pectin
76
6.2 Introduction
[195]. The process ability is affected and degraded by the sugar content. Sugar is a naturally ex-
cretion of aphids and whiteflies on the cotton through metabolic processes and is specifically
called honeydew [196]. This contamination on the raw cotton causes stickiness, which causes
problems in the processing stage. This leads to economic loss because the sticky raw cotton is
rejected during quality control [195,197]. In-line detection and subsequent removal of sticky cot-
ton would lead to an uninterrupted production and thus higher profit [191,198,199].
Several detection methods have been developed and applied to identify cotton contaminations
and stickiness in ultraviolet (UV)-Vis/NIR spectroscopic applications as well as Vis/NIR hyper-
spectral imaging [103,115,141,200]. Most of them are off-line such as optical spectroscopy
[103,201]. Identification of cotton and cotton trash components was studied by Fortier et al. [202]
using FT-NIR spectroscopy. Mustafic et al. [161] examined the applicability of hyperspectral im-
aging to detect and classify cotton foreign matter in the visible spectral region, whereas the
Vis/NIR region was studied by Jiang et al. [203]. Other methods, such as thermogravimetric anal-
ysis [162], high-pressure liquid chromatography (HPLC) and minicard [204], require an elaborate
sample preparation, are time-consuming and expensive compared to optical spectroscopy. In-line
detection and quantification of stickiness on cotton samples using NIR hyperspectral images was
investigated by Severino et al. [199]. They were able to discriminate between glucose on cellulose
and melezitose, trahalose, glucose, fructose, and sucrose at each pixel.
Tschannerl et al. [86] compared hyperspectral imaging in UV and NIR regions to precisely dis-
criminate between phenolic flavor concentrations in melted barley. The rarely used UV region
showed interesting results despite the illumination wasn't optimal. Previously, our group reported
a hyperspectral imaging setup for the UV spectral region, this setup was used to distinguishing
between different pharmaceutical drugs [19] as well as, for characterizing oxide layers thickness
and copper states on direct bonded copper [53]. The results clearly showed that a spectral imager
based on pushbroom technology has many advantages in terms of achieving fairly short UV op-
erational wavelengths and a high spectral resolution. However, hyperspectral imaging rapidly
scans samples resulting in a large amount of spectral data within a short time period. Therefore,
multivariate data analysis, such as principal component analysis (PCA) and partial squares re-
gression (PLS-R), are required to reduce the amount of data without losing important information.
PCA reveals the most relevant information of a data matrix [38]. A combination between PCA
and quadratic discriminant analysis (QDA) enables data classification and investigation of model
quality parameters [65,205]. The information about the relation between a number of predictor
variables and independent variables can be extracted by using PLS-R [206].
77
The aim of this work is to develop a chemometric model able to identify and quantify the amount
of honeydew on real cotton samples based on UV hyperspectral imaging. For this approach a
reference sample set, which consists of honeydew typical sugars and proteins, was prepared. Me-
chanically cleaned cotton was soaked with solutions with different sugar concentrations. Chemo-
metric models, especially PCA and PLS-R, were developed based on UV hyperspectral imaging.
PCA is used to classify the cotton samples according to their sugar concentration and PLS-R is
applied to correlate the UV spectra with the sugar concentration. This PLS-R model successfully
predicts the amount of honeydew in gram on real cotton samples. This work is considered as the
first scientific work to identify and quantify the amount of honeydew contents at each pixel.
Therefore, hyperspectral imaging is a suitable technique for in-line environment applications in a
rapid and non-destructive manner.
6.3.1 Chemicals and preparation of solutions

0.2 g of each macronutrient 1-6 was weighted and dissolved in 10 mL of deionized water (see
Table 6.1). A six-fold serial dilution was prepared in 50 mL volumatric flasks. For each diluting
step 25 mL of the previous solution and 25 mL of deionized water were mixed for 2 min (see
Table 6.2).
78
Table 6.1: Description of the macronutrients and natural materials.

Macronutrients
CAS Num-
and natural mate- Samples Description Manufacture
ber
rials
D-Glucose anhydrous
Fisher Scientific
1 Glucose Laboratory reagent 50-99-7
GmbH, Leics, UK
grade
ThermoFisher
2 Fructose D-Fructose, 99.0 % GmbH, Kandel, Ger- 57-48-7
many
D-Sucrose, >=99.9 % Fisher Scientific
3 Sucrose GmbH, New Jersey, 57-50-1
For Molecular Biology USA
Sigma-Aldric Chemie
D-(+)-Melezitose mo-
4 Melezitose GmbH, Steinheim, 10030-67-8
nohydrate, >=99.0 %
Germany
D- Trehalose anhyd- Acros Organics, New
5 Trehalose 99-20-7
rous, 99.0 % Jersy, USA
Bovine Serum Albumin
PAN-Biotech GmbH,
6 Protein (BSA) fraction V, ly- 9048-46-8
Aidenbach, Germany
ophilized powder
6.3.2 Sample set and sample preparation

The sample set consists of cleaned cotton to build the model and cotton samples contaminated by
honeydew to test the model.
The cotton samples for model building were collected from a bulk of cotton mechanically cleaned
[207,208] from the Texoversum Faculty Textile at Reutlingen University. The cotton is a blend of
different long staple Pima qualities. In total, 21 cotton samples were prepared with a weight of
0.3 g ± 0.0001 g (XSE205 DualRange, Mettler Toledo GmbH, Switzerland). The samples were
dried in a vacuum oven (Vacutherm VT 6130 M, Termo Fisher Scientific Inc., Waltham, Massa-
chusetts, USA) at 30 °C and 50 mbar for 8 h to remove absorbed humidity. The humidity was
estimated by a commercially available sensor (Humidity-Detector MD, H. Brennenstuhl GmbH
& Co. KG, Tübingen, Germany ). The weight loss is documented in Appendix Table 10.2. 4 mL
macronutrients solution were used for each sample. Three samples per concentration were made
(see Table 6.2). The samples were soaked in an aluminum plate (28 mL, Carl Roth GmbH+Co.
KG, Karlsruhe, Germany). The samples were dried again in a vacuum oven at 30°C and 50 mbar
for 44 h. By determining the remaining weight the average macronutrient content can be calcu-
lated for each sample (see Appendix Table 10.2). The humidity and temperature of the labatory
were monitored (BL30, Klima-Datenlogger, Trotec GmbH, Heinsberg, Germany) during the
whole workflow.
79
Comparable real cotton samples were collected by ICA Bremen GmbH (Bremen, Germany) to
test the predictive power of the model. The samples were chosen according to their honeydew
content in the steps light, strong and very strong. The samples origin is Sudan Acala (see Table
6.3).
Figure 6.1 shows the samples pressed in the sample holder prepared for measuring. The sample
types are named from A to F and one mechanically cleaned (CLN) sample, where A has the
highest concentration (2 wt%), and F has the lowest concentration (0.0625 wt%) (see Table 6.2).
The average sugar content remaining on the samples after 44 h was calculated (see Table 6.2).
For ease of reading, we omit the term macronutrients for the description of the solution of various
sugars and the protein in the following, and replace it with the short-term "sugar" for the sample
nomenclature.
A B C D E F CLN
Light Strong Very Strong

Figure 6.1: Overview of the samples pressed in the sample holder. For each concentration, three samples were prepared
and measured at once (A to F and CLN). Real cotton samples with different honeydew contents (light,
strong and very strong.
80
Table 6.2: The concentration of the sugar solutions and the weighted averaged sugar amount on cotton samples.
Sample type Sugar concentration / wt % Ratio of: sugar / g dried cotton / g
A 2 0.2593
B 1 0.1331
C 0.5 0.0743
D 0.25 0.0386
E 0.125 0.0326
F 0.0625 0.0322
CLN - -
Table 6.3: The number of honeydew stickiness points on cotton samples.

Average number of
Stickiness Type Single measurments Sample
sticky points
Light 2, 11, 5 6 4301
Strong 47, 45, 47 46 Sudan Girba Acala 3SG
Very strong 60, 69, 80 70 Sudan Gezira Acala type 3SG
6.3.3 UV hyperspectral imaging setup

Figure 6.2a shows a scheme of the hyperspectral imaging setup. The pushbroom imager consists
of a back-illuminated CCD camera (Apogee Alta F47: Compact, inno-spec GmbH, Nürnberg,
Germany) and a spectrograph (RS 50-1938, inno-spec GmbH, Nürnberg, Germany) with a slit
width of 30 µm. The CCD camera has a resolution of 1024 x 1024 pixel (spatial x spectral) and a
pixel size of 13 µm x 13 µm. The optimal integration time was 300 ms. The conveyor belt (700
mm × 215 mm × 60 mm, Dobot Magician, Shenzhen Yuejiang Technology Co., Ltd., Shenzhen,
China) moves with a constant speed of 0.15 mm/s. The conveyor belt was located totally in a
polytetrafluoroethylene (PTFE) (Sphereoptics GmbH, Herrsching, Germany) tunnel. The illumi-
nation was achieved by two Xenon lamps (XBO, 14 V, 75 W, OSRAM, München, Germany).
Figure 6.2d shows a sample holder developed to reduce the influence of the topography of the
samples. Therefore, a quartz glass made of suprasil 2 grade B with the dimension of 140 mm x
80 mm x 1 mm (Aachener Quarzglas-Technologie Heinrich GmbH & Co.KG, Aachen, Germany)
was used. PTFE was used as white reference.
Figure 6.2(c-e) illustrates the principle and workflow of the data acquisition. Figure 6.2c presents
the principle of hyperspectral imaging line scanning method which collects one line at a time,
with all of the pixels in a line being measured simultaneously. The continuous line by line collec-
tion of spectral information results in a lateral (x, y) 2D image as shown in Figure 6.2d, whereas
each location contains a further spectroscopic dimension (λ) as shown in Figure 6.2e. Thus, a 3D
data matrix (hypercube) was recorded.
81
(a) Pushbroom imager
Spectrograph CMOS camera
Tunnel
PTFE
Objective
Xenon lamp
Power supply
Xenon lamp
Power supply
(b) (c) (d) (e)

Quartz suprasil 2 grade B 0.30
PTFE
 0.25
- log (R/R0)
x 0.20
0.15
0.10
y 240 280 320 360 400
Wavelegth / nm
Sample holder Moving direction Hyperspectral image Spectrum at one point
Figure 6.2: (a) Setup of a hyperspectral imaging system based on the pushbroom concept (the tunnel in the scheme was
cut to show the inside). (b) Custom made sample holder consisting of quartz glass as sample cover and
PTFE as reference. (c) Pushbroom imager scanning principle. (d) Hyperspectral image generated immedi-
ately from the scanning of a sample. (e) UV spectrum after preprocessing for one point extracted from the
image given in (d).
6.3.4 Data collection and preprocessing

The UV hyperspectral imaging data were acquired by the SI-Cap-GB version V3.3.x.0 software
(inno-spec GmbH, Nürnberg, Germany). The reflectance was calculated by the SI-Cap-GB auto-
matically after recording Ireference and Idark. Illumination conditions will vary between samples and
even within samples across the scan line, especially for heterogeneous samples such as cotton
with high scattering due to sample topography. A common method to reduce the influence of the
sample topography is to convert the raw spectra of each pixel into reflectance spectra (radiometric
calibration) using the following formula [19,21,37-39]:
Isample − Idark
82
R and R0 are the intensities reflected from the sample and a specific reference material with high
reflectivity, in this case PTFE. The intensity of the original image is represented in Isample. Accord-
ingly, the intensity of the dark current image is given by Idark and the intensity of the PTFE image
is Ireference [20]. In order to enhance the absorption bands the negative decadic logarithm is calcu-
lated as -log (R/R0).
Hyperspectral data matrices were analyzed by Evince version 2.7.13 (Prediktera AB, Tvistevä-
gen, Sweden). It is used for data handling and extracting the spectra of each pixel. For model
building the sample set contains of 4896 pixel × 1024 pixel which represents approximately 5.0
million spectra. One spectrum ranges from 225 nm to 410 nm including 1024 variables. With
4464 pixel × 1024 pixel which represents approximately 4.6 million spectra the model was tested.
In total, this results in approximately 15 GB of data size. Figure 6.3a shows an example of RGB
hyperspectral images of cotton samples sprayed with different concentrations of sugar solution.
A region of interest was selected by using a rectangular shape to extract the spectra (see Figure
6.3b).
(a) (b)
Figure 6.3: Example of data extraction. (a) Hyperspectral raw images of 18 cotton samples sprayed with different
concentrations of sugar (A highest to F lowest) and one cleaned cotton sample (CLN). For model building,
all spectra were extracted manually. (b) Zoom-in-image of a cotton sample with the region of interest
marked by a black rectangle.
6.3.5 Multivariate data analysis and model building

Multivariate data analysis (MVA) was performed with “Aspen UnscramblerTM, version 10.5.1”
(Aspen Technology Inc., Bedford, MA, USA). The UV spectra were pretreated prior to the mul-
tivariate data analysis in the following way: Base line correction followed by a Savitzky-Golay
smoothing (8 points, symmetric, 2nd polynomial order). The principal component analysis (PCA)
models were calculated with mean centering, cross-validation, and NIPALS-algorithm. A partial
least square regression (PLS-R) model for the sugar concentrations was created with mean cen-
tering, full cross-validation and Kernel-algorithm. All cotton samples of each concentrations have
been used to develop the PLS-R model. The PLS-R model was tested by predicting the honeydew
83
content on the real cotton samples. Three different areas from these real cotton samples were also
investigated by PCA with aforementioned settings. To show the quality of the model, PCA was
combined with a quadratic discriminant analysis (QDA, 5 PCs). A fourth area was predicted by
the PCA-QDA model.
MATLAB (MATLAB 9.2.0, Mathworks, MA, USA) and PLS_Toolbox (PLS Toolbox 8.5.1, Ei-
genvector Research, Inc., Wenatchee, WA, USA) were used for presenting the data.
6.4.1 Cotton samples impregnated with sugar

Cotton samples were investigated using hyperspectral imaging in the UV region (225 nm – 410
nm). In total, 21 samples were measured. The reference sample set was created to get a proper
model to predict the honeydew content on real cotton samples. Figure 6.4a shows the averaged
absorbance spectra in terms of reflectance. A baseline correction was applied to eliminate the
spectral offset due to scattering. In general, the spectral shapes of all samples are quite similar.
The most dominant band is pronounced approximately at 332 nm and weak shoulder can be rec-
ognized at 346 nm (sh). An absorption band is observed at 261 nm. Two weak shoulders are
remarked at 291 nm. Despite the efficiency of the detector and the weak intensity light source in
the spectral region between 250-270 nm [19], the signal is less intense, but nevertheless contain
useful spectroscopic information for the actual application. A prove of the remaining performance
of the setup in this range is given in Appendix Figure 10.3.
(a) (b) (c)

0.10 0.08
0.09 A
—A PC1 [68.0%]
B 0.06 PC2 [22.0%]
0.08 — B C PC4 PC4 [1.0%]
—C
0.04  D 0.04
PC4 [1.0%]
—D
E
- log (R/R0)
0.06
—E 0.02 PC2
F
—F 0.09
0.04 — CLN 0.0  CLN

