Hyperspectral Imaging for Landmine Detection

Ihab Makki

To cite this version:

Ihab Makki. Hyperspectral Imaging for Landmine Detection. Optimization and Control [math.OC].
POLITECNICO DI TORINO, 2017. English. �NNT : �. �tel-01706356�

HAL Id: tel-01706356

https://theses.hal.science/tel-01706356
Submitted on 11 Feb 2018

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 PhD Thesis
Submitted in fulfillment of the requirements for the PhD degree from

LEBANESE UNIVERSITY AND POLITECNICO DI TORINO

Major: Electrical and Electronics Engineering

Presented and defended by:

Makki Ihab
12 Decembre 2017
Hyperspectral Imaging for Landmine Detection
Supervised by:
Prof. Rafic Younes Prof. Clovis Francis
Prof. Tiziano Bianchi Prof. Massimo Zucchetti

Members of the Jury:

Mme Maria Sabrina Greco, Prof, University of Pisa, Reviewer
M. Ali El-Zaart,Prof, Beirut Arab University Reviewer
Mme Luisa Verdoliva, Dr, University of Naples Federico II Examiner
M. Georges Sakr, Dr, University Saint Joseph Examiner
M. Rafic Younes, Prof, Lebanese University Director
M. Clovis Francis, Prof, Lebanese University Co-Director
M. Massimo Zucchetti, Prof, Politecnico di Torino Director
M. Tiziano Bianchi, Dr, Politecnico di Torino Co-Director
M. Abdallah Fahed, Prof, Lebanese University Invited
M. Bou Maroun, Pierre, Col., Lebanese army Invited
 Acknowledgment
I would like to express my deep gratitude to my supervisors Prof. Rafic Younes, Prof. Clovis
Francis, Prof Tiziano Bianchi and Prof. Massimo Zucchetti for their continuous support, kindness
and availability during the three years of work on the thesis.
I would like to thank the external reviewers Prof. Maria Greco and Prof. Ali El-Zaart for accepting
the participation in the final dissertation and for their valuable comments that helped me in
improving the quality of the report. I want to thank also the examiners Dr. Luisa Verdoliva and
Dr. George Sakr for their positive contribution in the jury. A special thank goes to Prof. Fahed
Abdallah and Colonnel Pierre Bou Maroun that honored me with their participation in the final
defense.
Also, I would like to thank the Lebanese army for providing real samples of landmines extracted
from minefields in Lebanon. These samples were used in the collection of spectra of landmines in
different conditions. Also I would like to thank Dr. Mohammad Awad for his help in the
acquisition of the landmines’ spectra using the spectroradiometer.
A special thanks goes to Lebanese University and Erasmus Committee for the financial support in
Lebanon and in Italy.
Finally, the constant support of my family, friends and colleagues allowed me to successfully
conclude this PhD.

Thank you

2
 Abstract
This PhD thesis aims at investigating the possibility to detect landmines using hyperspectral
imaging. Using this technology, we are able to acquire at each pixel of the image spectral data in
hundreds of wavelengths. So, at each pixel we obtain a reflectance spectrum that is used as
fingerprint to identify the materials in each pixel, and mainly in our project help us to detect the
presence of landmines.
The proposed process works as follows: a preconfigured drone (hexarotor or octorotor) will carry
the hyperspectral camera. This programmed drone is responsible of flying over the contaminated
area in order to take images from a safe distance. Various image processing techniques will be
used to treat the image in order to isolate the landmine from the surrounding. Once the presence
of a mine or explosives is suspected, an alarm signal is sent to the base station giving information
about the type of the mine, its location and the clear path that could be taken by the mine removal
team in order to disarm the mine.
This technology has advantages over the actually used techniques:
• It is safer because it limits the need of humans in the searching process and gives the
opportunity to the demining team to detect the mines while they are in a safe region.
• It is faster. A larger area could be cleared in a single day by comparison with demining
techniques
• This technique can be used to detect at the same time objects other than mines such oil or
minerals.
First, a presentation of the problem of landmines that is expanding worldwide referring to some
statistics from the UN organizations is provided. In addition, a brief presentation of different types
of landmines is shown. Unfortunately, new landmines are well camouflaged and are mainly made
of plastic in order to make their detection using metal detectors harder. A summary of all landmine
detection techniques is shown to give an idea about the advantages and disadvantages of each
technique.
In this work, we give an overview of different projects that worked on the detection of landmines
using hyperspectral imaging. We will show the main results achieved in this field and future work
to be done in order to make this technology effective.
Moreover, we worked on different target detection algorithms in order to achieve high probability
of detection with low false alarm rate. We tested different statistical and linear unmixing based
methods. In addition, we introduced the use of radial basis function neural networks in order to
detect landmines at subpixel level. A comparative study between different detection methods will
be shown in the thesis.
A study of the effect of dimensionality reduction using principal component analysis prior to
classification is also provided. The study shows the dependency between the two steps (feature

3
extraction and target detection). The selection of target detection algorithm will define if feature
extraction in previous phase is necessary.
A field experiment has been done in order to study how the spectral signature of landmine will
change depending on the environment in which the mine is planted. For this, we acquired the
spectral signature of 6 types of landmines in different conditions: in Lab where specific source of
light is used; in field where mines are covered by grass; and when mines are buried in soil. The
results of this experiment are very interesting. The signature of two types of landmines are used in
the simulations. They are a database necessary for supervised detection of landmines. Also we
extracted some spectral characteristics of landmines that would help us to distinguish mines from
background.

4
Contents
Acknowledgment ............................................................................................................... 2
Abstract .............................................................................................................................. 3
List of Figures .................................................................................................................... 9
List of Tables .................................................................................................................... 11
1. Introduction ............................................................................................................... 12
2. Problem of landmines and existing solutions ........................................................... 14
2.1. Problem of Landmines ................................................................................................. 14
Landmine contamination and impact ................................................................... 14
Types of landmines .............................................................................................. 17
2.2. Landmine detection techniques .................................................................................... 18
2.2.1. Electromagnetic Methods ..................................................................................... 19
2.2.2. Ground Penetrating Radar (GPR) ......................................................................... 19
2.2.3. Infrared/Hyperspectral Systems ........................................................................... 20
2.2.4. Acoustic/Seismic method ..................................................................................... 20
2.2.5. Nuclear Quadruple Resonance (NQR) ................................................................. 20
2.2.6. Vapor sensors ....................................................................................................... 21
2.2.7. Mechanical methods ............................................................................................. 21
3. Hyperspectral Imaging: Introduction to landmine detection and processing
techniques ......................................................................................................................... 22
3.1. Introduction to Hyperspectral imaging ......................................................................... 22
Broadband, Multispectral, Hyperspectral and Ultraspectral Imaging .................. 22
Hyperspectral Image Scanning Modes ................................................................. 23
3.1.2.1. Whiskbroom or Across Track scanner ......................................................... 23
3.1.2.2. Pushbroom or Along Track scanner ............................................................. 24
Important hyperspectral camera ........................................................................... 25
3.2. Previous projects on landmine detection using HI ....................................................... 26
3.2.1. Defence Research and Development Canada projects ......................................... 26
3.2.2. Equinox Corporation fusion test ........................................................................... 30
3.2.3. Hyperspectral Mine Detection program HMD ..................................................... 30
3.2.4. Hyperspectral Mine Detection Phenomenology program .................................... 30
3.2.5. Joint Multispectral Sensor Program (JMSP) ........................................................ 31

5
 3.2.6. Night Vision and Electronics Systems Directorate (NVESD) ............................. 31
3.2.7. Defense Science and Technology Laboratory DSTL countermine project .......... 32
3.2.8. Indian Test to detect landmines using infrared images ........................................ 32
3.2.9. NATO project ....................................................................................................... 33
3.2.10. Humanitarian DEMining (HUDEM) and Belgian Mine Action Technology
(BEMAT) ............................................................................................................................. 34
3.2.11. FOI Multiple-Optical Mine detection System (MOMS) project .............................. 34
3.2.12. TELOPS test to detect buried object using airborne thermal hyperspectral images 34
3.3. Hyperspectral Image Processing .................................................................................. 37
3.3.1. Contrast enhancement .......................................................................................... 37
3.3.1.1. Histogram equalization................................................................................. 38
3.3.1.2. Morphological Contrast Enhancement ......................................................... 38
3.3.2. Filtering ................................................................................................................ 39
3.3.2.1. Wiener filter ................................................................................................. 39
3.3.2.2. Adaptive 3D Wiener filter ............................................................................ 40
3.3.2.3. Multiway filtering......................................................................................... 40
3.3.3. Segmentation ........................................................................................................ 42
3.3.3.1. Watershed Algorithm ................................................................................... 42
3.3.3.2. Hierarchical segmentation ............................................................................ 42
3.3.4. Feature extraction ................................................................................................. 43
3.3.4.1. Principal Component Transformation (PCT) ............................................... 43
3.3.4.2. Linear Discriminant Analysis (LDA) ........................................................... 44
3.3.5. Classification ........................................................................................................ 44
3.3.5.1. Support vector machine (SVM).................................................................... 45
3.3.5.2. K means clustering ....................................................................................... 46
3.3.5.3. Orthogonal subspace projection (OSP) ........................................................ 46
3.3.5.4. Matched Filter (MF) ..................................................................................... 46
3.3.5.5. Constrained energy minimization (CEM) .................................................... 48
3.3.5.6. Multiple Target CEM (MTCEM) ................................................................. 48
3.3.5.7. Winner take all CEM (WTACEM) and Sum CEM (SCEM) ....................... 49
3.3.5.8. Adaptive Coherent/Cosine estimator (ACE) ................................................ 49
3.3.5.9. Fully constrained least square (FCLS) ......................................................... 50
3.3.5.10. Adaptive Matched Subspace Detector (AMSD) .......................................... 50
3.3.5.11. Hybrid Unstructured Detector (HUD) .......................................................... 51
3.3.5.12. Spectral Angular Mapper (SAM) ................................................................. 51
3.3.5.13. Spectral Information divergence (SID) ........................................................ 51