0.00
- log (R/R0)
0.08
0.02 -0.02
0.07
-0.04
0.06
-0.04
0.00 290 300 310
Wavelength / nm
320
-0.07 PC1
-0.06
240 260 280 300 320 340 360 380 400 240 260 280 300 320 340 360 380 400
-0.5 Wavelength / nm
Wavelength / nm
0.0
0.0 0.2 0.4
0.5 -0.4 -0.2
Figure 6.4: (a) Averaged UV spectra of cotton samples with sugar solutions in different concentrations: A (2 wt %,
red), B (1 wt %, light green), C (0.5 wt %, blue), D (0.25 wt %, light blue), E (0.0125 wt %, pink), F
(0.0625 wt %, yellow) and CLN (mechanically cleaned, dark green). PCA sugar model for the cotton
samples with (b) scores and (c) corresponding loadings (PC1 black, PC2 red and PC4 blue).
84
Figure 6.4b and Figure 6.4c present the PCA model of the cotton samples with different concen-
trations of sugar. Figure 6.4b shows the scores plot for the first (68.0 %), second (22.0 %) and
fourth (1.0 %) principal component (PCs). These PCs explain nearly 91.0 % of the total variance.
The variance on PC3 is not necessary to distinguish between different sugar concentrations. For
completeness, PC3 is displayed in supplementary materials Appendix Figure 10.2. Different sugar
concentrations on cotton can clearly be distinguished by the PCA scores. The mechanically
cleaned sample (CLN) is separated on PC2. PC4 shows the separation between the highest sugar
concentration and the lowest concentration. Slight overlapping is observed due to inhomogeneity
of the impregnation procedure for the samples with sugar. The overlap tendency increases from
higher to lower concentrations as well as the variance within a sample increases with the concen-
tration. Each cluster overlaps with the two closest sugar concentrations (higher and lower).
Figure 6.4c shows the loadings plot for PC1, PC2 and PC4. The strongest influence on PC1 is at
330 nm. Most of PC1 describes the morphology of the fiber itself. PC2 has a minimum at 280 nm
and a maximum at 380 nm. These bands are responsible for the separation of the CLN sample
from the others and are distinguishing between the different concentrations. For PC4 a maximum
contribution is observed at 285 nm. These bands can be assigned to the presence of protein (see
also Appendix Figure 10.3) [141]. The most significant differences between those loadings are
found in the spectral region between 290 nm and 380 nm.
PLS-R is used for quantitative spectroscopic analysis. A PLS-R model was developed with a
calibration sample set n = 21 to correlate the spectral information with the sugar concentration.
Cotton samples (Table 6.2) were used for testing the performance of the model with a cross vali-
dation.
The variance explained by the model for the X- and Y-variables was 84 % by using five factors.
Accordingly, the five PLS factors were sufficient to describe the correlation between the spectra
and the sugar content. The accuracy of the calibration and validation were evaluated using the
coefficient of determination (R2) for the calibration (R2c = 0.84) and cross validation (R2cv = 0.84)
model. The root mean square error of calibration (RMSEC = 0.009 g) and cross validation
(RMSECV = 0.009 g) indicates the model performance. A high R2c and R2cv was achieved with
an extremely low RMSEC and RMSECV.
Figure 6.5 shows the PLS-R model for the cotton soaked in different concentrations of sugar in
the UV region (200 nm – 380 nm). Figure 6.5a presents the correlation between reference vs.
predicted, while the regression coefficients for the five factor model is shown in Figure 6.5b. For
85
model building and understanding the PLS-R factor loadings, loading weights for all five factors
are displayed in the Appendix Figure 10.4. Sample E and sample F have similar ratios 0.0326 and
0.0322 (sugar / g per dried cotton / g), respectively, due to the preparation procedure’s limit.
Therefore, they are overlapping in the reference vs. predicted plot. A negative band at 263 nm
and a positive band at 284 nm can be assigned to protein absorbance. An average spectrum of
pure dried protein is shown in the supplementary materials (Appendix Figure 10.3). The protein
information is pronounced in the spectra and mandatory for the model, even though the illumina-
tion and detector should be optimized [19]. From 300 nm – 400 nm several features are observed
that cannot be related to a common reason.
(a) (b)
Figure 6.5: PLS-R model for different sugar concentrations in the UV region (225 nm – 410 nm). (a) Predicted vs.
reference plot and (b) corresponding regression coefficients for the sugar content with a five factor PLS-
R model.
6.4.2 Predicting the amount of sugar and honeydew based

on the sugar PLS-R model
The performance of the PLS-R model was tested by two methods. First, cleaned cotton samples
were manually sprayed with aforementioned sugar concentrations to get a distribution of sugar
droplets on the cotton surfaces. One benefit of hyperspectral imaging is to get the lateral infor-
mation. Therefore, the PLS-R model was used to predict the sugar content on the different sam-
ples, the result is shown in Figure 6.6. In the distribution map a clear lateral classification of the
different sugar concentrations resulting in different ratios of sugar / g per dried cotton / g was
achieved. From sample A to sample F the ratios decrease, CLN samples are not soaked in sugar.
Each sample was prepared three times which are shown in the rows. Again, sample E and sample
F are indistinguishable, due to the preparation procedure’s limit. Overall, the averaged predicted
ratios decrease from samples A to samples CLN. The lateral inhomogeneities become visible
through hyperspectral imaging.
86
Figure 6.6: Distribution maps of the sugar content predicted on the mechanically cleaned cotton samples, which are
manually sprayed by sugar solution. The prediction of each pixel is based on the PLS-R sugar model. Each
rectangle represents a single cotton sample: A (2 wt %), B (1 wt %), C (0. 5 wt %), D (0. 25 wt %), E
(0.125 wt %), F (0.0625 wt %) and CLN (mechanically cleaned). The colored pixels (see the score value
range) represent the sugar content, from low (blue) to high (red).
Second, the PLS-R model was used to predict the honeydew content for each pixel of the real
cotton types labeled: light, strong and very strong. In Figure 6.7, the resulting distribution maps
are shown. As described for Figure 6.6, the distribution map shows a clear lateral classification
of different ratios of sugar / g per dried cotton / g. The amount of sugar highly correlates with the
amount of honeydew. Honeydew consists of different types of sugars and proteins [209]. From
the very strong samples to the light samples the ratios decrease. Three samples per type were
collected are given in the rows.
Therefore, in Figure 6.7 red and blue pixels represent high and low concentrations of honeydew,
respectively. As expected, the light sample displays a low honeydew concentration, while the
other two samples show an increase in the laterally resolved honeydew concentration. Compared
to the samples shown in Figure 6.6 the real samples show a more heterogeneous distribution of
honeydew on the samples. Even in the strong and very strong samples regions can be found where
almost no honeydew is present. This can be seen in the presence of blue pixels on the distribution
map of the very strong and strong samples.
87
(a) (b) (c)
Figure 6.7: Distribution maps of the sugar content predicted on the real cotton samples, which are contaminated by
honeydew. The prediction of each pixel is based on the PLS-R sugar model. Each rectangle represents a
single cotton sample ((a) very strong, (b) strong, (c) light). The colored pixels (see the score value range)
represent the sugar content, from low (blue) to high (red).
6.5 Conclusions
In summary, this proof of principle study has successfully demonstrated the identification and
quantification of honeydew on real cotton samples by combining UV hyperspectral imaging
(225 nm – 410 nm) with multivariate data analysis. For this novel approach, a reference sample
set was created based on mechanically cleaned cotton which has been impregnated with honey-
dew typical sugar and protein solutions for further UV hyperspectral imaging investigations. The
PCA model enables to classify the cotton samples according to their sugar concentration. A PLS-
R model was created that is able to predict laterally resolved the sugar/honeydew content pixel
by pixel. This is shown for reference samples and for real cotton samples that were labeled as
light, strong and very strong contaminated by honeydew. The lateral distribution of the ratio of
sugar / g per dried cotton / g per pixel giving a deeper insight into the distribution of honeydew
on real cotton samples.
To the best of our knowledge, this is the first scientific work reporting the identification, quanti-
fication and distribution of the amount of honeydew content by UV hyperspectral imaging. This
approach may provide an advantage in the industrial environment in the practical process appli-
cation and commercialization in the future. It enables to control the honeydew contamination in
the industrial processing of cotton fibers in real time. Hence, each cotton batch, independent of
the honeydew amount, can be manufactured to minimize waste and costs.
88
6.5 Conclusions
Supplementary Materials: The following supporting information can be downloaded at:

www.mdpi.com/xxx/s1, Appendix Table 10.2: Description of the cotton sample preparation; Ap-
pendix Figure 10.2: PCA sugar model for the cotton samples with (a) scores and (b) corresponding
loadings (PC1 black, PC3 red); Appendix Figure 10.2: Mean spectrum of pure dried protein on
PTFE.
Author Contribution: Conceptualization, M.A.K, M.S.; methodology, M.A.K and M.S.; soft-
ware, M.A.K, M.S. and F.W.; validation, M.A.K and M.S.; formal analysis, M.A.K., M.S. and
F.W.; investigation, M.A.K and M.S.; resources, E.O., K.R., and M.B.; data curation, M.A.K.,
M.S. and F.W.; writing—original draft preparation, M.A.K and M.S.; writing—review and edit-
ing, M.A.K, M.S., F.W., E.O., V.J., K.R. and M.B.; visualization, M.A.K and M.S.; supervision,
K.R. and M.B.; project administration, M.B.; All authors have read and agreed to the published
version of the manuscript.
Acknowledgments: We especially thank Thomas Blum (Reutlingen University) for his advises
and valuable discussion for this study. Additionally, our thanks goes to the company innospec
GmbH (Nürnberg, Germany) for helpful discussions and their expertise in this field. We would
like to thank Gunther Pehl (Reutlingen University) for manufacturing the sample holder. Last but
not least, we thank Barbara Boldrini and Tobias Drieschner (Reutlingen University) for her sup-
porting contributions as an expert in hyperspectral imaging.
89
7 Paper V: Applying UV
Hyperspectral Imaging for
Quantification of Honeydew
Content on Raw Cotton via PCA
and PLS-R Models
Mona Knoblich1,2,†, Mohammad Al Ktash1,2,†, Frank Wackenhut1, Volker Jehle3, Edwin Ostertag1
and Marc Brecht1,2,*
1
Center of Process Analysis and Technology (PA&T), School of Life Sciences, Reutlingen Uni-
versity, Alteburgstraße 150, 72762 Reutlingen, Germany, mona.knoblich@reutlingen-univer-
sity.de (M.K.); [email protected] (M.A.K.); frank.wackenhut@reut-
lingen-university.de (F.W.); [email protected] (E.O.)
2
Institute of Physical and Theoretical Chemistry, Eberhard Karls University Tübingen, Auf der
Morgenstelle 18, 72076 Tübingen, Germany
3
Texoversum Faculty Textile, Reutlingen University, Alteburgstraße 150, 72762 Reutlingen, Ger-
many, [email protected]
*
Correspondence: [email protected].
†
This is originally published in Textiles (https://doi.org/10.3390/textiles3030019) as
“Knoblich, M.; Al Ktash, M.; Wackenhut, F.; Jehle, V.; Ostertag, E. and Brecht, M. Applying UV
Hyperspectral Imaging for the Quantification of Honeydew Content on Raw Cotton via PCA and
PLS-R Models. Textiles 2023, 23, 287-293. https://doi.org/10.3390/textiles3030019”
91
7 Paper V: Applying UV Hyperspectral Imaging for Quantification of Honeydew Content on Raw Cotton via
PCA and PLS-R Models
7.1 Abstract
Cotton contamination by honeydew is considered one of the significant problems for quality in
textiles as it causes stickiness during manufacturing. Therefore, millions of dollars in losses are
attributed to honeydew contamination each year. This work presents the use of UV hyperspectral
imaging (225–300 nm) to characterize honeydew contamination on raw cotton samples. As ref-
erence samples, cotton samples were soaked in solutions containing sugar and proteins at different
concentrations to mimic honeydew. Multivariate techniques such as a principal component anal-
ysis (PCA) and partial least squares regression (PLS-R) were used to predict and classify the
amount of honeydew at each pixel of a hyperspectral image of raw cotton samples. The results
show that the PCA model was able to differentiate cotton samples based on their sugar concen-
trations. The first two principal components (PCs) explain nearly 91.0% of the total variance. A
PLS-R model was built, showing a performance with a coefficient of determination for the vali-
dation (R2cv) = 0.91 and root mean square error of cross-validation (RMSECV) = 0.036 g. This
PLS-R model was able to predict the honeydew content in grams on raw cotton samples for each
pixel. In conclusion, UV hyperspectral imaging, in combination with multivariate data analysis,
shows high potential for quality control in textiles
7.2 Introduction
Cotton is widely regarded as an essential natural material in various textile products, from fabrics
to clothing [210,211]. It is considered one of the most imported and exported materials worldwide
[141]. Therefore, the assessment of the cotton quality is needed. Cotton contamination is one of
the most significant problems for quality [93,95,97,108,114,115,210,212]. The most relevant impu-
rities in raw cotton arise from insects producing honeydew. Honeydew is sugar-rich, excreted by
whiteflies and aphids, causing stickiness during manufacturing [213,214]. Therefore, it can cause
problems during processing, and the final product shows low quality. Modern techniques and
methods have appeared due to the increasing demand for higher processing and quality control.
These include off-line methods such as thermogravimetric analysis and single point spectroscopy.
However, these techniques are slow and time-consuming [36,91,103,162,168,173,201]. In contrast,
in- and on-line methods, such as hyperspectral imaging, are non-destructive and rapid, enabling
real-time data acquisition and analysis [141]. Hyperspectral imaging is a type of spectroscopic
imaging that allows for collecting and analyzing massive data spanning a wide wavelength range.
It involves both spectral and spatial information with high resolution. Hyperspectral imaging gen-
erates large amounts of data, requiring multivariate data analysis techniques such as principal
92
component analysis (PCA) and partial least squares regression (PLS-R) [54]. PCA can identify
and visualize groups within data clusters, while PLS-R is used to build quantitative models and
generate data clusters. It is also helpful for evaluating the robustness of these models, making it a
powerful tool for data analysis. Combining these two techniques is often required to analyze and
interpret the results of high-resolution hyperspectral imaging effectively [19,21,22,53,115,141]. In
a previous study, we developed a method using UV imaging to predict honeydew quantity on
cotton samples. The approach involves using a xenon-arc lamp to quantify the amount of honey-
dew in the UV-A and UV-B ranges. However, it could not accurately detect it in UV-C due to the
lamp's intensity limitations [54]. In this study, we overcome this limitation by using a deuterium
lamp as light source. Mechanically cleaned cotton was soaked with a sugar and protein containing
solution at different concentrations that are typical for honeydew. Chemometric models such as
PCA and PLS-R were established using UV hyperspectral images. Cotton samples were catego-
rized by sugar concentration using PCA, while PLS-R was used to correlate UV spectra with
sugar concentration. The PLS-R model accurately predicted the amount of honeydew in grams
on the raw cotton samples.
7.3.1 Chemicals and preparation of solutions and samples

The sugar and protein solutions applied to the cotton samples were formulated to mimic natural
honeydew [208,215,216]. 0.2 g of each macronutrient (glucose, fructose, sucrose, melezitose, tre-
halose, and protein) was weighed and dissolved in 10 mL of deionized water. A six-fold serial
dilution was prepared in 50 mL volumetric flasks by mixing 25 mL of the previous solution with
25 mL of deionized water for 2 minutes at each dilution step (Table 7.1).
In total, 24 mechanically cleaned cotton samples were prepared with a weight of 0.3 g ± 1 mg of
each sample. The samples were dried in a vacuum oven (Vacutherm VT 6130 M, Thermo Fisher
Scientific Inc., Waltham, MA, USA) at 30 °C and 50 mbar for 8 h to remove absorbed humidity.
4 mL of the aforementioned solution was used for soaking three samples per concentration. The
samples were dried again in a desiccator at room temperature for one month.
Raw cotton samples were collected by ICA Bremen GmbH (Bremen, Germany) to test the model's
predictive power. The samples were chosen according to their honeydew content in the steps light,
strong, and very strong [54,211]. The sample types are named from A to F, and one mechanically
cleaned (CLN) sample, where A has the highest concentration of sugar and protein solution
93
(4 wt%), and G has the lowest concentration (0.0625 wt%) (Table 7.1). The average ratio of sugar
mass to dried cotton mass (msugar/mcotton) remaining on the samples was calculated after drying the
sample for one month (Table 7.1). The term macronutrients is omitted to describe the solution
and replaced with the short-term "sugar" for the sample nomenclature.
Table 7.1: The sugar solution concentration and the weighted average sugar on cotton samples.
Sample type Sugar concentration / wt % msugar/mcotton

A 4 0.4249
B 2 0.2413
C 1 0.1194
D 0.5 0.0609
E 0.25 0.02313
F 0.125 0.0126
G 0.0625 0.0143
CLN - -
7.3.2 UV hyperspectral imaging setup and data processing

Compared to our previous studies [19,53,54], the illumination of the hyperspectral imaging setup
was modified; now a deuterium lamp (SL 3, StellarNet Inc, 24 V, 65.04 W, Tampa, Florida, USA)
is used, providing a higher illumination strength in UV-C region compared to the xenon-arc lamp
(e.g around 230 nm the intensity difference for deuterium illumination higher than xenon-arc
lamp ). Thus, the PTFE tunnel covering the convey belt for increasing the illumination strength
was no longer necessary.
Multivariate data analysis was acquired with “Aspen UnscramblerTM, version 10.5.1” (Aspen
Technology Inc., Bedford, MA, USA). The PCA model was calculated with mean centering,
cross-validation, and the NIPALS algorithm. A PLS-R model for the sugar concentrations was
processed with mean centering, category variable with eight segmented cross-validations, and the
Kernel algorithm.

The averaged absorbance spectra in terms of reflectance after a linear baseline correction are
shown in Figure 7.1a. The spectra show an almost linear decrease in the reflectivity for all sugar
concentrations. In the range of 275 nm and 295 nm are broad bands showing clear dependences
on the sugar concentration. These bands can be assigned to protein, cellulose, and lignin. A much
weaker band between 230–255 nm corresponds to the presence of pectin and DNA [217-221].
94
(a) (b) (c)

0.4 0.08
A
0.20 B
C
PC1
D 0.2 0.04
PC2 [12.0%]
E
0.15
- log (R/R0)
F
G
0.0 0.00
CLN
0.10
-0.04
-0.2
0.05
PC2
PC1[76.0%]
-0.08 PC2[12.0%]
0.00 -0.4
230 240 250 260 270 280 290 300 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 230 240 250 260 270 280 290 300
Wavelength / nm PC1 [79.0%] Wavelength / nm
Figure 7.1: (a) Averaged spectra recorded via UV hyperspectral imaging of raw cotton samples with sugar solutions in
different concentrations: A (4 wt %), B (2 wt %), C (1 wt %), D (0.5 wt %), E (0.25 wt %), F (0.125 wt
%), G (0.0625 wt %), and CLN (mechanically cleaned). PCA sugar model for the cotton samples with (b)
scores on the first principal component (PC1) and second principal component (PC2) and (c) correspond-
ing loadings.
Figure 7.1b,c present the cotton samples' PCA model at each sample pixel with different sugar
concentrations. Figure 1b shows the scores plot for the first (79.0%) and second (12.0%) principal
components (PCs). These PCs explain nearly 91.0% of the total variance.
The PCA scores enable to distinguish different sugar concentrations on cotton. On PC1, high
sugar concentrations are separated from low concentrations, while on PC2, the mechanically
cleaned sample (CLN) shows distinct separation from the sample with high sugar concentrations.
Moreover, different sugar concentrations on cotton can clearly be distinguished by the PC2. An
overlap naturally results from the preparation method chosen which results in a certain inhomo-
geneity. With decreasing concentration, the degree of overlap between samples increases together
with the variance within the sample. Each cluster shows overlap with the two nearest sugar con-
centrations (higher and lower). Figure 7.1c shows the loadings plots for PC1 and PC2. The most
significant differences between those loadings are found between 250 nm - 280 nm in the spectral
region. The maximum influence on PC1 at 250 nm and the minimum at 283 nm. Most of PC1
describes a clear dependence on the concentrations of sugar on the cotton samples. PC2 has a
maximum at 249 nm and a minimum at 282 nm. These bands represent the chromophores, pectin,
and DNA in the cotton fibers [217,222].
PLS-R was utilized as a technique for quantitative spectroscopic analysis. A PLS-R model was
developed using a calibration sample set of 24 samples to establish a correlation between the
spectral information and the sugar content. The PLS-R model's performance was tested using
cotton samples (Table 7.1) with different concentrations of sugar solutions.
The PLS-R model for the X- and Y-variables explained 91% of the variance. Five PLS-R factors
were sufficient to describe the correlation between the spectra and sugar content. The accuracy of
95
the calibration and validation were evaluated using the coefficient of determination (R2) for the
calibration (R2c = 0.9) and validation (R2cv = 0.91) model. The quality of the models were evalu-
ated according to values of the error of calibration, RMSEC = 0.03 g, and the error of cross-
validation, RMSECV = 0.036 g. High R2c and R2cv values are achieved with extremely low
RMSEC and RMSECV values.
Figure 7.2 presents the PLS-R model for cotton samples soaked with different concentrations of
sugar. For model building and understanding the PLS-R factor loadings, loading weights for all
three factors are displayed in the supplementary materials Appendix Figure 10.5. Figure 7.2a dis-
plays the correlation between the predicted and reference values, whereas the regression coeffi-
cients for the three-factor model are illustrated in Figure 7.2b. Samples F and G have similar
ratios, 0.0324 and 0.0321 (msugar/mcotton), hence they overlap in the regression coefficients plot.
Two negative bands at 235 nm and 282 nm and one positive band at 250 nm can be assigned to
protein and pectin absorbance [141,217].
(a) (b)
Figure 7.2: Five-factor PLS-R model for different sugar contents in the UV region (225 nm – 300 nm). (a) Predicted
vs. reference plot and (b) corresponding regression coefficients.
The PLS-R model was used to predict the honeydew content for each pixel of a hyperspectral
image. Three raw cotton samples of three grades of honeydew contamination (very strong, strong,
and light) were collected, and the resulting distribution maps are shown in Figure 7.3. The distri-
bution maps present a clear lateral classification of different ratios of msugar/mcotton, and the pre-
dicted ratios decrease from the very strong samples to the light samples. The sugar content is
highly correlated with the honeydew amount [217]. The analysis reveals a highly variable distri-
bution of honeydew across all samples. Some regions present minimal contamination, while oth-
ers, including areas/pixels in the light samples, exhibit up to 0.1 msugar/mcotton ratio, comparable to
those found in very strong samples. The observed inhomogeneity in honeydew distribution sug-
96
7.5 Conclusions
gests that our soaking method for the sugar solution is a realistic approach, as it induces a com-
parable level of variability. However, the inhomogeneity seems to be even higher in the raw sam-
ples, as shown in Figure 7.3.
(a) (b) (c)
Figure 7.3: Distribution maps of the sugar content predicted for each pixel of the UV hyperspectral imaging data from
the five-factor PLS-R model on the raw cotton samples contaminated by honeydew. Each rectangle repre-
sents a single cotton sample ((a) very strong, (b) strong, (c) light). The colored pixels (see the score value
range) represent the sugar content, from low (blue) to high (red).
7.5 Conclusions
UV hyperspectral imaging (225–300 nm) was combined with multivariate data analysis to suc-
cessfully identify and quantify honeydew on raw cotton samples. Therefore, a reference sample
set based on cotton samples was prepared and imaged in UV.
The samples were soaked with solutions containing sugar and proteins at different concentrations
to mimic honeydew. A PCA model enabled the classification of the cotton samples according to
their sugar concentrations. The PLS-R model was able to predict laterally resolved honeydew
content pixel by pixel in grams on raw cotton samples. The analysis reveals that the raw cotton
samples have an inhomogeneous distribution of honeydew. Therefore, the chosen soaking method
closely approximates the distribution patterns observed in the raw samples. The results were ob-
tained by analyzing samples labeled as light, strong, and very strong contaminated with honey-
dew. This combination of hyper-spectral imaging with multivariate data analysis represents a high
potential technique for detecting honeydew contamination in real-time.
97
Supplementary Materials: The following supporting information can be downloaded at:

www.mdpi.com/xxx/s1, Figure S1. X-loading weights and x-loadings for factor 1 (a, b), factor 2
(c, d), and factor 3 (e, f), respectively.
Author Contribution: Conceptualization, M.A.K, M.S.; methodology, M.A.K and M.S.; soft-
ware, M.A.K, M.S. and F.W.; validation, M.A.K and M.S.; formal analysis, M.A.K., M.S. and
F.W.; inves-tigation, M.A.K and M.S.; resources, E.O. and M.B.; data curation, M.A.K., M.S.
and F.W.; writ-ing—original draft preparation, M.A.K and M.S.; writing—review and editing,
M.A.K, M.S., F.W., E.O., V.J. and M.B.; visualization, M.A.K and M.S.; supervision M.B.; pro-
ject administration, M.B.; All authors have read and agreed to the published version of the man-
uscript.
Funding: This research received no external funding.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are
solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s).
MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting
from any ideas, methods, instructions or products referred to in the content.
98
8 Conclusion and Summary
In the first part of this thesis presented the application of UV hyperspectral imaging for the char-
acterization of active pharmaceutical ingredients (APIs) in tablets. Two sets of samples were an-
alyzed: one set containing tablets with 100% API content and another set consisting of commer-
cially available painkiller tablets. The results demonstrated that UV hyperspectral imaging,
combined with PCA, is a promising approach for detecting and differentiating different drug sam-
ples. The PCA models generated were able to effectively separate and classify the various drug
types based on their spectral characteristics. In particular, the first two PCs captured most of the
spectral variance and allowed for clear discrimination between the samples. The advantage of the
home-built UV hyperspectral imaging setup used in the study was its spatial and spectral resolu-
tion and data acquisition speed sufficient for scientific purposes. The results obtained from this
setup suggest that it can be easily adapted to meet the requirements of real industrial processes.
The comparison between the hyperspectral imaging results and reference measurements per-
formed using a commercial UV spectrometer indicated that while the hyperspectral imaging data
provided valuable information, there were some limitations. The sensitivity and efficiency of the
hyperspectral imaging setup were lower in the wavelength range below 275 nm, resulting in less
reliable spectroscopic information in that region. However, despite this limitation, combining UV
hyperspectral imaging and chemometric modeling enabled accurate classification and separation
of the different drug samples. Overall, the study demonstrated the potential of UV hyperspectral
imaging as a tool for rapid and non-destructive characterization of pharmaceutical tablets. The
technique has implications for quality control, process monitoring, and real-time release testing
in industrial settings. Further advancements in UV hyperspectral imaging technology and data
analysis methods are expected to enhance its applications in the pharmaceutical industry.
To confirm the effectiveness of hyperspectral imaging, 28 direct bonded copper samples were
characterized to determine the oxidation states of copper and measure the thickness of the oxide
layer in the UV region. Single-point UV spectroscopy was used as a basis of comparison for the
results obtained from the UV hyperspectral imaging setup. The results showed that UV hyper-
spectral imaging, in combination with PCA and PLS-R, offered a promising approach for detect-
ing and differentiating copper states and measuring oxide layer thickness with lateral resolution.
PCA models effectively separated the different types of direct bonded copper based on their cop-
per states and oxide layer thicknesses using only the first two PCs. With three factors, the PLS-R
99
models exhibited high coefficients of determination (R2) and low RMSE for calibration, cross-
validation, and prediction. The models successfully correlated the spectral features with the oxide
layer thickness. Notably, this study is the first to report the identification and quantification of
copper oxide thin films using UV hyperspectral imaging. The UV hyperspectral imaging setup
utilized in this study offered a high spatial and spectral resolution, along with relatively fast data
acquisition under laboratory conditions. The design and data presented in this study provide a
foundation for developing a UV hyperspectral imaging setup that meets the requirements of real
industrial processes. Overall, this study demonstrated the efficacy of UV hyperspectral imaging
and UV reflectance spectroscopy in characterizing direct bonded copper samples, paving the way
for future applications in industrial processes for real-time data acquisition, process control, and
in-line classification.
In the second part of this thesis, it was demonstrated that the integration of UV-Vis/NIR reflec-
tance spectroscopy and hyperspectral imaging, coupled with PCA, offers a powerful technique
for detecting and differentiating raw cotton types. It was determined that the most crucial infor-
mation for distinguishing cotton types was present in the UV and NIR spectra. In the UV range,
the most significant factor was identified as the protein contribution at 280 nm, while the key
contributors to absorbance in the NIR range were CH3 vibrations at 1775 nm and ROH vibrations
at 1500 nm. PCA analysis of the spectral data successfully classified all cotton varieties, with the
first two PCs accounting for the maximum data variance. Moreover, the investigation revealed
that the most dominant absorbance corresponding to CH3 and ROH in the case of NIR hyperspec-
tral imaging occurred at 2270 nm and 1525 nm, respectively. Two methods were employed to
process the substantial data generated, facilitating the differentiation of all cotton varieties. The
rugged online home-built setup used in the study presented benefits such as spatial/spectral reso-
lution and rapid data collection, enabling the assessment of multiple samples in a brief period and
at a low cost. In conclusion, based on the provided data, it is feasible to develop a streamlined
chemometric model that adheres to the demands of real-world industrial processes while main-
taining suitable standards and accuracy. This approach could offer a practical and cost-effective
means of distinguishing between various raw cotton types, potentially having considerable impli-
cations for the textile industry.
In the third part, UV hyperspectral imaging and multivariate data analysis techniques were suc-
cessfully utilized to identify, quantify, and spatially resolve honeydew content on cotton samples.
The aim is to address the need for real-time control of honeydew contamination in the industrial
processing of cotton fibers. A total of 21 cotton samples labeled A to F and one mechanically
100
7.5 Conclusions
cleaned (CLN) were prepared by soaking in sugar with different concentrations ranging from 2
wt % to 0.0625 wt %. These samples were used as reference samples to create a dataset for the
UV hyperspectral imaging investigations. Additionally, real cotton samples were chosen accord-
ing to honeydew contents in the steps of light, strong, and very strong. These samples were uti-
lized to assess the performance of the developed models. The honeydew contamination was
achieved by exposing the cotton samples to honeydew typical sugar and protein solutions in real-
world simulating conditions. The results showed that UV hyperspectral imaging, in combination
with PCA and PLS-R, offered a promising approach for predicting the amount of honeydew con-
taminated on real cotton. PCA is applied to classify the cotton samples based on their sugar con-
centrations. The PCA model effectively separates samples with different sugar concentrations,
allowing for accurate classification.
Additionally, the model reveals the spectral regions that significantly contribute to the classifica-
tion, such as those associated with fiber morphology and protein presence. To achieve quantitative
analysis, a PLS-R model is developed. The PLS-R model establishes a strong correlation between
spectral information and sugar concentration, enabling accurate sugar content prediction on ref-
erence and real cotton samples. This predictive capability provides valuable insights into the dis-
tribution of honeydew on cotton samples.
The study emphasizes the novelty of its findings, as it is the first scientific work to report the
identification, quantification, and distribution mapping of honeydew content using UV hyper-
spectral imaging. The spatial mapping of sugar concentrations and honeydew contamination on
cotton samples provides a deeper understanding and control over honeydew contamination in the
industrial processing of cotton fibers. Overall, this research holds promise for practical applica-
tions in the industry by enabling real-time control of honeydew contamination, optimizing man-
ufacturing processes, and minimizing waste. The combination of UV hyperspectral imaging and
multivariate data analysis techniques offers a powerful approach for efficiently and accurately
assessing honeydew contamination on cotton samples. Further exploration of this methodology
on a larger scale and its potential application to other agricultural products could enhance produc-
tivity and sustainability across various industries.
Finally, The UV hyperspectral imaging prototype underwent modifications compared to previous