6
 3.4. Recent developments in target detection using hyperspectral images.......................... 52
4. Experiments and Results ........................................................................................... 53
4.1. Preliminary test of applicability of hyperspectral images for landmine detection ....... 53
4.1.1. Detection Using VNIR, SWIR AND TIR ............................................................ 53
4.1.2. Supervised and unsupervised classification ......................................................... 55
4.1.3. Experiments .......................................................................................................... 55
4.2. Full pixel and subpixel mine detection ......................................................................... 58
4.2.1. Data description .................................................................................................... 58
4.2.2. Classification Results ........................................................................................... 60
4.2.3. Discussion ............................................................................................................ 62
4.2.4. Conclusions .......................................................................................................... 63
4.3. Effect of PCA Feature Selection Prior To Detection ................................................... 63
4.3.1. Data Description ................................................................................................... 63
4.3.2. Results .................................................................................................................. 64
4.3.3. Conclusions .......................................................................................................... 66
4.4. Effect of spectral variability on landmine detection ..................................................... 66
4.4.1. Data description .................................................................................................... 66
4.4.2. Results .................................................................................................................. 66
4.4.3. Conclusions .......................................................................................................... 68
4.5. MLP Neural network for landmine detection using hyperspectral imaging................. 68
4.5.1. Multi-Layer Perceptron (MLP) Neural networks ................................................. 69
4.5.2. MLP training and application ............................................................................... 69
4.5.3. Conclusions .......................................................................................................... 74
4.6. Multi Target Detection Using Neural Networks .......................................................... 74
4.6.1. Neural Networks based Target detection ............................................................. 75
4.6.2. Experiment on simulated data .............................................................................. 77
4.6.2.1. Data description ............................................................................................ 77
4.6.2.2. Results .......................................................................................................... 78
4.6.3. Real target experiment.......................................................................................... 81
4.6.3.1. Test Image .................................................................................................... 81
4.6.3.2. Results .......................................................................................................... 82
4.6.4. Conclusions .......................................................................................................... 83
4.7. Created Spectra method................................................................................................ 84
4.7.1. Spectrum creation ................................................................................................. 84
4.7.2. Conclusions .......................................................................................................... 86
4.8. Field Experiment .......................................................................................................... 87

7
 4.8.1. Reflectance spectra of landmines acquired in the lab........................................... 89
4.8.2. Reflectance spectra of landmines acquired in grass Field .................................... 95
4.8.3. Reflectance spectra of landmines acquired in soil Field .................................... 100
4.8.4. Conclusions ........................................................................................................ 105
5. Conclusions and Future Work ................................................................................107
References ......................................................................................................................110

8
List of Figures

Figure 1: Mines/erw casualties by civilian/military status in 2014 .............................................. 15

Figure 2: Mines/erw casualties by age in 2014 ............................................................................. 15
Figure 3: Comparison between AT and AP landmines ................................................................ 17
Figure 4: Landmine detection with metal detector ....................................................................... 19
Figure 5: GPR principle ................................................................................................................ 19
Figure 6: Amplitude of Surface Vibration of Ground in response to sound waves: over a Mine
(solid line) and a Blank (dashed line) ........................................................................................... 20
Figure 7: Difference between broadband, multispectral, hyperspectral and Ultraspectral Imaging
....................................................................................................................................................... 23
Figure 8: Wiskbroom scanning principle [132] ............................................................................ 24
Figure 9: Pushbroom scanner principle ........................................................................................ 24
figure 10: VNIR reflectance spectra of mines and background materials [103]. ......................... 54
Figure 11: VNIR and SWIR reflectance spectra of mines and background materials [105] ........ 54
Figure 12: The reflectivity spectrum of one pixel of Salinas ground, mine1, mine2. .................. 56
Figure 13: Detection performance of supervised methods: Normalized Cross Correlation (left)
and Orthogonal Subspace projection (right) ................................................................................. 56
Figure 14: Detection performance of unsupervised methods: Kmeans (left) and Fuzzy Cmeans
(right) ............................................................................................................................................ 57
Figure 15: Kmeans clustering after several run (left) and FCM clustering (right) in case of
subpixel target ............................................................................................................................... 58
Figure 16: Effect of PCA on FAR ................................................................................................ 64
Figure 17: Effect of PCA on computation time ............................................................................ 65
Figure 18: Example of multilayer perceptron NN ....................................................................... 69
Figure 19: Average Probability of detection ................................................................................. 72
Figure 20: Average False Alarm Rate .......................................................................................... 73
Figure 21: Average computational time ....................................................................................... 73
Figure 22: Multi-layer RBF Neural network ................................................................................ 76
Figure 23: Reflectance spectrum of the vs-2.2 mine (target) inserted in the image ..................... 77
Figure 24: Reflectance spectrum of the pmn mine (target) inserted in the image ........................ 77
Figure 25: Average computational time /algorithm ...................................................................... 79
Figure 26: average FAR/ algorithm .............................................................................................. 80
Figure 27: Created reflectance spectrum ...................................................................................... 85
Figure 28: Created spectrum performance .................................................................................... 85
Figure 29: Performance of created spectrum method when applied on another image ................ 86
Figure 30: Samples of landmines used for acquiring their reflectance spectra ............................ 87
Figure 31: Acquisition of the reflectance spectrum of TM-46 landmine ..................................... 90
Figure 32: trying different incident angle ..................................................................................... 91
Figure 33: PMN reflectance spectrum taken in LAB ................................................................... 92
Figure 34: VS 50 reflectance spectrum taken in lab ..................................................................... 92

9
Figure 35: PMD-6 reflectance spectrum taken in lab ................................................................... 93
figure 36: M411 reflectance spectrum taken in lab ...................................................................... 93
Figure 37: VS 2.2 reflectance spectrum taken in lab .................................................................... 94
Figure 38: TM 46-reflectance spectrum taken in lab .................................................................... 94
Figure 39: AP mines planted in grass ........................................................................................... 95
Figure 40: Grass reflectance spectrum .......................................................................................... 96
Figure 41: grass reflectance spectrum including in water absorption bands ................................ 96
Figure 42: PMN reflectance spectrum when covered by grass ..................................................... 97
Figure 43: VS50 reflectance spectrum when covered by grass .................................................... 97
Figure 44: PMD 6 reflectance spectrum when covered by grass.................................................. 98
Figure 45: M411 reflectance spectrum when covered by grass .................................................... 98
Figure 46: VS 2.2 reflectance spectrum when covered by grass .................................................. 99
Figure 47: TM-46 reflectance spectrum when covered by grass .................................................. 99
Figure 48: four AP mines exist in this scene. Could you localize them all? .............................. 100
Figure 49: Holding the device on my back, we acquired the spectra of the landmines after
burying them in the soil .............................................................................................................. 101
Figure 50: Landmines buried in soil ........................................................................................... 102
Figure 51: Untouched soil reflectance spectrum ........................................................................ 103
Figure 52: PMN reflectance spectrum when buried in soil ........................................................ 103
Figure 53: VS 50 reflectance spectrum when buried in soil ....................................................... 104
Figure 54: PMD 6 reflectance spectrum when buried in soil ..................................................... 104
Figure 55: M411 reflectance spectrum when buried in soil ....................................................... 105

10
List of Tables

Table 1: States/Areas With Mine/Erw Casualties In 2014 ........................................................... 16

Table 2: Summary of projects studied landmine detection using infrared and hyperspectral
imaging. ........................................................................................................................................ 35
Table 3: FAR (nb of false alarms/m2) .......................................................................................... 61
Table 4: Computation time in seconds.......................................................................................... 61
Table 5: Numberb of false alarms and computation time obtained when applying each algorithm
....................................................................................................................................................... 82
Table 6: FieldSpec 4 Hi-res spectroradiometer specifications ..................................................... 88

11
Chapter One
1.Introduction
Landmines and cluster munition constitute a main obstacle against the development of the
societies and return to normal life after the ceasefire is achieved. The fear of death and the
destruction won’t stop with the end of war, but will continue with the existence of threat of
cluster munitions, landmines, unexploded ordnance and improvised explosive devices. This
type of weapons doesn’t know when the war is ended and remain active for years or even
decades menacing innocent people in their everyday life
According to recent statistics [1], 80% of the casualties of landmines are from children that
have nothing to do with the war or its causes. Therefore, there is a need to ban the use of this
type of blind weapons. The efforts to ban landmines has started and we have an international
campaign to ban landmines and cluster munitions with numeral signees countries [130].
However, nobody can control the situation during the war and the landmines are being used
during recent conflicts (ex. In Libya and Syria 2016). Therefore, there is a need to find
detection techniques that are fast and reliable.
Different techniques have been addressed in order to detect landmines. Each method has its
advantages and inconveniences. One of the earlier and most used methods is the metal detector.
Due to electrical induction phenomena, this type of detectors is able to detect the objects that
contains metal under the soil. Although this technique is cheap, it has several drawbacks: it
detects all metals, either landmines or inert metals so it has very high false alarm rate; new
landmines contains less metals so they are harder to be detected. Other techniques used for
landmine detection will be mentioned in the next chapters.
In this thesis, we are addressing this problem with a new technique named Hyperspectral
Imaging or Imaging spectroscopy. This technique gives the ability to measure at each image
unit (pixel) the portion of light reflected in hundreds of wavelengths. Thus, we will obtain a
hypercube composed of two spatial dimensions and a third dimension that contains spectral
information. This technique is well used in remote sensing field for different purposes like
mapping, agriculture, astronomy, food monitoring, surveillance and others.
When light hits an object, it is either absorbed or reflected. The portion of light that is reflected
depends on the size of the molecules of the object that is reflecting on, the intermolecular

12
distances in addition to the wavelength of the radiation. Each material composed of different
components could reflect light of various wavelengths in a different manner. Therefore, we
have the ability using this technology to identify the materials remotely. In our case, we will
use the spectral and spatial information of the hyperspectral images to detect landmines and
cluster munitions without the presence of deminers on field.
Several approaches exist for target detection using hyperspectral imaging: some are supervised
where the spectrum of the data is known before; other are unsupervised based on searching for
targets that are spectrally different from their surroundings. The latter type of information does
not necessitate the knowledge of the target spectrum in advance. However, this type of
detectors is characterized by a high false alarm rate as we will see in the next chapters because
rare events in the image different from their background will be marked as targets.
During the work on this PhD thesis, we studied different scenarios of supervised and
unsupervised detection, taking into consideration image preprocessing techniques like feature
selection and dimensionality reduction. Knowing the target reflectance spectra will not make
its detection a straightforward process due to several reasons:
1) Spectral variability: the spectrum registered in lab conditions will not be necessary the
same in field condition due to effect of weather and illumination conditions.
2) Noisy images.
3) Low spatial resolution that make the reflectance detected in a pixel composed by a
mixture of endmembers reflectance spectra.
We worked on different types of supervised and unsupervised detection algorithms based
on probabilistic and linear mixture models. In addition, we worked on artificial neural
networks in order to detect landmines using hyperspectral images in a fast and more
accurate way.
In Chapter 2, we analyze the problem of landmines and show the existing methods currently
used to address this issue. An introduction to hyperspectral imaging with stat of art of landmine
detection using hyperspectral imaging and different tools used in hyperspectral image
treatment are shown in Chapter 3. In Chapter 4, we present all experiments done during my
work on the thesis with the results achieved. The conclusions are drawn in Chapter 5.

13
Chapter Two:
2.Problem of landmines and existing
solutions

2.1. Problem of Landmines

This chapter presents an introduction concerning landmines contamination issues, crisis, legacy,
and action. We conduct an analysis of mine action, national and international programs. It is
consequential to survey the foundation of the mine activity area since its start. We introduce the
problem of landmines and how action is taken to face these problems. A concise abridgement of
the expedition to ostracize killing mines is additionally included. In addition, we show the main
types of landmines.
Part of the information shown here is published in [129].

Landmine contamination and impact

Several countries suffer from the existence of millions of landmines in their territories. These
landmines have indefinite life, and may still cause horrific personal injuries and economic
dislocation for decades after a war has finished. Therefore, there is a growing demand by these
countries for reliable landmine inspection systems.
This problem affects the social and economic development of the regions, diminishes the areas to
be cultivated [2], [3], [4], [5]], and also risks killing innocent people; triggered by the fact that
mines do not know truce [[6], [7], [8], [9]]. As known, mines lead to hundreds of thousands of
deaths or to amputation of limbs. For instance, in Cambodia there are more than 35000 amputees
affected by landmine explosion [[2], [5]]. Some of the injured people die in the fields from bleeding
14
or lack of transport to reach the hospital [5]. Mines can decrease the area to be cultivated, also
prevent the income of valuable foreign currency coming from tourist’s visit, which lead to
economic regression.
Various obstacles are faced in removing these landmines, such as the loss or absence of maps or
information about the landmine types used or the areas where they were originally emplaced, the
change of landmines locations due to climatic and physical factors, the large variety of types of
landmines, and the high cost of locating and removing landmines. The landmines sensitivity to
explosion with time or atmospheric factors also poses a great danger to individuals.
Landmines are victim-activated and indiscriminate. Mines emplaced during a conflict against
enemy forces can still kill or injure civilians decades later. Land mines, cluster munitions, and
other explosive remnants of war (ERW) continue to kill or injure at least 4,300 people every
year [3]. The vast majority of recorded casualties are civilians (80% in 2015) as shown in Fig. 1
of which 39% are children (Fig 2). Between 1999 and 2012, more than 1,000 deminers have been
killed or injured while undertaking demining operations [10].

Figure 1: Mines/erw casualties by civilian/military

Figure 2: Mines/erw casualties by age in 2014
status in 2014

States with causalities in 2015 are shown in the next table:

15
Table 1: States/Areas With Mine/Erw Casualties In 2014

The total casualties in 2015 denoted the most yearly recorded losses since 2006. Year 2015
additionally denoted the most noteworthy number of yearly losses by extemporized mines
recorded by the Monitor [5].
When a landmine explodes, the impact of the explosion weakens as the distance increases from
the mine. The blast wave generated by to explosion has a peak power at the beginning and loses
its power while moving in the atmosphere. Accordingly, it is possible that get a high killing power
from a mine containing small amount of explosives in close contact, (for example, a mine under
the foot) while encountering considerably less damage from a significantly bigger dangerous
charge a few meters away. There is consequently a colossal assortment in the scope of touchy
wounds from landmines and UXO. Mine/UXO wounds have two fundamental effects. Firstly, they
influence the lives of the wounded and their family; furthermore, they have impacts on the
medicinal foundation of the influenced nation. The fundamental monetary impact on the casualty
is the constraining of capacity to acquire wage to bolster themselves and their family. In addition
to evident physical wounds, the setback may endure mental harm. Female setbacks are viewed as
being especially helpless as the broad physical harm can seriously restrict their odds of marriage.
The impacts are not constrained to the setback or their close families. Treating mine wounds
depletes the neighborhood therapeutic foundation of developing nations, as these sorts of wounds
unavoidably wind up noticeably tainted and typically requires 2-3 operations to debride the
wounds. Every loss will require prosthesis or a wheelchair on the chance to recover portability.
The prostheses will likewise require concentrated physiotherapy to figure out how to utilize the
counterfeit appendage. Moreover, most amputees will require another appendage ever 2-3 years
as the old ones destroy.

16
 Types of landmines

Mines can be outlined either as ‘anti-personnel’ or as ‘anti-tank'. Anti-personnel (AP) mines are
intended to be actuated by individuals, while anti-tank (AT) mines are expected to thrash tanks or
other shielded vehicles [11].
Anti-tank mines are designed to be triggered by heavy vehicles such as tanks. They are large
(usually bigger than a person’s shoe) and heavy (weighing more than 5 kilos). These mines contain
enough explosives to destroy the vehicle that runs over them and as a result also frequently kill
people in or near the vehicle. Anti-tank mines are laid where enemy vehicles are expected to travel:
on roads, bridges and tracks.
Anti-personnel mines are triggered much more easily and are designed to wound people. They
have less explosives and are much smaller and lighter than anti-tank mines—they could be as small
as a packet of cigarettes, weighing as little as 50 grams. Anti-personnel mines come in all shapes
and colors and are made from a variety of materials.
Although AP mines may kill a person, they are primarily designed to cause severe injury—a
wounded person must be assisted and this takes more of the enemy’s time and resources. Anti-
personnel mines can be laid anywhere and can be set off in a number of ways—stepping on them,
pulling on a wire or simply shaking them. Anti-personnel mines may also explode when an object
placed over them is removed. [12]
type AP landmine AT landmine
weight Light(100g-4Kg) Heavy(6Kg-11Kg)
size 6-20cm 20-50cm
target Human Vehicle
Case material Plastic,metal,wood Plastic,metal
Operating 5Kg 120Kg
pressure
Figure 3: Comparison between AT and AP landmines

Generally, there are two types of AP mines: blast mines and fragmentation mines.
BLAST MINES
Blast landmines are buried close to the surface of the soil and are generally triggered by
pressure. When a person steps on a blast mine and activates it, the mine's main charge
detonates, creating a blast shock wave consisting of hot gases travelling at extremely high
velocity. A famous type of blast mines is scatterable mines.[13]

17
FRAGMENTATION MINES
This type of landmines release fragments in all directions, or can be arranged to send fragments in
one direction. These landmines can cause injuries up to 200m away and kill at closer distances.
The fragments used in these landmines are either metal or glass. [13]
Anti-tank mines are designed to immobilize or destroy vehicles. All anti-tank mines are blast
mines, because the goal of the anti-tank mine is to destroy the tank's tracks and body. There's no
need for a fragmentation anti-tank mine.
Most countries and armies try to possess landmines to protect the main installations and key basis
from enemy intruders. When a military base is constructed in an open area, it will be vulnerable to
attacks by the enemy from all sides. In such cases, the landmines are used to limit the reachable
zones and focus the defensive forces in one side. Mines can also be used as part of the support
system for heavy artillery.
However, while landmines may have readily identifiable military applications, the nature, design,
and deployment of large numbers of mines will necessarily lead to civilian casualties
The neutralization of mines requires specialized training and remains a tedious and dangerous
process. Mines are often designed and deployed in order to make their detection as difficult as
possible. Furthermore, advances in technology are exacerbating the problem because most
modern mines are now made with plastics and may contain only traces of metal, if any. Newer
models may also contain sophisticated electronic fuses that make them more hazardous to remove.