studies. Specifically, a deuterium lamp was utilized as the illumination source, providing higher
illumination strength in the UV-C region compared to the previous xenon-arc lamp, eliminating
the need for a PTFE tunnel to increase illumination strength. These modifications enhanced the
performance and accuracy of the hyperspectral imaging system for honeydew detection on cotton
101
samples. Therefore, applying UV hyperspectral imaging combined with multivariate data analysis
techniques, PCA and PLS-R, proved an effective method for identifying and quantifying honey-
dew contamination on raw cotton samples. By creating a reference sample set that mimicked
honeydew using sugar and protein solutions at different concentrations, the UV hyperspectral
imaging system was able to capture and analyze the spectral information of the cotton samples.
The PCA model showed the ability to differentiate cotton samples based on their sugar concen-
trations, providing a means for categorizing the contamination levels. On the other hand, the PLS-
R model demonstrated strong performance in predicting and quantifying the honeydew content
in grams on a pixel-by-pixel basis. The model achieved a high accuracy with an R2 of 0.91 and
low RMSECV = 0.036 g.
The results revealed the presence of distinct spectral bands within the UV range (225 nm – 300
nm) that were correlated with sugar concentration and indicative of honeydew contamination. The
distribution maps generated by the PLS-R model showed lateral classification of different sugar
ratios and revealed the inhomogeneity of honeydew distribution on the raw cotton samples. The
spraying and soaking method employed to simulate honeydew contamination proved realistic and
induced comparable levels of variability in the samples.
Overall, this study demonstrates the potential of UV hyperspectral imaging combined with mul-
tivariate data analysis as a rapid and non-destructive technique for quality control in textiles. Ac-
curately identifying and quantifying honeydew contamination on raw cotton samples can signifi-
cantly improve the quality and processing of cotton in the textile industry, thereby reducing the
financial losses associated with such contamination. Further research and optimization of the
technique may enhance its applicability and broaden its potential for other quality control appli-
cations in textiles and beyond.
102
9 Bibliography
1. Atascientific. Spectrometry and spectroscopy: what’s the difference? Available online:
https://www.atascientific.com.au/spectrometry/ (accessed on 17.01.2020).
2. Wikipedia. Spectroscopy. Available online: https://en.wikipedia.org/wiki/Spectroscopy
(accessed on 19.04.2023).
3. Penner, M.H. Basic principles of spectroscopy. Food analysis 2017, 79-88,
doi:10.1007/978-3-319-45776-5_6.
4. Smith, G.S. An introduction to classical electromagnetic radiation; Cambridge University
Press: 1997.
5. LibreTexts. The Electromagnetic Spectrum. Available online:
https://chem.libretexts.org/@go/page/2398 (accessed on 23.03.2023).
6. Skrabal, P.M. Spectroscopy: An interdisciplinary integral description of spectroscopy from
UV to NMR; vdf Hochschulverlag AG: 2012.
7. Sindhu, P. Fundamentals of Molecular Spectroscopy; New Age International: 2006.
8. Gribov, L. Introduction to molecular spectroscopy. Moscow Izdatel Nauka 1976.
9. Hardesty, J.H.; Attili, B. Spectrophotometry and the Beer-Lambert Law: An important
analytical technique in chemistry. Collin College Department of Chemistry 2010.
10. Parnis, J.M.; Oldham, K.B. Beyond the Beer–Lambert law: The dependence of absorbance
on time in photochemistry. Journal of Photochemistry Photobiology 2013, 267, 6-10,
doi:10.1016/j.jphotochem.2013.06.006.
11. Fellers, T.; Davidson, M. Introduction to the Reflection of Light. Available online:
https://www.olympus-lifescience.com/en/microscope-
resource/primer/lightandcolor/reflectionintro/ (accessed on 01.04.2023).
12. Blanco, M.; Villarroya, I. NIR spectroscopy: a rapid-response analytical tool. TrAC Trends
in Analytical Chemistry 2002, 21, 240-250, doi:10.1016/S0165-9936(02)00404-1.
13. Manso, M.; Carvalho, M.L. Application of spectroscopic techniques for the study of paper
documents: A survey. Spectrochimica Acta 2009, 64, 482-490,
doi:10.1016/j.sab.2009.01.009.
14. Haroon, K.; Arafeh, A.; Martin, P.; Rodgers, T.; Mendoza, Ć.; Baker, M. Use of inline
near‐infrared spectroscopy to predict the viscosity of shampoo using multivariate
analysis. International journal of cosmetic science 2019, 41, 346-356,
doi:10.1111/ics.12536.
15. Ni, C.; Li, Z.; Zhang, X.; Sun, X.; Huang, Y.; Zhao, L.; Zhu, T.; Wang, D. Online Sorting
of the Film on Cotton Based on Deep Learning and Hyperspectral Imaging. IEEE Access
2020, 8, 93028-93038, doi:10.1109/ACCESS.2020.2994913.
16. Haven, J.J.; Junkers, T. Online monitoring of polymerizations: current status. European
Journal of Organic Chemistry 2017, 2017, 6474-6482, doi:10.1002/ejoc.201700851.
17. Lewis, E.; Schoppelrei, J.; Lee, E.; Kidder, L. Near-infrared chemical imaging as a process
analytical tool. Process analytical technology 2008, p.187-225.
18. Li, Q.; He, X.; Wang, Y.; Liu, H.; Xu, D.; Guo, F. Review of spectral imaging technology
in biomedical engineering: achievements and challenges. Biomedical optics 2013, 18,
100901, doi:10.1117/1.JBO.18.10.100901.
19. Al Ktash, M.; Stefanakis, M.; Boldrini, B.; Ostertag, E.; Brecht, M. Characterization of
Pharmaceutical Tablets Using UV Hyperspectral Imaging as a Rapid In-Line Analysis
Tool. Sensors 2021, 21, 4436, doi:10.3390/s21134436.
IX
9 Bibliography
20. Gowen, A.A.; O'Donnell, C.P.; Cullen, P.J.; Downey, G.; Frias, J.M. Hyperspectral
imaging–an emerging process analytical tool for food quality and safety control. Trends
in food science & technology 2007, 18, 590-598.
21. Boldrini, B.; Kessler, W.; Rebner, K.; Kessler, R.W. Hyperspectral imaging: a review of
best practice, performance and pitfalls for in-line and on-line applications. Journal of
near infrared spectroscopy 2012, 20, 483-508, doi:10.1255/1003.
22. Gowen, A.; Odonnell, C.; Cullen, P.; Downey, G.; Frias, J. Hyperspectral imaging – an
emerging process analytical tool for food quality and safety control. Trends in Food
Science & Technology 2007, 18, 590-598, doi:10.1016/j.tifs.2007.06.001.
23. Al Ktash, M.; Hauler, O.; Ostertag, E.; Brecht, M. Ultraviolet-visible/near infrared
spectroscopy and hyperspectral imaging to study the different types of raw cotton.
Journal of Spectral Imaging 2020, 9, a1, doi:10.1255/jsi.2020.a18.
24. Lu, G.; Fei, B. Medical hyperspectral imaging: a review. Biomed Opt 2014, 19, 10901,
doi:10.1117/1.JBO.19.1.010901.
25. Manolakis, D.; Shaw, G. Detection algorithms for hyperspectral imaging applications.
IEEE signal processing magazine 2002, 19, 29-43, doi:10.1109/79.974724.
26. Stiedl, J.; Boldrini, B.; Green, S.; Chassé, T.; Rebner, K. Characterisation of oxide layers
on technical copper based on visible hyperspectral imaging. Journal of Spectral Imaging
2019, 8, doi:10.1255/jsi.2019.a10.
27. Grahn, H.; Geladi, P. Techniques and applications of hyperspectral image analysis; John
Wiley & Sons: 2007.
28. Amigo, J.M.; Santos, C. Preprocessing of hyperspectral and multispectral images. In Data
Handling in Science and Technology; Elsevier: 2020; Volume 32, pp. 37-53.
29. Vasefi, F.; MacKinnon, N.; Farkas, D. Hyperspectral and multispectral imaging in
dermatology. In Imaging in Dermatology; Elsevier: 2016; pp. 187-201.
30. Gao, L.; Smith, R.T. Optical hyperspectral imaging in microscopy and spectroscopy–a
review of data acquisition. Journal of biophotonics 2015, 8, 441-456,
doi:10.1002/jbio.201400051.
31. Bassler, M.C.; Stefanakis, M.; Sequeira, I.; Ostertag, E.; Wagner, A.; Bartsch, J.W.;
Roessler, M.; Mandic, R.; Reddmann, E.F.; Lorenz, A.; et al. Comparison of Whiskbroom
and Pushbroom darkfield elastic light scattering spectroscopic imaging for head and neck
cancer identification in a mouse model. Anal Bioanal Chem 2021, 413, 7363-7383,
doi:10.1007/s00216-021-03726-5.
32. Kamruzzaman, M.; ElMasry, G.; Sun, D.-W.; Allen, P. Application of NIR hyperspectral
imaging for discrimination of lamb muscles. Journal of food engineering 2011, 104, 332-
340, doi:10.1016/j.jfoodeng.2010.12.024.
33. Türker-Kaya, S.; Huck, C.W. A review of mid-infrared and near-infrared imaging:
principles, concepts and applications in plant tissue analysis. Molecules 2017, 22, 168,
doi:10.3390/molecules22010168.
34. Ariana, D.P.; Lu, R. Quality evaluation of pickling cucumbers using hyperspectral
reflectance and transmittance imaging: Part I. Development of a prototype. Sensing
Instrumentation for Food Quality Safety 2008, 2, 144-151, doi:10.1007/s11694-008-
9057-x.
35. Barnaby, J.Y.; Huggins, T.D.; Lee, H.; McClung, A.M.; Pinson, S.R.; Oh, M.; Bauchan,
G.R.; Tarpley, L.; Lee, K.; Kim, M.S. Vis/NIR hyperspectral imaging distinguishes sub-
population, production environment, and physicochemical grain properties in rice.
Scientific Reports 2020, 10, 1-13.
36. Rodgers, J.; Beck, K. NIR characterization and measurement of the cotton content of dyed
blend fabrics. Textile research journal 2009, 79, 675-686,
doi:10.1177/0040517508090884.
X
9 Bibliography
37. Schlapfer, D.R.; Kaiser, J.W.; Brazile, J.; Schaepman, M.E.; Itten, K.I. Calibration concept
for potential optical aberrations of the APEX pushbroom imaging spectrometer. In
Proceedings of the Sensors, Systems, and Next-Generation Satellites VII, 2004; pp. 221-
231.
38. Bro, R.; Smilde, A.K. Principal component analysis. Analytical methods 2014, 6, 2812-
2831.
39. Calvini, R.; Ulrici, A.; Amigo, J. Sparse-Based Modeling of Hyperspectral Data. In Data
Handling in Science and Technology; Elsevier: 2016; Volume 30, pp. 613-634.
40. DESY, D.E.-S. P66 Time-resolved luminescence spectroscopy. Available online:
https://photon-
science.desy.de/facilities/petra_iii/beamlines/p66_superlumi/index_eng.html (accessed
on 06.09.2023).
41. Wiedemann, H.; Wiedemann, H. Synchrotron radiation; Springer: 2003.
42. Song, Z.; Li, J.; Davis, K.D.; Li, X.; Zhang, J.; Zhang, L.; Sun, X. Emerging Applications
of Synchrotron Radiation X‐Ray Techniques in Single Atomic Catalysts. Small Methods
2022, 6, 2201078, doi:10.1002/smtd.202201078.
43. Kim, K.J. Characteristics of synchrotron radiation. In Proceedings of the AIP conference
proceedings, 1989; pp. 565-632.
44. Chawla, A.; Lobacz, A.; Tarapata, J.; Zulewska, J. UV light application as a mean for
disinfection applied in the dairy industry. Applied Sciences 2021, 11, 7285,
doi:10.3390/app11167285.
45. Tokode, O.; Prabhu, R.; Lawton, L.A.; Robertson, P.K. UV LED sources for heterogeneous
photocatalysis. Environmental Photochemistry Part III 2015, 159-179,
doi:10.1007/698_2014_306.
46. Ditmire, T.; Gumbrell, E.; Smith, R.; Mountford, L.; Hutchinson, M. Supersonic ionization
wave driven by radiation transport in a short-pulse laser-produced plasma. Physical
review letters 1996, 77, 498, doi:10.1103/PhysRevLett.77.498.
47. Saunders, R.; Ott, W.; Bridges, J. Spectral irradiance standard for the ultraviolet: the
deuterium lamp. Applied Optics 1978, 17, 593-600, doi:10.1364/AO.17.000593.
48. Zama, T.; Awazu, K.; Onuki, H. Improvement of the aging characteristics of deuterium
lamp. Journal of electron spectroscopy related phenomena 1996, 80, 493-496,
doi:10.1016/0368-2048(96)03024-1.
49. Lin, B.J. Deep UV lithography. Journal of vacuum science technology 1975, 12, 1317-
1320, doi:10.1116/1.568527.
50. Davidson, M.W. Fundamentals of Mercury Arc Lamp. Available online: https://zeiss-
campus.magnet.fsu.edu/articles/lightsources/mercuryarc.html (accessed on 15.05.2023).
51. Davidson, M.W. Fundamentals of Xenon Arc Lamp. Available online: https://zeiss-
campus.magnet.fsu.edu/articles/lightsources/xenonarc.html (accessed on 15.05.2023).
52. Andreas Nolte, L.H., Michael W. Davidson. Fundamentals of Illumination Sources for
Optical Microscopy. Available online: https://zeiss-
campus.magnet.fsu.edu/articles/lightsources/lightsourcefundamentals.html (accessed on
15.05.2023).
53. Al Ktash, M.; Stefanakis, M.; Englert, T.; Drechsel, M.S.; Stiedl, J.; Green, S.; Jacob, T.;
Boldrini, B.; Ostertag, E.; Rebner, K.; et al. UV hyperspectral imaging as process
analytical tool for the characterization of oxide layers and copper states on direct bonded
copper. Sensors 2021, 21, 7332, doi:10.3390/s21217332.
54. Al Ktash, M.; Stefanakis, M.; Wackenhut, F.; Jehle, V.; Ostertag, E.; Rebner, K.; Brecht,
M. Prediction of Honeydew Contaminations on Cotton Samples by In-Line UV
Hyperspectral Imaging. Sensors 2023, 23, 319, doi:10.3390/s23010319.
XI
9 Bibliography
55. Knoblich, M.; Al Ktash, M.; Wackenhut, F.; Jehle, V.; Ostertag, E.; Brecht, M. Applying
UV Hyperspectral Imaging for the Quantification of Honeydew Content on Raw Cotton
via PCA and PLS-R Models. Textiles 2023, 3, 287-293, doi:10.3390/textiles3030019.
56. Trygg, J.; Holmes, E.; Lundstedt, T. Chemometrics in metabonomics. Journal of proteome
research 2007, 6, 469-479, doi:10.1021/pr060594q.
57. Wold, S. Chemometrics; what do we mean with it, and what do we want from it?
Chemometrics Intelligent Laboratory Systems 1995, 30, 109-115, doi:10.1016/0169-
7439(95)00042-9.
58. Kumar, N.; Bansal, A.; Sarma, G.; Rawal, R.K. Chemometrics tools used in analytical
chemistry: An overview. Talanta 2014, 123, 186-199, doi:10.1016/j.talanta.2014.02.003.
59. Rebner, K. Ortsaufgelöste Streulichtspektroskopie an mikrostrukturierten Systemen.
Eberhard-Karls-Universität Tübingen, 2010.
60. Wold, S.; Esbensen, K.; Geladi, P. Principal component analysis. Chemometrics Intelligent
Laboratory Systems 1987, 2, 37-52, doi:10.1016/0169-7439(87)80084-9.
61. Geladi, P.; Isaksson, H.; Lindqvist, L.; Wold, S.; Esbensen, K. Principal component
analysis of multivariate images. Chemometrics Intelligent Laboratory Systems 1989, 5,
209-220, doi:10.1016/0169-7439(89)80049-8.
62. Esbensen, K.H.; Guyot, D.; Westad, F.; Houmoller, L.P. Multivariate data analysis: in
practice: an introduction to multivariate data analysis and experimental design, 6 ed.;
Multivariate Data Analysis: 2002.
63. Kazakevich, Y.V.; Lobrutto, R. HPLC for pharmaceutical scientists; John Wiley & Sons:
2007.
64. Stefanakis, M.; Lorenz, A.; Bartsch, J.W.; Bassler, M.C.; Wagner, A.; Brecht, M.;
Pagenstecher, A.; Schittenhelm, J.; Boldrini, B.; Hakelberg, S. Formalin fixation as tissue
preprocessing for multimodal optical spectroscopy using the example of human brain
tumour cross sections. Journal of spectroscopy 2021, 2021, 1-14,
doi:10.1155/2021/5598309.
65. Mika, S.; Ratsch, G.; Weston, J.; Scholkopf, B.; Mullers, K.-R. Fisher discriminant analysis
with kernels. In Proceedings of the Neural networks for signal processing IX: Proceedings
of the 1999 IEEE signal processing society workshop (cat. no. 98th8468), 1999; pp. 41-
48.
66. Bassler, M.C. Spectroscopic and multivariate approaches for tumor diagnostics and-
therapy. Dissertation, Tübingen, Germany, 2022.
Pagenstecher, A.; Schittenhelm, J.; Boldrini, B.; Hakelberg, S. Formalin Fixation as
Tissue Preprocessing for Multimodal Optical Spectroscopy Using the Example of Human
Brain Tumour Cross Sections. Journal of Spectroscopy 2021, 2021, 14,
doi:10.1155/2021/5598309.
68. Kessler, W. Multivariate datenanalyse: für die pharma, bio-und Prozessanalytik; John
Wiley & Sons: 2011.
69. Esbensen, K.H.; Swarbrick, B. Multivariate data analysis: an introduction to multivariate
analysis, process analytical technology and quality by design; Camo: 2018.
70. Ozer, D.J. Correlation and the coefficient of determination. Psychological bulletin 1985,
97, 307, doi:10.1037/0033-2909.97.2.307.
71. Chai, T.; Draxler, R.R. Root mean square error (RMSE) or mean absolute error (MAE).
Geoscientific model development 2014, 7, 1525-1534, doi:10.5194/gmdd-7-1525-2014.
72. Wikipedia. Root-mean-square deviation. Available online:
https://en.wikipedia.org/wiki/Root-mean-square_deviation (accessed on 20.03.2023).
73. Lee, D.K.; In, J.; Lee, S. Standard deviation and standard error of the mean. Korean journal
of anesthesiology 2015, 68, 220-223, doi:10.4097/kjae.2015.68.3.220.
XII
9 Bibliography
74. Saffarian, S.; Elson, E.L. Statistical analysis of fluorescence correlation spectroscopy: the
standard deviation and bias. Biophysical journal 2003, 84, 2030-2042,
doi:10.1016/S0006-3495(03)75011-5.
75. Organization, W.H. March 2014 supplement to the 2013 consolidated guidelines on the use
of antiretroviral drugs for treating and preventing HIV infection: recommendations for a
public health approach. 2014.
76. Abebe, W. Herbal medication: potential for adverse interactions with analgesic drugs.
Journal of clinical pharmacy
therapeutics 2002, 27, 391-401, doi:10.1046/j.1365-2710.2002.00444.x.