2.2. Landmine detection techniques

In this section, we show the main detection techniques used in the detection of landmines. The
goal of this section is to show the techniques already used, their pros and cons and compare their
performance with the hyperspectral image technique.
The most widely used method for detecting mines follows the same techniques developed during
the Second World War, and directly involves human beings. The typical deminer’s tool kit today
largely resembles those used more than 50 years ago (It consists of a metal detector and a prodding
instrument).
Several techniques have been designed and developed for demining. Each technique is suitable
for detection under some conditions depending on the type of the mine case, the explosive material
and the soil.
As we are studying the potential of hyperspectral imaging for landmine detection, we will
describe the detection techniques currently used. The rationale of this section is not to go through
the complicated physics principles of how the sensors work but to give some brief information
about these techniques stating their strengths and limitations, in order to highlight the advantages
of hyperspectral imaging over other techniques.

18
 Generally, most of the landmine detection techniques consist of three main units; a sensor to
capture a signature of the landmine, a signal or image processing unit to arrange the acquired data
and a decision making unit to decide whether a landmine exists or not.
The sensor may be electromagnetic, acoustic, nuclear, biological, chemical or mechanical.

2.2.1. Electromagnetic Methods

The deminer holds the handle of Electro-Magnetic Induction

(EMI) detector (Fig.4) close to the ground and sweeps it slowly
around the area being investigated. Electrical current flowing
through the first coil, the “transmit coil,” induces a time-varying
magnetic field in the ground. This primary magnetic field, in turn,
induces electrical (eddy) currents in buried metal objects. The
currents from the buried objects create a weaker, secondary
Figure 4: Landmine detection with
magnetic field. The second coil, the “receiver coil,” detects metal detector
changes in voltage induced by the secondary magnetic field as
shown. The detector then converts these changes in the electric potential to an audible signal. [13]

2.2.2. Ground Penetrating Radar (GPR)

Difficulty in detecting tiny amounts of metal in a plastic land

mine with a metal detector has led to the development of this
technique. GPR detects buried objects by emitting radio waves
(ranging from about 10 MHz to a few GHz) into the ground and
then analyzing the return signals generated by reflections of the
waves at any subsurface discontinuity with different indexes of
refraction such as at the boundary between soil and a landmine or
between soil and a large rock. The GPR technique (Fig.5) uses an
antenna pair (transmitter and receiver separated by a small fixed
distance) to send short pulses of electromagnetic energy into the
subsurface and then record the returning signals. The return signal
is interpreted using a computerized signal processing system that Figure 5: GPR principle
gives an audio image to determine the object’s shape and position.
[[13],[14]]

19
 2.2.3. Infrared/Hyperspectral Systems

Infrared radiation consists of wavelength of 0.7𝜇𝑚 to 1𝑚𝑚 in microwave regions.

Infrared/hyperspectral methods detect anomalous variations in electromagnetic radiation reflected
or emitted by either surface mines or the soil and vegetation immediately above buried mines. Two
modes of action, including active and passive irradiation using a broad range of electromagnetic
wavelengths: A passive IR system detects natural radiation from the object whereas active systems
are provided with heat source and detects radiation from heated object. Thermal detection methods
exploit diurnal variations in temperatures of areas near mines relative to surrounding areas. The
physical activity of emplacing mines changes the natural soil particle distribution by bringing small
particles to the surface, which in turn affects the way in which the soil scatters light. Systematic
changes in vegetation moisture levels immediately above buried mines also may have
influence [13].

2.2.4. Acoustic/Seismic method

These methods are unique among detection methods

that identify the mine casing based on the
mechanical properties and are not based on
electromagnetic properties. The A/S technique is
used for the detection of landmines by vibrating
them with acoustic or seismic waves that are
generated and received by non-contact (acoustic)
and contact (seismic) transducers, respectively. The
transmitting system may be composed of acoustic
loudspeakers or electrodynamic shakers. When the Figure 6: Amplitude of Surface Vibration of Ground in response
receiver senses a reflected energy that means an to sound waves: over a Mine (solid line) and a Blank (dashed
line)
object possibly a landmine is buried. (Fig.6) [15].

2.2.5. Nuclear Quadruple Resonance (NQR)

This is a radiofrequency-based technique used to detect specific chemical compound like

explosives. It is composed of an emitter that sends a radiation with a frequency that corresponds
to the frequency of resonance of the explosive material. By this, the nuclei of the component is
excited and when it returns to the stable state, it emits another radiation that induces an electric
potential at the receiver coil. By this, the presence of a landmine is noticed by the detection of the
presence of explosives [16].

20
 2.2.6. Vapor sensors

A small percentage of the explosive manages to get out, as vapor, through fissures and shield
structures of mines. The idea is to detect the presence of vapor from explosives. There are two
research lines in this topic: biological and chemical.
Biological detection methods involve the use of mammals, insects, microorganisms, or plants to
detect explosives. Each of the different methods operates on a different set of principles and is at
a different stage of development.
A variety of possible non-biological mechanisms for detecting low concentrations of explosives
in air or in soil samples have been investigated in recent years leading to the development of highly
sensitive odor detection devices. When a sample of air containing explosives passes between the
slides of the sensor, some of the explosive binds to the polymer and reduces the amount of
fluorescent light that one slide emits. This reduction in the intensity of radiation received is
detected by a small photomultiplier device giving notice the existence of explosive
material.[13],[15]

2.2.7. Mechanical methods

In some cases, if the terrain and soil conditions are suitable, it is possible to use large armored
vehicles in order to clean the minefields. This method is preferred by the army during the time of
conflict as there is no much time to localize, identify and isolate the mines. It necessitates the use
of large and expensive vehicles. The risk is minimized as the demining personnel are either in a
well shielded place or are remotely controlling the vehicles. However, this technique leaves the
area virtually destroyed. In addition, a landmine may be buried deeper or partly damaged making
it more dangerous. [15]

21
Chapter Three
3.Hyperspectral Imaging: Introduction to
landmine detection and processing
techniques
3.1. Introduction to Hyperspectral imaging

Hyperspectral imaging is a trending technique in the field of remote sensing. It is based on

acquiring images in quasi-continuous bands in the visible and infrared domain. By this, we get
at each pixel a reflectance spectrum that help us to identify the constituents of the materials in
the image. This type of imaging is a developed version of the multispectral imaging technique.
In this section, we would like to show the origin of hyperspectral imaging, different data
acquisition processes used to acquire the hypercube in addition to main hyperspectral imaging
cameras used in this field.

Broadband, Multispectral, Hyperspectral and Ultraspectral

Imaging

Hyperspectral imaging is the result of development in the field of electro optics. The
development in sensor manufacturing made the image acquisition in hundreds of wavelengths
possible. The first imagers were broadband imagers that sense the light intensity in a wide

22
 range of spectrum. These imagers detect the light intensity in a wide range of frequencies in
the visible or infrared domain. After that, the eighties and nineties became the era of
multispectral imagers were the imagers were able to acquire image slices in tens of frequencies
(Fig.7). The development of photodetectors made possible to acquire image slices in even
narrower bands. This have increased the spectral resolution of the imager and improved the
possibility to distinguish more materials.

Figure 7: Difference between broadband, multispectral, hyperspectral and Ultraspectral Imaging

Hyperspectral Image Scanning Modes

Generally, to acquire a multispectral or hyperspectral images, two types of scanners are usually
used: Wiskbroom scanners and Pushbroom scanners. These two scanners differs in the technology
used to detect the light of different wavelengths. In both techniques, we use the forward motion of
the platform to record successive scan lines and build up the 2 dimensional images. In the
following, we will some of the characteristics of each technology.
3.1.2.1. Whiskbroom or Across Track scanner
Whiskbroom scanners collect measurements from one pixel in the image at a time. A rotating
reflecting device moves forth and back to reflect the incident light from different angles to the
single sensor that the scanner have. This allows the scanner to measure the energy from one side
of the aircraft to the other. The Instantaneous Field of View (IFOV) is scanned perpendicular to
the direction of motion of the sensor to form one spatial rom of the spectral image [132]. The
incoming energy is separated into several spectral components that are independently sensed.

23
 Figure 8: Wiskbroom scanning principle [132]

3.1.2.2. Pushbroom or Along Track scanner

A push broom scanner collects data along track using a row of sensors arranged perpendicular to
the direction of travel. The data are collected row by row. The imager scans a slit on the ground in
across-track direction. The slit image is focused and spectrally dispersed onto a two dimensional
array of image sensors.

Figure 9: Pushbroom scanner principle

24
Whiskbroom imager have inherently inferior spatial resolution compared to pushbroom imagers.
In addition, pushbroom imager has better geometry than wiskbroom imagers due to fixed distance
among detector elements. However, in pushbroom imager more detectors need to be calibrated
before use.
In new hyperspectral imagers, instead of using the dispersive elements to measure the energy in
different wavelengths, spectral filters in the fore optics at the focal plan are used to switch between
wavebands.

Important hyperspectral camera

Several hyperspectral imagers gained a large reputation in the field of hyperspectral imaging. The
number of companies that manufacture hyperspectral cameras is increasing as the use of this
technology is expanding to new domains. Now this technique is used in different unrelated
domains like chemistry, mapping, military, food quality monitoring, agriculture and other.
Therefore, the number of buyers is increasing and the number of the manufacturers so. Some of
the main hyperspectral cameras are:
 AVIRIS: Airborne Visible Infrared Imaging Spectrometer is operated by Jet Propulsion Lab
(JPL) of NASA. This sensor acquires hyperspectral images in 224 bands between 0.4 and 2.5
µm. the spectral resolution is about 10nm. It is a pushbroom sensor that have field of View
FOV=30° distributed on 614 pixels. The instantaneous field of view IFOV equal to 1mrad and
could be calibrated to 0.1mrad.
 HyMap sensor: it works from visible to thermal infrared range. The spectral resolution is about
10-20nm in the VNIR and SWIR regions and about 100-200nm in TIR. The FOV vary between
30 and 65° distributed on 512 sample with IFOV between 1 and 3 mrad. This sensor is
fabricated by Integrated Spectronics and mainly used for earth observation.
 COMPASS: this sensor is developed by the Night Vision and Electronic Sensors Directorate
(NVESD) of the US army. It works between 400 and 2350nm in 256 samples.
 HYDICE: Hyperspectral Digital Imagery Collection Experiment. It acquires hyperspectral
images in 210 bands between 400 and 2500 nm. It is manufactured by the NAVAL Research
Lab.
 CASI: Compact Airborne Spectrographic Imager. It is one of series of sensors manufactured
by ITRES research limited in collaboration with Defence Research and Development Canada
(DRDC). It detects image slices between 400 and 1000nm with 10 nm spectral resolution.
Other imagers also manufactured by the same company cover other ranges like SASI (1000 to
2500nm) and microTABI (3700 to 4800nm).
 NVIS: Night Vision Imaging Spectrometer is a pushbroom imager that uses two co-aligned
imaging spectrometers covering together the range between 400 and 2350 nm with 384 spectral
bands. The cross-track FOV is 13° composed of 256 pixels with IFOV about 0.9 mrad.

25
 EPS-H: It is a sensor manufactured by GER Corporation. It is composed of several imagers
that covers the range between visible 430nm to thermal infrared 12500nm. The spectral
resolution changes in each range.

3.2. Previous projects on landmine detection using HI

This research was published in the journal paper ISPRS journal of photogrammetry and remote
sensing [17].
Our goal in this section is to describe past projects that used infrared hyperspectral imaging for
landmine detection and that have been presented in conferences proceedings and journal articles.
Note that additional military research may exist in this field. Such projects, however, are not
described herein due to lack of information.

3.2.1. Defence Research and Development Canada projects

One of the earlier projects doing research on landmine detection using infrared wavelengths took
place at Defence Research & Development Canada (DRDC). DRDC started their research, in
support of the Canadian army on landmine and unexploded ordnance detection in 1978 and, in
collaboration with Itres Research, on hyperspectral imaging for landmine detection in 1989.
Detection of sparse targets using optical imaging was previously studied. Algorithms developed
during this project could be applied to preprocessed images of hyperspectral imagers. An early
project proposed a hierarchical image-processing algorithm to detect sparsely distributed bright
region of several pixels wide in a monochromatic image [18]. A preprocessing operation is
performed in order to remove distortions, dropouts, overlapping areas, misregistration, and any
other artifacts and imperfections. Non suspected areas are discarded to reduce the data size. Then,
suspected regions are segmented into homogeneous sub-regions and the morphological features of
the sub-regions are extracted. Based on the extracted features, sub regions are classified. Finally,
the spatial relationships between mine-like objects are determined. A supervised method analyzes
these relationships and classifies the areas as a minefield providing a specific likelihood ratio. This
hierarchical method can potentially achieve real-time detection of surface-laid mines. With the
aim of improving the detection system, scientific research was focused on two topics: the first one
dealt with the enhancement of the detection algorithms in order to achieve real-time detection,
while the second one was related to the improvement of proper imaging technologies in order to
obtain a higher image quality.
After the development of Visible and Near Infrared (VNIR) hyperspectral imagers (400-1000 nm),
several experiments showed their compatibility with the detection of surface-laid and buried
landmines. While testing the possibility to detect surface-laid mines, it was found that their spectral
reflectance has similar behavior under different illumination conditions with different scaling
factors and offsets. More precisely, a linear correlation exists between the mine spectra under
different incident illuminations if the spectral vector is confined between 500nm and 680nm [21].

26
For classification purposes, the authors tested two methods: Linear Cross Correlation (LCC) and
linear spectral unmixing. LCC is better in the case of high spatial resolution images. The linear
unmixing method has a higher Probability of detection in the case of subpixel sized mines; but has
also a higher false alarm rate.
Other tests led to study the possibility of detecting buried landmines using a VNIR imager. It was
noticed that buried mines could not be detected by calculating the shift of the red edge of vegetative
spectra. However, by using linear correlation, some mines with low vegetative cover were
detected [18]. It was also noticed that Anti-Tank surrogates were more detectable than
Antipersonnel surrogates, presumably due to the increased area of disturbance required to bury the
former [20]. The probability of detection (PD), intended as the number of mines detected over all
existing mines in the image, obtained during the experiment varies between 33% and 100% and
the False Alarm Rate (FAR), measured as the number of falsely detected mines per unit area, varies
between 0.1 and 0.52/m2. According to the authors of [20], improving the classification algorithms
and optimizing the spectral vectors, involving a systematic pattern classification study and
emphasizing discriminant analysis and feature analysis, are possible steps to achieve better PD and
lower FAR.
The spatial resolution of the image affects the performance of the detection algorithm [22]. As the
pixel size gets closer to the size of the mine, the possibility to isolate landmines increases. This
has been proven by the research team of DRDC in [23]. The authors acquired two types of images
using a VNIR imager: Medium resolution images at the altitude of 300m and high-resolution
images at the altitude of 6m in a different place. In the medium resolution experiment, they
obtained a 100% PD and 0.00034/m2 FAR. In the high-resolution experiment, all mines were
detected with a false alarm rate of 0.0043/m2. Linear Cross Correlation (LCC) and Orthogonal
subspace projection (OSP) were used in classification. The best detection is achieved when taking
the result of the combination of the two techniques.
In order to have quasi real-time detection of surface-laid mines using a VNIR imager, the authors
in [24] proposed a system consisting of two modes: in the first mode, the system learns the target
spectra. In the second mode, the system looks for the targets by acquiring spectral data for each
pixel and then applying comparative algorithms to the candidate pixels, using the stored reference
spectra. The processing platform involves a system that generates the results of data acquisition
and target analysis to an operator by displaying probability information alongside the base
imagery. The entire process (data acquisition - radiometric correction - data fusion from different
systems) finishes within few time frames of acquisition (a time frame is approximately 15-35 ms).
The radiometric and target identification processes can be applied independently to each frame, so
the processing of a frame will not affect the results related to the processing of other frames [24].
In [25], which is a continuation of the research in [24], we find the first experiment that aims at
detecting landmines from an airborne hyperspectral imaging system in real time. The above paper
describes how software and hardware improvements can achieve real time detection from an
airborne platform. First, radiometric correction is applied on raw data, then custom classification
algorithms are applied to the corrected data. A spectral signature library provides reference spectral
vectors. The classification results are stored and displayed in real time. The first real time landmine

27
detection system was mounted on a slow vehicle (1-2 km/h) [24]. A display system shows selected
bands including corrected spectral bands, partial data results or final target bands. The second real-
time detection system was an improvement of the first system to be compatible with airborne
imaging data rates. A hardware/software system was implemented measuring the change in slit
contamination (filings, dust, paint flecks) relative to the slit performance during calibration and
modifying the correction matrix accordingly during radiometric conversion. Detection rates were
not the prime concern of the test. The authors wanted to test the ability to detect landmines from
an airborne platform in real time. There are no indications regarding the algorithms used for data
correction, band selection, and classification.
Short wave infrared (SWIR) bands (1000-2500nm) have also been considered to detect landmines.
As the spectrum is wider with the inclusion of SWIR bands, the possibility to distinguish
landmines is higher. A simple classification boundary should be able to distinguish surface-laid
mines from many human-made artifacts and natural materials. However, old buried landmines are
hard to be detected using SWIR [26].
A project studying Long Wavelength Infrared (LWIR) hyperspectral imaging of landmines led to
the development of a commercially available LWIR hyperspectral imager suitable for airborne
landmine detection [27]. The instrument was used to collect imagery of surface and buried mines
and improvised explosive devices over full diurnal cycles in arid, desert-like conditions and was
found to provide some advantages over broad-band imaging in the detection of buried threat
objects [28].
The team of DRDC started in 1997 a project testing the combination of various detection
technologies called Improved Landmine Detector Project ILDP. Since a single detection technique
will not be able to detect all types of landmines in all conditions, the fusion of various techniques
can be more effective [29],[30]. The authors tested a small teleoperated vehicle carrying four types
of detectors: Forward Looking Infrared imager, down looking electromagnetic induction detector,
down-looking Ground Penetrating Radar (GPR) and finally a thermal neutron activation detector
used as confirmatory detector of suspected targets. In order to apply sensor data fusion, several
methodologies were used, including spatial correspondence and custom designed navigation. The
above system was intended for anti-vehicle landmines, but not for anti-personnel mines. In order
to address the latter, a smaller system with different sensors was proposed. Therefore, using a high
mobility robotic platform, the authors proposed a system that contains five separate technologies:
2 hyperspectral cameras (thermal infrared (TIR) and VNIR), a scanning sensor imaging system
which is mounted on a custom built articulated robotic scanner, and a nuclear confirmation
sensor [31]. The role of each technique is as follows:
 Forward looking SWIR or TIR cameras should detect thermal contrast between a landmine
and its surroundings.
 VNIR camera should detect spectral reflectance differences between disturbed and
undisturbed soil and the presence of a trip wire.
 Articulated Robotic Scanner affords the mechanical precision to provide images from scans
of a lightweight non imaging sensor.
 Nuclear imaging is used for confirmation.