77. Amani, M.B.; Kindenge, J.M.; Baruti, E.T.; Bakiantima, E.N.; Agasa, S.B.; Hubert, P.;
Marini Djang’eing’a, R. Quality control of tramadol in kisangani: development,
validation, and application of a uv-vis spectroscopic method. American Journal of
Analytical Chemistry 2021, 12, 295-309, doi:10.4236/ajac.2021.128018
78. Sankar, R.; Snehalatha, K.S.; Firdose, S.T.; Babu, P.S. Applications in HPLC in
pharmaceutical analysis. International Journal of Pharmaceutical Sciences 2019, 59,
117-124.
79. Saeed, M.; Ahmed, Q. Estimation of paracetamol, aspirin, ibuprofen, codeine and caffeine
in some formulated commercial dosage using UV–spectroscopic method. Eur. J. Pharm.
Med. Res 2017, 4, 33-38.
80. Akhtar, S.; Kareem, L.; Arif, A.; Siddiqui, M.; Hakeem, A. Development of a ceramic-
based composite for direct bonded copper substrate. Ceramics International 2017, 43,
5236-5246, doi:10.1016/j.ceramint.2017.01.049.
81. Schulz-Harder, J. Advantages and new development of direct bonded copper substrates.
Microelectronics Reliability 2003, 43, 359-365, doi:10.1016/S0026-2714(02)00343-8.
82. Gong, Y.; Lee, C.; Yang, C. Atomic force microscopy and Raman spectroscopy studies on
the oxidation of Cu thin films. Journal of Applied Physics 1995, 77, 5422-5425,
doi:10.1063/1.359234.
83. Lenglet, M.; Kartouni, K.; Delahaye, D. Characterization of copper oxidation by linear
potential sweep voltammetry and UV-Visible-NIR diffuse reflectance spectroscopy.
Journal of applied electrochemistry 1991, 21, 697-702, doi:10.1007/BF01034048.
84. Graham, M. Recent advances in oxide film characterization. Pure applied chemistry 1992,
64, 1641-1645, doi:10.1351/pac199264111641.
85. Liu, Y.; Gamble, G.; Thibodeaux, D. UV/visible/near-infrared reflectance models for the
rapid and non-destructive prediction and classification of cotton color and physical
indices. Transactions of the ASABE 2010, 53, 1341-1348, doi:10.13031/2013.32584.
86. Tschannerl, J.; Ren, J.; Jack, F.; Krause, J.; Zhao, H.; Huang, W.; Marshall, S. Potential of
UV and SWIR hyperspectral imaging for determination of levels of phenolic flavour
compounds in peated barley malt. Food chemistry 2019, 270, 105-112,
doi:10.1016/j.foodchem.2018.07.089.
87. Stiedl, J.; Green, S.; Chassé, T.; Rebner, K. Characterization of oxide layers on technical
copper material using ultraviolet visible (UV–Vis) spectroscopy as a rapid on-line
analysis tool. Applied Spectroscopy 2019, 73, 59-66, doi:10.1364/AS.73.000059.
88. Wang, H.; Memon, H.J.P.S., Properties. Cotton science and processing technology;
Springer: Singapore, 2020; Volume 5, pp. 79-98.
89. Tokel, D.; Dogan, I.; Hocaoglu-Ozyigit, A.; Ozyigit, I.I. Cotton agriculture in Turkey and
worldwide economic impacts of Turkish cotton. Journal of Natural Fibers 2022, 19,
10648-10667, doi:10.1080/15440478.2021.2002759.
90. Karadag, R. Cotton Dyeing with Cochineal by Just in Time Extraction, Mordanting,
Dyeing, and Fixing Method in the Textile Industry. Journal of Natural Fibers 2023, 20,
1-11, doi:10.1080/15440478.2022.2108184.
XIII
9 Bibliography
91. Fuhrer, L. Mapping of In-Field Cotton Fiber Quality Utilizing John Deere’s Harvest
Identification System (Hid). University of Georgia, 2022.
92. Ge, Y. Mapping in-field cotton fiber quality and relating it to soil moisture; Texas A&M
University: 2007.
93. Campbell, B.T.; Hinze, L. Cotton production, processing and uses of cotton raw material.
In Industrial crops and uses; CABI Wallingford UK: 2010; pp. 259-276.
94. BlouseRoumaine. What is slowfashion movement? Available online:
https://www.blouseroumaine-shop.com/en/blog/what-is-slowfashion-movement
95. Barotova, A.; Xurramov, A.; Raxmatullayev, S.; Ismoilova, A. Evaluation of fiber quality
indexes in different varieties of cotton plants. Journal of Agriculture Horticulture 2023,
3, 41-46, doi:10.5281/zenodo.7655573.
96. Ravandi, S.H.; Valizadeh, M. Properties of fibers and fabrics that contribute to human
comfort. In Improving comfort in clothing; Elsevier: 2011; pp. 61-78.
97. Chand, N.; Fahim, M. of NFPC, editors. Cotton reinforced polymer composites. 2008,
doi:10.1533/9781845695057.129.
98. Yu, C. Natural textile fibres: vegetable fibres. In Textiles and fashion; Elsevier: 2015; pp.
29-56.
99. Hequet, E.F.; Abidi, N.; Ethridge, D. Processing sticky cotton: Effect of stickiness on yarn
quality. Textile Research Journal 2005, 75, 402-410, doi:10.1177/0040514505053953.
100. Chung, C.; Lee, M.; Choe, E.K. Characterization of cotton fabric scouring by FT-IR ATR
spectroscopy. Carbohydrate Polymers 2004, 58, 417-420,
doi:10.1016/j.carbpol.2004.08.005.
101. Behera, P.; Aravind, S.; Seetharaman, B. Honeydew contaminated cotton: a sticky problem
needs a solution. Research Journal of Textile and Apparel 2022, doi:10.1108/RJTA-05-
2022-0053.
102. Hequet, E.; Abidi, N. Processing sticky cotton: implication of trehalulose in residue build-
up. Journal of Cotton Science 2002, 6, 77-90.
103. Abidi, N.; Hequet, E. Fourier transform infrared analysis of cotton contamination. Textile
Research Journal 2007, 77, 77-84, doi:10.1177/0040517507074624.
104. Amada44. Honeydew. Available online: https://www.thedailygarden.us/garden-word-of-
the-day/honeydew (accessed on 02.09.2023).
105. Vazquez, J.R. Bemisia tabaci. Available online:
https://www.biodiversidadvirtual.org/insectarium/Bemisia-tabaci-img925792.html
106. Hequet, E.; Abidi, N. Effects of the origin of the honeydew contamination on cotton
spinning performances. Textile Research Journal 2005, 75, 699-709,
doi:10.1177/0040517505053909.
107. Perkins Jr, H.H. Identification and processing of honeydew-contaminated cottons. Textile
Research Journal 1983, 53, 508-512.
108. Afzal, M.I.J.I.o.t.M.o.C.P.i.Z.a.b.S. Cotton stickiness–A marketing and processing
problem. 2001, 105-111.
109. Domelsmith, L.; Berni, R. Potassium: A new marker for washed cotton. Textile Research
Journal 1984, 54, 210-214.
110. Fischer, J. Evaluation of Cleaning and Washing Processes for Cotton Fiber: Part VII.
Microbiological Evaluation 1. Textile Research Journal 1980, 50, 93-95.
111. Tzanov, T.; Calafell, M.; Guebitz, G.M.; Cavaco-Paulo, A. Bio-preparation of cotton
fabrics. Enzyme Microbial Technology 2001, 29, 357-362, doi:10.1016/S0141-
0229(01)00388-X.
112. Shams, N.A.; Mohajerani, M. Evaluation of cotton fibers stickiness by colorimetric
method. 2013.
XIV
9 Bibliography
113. Luttrell, R.; Fitt, G.; Ramalho, F.; Sugonyaev, E. Cotton pest management: Part 1. A
worldwide perspective. Annual Review of Entomology 1994, 39, 517-526.
114. Bradow, J.M.; Davidonis, G.H. Quantitation of fiber quality and the cotton production-
processing interface: a physiologist’s perspective. Cotton Science 2000, 4, 34-64.
115. Jiang, Y.; Li, C. Detection and discrimination of cotton foreign matter using push-broom
based hyperspectral imaging: System design and capability. PloS one 2015, 10,
e0121969, doi:10.1371/journal.pone.0121969.
116. Knowlton, J.L. USDA Cotton Classing and the Need for a Universal Cotton Quality
Evaluation System. In Proceedings of the EFS® Systems Conference, Turkey,
http://www. cottoninc. com/EFSConference, 2005.
117. Yang, S.; Gordon, S. Accurate prediction of cotton ring-spun yarn quality from high-
volume instrument and mill processing data. Textile Research Journal 2017, 87, 1025-
1039, doi:10.1177/0040517516646051.
118. Sayeed, M.A.; Schumann, M.; Wanjura, J.; Kelly, B.R.; Smith, W.; Hequet, E.F.
Characterizing the total within-sample variation in cotton fiber length using the High
Volume Instrument fibrogram. Textile Research Journal 2021, 91, 175-187,
doi:10.1177/0040517520935212.
119. Steven Brown, T.S. How to Think About Fiber Quality in Cotton. Available online:
https://www.aces.edu/blog/topics/crop-production/how-to-think-about-fiber-quality-in-
cotton/ (accessed on 05.08.2022).
120. Barnhardt-Purified-cotton. Properties of Cotton. Available online:
https://barnhardtcotton.net/technology/cotton-properties/ (accessed on 01.04.2023).
121. Zhou, J.; Xu, B. Reliability of cotton fiber length distributions measured by dual-beard
fibrography and advanced fiber information system. Cellulose 2021, 28, 1753-1767,
doi:10.1007/s10570-020-03611-x.
122. Abd El-Ghany, N.M.; Abd El-Aziz, S.E.; Marei, S.S. A review: application of remote
sensing as a promising strategy for insect pests and diseases management. Environmental
Science and Pollution Research 2020, 1-13, doi:10.1007/s11356-020-09517-2.
123. Rego, C.H.Q.; França-Silva, F.; Gomes-Junior, F.G.; Moraes, M.H.D.d.; Medeiros, A.D.d.;
Silva, C.B.d. Using Multispectral Imaging for Detecting Seed-Borne Fungi in Cowpea.
Agriculture 2020, 10, 361, doi:10.3390/agriculture10080361.
124. Bannon, D. Hyperspectral imaging: Cubes and slices. Nature photonics 2009, 3, 627,
doi:10.1038/nphoton.2009.205.
125. Tschannerl, J.; Ren, J.; Jack, F.; Krause, J.; Zhao, H.; Huang, W.; Marshall, S. Potential of
UV and SWIR hyperspectral imaging for determination of levels of phenolic flavour
compounds in peated barley malt. Food chemistry 2019, 270, 105-112,
doi:10.1016/j.foodchem.2018.07.089.
126. Lodhi, V.; Chakravarty, D.; Mitra, P. Hyperspectral imaging system: Development aspects
and recent trends. Sensing Imaging 2019, 20, 1-24, doi:10.1007/s11220-019-0257-8.
127. Lu, G.; Fei, B. Medical hyperspectral imaging: a review. Biomedical optics 2014, 19,
010901, doi:10.1117/1.JBO.19.1.010901.
128. Rebner, K. Hyperspectral Imaging for Quality Analysis and Control. In Proceedings of the
Applied Industrial Optics: Spectroscopy, Imaging and Metrology, 2016; p. AITh2B. 1.
129. Biancolillo, A.; Marini, F. Chemometric methods for spectroscopy-based pharmaceutical
analysis. Frontiers in chemistry 2018, 6, 576, doi:10.3389/fchem.2018.00576.
130. Tonnesen, H.H. Photostability of drugs and drug formulations, 2nd ed.; CRC Press: Boca
Raton, 2004; p. 448.
131. Murtaza, G.; Hussain, I.; Khan, S.A.; Shabbir, A.; Mahmood, A.; Asad, M.H.H.B.;
Farzanal, K.; Malik, N.S. Development of a UV-spectrophotometric method for the
simultaneous determination of aspirin and paracetamol in tablets. Scientific research and
Essays 2011, 6, 417-421, doi:10.5897/SRE10.925.
XV
9 Bibliography
132. Saeed, M.; Ahmed, Q. Estimation of paracetamol, aspirin, ibuprofen, codeine and caffeine
in some formulated commercial dosage using UV–spectroscopic method. Eur J Pharm
Med Res 2017, 4, 33-38.
133. Rote, A.R.; Kumbhoje, P.A.; Bhambar, R.S. UV-visible spectrophotometric simultaneous
estimation of paracetamol and nabumetone by AUC method in combined tablet dosage
form. Pharmaceutical methods 2012, 3, 40-43, doi:10.4103/2229-4708.97722.
134. Jolliffe, I.T.; Cadima, J. Principal component analysis: a review and recent developments.
Philosophical Transactions of the Royal Society A 2016, 374, 20150202,
doi:10.1098/rsta.2015.0202.
135. Bianchi, F.; Riboni, N.; Trolla, V.; Furlan, G.; Avantaggiato, G.; Iacobellis, G.; Careri, M.
Differentiation of aged fibers by Raman spectroscopy and multivariate data analysis.
Talanta 2016, 154, 467-473, doi:10.1016/j.talanta.2016.04.013.
136. Lawson-Wood, K.; Robertson, I. Pharmaceutical Assay and Multicomponent Analysis
using the LAMBDA 365 UV/Vis Spectrophotometer. 2016.
137. Barnaby, J.Y.; Huggins, T.D.; Lee, H.; McClung, A.M.; Pinson, S.R.; Oh, M.; Bauchan,
G.R.; Tarpley, L.; Lee, K.; Kim, M.S. Vis/NIR hyperspectral imaging distinguishes sub-
population, production environment, and physicochemical grain properties in rice.
Scientific reports 2020, 10, 9284, doi:10.1038/s41598-020-65999-7.
138. Atif, M.; Farooq, W.; Fatehmulla, A.; Aslam, M.; Ali, S.M. Photovoltaic and impedance
spectroscopy study of screen-printed TiO2 based CdS quantum dot sensitized solar cells.
Materials 2015, 8, 355-367, doi:doi.org/10.3390/ma8010355.
139. Esa, S.R.; Yahya, R.; Hassan, A.; Omar, G. Nano-scale copper oxidation on leadframe
surface. Ionics 2017, 23, 319-329, doi:10.1007/s11581-016-1894-8.
140. Council, N.R. Expanding the vision of sensor materials; National Academies Press: 1995.
141. Al Ktash, M.; Hauler, O.; Ostertag, E.; Brecht, M. Ultraviolet-visible/near infrared
spectroscopy and hyperspectral imaging to study the different types of raw cotton.
Journal of Spectral Imaging 2020, 9, doi:10.1255/jsi.2020.a18.
142. Lodhi, V.; Chakravarty, D.; Mitra, P. Hyperspectral imaging system: Development aspects
and recent trends. Sensing and Imaging 2019, 20, 1-24, doi:10.1007/s11220-019-0257-8.
143. Jin, S.; Hui, W.; Wang, Y.; Huang, K.; Shi, Q.; Ying, C.; Liu, D.; Ye, Q.; Zhou, W.; Tian,
J. Hyperspectral imaging using the single-pixel Fourier transform technique. Scientific
reports 2017, 7, 45209, doi:10.1038/srep45209.
144. Willoughby, C.T.; Folkman, M.A.; Figueroa, M.A. Application of hyperspectral-imaging
spectrometer systems to industrial inspection. In Proceedings of the Three-Dimensional
and Unconventional Imaging for Industrial Inspection and Metrology, 1996; pp. 264-272.
145. Lu, G.; Fei, B. Medical hyperspectral imaging: a review. Journal of biomedical optics
2014, 19, 010901, doi:10.1117/1.JBO.19.1.010901.
146. Stiedl, J.; Green, S.; Chassé, T.; Rebner, K. Auger electron spectroscopy and UV–Vis
spectroscopy in combination with multivariate curve resolution analysis to determine the
Cu2O/CuO ratios in oxide layers on technical copper surfaces. Applied Surface Science
2019, 486, 354-361, doi:10.1016/j.apsusc.2019.05.028.
147. Mazzeo, G.; Prestopino, G.; Conte, G.; Salvatori, S. Metal-diamond-metal planar structures
for off-angle UV beam positioning with high lateral resolution. Sensors Actuators A:
Physical 2005, 123, 199-203, doi:10.1016/j.sna.2005.02.016.
148. Ojeda, C.B.; Rojas, F.S. Process analytical chemistry: applications of ultraviolet/visible
spectrometry in environmental analysis: an overview. Applied Spectroscopy Reviews
2009, 44, 245-265, doi:10.1080/05704920902717898.
149. Mazzeo, G.; Prestopino, G.; Conte, G.; Salvatori, S. Metal-diamond-metal planar structures
for off-angle UV beam positioning with high lateral resolution. Sensors and Actuators A:
Physical 2005, 123, 199-203, doi:10.1016/j.sna.2005.02.016.
XVI
9 Bibliography
150. Reyes, G.; Diaz, W.; Toro, C.; Balladares, E.; Torres, S.; Parra, R.; Vásquez, A. Copper
Oxide Spectral Emission Detection in Chalcopyrite and Copper Concentrate Combustion.
Processes 2021, 9, 188, doi:10.3390/pr9020188.
151. Obeidat, S.M.; Al-Ktash, M.M.; Al-Momani, I.F. Study of fuel assessment and adulteration
using EEMF and multiway PCA. Energy & Fuels 2014, 28, 4889-4894,
doi:10.1021/ef500718e.
152. Geladi, P.; Kowalski, B.R. Partial least-squares regression: a tutorial. Analytica chimica
acta 1986, 185, 1-17, doi:10.1016/0003-2670(86)80028-9.
153. inno-spec GmbH. BlueEye UV Hyperspectral Imaging Camera (220 – 380 nm). Available
online: https://inno-spec.de/wp-content/uploads/2021/10/210928_BlueEye.pdf (accessed
on 29.10.2021).
154. PCO AG. pco.edge 4.2 bi cooled sCMOS camera. Available online:
https://www.pco.de/fileadmin/user_upload/pco-
product_sheets/DS_PCOEDGE42BI_V104.pdf (accessed on 29.10.2021).
155. Quantum Desgin Europe GmbH. Lamp Spectra and Irradiance. Available online:
https://qd-
europe.com/fileadmin/Mediapool/products/lightsources/en/LQ_Lamp_spectra_and_irra
diance_en.pdf (accessed on 27.09.2021).
156. Zhang, H.; Li, D. Applications of computer vision techniques to cotton foreign matter
inspection: A review. Computers and Electronics in Agriculture 2014, 109, 59-70,
doi:10.1016/j.compag.2014.09.004.
157. Morais, J.P.S.; de Freitas Rosa, M.; Nascimento, L.D.; do Nascimento, D.M.; Cassales,
A.R. Extraction and characterization of nanocellulose structures from raw cotton linter.
Carbohydrate polymers 2013, 91, 229-235, doi:10.1016/j.carbpol.2012.08.010.
158. Dohlman, E.; Johnson, J.; MacDonald, S.; Meyer, L.; Soley, G. The world and United
States cotton outlook. In Proceedings of the 2015-02-20)[2015-04-01]. http: www. usda.
gov/oce/forum, Arlingron/Virginia, 21.-22.02.2019.
159. Pray, C.E.; Huang, J.; Hu, R.; Rozelle, S. Five years of Bt cotton in China–the benefits
continue. The Plant Journal 2002, 31, 423-430, doi:10.1046/j.1365-313X.2002.01401.
160. Heuvels, S.; Molenaar, J.; Raap, N.; Petit, C.; Shanmugavel, A.; Li, W. Sustainable cotton
ranking 2017; Assesing company performance: October 2017 2017; p. 28.
161. Mustafic, A.; Jiang, Y.; Li, C. Cotton contamination detection and classification using
hyperspectral fluorescence imaging. Textile Research Journal 2016, 86, 1574-1584,
doi:10.1177/0040517515590416.
162. Ghule, A.V.; Chen, R.K.; Tzing, S.H.; Lo, J.; Ling, Y.C. Simple and rapid method for
evaluating stickiness of cotton using thermogravimetric analysis. Analytica Chimica Acta
2004, 502, 251-256, doi:10.1016/j.aca.2003.10.021.
163. Abidi, N.; Hequet, E. Fourier Transform Infrared Analysis of Cotton Contamination.
Textile Research Journal 2016, 77, 77-84, doi:10.1177/0040517507074624.
164. Was-Gubala, J.; Starczak, R. Nondestructive identification of dye mixtures in polyester and
cotton fibers using raman spectroscopy and ultraviolet-visible (UV-Vis)
microspectrophotometry. Appl. Spectrosc. 2015, 69, 296-303, doi:10.1366/14-07567.
165. Goldfarb, A.R.; Saidel, L.J.; Mosovich, E. The ultraviolet absorption spectra of proteins.
Journal of Biological Chemistry 1951, 193, 397-404.
166. Zhou, F.; Ding, T. Detection of Cotton Lint Trash within the Ultraviolet—Visible Spectral
Range. J Applied spectroscopy 2010, 64, 936-941, doi:10.1366/000370210792081091.
167. Peets, P.; Leito, I.; Pelt, J.; Vahur, S. Identification and classification of textile fibres using
ATR-FT-IR spectroscopy with chemometric methods. Spectrochimica Acta 2017, 173,
175-181, doi:10.1016/j.saa.2016.09.007.
XVII
9 Bibliography
168. Chung, C.; Lee, M.; Choe, E. Characterization of cotton fabric scouring by FT-IR ATR
spectroscopy. Carbohydrate Polymers 2004, 58, 417-420,
doi:10.1016/j.carbpol.2004.08.005.
169. Blanco, M.; Coello, J.; Iturriaga, H.; Maspoch, S.; PageÁs, J. Use of near-infrared
spectrometry in control analyses of acrylic ®bre manufacturing processes. Analytica
Chimica Acta 1999, 383, 291±298, doi:10.1016/S0003-2670(98)00804-6.
170. Cleve, E.; Bach, E.; Schollmeyer, E. Using chemometric methods and NIR
spectrophotometry in the textile industry. Analytica Chimica Acta 2000, 420, 163–167,
doi:10.1016/S0003-2670(00)00888-6.
171. Fortier, C.A.; Rodgers, J.E.; Cintrón, M.S.; Xiaoliang, C.; Foulk, J.A. Identification of
cotton and cotton trash components by Fourier transform near-infrared spectroscopy.
Textile Research Journal 2010, 81, 230-238, doi:10.1177/0040517510383620.
172. Ruckebusch, C.; Orhan, F.; Durand, A.; Boubellouta, T.; Huvenne, J.P. Quantitative
Analysis of Cotton–Polyester Textile Blends from Near-Infrared Spectra.pdf>. Applied
Spectroscopy 2006, 60, 539 - 544, doi:10.1366/000370206777412194.
173. Liu, Y. Recent Progress in Fourier Transform Infrared (FTIR) Spectroscopy Study of
Compositional, Structural and Physical Attributes of Developmental Cotton Fibers.
Materials 2013, 6, 299-313, doi:10.3390/ma6010299.
174. Ríos-Reina, R.; García-González, D.L.; Callejón, R.M.; Amigo, J.M. NIR spectroscopy
and chemometrics for the typification of Spanish wine vinegars with a protected
designation of origin. Food Control 2018, 89, 108-116,
doi:10.1016/j.foodcont.2018.01.031.
175. Obeidat, S.M.; Khanfar, M.S.; Obeidat, W.M. Classification of edible oils and uncovering
adulteration of virgin olive oil using FTIR with the aid of chemometrics. Australian
Journal of Basic and Applied Sciences 2009, 3, 2048-2053.
176. Schneider, A.; Feussner, H. Biomedical engineering in gastrointestinal surgery; Academic
Press: 2017.
177. Colarusso, P.; Kidder, L.H.; Levin, I.W.; Fraser, J.C.; Arens, J.F.; Lewis, E.N. Infrared
spectroscopic imaging: from planetary to cellular systems. Applied Spectroscopy 1998,
52(3), 106A–120A, doi:10.1366/0003702981943545.
178. Mirschel, G.; Daikos, O.; Scherzer, T.; Steckert, C. Near-infrared chemical imaging used
for in-line analysis of functional finishes on textiles. Talanta 2018, 188, 91-98,
doi:10.1016/j.talanta.2018.05.050.
179. Zhu, S.; Zhou, L.; Gao, P.; Bao, Y.; He, Y.; Feng, L. Near-infrared hyperspectral imaging
combined with deep learning to identify cotton seed varieties. Molecules 2019, 24, 3268,
doi:10.3390/molecules24183268.
180. Martens, H.; Høy, M.; Wise, B.M.; Bro, R.; Brockhoff, P.B. Pre‐whitening of data by
covariance‐weighted pre‐processing. Journal of Chemometrics 2003, 17, 153-165,
doi:10.1002/cem.780.
181. Wise, B.M.; Martens, H.; Høy, M.; Bro, R.; Brockhoff, P.B. Calibration transfer by
generalized least squares. In Proceedings of the Proceedings of the Seventh Scandinavian
Symposium on Chemometrics (SSC7), Copenhagen, Denmark, 2001; pp. 19-23.
182. Jirata, M.T.; Chelule, J.C.; Odhiambo, R. Deriving some estimators of panel data regression
models with individual effects. International Journal of Science and Research 2012, 3,
53-59.
183. Liu, Y.; Delhom, C.; Campbell, B.T.; Martin, V. Application of near infrared spectroscopy
in cotton fiber micronaire measurement. Information Processing in Agriculture 2016, 3,
30-35, doi:10.1016/j.inpa.2016.01.001.
184. Femenias, A.; Marín, S. Hyperspectral Imaging; Vicente M. Gómez-López, R.B., Ed.;
2021; pp. 363-390.
XVIII
9 Bibliography
185. Park, B.; Lu, R. Hyperspectral imaging technology in food and agriculture; Barbosa-
Ca'novas, G.V., Ed.; Springer: Athens, GA,USA, 2015.
186. Chen, S.-Y.; Chang, C.-Y.; Ou, C.-S.; Lien, C.-T. Detection of insect damage in green
coffee beans using VIS-NIR hyperspectral imaging. Remote Sensing 2020, 12, 2348,
doi:10.3390/rs12152348.
187. Devassy, B.M.; George, S. Estimation of strawberry firmness using hyperspectral imaging:
a comparison of regression models. Journal of Spectral Imaging 2021, 10,
doi:10.1255/jsi.2021.a3.
188. Daikos, O.; Scherzer, T. In-line monitoring of the residual moisture in impregnated black
textile fabrics by hyperspectral imaging. Progress in Organic Coatings 2021, 106610,
doi:10.1016/j.porgcoat.2021.106610.
189. Wang, C.; Xu, M.; Jiang, Y.; Zhang, G.; Cui, H.; Deng, G.; Lu, Z. Toward Real
Hyperspectral Image Stripe Removal via Direction Constraint Hierarchical Feature
Cascade Networks. Remote Sensing 2022, 14, 467, doi:doi.org/10.3390/rs14030467.
190. Blanch-Perez-del-Notario, C.; Saeys, W.; Lambrechts, A. Hyperspectral imaging for textile
sorting in the visible–near infrared range. Journal of Spectral Imaging 2019, 8,
doi:10.1255/jsi.2019.a17.
191. Mirschel, G.; Daikos, O.; Scherzer, T. In-line monitoring of the thickness distribution of
adhesive layers in black textile laminates by hyperspectral imaging. Computers &
Chemical Engineering 2019, 124, 317-325, doi:10.1016/j.compchemeng.2019.01.015.
192. Feng, X.; Cheng, H.; Zuo, D.; Zhang, Y.; Wang, Q.; Lv, L.; Li, S.; Yu, J.Z.; Song, G.
Genome-wide identification and expression analysis of GL2-interacting-repressor (GIR)
genes during cotton fiber and fuzz development. Planta 2022, 255, 1-18,
doi:10.1007/s00425-021-03737-7.
193. Anthony, W.S. Improvement of the Marketability of Cotton Produced in Zones Affected by
Stickiness, GOURLOT J.-P. ed.; Common Fund for Commodities: France, 2001; p. 99.
194. Rony, A.N.U. Technical Properties of Cotton Fiber, Textile Learner GmbH. Available
online: https://textilelearner.net/technical-properties-of-cotton-fiber/ (accessed on 26
December 2020).
195. Abidi, N.; Hequet, E.; Cabrales, L. Changes in sugar composition and cellulose content
during the secondary cell wall biogenesis in cotton fibers. Cellulose 2010, 17, 153-160,
doi:doi.org/10.1007/s10570-009-9364-3.
196. Calvo‐Agudo, M.; Tooker, J.F.; Dicke, M.; Tena, A. Insecticide‐contaminated honeydew:
risks for beneficial insects. Biological Reviews 2021, doi:10.1111/brv.12817.
197. Balasubramanya, R.; Bhatawdekar, S.; Paralikar, K. A new method for reducing the
stickiness of cotton. Textile Research Journal 1985, 55, 227-232,
doi:10.1177/004051758505500405.
198. Jumaniyazov, K.; Egamberdiev, F.; Abbazov, I. The Effect of Crop Type on Cotton Quality
Indicators. In Proceedings of the International Journal of Advanced Research in Science,
Engineering and Technology, 2020; pp. 13510-13518.
199. Severino, L.; Leite, B.; Gambarra-Neto, F.; Araújo, J.; Medeiros, E. Detection and
Quantification of Stickiness on Cotton Samples Using Near Infrared Hyperspectral
Images Bremen Baumwollboerse; p. 8.
200. Gamble, G.R. Evaluation of cotton stickiness via the thermochemical production of volatile
compounds. Journal of Cotton Science 2003, 45-50.
201. Was-Gubala, J.; Starczak, R. Nondestructive identification of dye mixtures in polyester and
cotton fibers using Raman spectroscopy and ultraviolet-visible (UV-Vis)
microspectrophotometry. Applied spectroscopy 2015, 69, 296-303, doi:10.1366/14-
07567.
XIX
9 Bibliography
202. Fortier, C.A.; Rodgers, J.E.; Cintron, M.S.; Cui, X.; Foulk, J.A. Identification of cotton and
cotton trash components by Fourier transform near-infrared spectroscopy. Textile
Research Journal 2011, 81, 230-238, doi:10.1177/0040517510383620.
203. Miller, W.B.; Peralta, E.; Ellis, D.R.; Perkins Jr, H.H. Stickiness potential of individual
insect honeydew carbohydrates on cotton lint. Textile research journal 1994, 64, 344-
350, doi:10.1177/004051759406400606.
204. Barton, F.; Bargeron III, J.; Gamble, G.; McAlister, D.; Hequet, E. Analysis of sticky cotton
by near-infrared spectroscopy. Applied spectroscopy 2005, 59, 1388-1392,
doi:10.1366/000370205774783214.
Pagenstecher, A.; Schittenhelm, J.; Boldrini, B.; Hakelberg, S. Formalin Fixation as
Tissue Preprocessing for Multimodal Optical Spectroscopy Using the Example of Human
Brain Tumour Cross Sections. Journal of Spectroscopy 2021, 2021,
doi:10.1155/2021/5598309.
206. van Kollenburg, G.; Bouman, R.; Offermans, T.; Gerretzen, J.; Buydens, L.; van Manen,
H.-J.; Jansen, J. Process PLS: Incorporating substantive knowledge into the predictive
modelling of multiblock, multistep, multidimensional and multicollinear process data.
Computers & Chemical Engineering 2021, 154, 107466,
doi:10.1016/j.compchemeng.2021.107466.
207. van Kollenburg, G.; Bouman, R.; Offermans, T.; Gerretzen, J.; Buydens, L.; van Manen,
H.-J.; Jansen, J. Process PLS: Incorporating substantive knowledge into the predictive
modelling of multiblock, multistep, multidimensional and multicollinear process data.
Computers Chemical Engineering 2021, 154, 107466,
doi:10.1016/j.compchemeng.2021.107466.
208. Victorita, B.; Marghitas, L.A.; Stanciu, O.; Laslo, L.; Dezmirean, D.; Bobis, O. High-
performance liquid chromatographic analysis of sugars in Transylvanian honeydew
honey. Entomologia Experimentalis et Applicata 2008, 65, 229-232, doi:10.1111/j.1570-
7458.2006.00505.x.
209. Lottspeich, F.; Zorbas, H. Bioanalytik, 4 ed.; Spektrum, Akad. Verlag: Springer Spektrum
Berlin, Heidelberg, 2022.
210. Wang, H.; Memon, H. Cotton science and processing technology; Springer Nature
Singapore: China, 2021; Volume 5, p. 565.
211. DIN Deutsches Institut für Normung e. V., D.G.I.f.S. Textiles - Determination of cotton
fibre stickiness - Part 2: Method using an automatic thermodetection plate device. 2004.
212. k Mansuri, A.; Somani, S.; Pathak, R. Effect of Waste Control on Yarn Parameters and
Yield Improvement in Spinning Mill. Journal of critical reviews 2020, 07.
213. DIN Deutsches Institut für Normung e. V., D.G.I.f.S. Textiles - Determination of cotton
fibre stickiness - Part 2: Method using an automatic thermodetection plate device;
German version EN 14278-2:2004. 2004, 11.
214. Bi, J.; Ballmer, G.; Hendrix, D.; Henneberry, T.; Toscano, N. Effect of cotton nitrogen
fertilization on Bemisia argentifolii populations and honeydew production. Entomologia
Experimentalis et Applicata 2001, 99, 25-36, doi:10.1046/j.1570-7458.2001.00798.x.
215. Fischer, M.K.; Völkl, W.; Hoffmann, K.H. Honeydew production and honeydew sugar
composition of polyphagous black bean aphid, Aphis fabae (Hemiptera: Aphididae) on
various host plants and implications for ant-attendance. European Journal of Entomology
2005, 102, 155-160, doi:10.14411/eje.2005.025.
216. Hogervorst, P.A.; Wäckers, F.L.; Romeis, J. Effects of honeydew sugar composition on the
longevity of Aphidius ervi. Entomologia Experimentalis et Applicata 2007, 122, 223-
232, doi:10.1111/j.1570-7458.2006.00505.x.
217. Lottspeich, F.; Zorbas, H. Bioanalytik, 4th ed.; Spektrum, Akad. Verlag, Berlin/Heidelberg,
Germany, 2022.
XX
9 Bibliography
218. Janshekar, H.; Brown, C.; Fiechter, A. Determination of biodegraded lignin by ultraviolet
spectrophotometry. Analytica Chimica Acta 1981, 130, 81-91, doi:10.1016/S0003-
2670(01)84153-2.
219. Sadeghifar, H.; Venditti, R.; Jur, J.; Gorga, R.E.; Pawlak, J.J. Cellulose-lignin
biodegradable and flexible UV protection film. ACS Sustainable Chemistry Engineering
2017, 5, 625-631, doi:10.1021/acssuschemeng.6b02003.
220. Mach, H.; Volkin, D.B.; Burke, C.J.; Russell Middaugh, C. Ultraviolet absorption
spectroscopy. Protein Stability Folding 1995, 91-114, doi:10.1385/0-89603-301-5:91.
221. Lottspeich, F.; Zorbas, H. Bioanalytik, 4th ed.; Spektrum, Akad. Verlag: Berlin/Heidelberg,
Germany, 1998.
222. Al Ktash, M.; Haular, O.; Ostertag, E.; Brecht, M. UV-Vis/NIR spectroscopy and
hyperspectral imaging to study the different types of raw cotton. Journal spectral imaging
2020, 14.
XXI
10 Appendix
10.1 Supplementary Information for Paper II: UV
Hyperspectral Imaging as Process Analytical
Tool for the Characterization of Oxide Layers
and Copper States on Direct Bonded Copper
Diffuse reflectance spectra of the copper powders were recorded in the range of 200 nm – 380 nm
using a commercial UV spectrometer (Lambda 1050+, PerkinElmer, Inc., Waltham, MA, USA).
The spectrometer was equipped with a 150 mm Spectralon® integrating sphere to acquire data in
reflection mode with an R6872-Photomultiplier (PMT). A deuterium lamp was used as light
source in the spectrometer. A 10 mm quartz SUPRASIL® cuvette (QS, 100-10-40, Hellma, Müll-
heim, Germany) was used for measuring the copper powder see Table 10.1. The filled cuvette
was placed at the reflectance port of the integrating sphere. The port measuring area is approxi-
mately 4.9 cm².
Figure 10.1: Reference spectra for the copper Cu0, Cu2O and CuO by using UV spectrometer).
XXI
10 Appendix
Table 10.1: Description of the direct bonded copper substrates and their sample preparation.
Article
Sample type Description Manufacturer
Number
Copper, powder,
Merck KGaA, Darmstadt,
Cu electrolytically pro- 2715
Germany
duced
Copper (I) oxide Riedel-de Haën AG, Seelze,
Cu2O 12841
powder, red Germany
Copper (II) oxide
Riedel-de Haën AG, Seelze,
CuO powder, heavy, 12867
Germany
powder, technical
10.2 Supplementary Information for Paper IV:

Prediction of Honeydew Contaminations on
Cotton Samples by In-Line UV Hyperspectral
Imaging
10.2.1 Cotton sample preparation

All details and measured values of the reference sample set are given in Table 10.2.
XXII
10 Appendix
Table 10.2: Description of the direct bonded copper substrates and their sample preparation.
Sample Sugar Cotton Cotton Hu- Soaked Humid- Amount Ratio
type concent- weight / g weight / g af- mid- cotton ity / % of sugar of
ration / ter dried at ity / weight / g on cotton sugar /
wt (± 0.0001 g) after /g g
30 °C, 8 h % after dried
dried dried
at 30 °C, cotton
(± 0.0001 g) after at
44 h /g
dried 30 °C,
at 44 h
(±
30 °C, 0.0001 g)
8h
A1 2 0.3 0.2879 53.7 0.3734 50.0 0.0855
A2 2 0.3 0.2925 53.5 0.3643 50.0 0.0718
A3 2 0.3 0.2914 53.3 0.3600 50.0 0.0686
B1 1 0.3 0.2917 54.2 0.3295 50.0 0.0378
B2 1 0.3 0.2931 54.7 0.3291 50.0 0.0360
B3 1 0.3 0.2913 54.2 0.3341 50.0 0.0428
C1 0.5 0.3 0.2916 55.2 0.3142 50.0 0.0226
C2 0.5 0.3 0.2929 55.6 0.3137 50.0 0.0208
C3 0.5 0.3 0.2992 55.7 0.3215 50.0 0.0223
D1 0.25 0.3 0.2925 56.0 0.3050 51.0 0.0125
D2 0.25 0.3 0.2928 56.2 0.3030 51.0 0.0102
D3 0.25 0.3 0.2921 56.2 0.3033 51.0 0.0112
E1 0.125 0.3 0.2908 56.5 0.2979 50.5 0.0071
E2 0.125 0.3 0.2819 56.9 0.2915 50.0 0.0096
E3 0.125 0.3 0.2899 57.2 0.3013 49.9 0.0114
F1 0.0625 0.3 0.2903 57.4 0.2997 50.7 0.0094
F2 0.0625 0.3 0.2899 57.6 0.2990 49.9 0.0091
F3 0.0625 0.3 0.2823 57.7 0.2915 49.7 0.0093
CLN1 0 0.3 0.2984 57.7 - - -
CLN2 0 0.3 0.2991 57.9 - - -
CLN3 0 0.3 0.2894 58.1 - - -
10.2.2 Additional figures of the principal component

analysis of the sugar cotton samples
For the PCA model of the cotton samples with different concentrations of sugar four PCs are
necessary. The variance on PC3 is not necessary to distinguish between different sugar concen-
trations. The information on PC3 might be related to the morphology of the fiber itself. PC1
against PC2, PC3 and PC4 are shown respectively in Figure 10.2 to complement the PCA sugar
model.
XXIII
10 Appendix
(a) (b)
0.4 PC1[68.0%]
0.04
PC2[22.0%]
PC2
0.2 0.02
PC2 [22.0%]
0.0 A
0.00
B
C
-0.2
D -0.02
E
-0.4  F PC1
-0.04
 CLN
-1.5 -1.0 -0.5 0.0 0.5 1.0 240 260 280 300 320 340 360 380 400
(c) (d)
0.15 0.04
PC1 [68.0%]
0.10 PC3 [2.0%]
0.02 PC3
0.05
PC3 [2.0%]
0.00 0.00
A
-0.05
B -0.02
-0.10 C
D -0.04
-0.15
E PC1
-0.20 F -0.06
 CLN
-0.25
-0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 240 260 280 300 320 340 360 380 400
(e) (f)
0.15
A 0.08 PC1[68.0%]
B 0.06
PC4[1.0%]
0.10
C
0.04
D
PC4
PC4 [1.0%]
0.05  E 0.02
F 0.00
 CLN
0.00 -0.02
-0.04
-0.05
-0.06 PC1
-0.08
-0.10
-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 240 260 280 300 320 340 360 380 400
Figure 10.2: PCA sugar model for the cotton samples with (a,c,e) scores and (b,d,f) corresponding loadings (PC1 black,
PC2, PC3 and PC4 red).
10.2.3 Pure dried protein spectrum

Protein spectra were acquired to identify the information in the range of 250 nm to 280 nm. The
protein was solved in distilled water and the solution was dropped on a piece of PTFE. After-
wards, the sample was dried in a vacuum oven (see 6.3.2). Data was acquired with the hyperspec-
tral imaging setup with the settings mentioned in 2.3 and 2.4. This experiment was necessary to
XXIV
10 Appendix
verify the spectral range between 250 nm and 280 nm contains true information and is not an
artifact due to the efficiency of the detector and the weak intensity light source in the UV range
[2].
Figure 10.3: Mean spectrum of pure dried protein on PTFE.
10.2.4 X-loadings weights and x-loadings of the PLS-R

model
For model building and understanding the PLS-R factor loadings and loading weights for all five
factors are displayed in the Figure 10.4.
(a) (b)
XXV
10 Appendix
(c) (d)
(e) (f)
(g) (h)
XXVI
10 Appendix
(i) (j)
Figure 10.4: X-loadings weights and x-loadings for factor 1 (a, b), factor 2 (c, d), factor 3 (e, f), factor 4 (g, h) and
factor 5 (i, j), respectively.
10.3 Supplementary Information for Paper V:

Applying UV Hyperspectral Imaging for
Quantification of Honeydew Content on Raw
Cotton via PCA and PLS-R Models
For model building and understanding the PLS-R factor loadings and loading weights for all three
factors are displayed in the Figure 10.5. X-loadings reflect the relationship between the predictor
variables and the latent variables, while X-loadings weights represent the relationship between
the predictor variables and the response variable. Both X-loadings and X-loadings weights play
important roles in PLS-R for variable selection, model interpretation, and prediction.
(a) (b)
XXVII
10 Appendix
(c) (d)
(e) (f)
Figure 10.5: X-loadings weights and x-loadings for factor 1 (a, b), factor 2 (c, d) and factor 3 (e, f), respectively.
XXVIII
List of Abbreviations and Symbols
Abbreviation Concept
EM Electromagnetic
UV-Vis Ultraviolet-Visible
NIR Near Infrared
NMR Nuclear Magnetic Resonance
MS Mass Spectroscopy
HSI Hyperspectral Imaging
MVA Multivariate Data Analysis
PCA Principal Component Analysis
PCs Principal Components
PLS-R Partial Least Square Regression
QDA Quadratic Discriminant Analysis
R2 High coefficient of determination
RMSEC Root Mean Square Error of Calibration
RMSECV Root Mean Square Error of Calibration and Cross-Validation
R2C High Coefficient of Calibration
R2CV High Coefficient of Cross-Validation
n Sample number
PAT Process Analysis and Technology
2D Two Dimensional
3D Three Dimensional
CCD Charge Couple Device
CMOS Complementary Metal Oxide Semiconductor
RGB Red Green Blue
PTFE Polytetrafluoroethylene
InGaAs Indium Gallium Arsenide
HPLC High-Performance Liquid Chromatography
MS Mass Spectrometry
PMT Photomultiplier
XBO Xenon-arc lamp
HBO Mercury-arc lamp
LEDs Light Emitting Diodes
FDA Food and Drug Administration
API Active Pharmaceutical Ingredients
RoB Organic raw cotton
HC Hemp plant
RcO Recycled cotton
RoSt Standard raw cotton
RcBH Recycled organic bright cotton
CLN Cleaned Cotton
IBUpure Ibuprofen, >98%, API
ASA Acetylsalicylic Acid, 99%, API
PARpure Paracetamol, 99%, API
THO Thomapyrin
IBUratio Ibuprofen, Ratiopharm GmbH
IBUbeTa Ibuprofen, Betapharm
ASPBAYER Acetylsalicylic Acid, Bayer vital GmbH
XXIX
List of Abbreviations and Symbols
PARratio Paracetamol, Ratiopharm GmbH