28
  High mobility platform helps in moving the sensor payload.
In order to handle the enormous volume of data generated by hyperspectral imaging, the authors
proposed to use real-time techniques and algorithms described in [24], [25] to compress the
hyperspectral images into single band images, which could then be processed by the minefield
detection algorithms described in [18]. The results of these projects were encouraging and show
that a teleoperated replacement of a human operator may be possible in the future.
A discussion of the results obtained after landmine detection tests using VNIR, SWIR, and TIR
imagers by DRDC and Itres was presented in [32]. Reliable surface-laid mine detection in various
weather conditions was achieved using VNIR and SWIR spectra, even if not in real time. Reliable
buried landmine detection was not achieved. There is no huge difference in the VNIR range
between the signatures of buried landmines and background materials, however they could be
indirectly detected by observing differences in reflectance between compact soil over mines and
background.
DRDC and Itres presented a review of the research on infrared and hyperspectral technologies for
landmine detection in [33]. Besides providing the theoretical background for the detection of
surface-laid and buried mines and the results of their experiments, the authors also described
examples of Hyperspectral Imagery (HSI) images of trace amounts trinitrotoluene (TNT) and
Cyclotrimethylenetrinitramine (RDX) distributed on the ground surface. The mechanism of the
distribution of the trace explosives by ants is further discussed in [34], [35].
The Canadian research and development conducted a project between 2004 and 2008 called Shield
ARP 12rl in order to develop and exploit optical imaging sensors for mine detection. Airborne
tests of real time hyperspectral imaging and a SWIR HSI imaging phenomenology study were
completed in October 2006. Tests on vehicle mounted optical tripwire imager and development of
Thermal infrared hyperspectral imager were completed on March 2008 [36]. After the realization
of simultaneous imaging in VNIR and SWIR bands, the ability of classifiers to separate
camouflage coatings from background improves when the VNIR and SWIR spectra are combined.
Simultaneous collection of SWIR and TIR images from an airborne platform in an environment
with minimal infrastructure has also been done. In vehicle-mounted trip wire detector tests, the
SWIR provided better wire/background contrast than the VNIR band. The above report describes
the tests and the results obtained during the project without mentioning the algorithms used or the
way the real time airborne detection is performed.
DRDC and Itres proposed in [37] a new design of hyperspectral camera with a range-gated
intensifier and combined the camera with selected pulsed lasers. The authors showed that it is
possible to relate the reflected signal to specific light matter interactions, like induced fluorescence.
This approach is independent of the ambient light conditions and can be customized to specific
wavelengths. In addition, it could help in surveying a specific area in order to increase the SNR.
The preliminary results indicate that the false alarm rate associated with this scenario might be too
high for ground area scanning speeds of practical interest.
DRDC also began a project in 2005 to demonstrate the military utility of space-based reflective
hyperspectral imagery (0.4-2.5 microns), especially in the domain of target detection and

29
identification for land and marine mapping applications. The results achieved are encouraging and
show that target abundance can be retrieved with high accuracy at the subpixel level using the
Constrained Energy Minimization (CEM) algorithm. The fact that the estimated abundances are
generally lower than the true abundances is consistent with an error introduced during the manual
delineation of targets area, by assigning to targets larger areas than their true area [38].

3.2.2. Equinox Corporation fusion test

The fusion of visible and SWIR bands could give better detection results. A basic fusion of two
spectrum bands produces acceptable segmentation of objects against background, irrespective of
illumination conditions. In other words, selecting a set of two or three spectral image bands has
been found to be just as effective in differentiating man-made objects from background as using
all spectral bands at once [39]. Such fusion has the potential to detect mine-like objects in an image
using an integrated camera with visible and SWIR sensors and more sophisticated and specialized
detection algorithms.

3.2.3. Hyperspectral Mine Detection program HMD

In [40], a Defense Advanced Research Project Agency (DARPA) sponsored experiment testing
the potential to detect buried landmines using hyperspectral Mid-wave Infrared (MWIR) (3 to 5
µm) and Long-wave Infrared (LWIR) (8 - 12 µm) bands is described. The project emphasizes the
detection of surface disturbances due to landmine burying. Previous experiments showed the
capability of VNIR and SWIR imagers to detect surface disturbances [19], [20],[26]. However,
the problem was the high false alarm rate induced by surrounding vegetation and rocks. According
to the authors, the main rationale behind the detection of buried landmines using the spectral
properties is that the surface proprieties are in some way different from the properties of subsurface
soil. The soil exposure at the surface changes some of its physical and chemical properties. These
experiments showed that spectral information are necessary for landmine detection.
In addition, the researchers of the Hyperspectral mine detection program HMD tried to detect
buried landmines by evaluating the contrast in thermal reflectivity between the mine and the soil
in just two bands of the thermal IR region [41]. They noticed that recently buried landmines could
be seen in thermal infrared imaging as bright spots because the disturbed soil has an apparent
temperature different from that of the surrounding undisturbed soil. In addition, they claimed that
even mines buried for a very long time could be detected in some types of soil as the subsurface
mine will have different thermal properties.

3.2.4. Hyperspectral Mine Detection Phenomenology program

The American army also started the project “Hyperspectral mine detection phenomenology
program” (HMDP). Their main objective was to determine the existence of spectral characteristics

30
that are useful for landmine detection [42]. Therefore, they collected high quality hyperspectral
signatures of background materials and mines, measured temporal effects on buried landmines and
measured a statistically significant set of hyperspectral signatures of surface and buried mines in
natural soils, under variations of controlled variables. The spectral analysis results obtained during
the HMDP project recordings are presented in [43]. The authors concluded that uncontrolled
variables, mainly wind and rainfall, usually affect the results. The mines affected by more rainfall
continue to produce a signature distribution that is different from the background. Also, it is
remarkable that the temporal evolution of vegetation around landmines is too complex and makes
the characterization of temporal signature evolution extremely difficult. The following general
observations were made: 1) A light shower won't significantly reduce the signature; 2) The
signature is reduced by one-half inch of rain, 3) One-inch of rain further reduces the signature, but
does not eliminate it, and 4) For some conditions, several inches of rain may not eliminate the
signature. Overall, the VNIR and LWIR spectral regions show the most consistent and highest
performance. SWIR and LWIR show good performance for some conditions. MWIR showed the
least consistent and lowest performance.

3.2.5. Joint Multispectral Sensor Program (JMSP)

The goal of the research presented in [44] is to test the design of multispectral and hyperspectral
imagers that are able to obtain better detection performance by respecting the requirements and
conditions of target detection. For target detection, it is necessary to detect targets both in daylight
and nighttime conditions. Panchromatic or multispectral images in VNIR and SWIR ranges give
this capability during daylight. However, for military use, the MWIR and LWIR ranges are
necessary for nighttime operation. Due to high correlation of spectral bands of background
materials in all background conditions, the possibility to detect targets is high using MWIR and
LWIR ranges. After testing dual bands in MWIR and LWIR ranges, the authors concluded that
thermal multispectral images would give a better target detection and false alarm rate than a single
band infrared sensor. Tests showed that appropriately chosen small bands could provide good
detection, the optimal bands range being between 8 and 10.5 micrometers. There is a significant
increased utility of using LWIR with MWIR compared to the use of MWIR alone. Thanks to the
obtained results, the authors manufactured a new hyperspectral imager called SEBASS that works
in the ranges 2.9 to 5.2 micron and 7.8 to 13.4 micron. The Aerospace Corporation is still using
this sensor to take remote hyperspectral images in MWIR and LWIR ranges.

3.2.6. Night Vision and Electronics Systems Directorate (NVESD)

Night Vision and Electronics Systems Directorate (NVESD) has conducted during the fall of
2002 and spring of 2003 a wide variety of tests to examine airborne sensors for landmine
detection [45]. The examined hyperspectral sensors were the Airborne Hyperspectral Imager
(AHI) of the University of Hawaii, which is a Long-wave Infrared (LWIR) imager, and the
Compact airborne hyperspectral sensor (COMPASS) which is an NVESD VNIR/SWIR sensor. In
addition, a high frequency Synthetic Aperture Radar (SAR) and GPR have been used. The authors
tested two methods for classification: Signature based and anomaly detection. Further, for anomaly

31
detection two approaches were considered: Local like Reed-Xioli method and Global like
NFINDR. The latter is an unmixing model method and alone is not sufficient for classification
since it produces only abundance fractions as output. For that purpose, the authors proposed to use
it with a Stochastic Target Detector (STD). The output of STD is a detection stochastic map that
can be thresholded. The tests showed the capability of LWIR and reflection bands to detect
landmines with the use of proper algorithms. The detection of landmines at subpixel level is
challenging, but indeed possible with the use of high quality hyperspectral instruments and
algorithms.
Using the LWIR hyperspectral images acquired by AHI, another test has been conducted by
researchers at the Georgia Institute of Technology to detect a grid pattern of landmines and to use
this information to improve the detection performance. First, an anomaly detector is applied to the
hyperspectral data; in this case, the authors used the Dual Window-based Eigen Separation
Transform (DWEST). Then, pattern parameters are extracted and used to form a pattern projection
image. Finally, a pattern-based false alarm reduction is performed [46]. Using this process, higher
probability of detection at lower false alarm rate is obtained. Therefore, the results prove that the
inclusion of spatial pattern information in anomaly detection improves the detection of landmines
in minefields [46].

3.2.7. Defense Science and Technology Laboratory DSTL countermine

project

A project similar to those of DRDC and DARPA was started in Britain with the goal to detect
landmines using a VNIR imager [47]. The program was called DSTL countermine project. Using
the VNIR hyperspectral camera SOC 700 mounted on a tripod, the team took high spatial
resolution images of landmines. However, the data is mainly used to investigate different
processing methods and not to evaluate the PD and the FAR of the sensor. For data processing,
the authors used Principal Component Analysis (PCA) for dimensionality reduction and anomaly
detection method for classification. The authors avoid the use of spectral comparisons between the
target and each pixel of the image, as it will be very time consuming due to the low
target/background ratio. The results were still preliminary, however the authors concluded that
VNIR has the potential to distinguish surface-laid landmines from background.

3.2.8. Indian Test to detect landmines using infrared images

In India, researchers proposed a hierarchical algorithm to detect landmines from infrared images
that consist of preprocessing (contrast enhancement- filtering- smoothing), segmentation, feature
extraction, and ANN based classification [48]. The authors tested the algorithm on surface-laid
mines in two types of soil: black cotton and sand. During the preprocessing, the image is converted
to gray level. The two most important preprocessing stages are the contrast enhancement and noise
removal. Segmentation is the process of grouping homogenous pixels sharing some common
attributes such as color, intensity or texture in an image. The aim is to separate the image into
regions of interest and background, in order to make further analysis easier. Clustering, edge

32
detection, and threshold based region growing are the main three categories encompassing the
various existing image segmentation techniques [48]. Therefore, feature extraction and further
processes are applied on the clusters that are deemed mine like. A Neural Network (NN) based
algorithm is used to classify the mine from the surrounding. During the tests, the authors used a
small NN of 1 hidden layer and 4 neurons. The results provided on a simple dataset are good,
however the algorithm is not expected to work well on another field or type of soil as the data used
during the phase of learning are not rich enough.

3.2.9. NATO project

In the Netherlands, a project took place in cooperation with NATO to make a remote detector of
landmines. The main objective was to obtain near real time minefield detection during a conflict
using an Unmanned Aerial Vehicle (UAV) at a typical altitude of 100 m. First, the authors
presented the imaging technologies available at that time: Radar, Microwave radiometers, visible
wavelengths, near, middle and far infrared. After that, the authors showed the principal signal
processing techniques used for mine detection at that time. The main steps involved can be
categorized as:

* image enhancement
* edge detection
* segmentation
* feature extraction and classification
* morphology
At the end of the report, the authors gave the following main recommendations based on various
experimental results [49]
1. Conventional medium-resolution imaging radars are less suitable for remote mine
detection.
2. Microwave radiometry detection principle is promising for remote mine detection.
3. The characteristics of visible and near infrared imaging are often requested. This is because
imaging systems in these bands are often low cost, compact, have a high spatial resolution
and can be used in real time detection.
4. The mid- or long-wave infrared wavelength band is a promising band for remote mine
detection.
5. As Meteorological conditions (such as rain showers) can make mine and minefield
detection in mid- and longwave infrared wavelength bands difficult, it is better to combine
several wavelength bands.
6. A study on the best processing techniques and a reliable and accurate interpretation of the
images of a remote mine detection system has to run in parallel with the development of a
mine (field) detection system.

33
 3.2.10.Humanitarian DEMining (HUDEM) and Belgian Mine Action
Technology (BEMAT)

In Belgium, a research project focused on using the fusion of data from multiple sensors (Ground
penetrating radar, metal detector and infrared sensor) [50]. In the above paper, the authors
presented their views regarding multi-sensor data fusion potentials in improving the close-in
detection of landmines and reduction of mined area. Modelling and fusion of the extracted features
are based on belief function theory and possibility theory. After modelling, the fusion part is
performed in two steps: the first step consists in analyzing all data measured by one sensor. The
second step combines the results of the three sensors. The final part of the fusion approach is the
decision. According to the authors, the final decision about the identity of the object should be left
to a human observer with field experience. Therefore, the fusion output is an informative decision.
The experience showed that the fusion gives better detection than any input sensor used alone.

3.2.11.FOI Multiple-Optical Mine detection System (MOMS) project

FOI, A Swedish defense research agency, worked on a project for the Swedish armed forces
called Multi-Optical Mine detection System (MOMS). The objective of the project was to provide
knowledge and competence for fast detection of surface-laid mines using multiple optical
sensors [50]. The authors conducted research to test the feasibility of detecting landmines using
optical sensors and the possibility to combine multiple sensors. According to the authors,
hyperspectral imaging is an encouraging candidate for automatic detection and recognition of
exposed and semi-hidden mines, when a priori knowledge of the target spectral signature is
available. However, the detection performance is limited when the targets are camouflaged by
natural vegetation or hidden under other objects. In addition, the authors claim that no single
detection architecture is able to meet the performance needed under all operating conditions; the
choice of the particular sensors and algorithms will depend on environmental and operations
conditions [51].

3.2.12.TELOPS test to detect buried object using airborne thermal

hyperspectral images

In 2015, a Canadian research company specialized in infrared and hyperspectral imaging named
TELOPS proved the possibility to detect buried objects using an airborne LWIR hyperspectral
imager [52]. From an aircraft platform, they acquired thermal hyperspectral images of areas that
contain man-made objects previously buried. They found that the disturbed soil right above a
buried target is warmer than the undisturbed soil area next to it [52]. By comparing the emissivity
data obtained through the Temperature-Emissivity separation, the buried target sites show up as
part of the hottest ground area within the scene but further classification or additional information
are needed to discriminate the buried objects from other naturally hot areas.
A summary of the above projects and of the results obtained is given in Table 2.

34
Table 2: Summary of projects studied landmine detection using infrared and hyperspectral
imaging.
Research Project Type of data Techniques Used Comments
Detection of Infrared Hierarchical image Method would be useful as follow-on
surface-laid monochrom processing stage to process airborne hyperspectral
minefields using atic Image imagery after preprocessing in order to
a hierarchical reduce the hyperspectral image to a
image processing single band.
algorithm
(DRDC)
Surface laid VNIR LCC & Linear Surface-laid mines have consistent shape
Landmine Unmixing in VNIR bands; LCC performs well in
detection using case of high spatial resolution images;
VNIR (DRDC) Unmixing techniques have higher PD in
the case of subpixel target at the price of
higher FAR
Buried VNIR LCC Using VNIR, buried mines are not
Landmines directly detected, however the change of
detection using soil characteristics and vegetative stress
VNIR (DRDC) due to mine burying is detectable.
Effect of Spatial VNIR LCC & OSP LCC performs better when the pixel size
resolution on is smaller than mine size. OSP is better
mines detection when mine size is smaller than pixel size.
(DRDC) Best detection is achieved when the result
of two methods are combined.
Surface-laid VNIR Pipeline image the proposed suite of algorithms proves
Landmine processing the possibility to detect landmines in
detection using quasi real time using an airborne platform
VNIR in real
time (DRDC)
Landmines SWIR LCC Similarly to VNIR bands, the use of
detection using SWIR is beneficial to detect surface-laid
SWIR bands mines and recently buried landmines.
(DRDC)
Landmines LWIR (TIR) Spectral LWIR hyperspectral imaging provides
detection using comparison advantages over broadband LWIR
LWIR bands
(DRDC)
Multiple sensors Fusion of Dynamic range A proposed system employing
mounted on a VNIR, detector and hyperspectral imagers for close-in anti-
robot (DRDC) SWIR, contrast personnel mine detection.
LWIR HSI enhancement
and other
sensors

35
Active VNIR Casi imager with With the addition of external
hyperspectral intensifier illumination, the FAR increases as
imaging reflectivity of background increases.
(DRDC/Itres)
Equinox Project Fusion of Thresholded Ratio Here a ratio between two or three bands
visible and vegetation index is used. More bands using other
SWIR approaches may improve the results.
DARPA project MWIR and spectral LWIR and MWIR are more suitable to
to detect buried LWIR comparison detect buried landmines.
landmines
Hyperspectral VNIR,SWI Data collection Weather conditions affect the intensity of
mine detection R,MWIR,L using the reflected spectra. The effect of rain is
phenomenology WIR spectrometers more important than other effects.
program
Joint VNIR,SWI Fourier Transform Thermal sensor are beneficial for target
Multispectral R,MWIR,L detection at nighttime. LWIR bands are
Sensor Program WIR more effective than MWIR
airborne sensors VNIR,SWI RX and NFINDR LWIR gives a good detection with the use
tests (NVESD) R,MWIR,L with STD anomaly of proper algorithms. The inclusion of
WIR detection. Grid
spatial pattern information in anomaly
pattern detection of
detection improves the detection
landmines performance.
DSTL VNIR PCA more tests and other algorithms shall be
countermine tested to evaluate the effectiveness of
project VNIR bands in landmine detection
Indian Test to Infrared Hierarchical image
More images are needed to train the
detect landmines Image processing Neural network based classifier. A more
using infrared complex one may be used in complex
image situations.
NATO project VNIR,SWI Hierarchical image Radars are less suitable for airborne mine
R,MWIR, processing detection. Combination of bands is
LWIR necessary to overcome the
meteorological effects. Improvement of
algorithms and techniques in parallel is
necessary.
Humanitarian GPR, metal belief and Fusion of sensors may give better results
demining detector, possibility theory than single sensor.
(HUDEM & infrared
BEMAT) sensor
FOI (MOMS) VNIR,SWI Anomaly Hyperspectral imaging is useful for
R,MWIR, detection, Support automatic detection of open and semi-
LWIR, 3D Vector Machines hidden mines.
LADAR. The choice of sensor suite and algorithms
depends on environmental and
operational conditions.