DBC Direct Bonded Copper
GLS Generalized Least Squares
IC Integrated Circuit
XPS X-ray Photoelectron Spectroscopy
USDA United States Department of Agriculture
HVI High Volume Instrument
AFIS Advanced Fiber Information System
FQI Fibre Quality Index
SCI Spinning Consistency Index
PDI Premium-Discount Index
XXX
List of Figures
Figure 1.1: Energy level diagram illustrates electronic, vibrational, and rotational
energy. .................................................................................................... 2
Figure 1.2: (a) Classical and (b) diffuse reflected light.................................................... 3
Figure 1.3: Schematic shows the difference between multispectral and hyperspectral. .. 4
Figure 1.4: Visualization of the different imaging technologies: (a) Whiskbroom
imaging (single point scanning) (b) Snapshot imaging (c) Staring
imaging (2D scanning) (d) Pushbroom imaging (line scanning). .......... 5
Figure 1.5: Scheme of Vis/NIR hyperspectral imaging (pushbroom).............................. 6
Figure 1.6: (a) Schematic shows the concept of hyperspectral imaging based on the
pushbroom (the tunnel in the scheme was cut to show the inside). (b)
Pushbroom Imager scanning principle. (c) Hyperspectral image
produced immediately during sample scanning. (d) UV spectrum for
one single pixel extracted from the image in (c). ................................... 7
Figure 1.7: Spectral radiance for a different light source in the UV region. (a)
Synchrotron radiation of P66 beamline (b) LED radiance (Roithner
LaserTechnik GmbH, Wien, Germany) (c) Plasma radiance (EQ-77,
Energetiq Technology LDLSTM, Wilmington, MA, USA) (d) Deuterium
radiance (SL 3, StellarNet Inc, 24 V, 65.04 W, Tampa, Florida, USA)
(e) Xenon-arc radiance (XBO, 14 V, 75 W, Osram, München,
Germany) (f) Mercury-arc radiance (HBO, 14 V, 75 W, OSRAM,
München, Germany). All light sources were tested by UV hyperspectral
imaging except Synchrotron radiation of P66 taken from reference
[40]. ........................................................................................................ 9
As a result, the synchrotron was excluded from this study because it did not serve our
purpose; it was deemed unsuitable for on-line measurements due to
lack of portability and expense. LED and mercury lamps were ruled out
due to their narrow wavelength bands, see Figure 1.7b and f. Using
plasma was limited due to cost considerations and the requirement for
cooling. In contrast, Deuterium and XBO-arc lamps were selected,
offering a continuous spectrum (Figures 1.7d and e) and low cost. ....... 9
Figure 1.8: Graphical representation of the principal components. The original data in
the original data space (x1, x2) are transformed into new principal axes
(PC1, PC2). .......................................................................................... 10
Figure 1.9: Schematic description of a decomposition of a matrix X with PCA using
two PCs. ............................................................................................... 11
Figure 1.10: Schematic shows different discriminant analysis (DA). DA function
creates a border of variable shape that optimally separates a training
data set into multiple groups. PCA can be used to reduce the dimension
XXXI
List of Figures
of the training set before creating the discriminant function. The three
most common types of DA separators are linear, quadratic, and
mahalanobis distance-based separators. Linear separators are
represented by straight orange lines (a), quadratic separators by orange
curves (b), and mahalanobis distance-based separators by ellipses (c).
This figure is taken and modified from reference [66]. ........................ 12
Figure 1.11: Partial least square regression method. The Y variables influence the X
variables. ............................................................................................... 14
Figure 1.12: Drug samples. Reference API samples and painkiller tablets. This figure is
taken from reference [19]. .................................................................... 18
Figure 1.13: Direct bonded copper Curamik®Power substrates. (1), (2), (3), (4), (5) and
(6) are different samples with different types of copper state and
thickness of the copper oxide on DBC. This figure is taken and
modified from reference [53]................................................................ 19
Figure 1.14: Cotton fibers. ............................................................................................. 19
Figure 1.15: Manufacture process for raw cotton; this photo is taken from reference
[94]. ...................................................................................................... 20
Figure 1.16: Schematic illustration of the structure of cotton fiber, showing its different
layers..................................................................................................... 21
Figure 1.17: (a) and (b) Cotton fiber contaminated by sugar cause of (c) aphid and (d)
Whiteflies insects these photos are taken from [104,105]. ................... 22
Figure 1.18: Honeydew chemical structure contents (a) Trehalulose (b) Trehalose (c)
Melezitose (d) Sucrose (e) Fructose (f) Glucose. ................................. 22
Figure 1.19: sticky cotton residue on a draw frame roll; this photo is taken from
reference [102]. ..................................................................................... 23
Figure 3.1: Drug samples. Reference API samples and painkiller table. ....................... 28
Figure 3.2: (a) Setup of a hyperspectral imaging system based on the pushbroom
concept (the tunnel in the scheme was cut to show the inside). (b)
Pushbroom Imager scanning principle. (c) Hyperspectral image
generated immediately from the scanning of a sample. (d) UV spectrum
for one single pixel extracted from the image given in (c). .................. 31
Figure 3.3: Hyperspectral raw image of nine drug samples on the left. Images after
subtraction of the background on the right. .......................................... 32
Figure 3.4: UV absorbance spectra of APIs ibuprofen (IBU), acetylsalicylic acid
(ASA), paracetamol (PAR) and a mixture of acetylsalicylic acid and
paracetamol (ASA+PAR) in liquid phase............................................. 33
Figure 3.5: UV total hemispherical reflectance spectra of drug samples in the solid
phase in the wavelength range 200–380 nm. (a) API drugs IBUpure,
ASApure, PARpure and a mixture of ASApure with PARpure. Upper right:
Fluorescence emission of IBU sample with excitation at 270 nm. (b)
Painkiller tablets IBUratio, IBUbeTa, ASPratio, PARratio and THO............ 34
XXXII
List of Figures
Figure 3.6: (a) Raw hyperspectral image for all API drug samples before and after
subtracting the background. (b) Spectrum recorded for a single pixel of
each of pure API samples in the UV range 225–400 nm. (c,d) Scores
and corresponding loadings plot........................................................... 36
Figure 3.7: (a) Raw hyperspectral image for all commercial painkiller tablets before
and after subtracting the background. (b) Spectrum recorded for a
single pixel of each painkiller tablet in the UV range 200–400 nm. (c,d)
Scores and corresponding loadings plot. .............................................. 38
Figure 4.1: Direct bonded copper Curamik®Power substrates. (1) is an example of
sample type 1, (2) sample type 2, (3) sample type 3, (4) sample type 4,
(5) sample type 5 and (6) sample type 6............................................... 46
Figure 4.2: An example of a direct bonded copper sheet rotated according to the three
different measurement angles (a) 0°, (b) 45° and (c) 90°. .................... 47
Figure 4.3: Hyperspectral raw images of 28 direct bonded copper samples on the left
(a). Images after subtraction of the background on the right (b). In total,
24 samples were used for building the PLS-R model and four samples
were used for prediction. ...................................................................... 49
Figure 4.4: (a) UV reflectance spectra of copper sheets. Copper with initial condition
type 1 (green), 2 (red), 3 (blue), 4 (light blue), 5 (pink) and 6 (yellow)
represent the oxidation layer thicknesses 0 nm, 4 nm, 8.3 nm, 14 nm
and 21.1 nm, respectively. (b) PCA with scores and (c) the
corresponding loadings plot. ................................................................ 50
Figure 4.5: (a) Average UV hyperspectral imaging spectra of copper sheets. Copper
with initial condition type 1 (green), 2 (red), 3 (blue), 4 (light blue), 5
(pink) and 6 (yellow) represent the oxidation layer thicknesses 0 nm, 4
nm, 8.3 nm, 14 nm and 21.1 nm, respectively. (b) PCA with scores and
(c) the corresponding loadings. ............................................................ 52
Figure 4.6: Distribution maps of the oxide layer PC1 (a) and PC2 (b). Each rectangle
represents a single copper sheet. The sample type for each row
corresponds to Table 1. The samples are divided into two sets: model
building and model prediction for PLS-R. The colored pixels (the score
value range) represent the oxide content, from low (blue) to high
(red). ..................................................................................................... 53
Figure 4.7: Three-factor PLS-R models for the oxide layer thicknesses of direct bonded
copper in the UV region (200–380 nm). (a) Predicted vs. reference of
UV spectra. (b) Predicted vs. reference of UV hyperspectral imaging.
(c) Regression coefficients of the UV spectra. (d) Regression
coefficients of the UV hyperspectral imaging. ..................................... 55
Figure 4.8: Distribution map predicted from the three-factor PLS-R model of the UV
hyperspectral imaging data. The oxide layer thicknesses for each pixel
of samples (a) sample type 1, (b) sample type 3, (c) sample type 4 and
(d) sample type 5 were calculated for model prediction. ..................... 57
XXXIII
List of Figures
concept. (b) hyperspectral imaging scanning principle. (c) hyperspectral
imaging generated immediately from the scanning of a cotton sample
disc. (d) NIR-Spectrum for one single pixel extracted from the
image. ................................................................................................... 63
Figure 5.2: Raw cotton sample discs. ............................................................................. 64
Figure 5.3: (a) hyperspectral imaging of a cotton sample disc with area of interest (dash
line) with a diameter of 2.5 cm. (b) Spectra extracted from the selected
area. (c) Average spectrum of all spectra shown in (b). ....................... 64
Figure 5.4: (a) Hyperspectral raw imaging of 18 cotton sample discs with a diameter
of 3.1 cm. (b) Image of the color channel with the highest variance
between cotton disks and background. (c) Images after subtraction of
the background, removal of outliers, and application of filters. (d)
Image of RGB value corresponds to scores of the first (R), second (G)
and third (B) components. .................................................................... 66
Figure 5.5: (a) UV-Vis/NIR spectra of cotton sample discs including one HC sample in
the wavelength range 210 nm – 2200 nm. Upper left: Image of a cotton
sample disc where the region of integration for determining the average
spectra is indicated by a black area with a diameter of 2.5 cm. (b)
Scores plot for the processed spectra in the UV-Vis. The 2D projection
of the 95 % confidence ellipse of the data collected from each type of
cotton is included to facilitate visualization of the obtained results. (c)
Loadings plot for the UV-Vis. (d) Scores plot for the NIR. (e) Loadings
plot for the NIR..................................................................................... 69
Figure 5.6: (a) Spectra recorded by hyperspectral imaging of cotton sample discs
including one HC sample in the NIR range from 1100nm - 2200nm.
Upper right: Image of a cotton sample disc where the region of
integration for determining the average spectra for each sample is
indicated by a black circle with a diameter of 2.5 cm. (b) Scores plot for
the processed spectra in NIR- hyperspectral imaging. The 2D projection
of the 70 % confidence ellipse of the data collected from each type of
cotton is included to facilitate visualization of the obtained results. (c)
Loadings plot for the NIR- hyperspectral imaging. .............................. 70
Figure 5.7: Hyperspectral of cotton sample discs including one HC sample in the NIR
range from 1100nm - 2200nm. (a) Six example spectra recorded at
individual pixels. (b) Scores plot calculated for the whole data set
including several thousand processed spectra. (c) Loadings plot for the
NIR-Hyperspectral imaging.................................................................. 71
Figure 6.1: Overview of the samples pressed in the sample holder. For each
concentration, three samples were prepared and measured at once (A to
F and CLN). Real cotton samples with different honeydew contents
(light, strong and very strong. ............................................................... 80
XXXIV
List of Figures
concept (the tunnel in the scheme was cut to show the inside). (b)
Custom made sample holder consisting of quartz glass as sample cover
and PTFE as reference. (c) Pushbroom imager scanning principle. (d)
Hyperspectral image generated immediately from the scanning of a
sample. (e) UV spectrum after preprocessing for one point extracted
from the image given in (d). ................................................................. 82
Figure 6.3: Example of data extraction. (a) Hyperspectral raw images of 18 cotton
samples sprayed with different concentrations of sugar (A highest to F
lowest) and one cleaned cotton sample (CLN). For model building, all
spectra were extracted manually. (b) Zoom-in-image of a cotton sample
with the region of interest marked by a black rectangle. ...................... 83
Figure 6.4: (a) Averaged UV spectra of cotton samples with sugar solutions in different
concentrations: A (2 wt %, red), B (1 wt %, light green), C (0.5 wt %,
blue), D (0.25 wt %, light blue), E (0.0125 wt %, pink), F (0.0625 wt
%, yellow) and CLN (mechanically cleaned, dark green). PCA sugar
model for the cotton samples with (b) scores and (c) corresponding
loadings (PC1 black, PC2 red and PC4 blue). ...................................... 84
Figure 6.5: PLS-R model for different sugar concentrations in the UV region (225 nm –
410 nm). (a) Predicted vs. reference plot and (b) corresponding
regression coefficients for the sugar content with a five factor PLS-R
model. ................................................................................................... 86
Figure 6.6: Distribution maps of the sugar content predicted on the mechanically
cleaned cotton samples, which are manually sprayed by sugar solution.
The prediction of each pixel is based on the PLS-R sugar model. Each
rectangle represents a single cotton sample: A (2 wt %), B (1 wt %), C
(0. 5 wt %), D (0. 25 wt %), E (0.125 wt %), F (0.0625 wt %) and CLN
(mechanically cleaned). The colored pixels (see the score value range)
represent the sugar content, from low (blue) to high (red). .................. 87
Figure 6.7: Distribution maps of the sugar content predicted on the real cotton samples,
which are contaminated by honeydew. The prediction of each pixel is
based on the PLS-R sugar model. Each rectangle represents a single
cotton sample ((a) very strong, (b) strong, (c) light). The colored pixels
(see the score value range) represent the sugar content, from low (blue)
to high (red). ......................................................................................... 88
Figure 7.1: (a) Averaged spectra recorded via UV hyperspectral imaging of raw cotton
samples with sugar solutions in different concentrations: A (4 wt %), B
(2 wt %), C (1 wt %), D (0.5 wt %), E (0.25 wt %), F (0.125 wt %), G
(0.0625 wt %), and CLN (mechanically cleaned). PCA sugar model for
the cotton samples with (b) scores on the first principal component
(PC1) and second principal component (PC2) and (c) corresponding
loadings. ............................................................................................... 95
XXXV
List of Figures
Figure 7.2: Five-factor PLS-R model for different sugar contents in the UV region
(225 nm – 300 nm). (a) Predicted vs. reference plot and (b)
corresponding regression coefficients. ................................................. 96
Figure 7.3: Distribution maps of the sugar content predicted for each pixel of the UV
hyperspectral imaging data from the five-factor PLS-R model on the
raw cotton samples contaminated by honeydew. Each rectangle
represents a single cotton sample ((a) light, (b) strong, (c) very strong).
The colored pixels (see the score value range) represent the sugar
content, from low (blue) to high (red). ................................................. 97
Figure 10.1: Reference spectra for the copper Cu0, Cu2O and CuO by using UV
spectrometer). .................................................................................... XXI
Figure 10.3: Mean spectrum of pure dried protein on PTFE..................................... XXV
Figure 10.4: X-loadings weights and x-loadings for factor 1 (a, b), factor 2 (c, d), factor
3 (e, f), factor 4 (g, h) and factor 5 (i, j), respectively. ..................XXVII
Figure 10.5: X-loadings weights and x-loadings for factor 1 (a, b), factor 2 (c, d) and
factor 3 (e, f), respectively. .......................................................... XXVIII
XXXVI
List of Tables
Table 1.1: Confusion Matrix. ......................................................................................... 13
Table 3.1: Types of drug samples .................................................................................. 29
Table 3.2: UV band maxima positions of liquid and solid phase samples
[131,132,137]. ...................................................................................... 34
Table 3.3: The confusion matrix of the pure API spectra. ............................................. 39
Table 3.4: Classification of the painkiller tablets based on the pure API model. .......... 40
Table 4.1: Sample preparation protocol for the direct bonded copper substrates .......... 46
Table 4.2: Model statistics for the calibration and full cross-validation models for oxide
layer thickness on the direct bonded copper......................................... 54
Table 4.3: Prediction of the oxide layer thicknesses for direct bonded copper from PLS-
R models. .............................................................................................. 56
Table 5.1: UV-Vis/NIR reflectance maxima [165],[85,183]. ........................................ 68
Table 5.2: NIR Hyperspectral imaging reflectance maxima [85,183]. .......................... 71
Table 6.1: Description of the macronutrients and natural materials. ............................. 79
Table 6.2: The concentration of the sugar solutions and the weighted averaged sugar
amount on cotton samples. ................................................................... 81
Table 6.3: The number of honeydew stickiness points on cotton samples. ................... 81
Table 7.1: The sugar solution concentration and the weighted average sugar on cotton
samples. ................................................................................................ 94
Table 10.1: Description of the direct bonded copper substrates and their sample
preparation. ....................................................................................... XXII
Table 10.2: Description of the direct bonded copper substrates and their sample
preparation. ..................................................................................... XXIII
XXXVII
List of Publications
Published Manuscripts
1. Al Ktash, M.; Stefanakis, M. et al. Characterization of Pharmaceutical Tablets
Using UV Hyperspectral Imaging as a Rapid In-Line Analysis Tool. Sen-
sors 2021, 21,4436. https://doi.org/10.3390/s21134436.
2. Al Ktash, M.; Stefanakis, M. et al. UV Hyperspectral Imaging as Process Ana-

lytical Tool for the Characterization of Oxide Layers and Copper States on Direct
Bonded Copper. Sensors 2021, 21, 7332. https://doi.org/10.3390/s21217332.
3. Al Ktash, M.; Hauler, O. et al. Ultraviolet-visible/near infrared spectroscopy and

hyperspectral imaging to study the different types of raw cotton. Journal of Spec-
tral Imaging 2020; 9: a18. http://dx.doi.org/10.1255/jsi.2020.a18
4. Al Ktash, M.; Stefanakis, M.; et al. Prediction of Honeydew Contaminations on

Cotton Samples by In-Line UV Hyperspectral Imaging. Sensors 2023, 23, 319.
https://doi.org/10.3390/s23010319.
5. Knoblich, M.; Al Ktash, M.; et al. Applying UV Hyperspectral Imaging for the
Quantification of Honeydew Content on Raw Cotton via PCA and PLS-R Mod-
els. Textiles 2023, 23, 287-293. https://doi.org/10.3390/textiles3030019.
Submitted Manuscript
6. Knoblich, M.; Al Ktash, M. et al. Rapid Detection of Cleanliness on Direct
Bonded Copper Substrate Using UV Hyperspectral Imaging.
XXXIX
Declaration of Contribution
Description of the meaning of the personal contribution according to German law § 6 Abs. 2 sentences 3 of the doctoral regulations
Position of the Number of Scientific ideas by Data generation by Interpretation and analy- Paper writing by
No. Published /
candidate in the authors the candidate (%) the candidate (%) sis by the candidate (%) the candidate (%)
submitted
list of authors
1 published 1 5 60 80 70 60
2* published 1 11 45 80 45 45
3 published 1 4 40 80 70 60
4* published 1 7 45 90 40 40
5* published 1 6 45 50 30 40
6* submitted 1 11 45 50 30 40
*Shared first author
XLI
Acknowledgments
There are a number of people without whom this thesis might not have been originated and
finished, and to whom I am very grateful. There are also a number of people without whom
the life of my family during these years in Germany had been very difficult, and to whom I
am greatly indebted.
To:
Prof. Dr. Marc Brecht, you provided excellent supervision and assistance during my research
work. Your gentlemanly conduct impressed me greatly, and your constructive criticism and ad-
vice were always delivered politely. You instilled confidence in me and transferred positive en-
ergy whenever I needed it. You have been a wonderful academic advisor, and I will be forever
indebted to you. Please continue to be the same amazing person that you are.!
Porf. Dr. Alfred Meixner, thank you for your support during my time in Germany. Your expertise
and ideas greatly contributed to the success of my study, and your impact on my scientific and
personal life has been wonderful. Thank you very much!
Prof. Dr. Hermann Mayer and Dr. Erik Schäffer for agreeing to participate in this work and for
attending as dissertation examiners.
Prof. Dr. Karsten Rebner, who push me in the best way and help me in all steps to get my schol-
arship for my doctoral studies. I am grateful for your support.
Dr. Edwin Ostertag, I would like to express my gratitude for my continuous support during my
Ph.D. studies and related research. I am thankful for patience, motivation, and immense
knowledge. His guidance has been invaluable to me throughout the research and writing of this
thesis.
Prof. Dr. Volker Jehle, who provided invaluable assistance in supplying the necessary samples
and engaging in a productive discussion that greatly aided my understanding.
XLIII
Acknowledgments
My Friends Mona Knobblich, Tim Bäuerle, Tobias Drieschner, Frank Wackenhut, Barbara
Blodrini, Miriam Bassler, Julia Steinbach, Alexandra Wagner and Ashatoush Mukherjee.
I would like to thank for sharing their time with me, for their support during all these years in
Germany, and for always being glad to help or simply listen.
Reutlingen University, the university where I worked in Germany, allowed me to qualify and
economically supported me all those years.
KAAD, the Katholischer Akademischer Ausländer-Dienst that gave me a scholarship for my doc-
toral studies, had an extraordinary fulfillment!
Lastly, I would like to express my gratitude to my family, including my parents Mahmoud and
Nawla, and to my lovely wife Bayan and my son Osama my brother Alaa and sisters Madleen,
Nivin, Nisrein, and Walaa for supporting me spiritually throughout this work and my life in
general. God save and bless them.
XLIV

UV Hyperspectral Imaging

Hochgeladen von

Copyright:

Verfügbare Formate

UV Hyperspectral Imaging

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

UV Hyperspectral Imaging

Hochgeladen von

Copyright:

Verfügbare Formate

Development of a UV Hyperspectral Imaging

Prototype for Industrial Applications

Tag der mündlichen Qualifikation: 16.05.2024

In den letzten Jahren haben hyperspektrale Bildgebungssysteme aufgrund ihrer Multimodalität,

4.3.4 Data Collection and Preprocessing ....................................................... 47

7.4 Results and Discussion ..................................................................................... 94

E = Eelectronic + Evibrational + Erotational (1.1)

1.2 Hyperspectral imaging

400 500 600 700

(a) Whiskbroom (b) Snapshot (c) Staring (d) Pushbroom

Figure 1.5: Scheme of Vis/NIR hyperspectral imaging (pushbroom).

(b) (c) (d)

 - log (R/R0) 1.0

Synchrotron radiation is a form of electromagnetic radiation emitted by a particle accelerator

of the most expensive radiation techniques, with approximately 70 synchrotrons worldwide.

(a) (b) (c)

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

(d) (e) (f)

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

1.3.1 Principal component analysis (PCA)

X = t1p1 + t2p2 + … + tapa + E (1.4)

(a) (b) (c) PC3

Table 1.1: Confusion Matrix.

Positive (1) Negative (0)

1.3.2 Partial least squares regression (PLS-R)

B = W(P T) -1Q T (1.8)

1.3.2.1 Coefficient of determination

1.3.2.2 Root mean square error (RMSE)

Where n is the number of samples.

1.3.2.3 Standard deviation

1.4 Model systems

1.4.1 Pharmaceutical tablets

1.4.2 Direct bonded copper

1.4.3 Cotton fiber

Figure 1.14: Cotton fibers.

(a) (b) (c)

(a) Trehalulose (b) Trehalose (c) Melezitose

(d) Sucrose (e) Fructose (f) Glucose

This is originally published in sensors (https://doi.org/10.3390/s21134436) as

3.3 Materials and Methods

Table 3.1: Types of drug samples

3.3.2 API’s in Solution

Fluorescence excitation spectra were recorded by using a commercial setup (Fluorolog–3,

3.3.4 UV Hyperspectral Imaging

(b) (c) (d)

x 250 300 350 400

3.3.5 Data Collection and Preprocessing

3.3.6 Data Handling and Software

3.4 Results and Discussion

3.4.1.1 APIs in Liquid Phase, Transmission Spectroscopy

(ASA+PAR)pure 240 260 280 300

3.4.1.2 API and Painkiller Tablets, Total Hemispherical Reflectance Spectroscopy

0.8 ASApure 0.8 ASPBAYER

3.4.2 UV Hyperspectral Imaging

-4 -2 0 2 4 225 250 275 300 325 350 375 400