36
 TELOPS LWIR Temperature- Soil above landmines is warmer than
Emissivity surrounding undisturbed soil.
separation, Linear Complementary information are needed
Unmixing to study to reduce the FAR.
the mineral
distribution

3.3. Hyperspectral Image Processing

In this section, we will introduce the detection algorithms used for target detection in
hyperspectral imagery. In addition, we will introduce several preprocessing steps and
hyperspectral data treatment usually used in a preliminary phase to simplify further detection or
classification. Most of these methods were developed during research on general problems
regarding the processing of hyperspectral images and are not specific for the landmine detection
problem. However, advances in that research will directly affect the success of landmine detection
using hyperspectral imaging. A review of different processing techniques used for data fusion,
spectral unmixing, classification and target detection could be found in [53].
After the acquisition of a hyperspectral image, the data pass through several steps. First, the image
is preprocessed to remove impurities, noise, and to reduce the size of the image. The main pre-
processing steps are contrast enhancement, filtering and smoothing. Then, segmentation is done to
separate useful data from background. After that, feature extraction is applied to extract the most
appropriate features for classification. Finally, classification or clustering methods are applied to
locate a target. In the following, we present the main algorithms used for target detection using
hyperspectral images. Many other methods may be used in each phase. However, in this chapter
we detail the most commonly used ones.

3.3.1. Contrast enhancement

The image enhancement process consists of a collection of techniques that try to improve the
visual appearance of an image or to convert the image into a better form suited for analysis by a
human or a machine [54]. Image enhancement methods are divided into two main categories:
spatial domain methods and frequency domain methods. Spatial domain methods are applied
directly on the pixels of the image. In frequency domain methods, the image is processed in the
frequency domain after applying the Fourier transform on the original data. Contrast enhancement
is one of the most commonly used image enhancement methods. For the mine detection case, the
role of contrast enhancement is to enhance the difference between the landmine and the
background materials [55]. The main contrast enhancement methods used are:

37
 3.3.1.1. Histogram equalization

Histogram Equalization (HE) is the most widely used contrast enhancement technique due to its
simplicity and effectiveness. The aim of HE is to make the probability distribution of gray levels
approximately uniform in the output image. It is a global method that flattens the histogram and
stretches the dynamic range using the cumulative density function of the image [55].
𝑛𝑘
The probability of the kth gray level in an image f can be described as 𝑝𝑓 (𝑓𝑘 ) =
𝑛

where k ∈ [0, L-1], L is the number of gray levels in an image, nk is the number of times the kth
level appears in the image, and n is the total number of pixels in the image. The histogram is the
plot of 𝑝𝑓 (𝑓𝑘 ) versus k, and the goal of the histogram equalization is to obtain an image with a
uniform histogram. The uniform histogram can be achieved by
𝑘 𝑘
𝑛𝑗
𝑔𝑘 = 𝑇(𝑓𝑘 ) = ∑ = ∑ 𝑝𝑓 (𝑓𝑗 )
𝑛
𝑗=0 𝑗=0

Keeping two conditions,

(a) T(fk) is single valued and monotonically increasing in the range k ∈ [0, L-1].
(b) T(fk) should be T(fk)∈ [0,L-1] for k ∈ [0,L-1].
The drawback of HE is that the brightness of the image is changed. To overcome this drawback
and improve the performance, many derivations of this method were proposed. Among them, we
list the following:
Brightness Bi-Histogram Equalization (BBHE)[56], Dualistic Sub Image Histogram Equalization
(DSIHE) [57], Minimum Mean Brightness Error Bi-Histogram Equalization (MMBEBHE)[58],
Recursive Mean Separate Histogram Equalization (RMSHE)[59], Multi Histogram Equalization
(MHE) [60], Brightness Preserving Dynamic Histogram Equalization (BPDHE) [61], Recursive
Separated and Weighted Histogram Equalization (RSWHE) [62], Global Transformation
Histogram Equalization (GHE)[63] and Local Transformation Histogram Equalization (LHE)[63].
3.3.1.2. Morphological Contrast Enhancement

Morphological theory has been introduced in image processing to overcome a number of

problems like image distortion due to noise. The first step in morphological contrast enhancement
is to find peaks and valleys in the original image. Peaks are light shades of gray tone image, while
valleys are dark ones. Peaks are obtained by subtracting the opening from the original image, and
valleys are obtained by subtracting the original image from the closing as
p(f) = f −γ ( f ),
v(f) = ϕ( f )− f ,

38
where p(f) denotes the peaks, v (f) denotes the valleys, γ(f) denotes the opening, and ϕ(f) denotes
the closing of an image function f. Basic definitions of morphological methods and operators
(erosion, dilation, opening and closing) could be found in [59]. To improve the contrast, the
peaks and valleys are multiplied by constants as follows:
max(𝑓)−max(𝐼) min(𝑓)−min(𝐼)
p′( f ) = p(f)× c1 , v ′( f ) = v (f)× c2 where: 𝑐1 = | | and 𝑐2 = | |
max[𝑝(𝑓)] max[𝑣(𝑓)]

where I indicates the gray level. In the case of 8 bit gray levels, max(I)=255 and min (I)=0.
The contrast-enhanced image is obtained as the summation of the original image, the peaks, and
the negative valleys f ′ = f + p′( f )− v ′( f ) [55].

3.3.2. Filtering

Filtering is an operation that allows to reduce the noise or to sharpen blurred areas in an image in
order to make it clearer and more suitable for further processes. In the filtering of hyperspectral
images, several techniques usually used in image processing have been upgraded to obtain
multichannel restoration. For example, the well-known Wiener filter used in image processing has
been extended to be used in hyperspectral images. There are two groups of filters: One is based on
the assumption that the within-channel information is separable from between-channel
information, i.e., spectral and spatial information are separable. These filters are called Hybrid
filters. In this case, the first step is to decorrelate channels using Fourier Transform or PCA and
then apply a classic 2D restoration method such as Wiener filter or Static Wavelet Transform. The
other group consists of a few proposed filters that do not rely on the assumption of spectral and
spatial separability [64].

3.3.2.1. Wiener filter

The Wiener filter is a widely used filter based on minimum mean square estimation. The original
image is obtained from the received image by minimizing the mean square error. It assumes that
the acquired image is composed of the original image and a white noise component that has a zero-
mean Gaussian distribution [65].
g (t)= f(t) + n(t) Where f(t) is the original image, g(t) the acquired image and n(t) the noise.

The estimation of f(t) is 𝑓̂(𝑡) = ∑𝑘=𝐿−1

𝑘=0 ℎ(𝑘)𝑔(𝑡 − 𝑘). It is estimated using L samples taken from
the received signal. h(k) is a variable independent of time to be found. It is calculated by
minimizing the approximation error
2
𝐽 = 𝐸(𝑒 2 (𝑡)) = 𝐸 [(𝑓(𝑡) − 𝑓̂(𝑡)) 2 ] = 𝐸[{𝑓(𝑡) − ∑𝑘=𝐿−1
𝑘=0 ℎ(𝑘)𝑔(𝑡 − 𝑘)} ]
𝜕𝐽 𝑑𝑒(𝑡)
The minimum is achieved by 𝜕ℎ(𝑖) = 𝐸 [2{𝑓(𝑡) − ∑𝑘=𝐿−1
𝑘=0 ℎ(𝑘)𝑔(𝑡 − 𝑘)} 𝑑ℎ(𝑖)] = 0

39
 𝑑𝑒(𝑡)
and = −𝑔(𝑡 − 𝑖)
𝑑ℎ(𝑖)

We can reformulate it in a matrix form:

H=[h0,h1, h2,…..,hL-1]T and G(k)=[g(k) g(k-1) … g(k-L+1)]T
𝜕𝐽(𝐻)
Thus = 2 RH-2P => H*=R-1P. This is called Wiener-Hopf equation.
𝜕𝐻

Note that R is the autocorrelation of G. It is a symmetric Toeplitz matrix and therefore it is positive
definite and non singular so R-1 has a solution. P is the cross-correlation between H and the input
image.
3.3.2.2. Adaptive 3D Wiener filter

As most of the filters used while preprocessing hyperspectral images are based on the assumption
of spectral and spatial separability, Gaucel et al [64]proposed a new filter for hyperspectral images
relying on spectral and spatial information simultaneously.
First the authors assume that the channel vector v(n1,n2) represents the zero-mean white Gaussian
noise, uncorrelated with the original image f(n1,n2). The received image is
g(n1,n2)=f(n1,n2)+v(n1,n2). Then, they apply the filter in local regions in which the signal-pixel
vector f(n1,n2) is assumed homogeneous. So f could be modelled as f(n1,n2)= mf +w(n1,n2), where
mf is the local mean of f(n1,n2) and w(n1,n2) a zero mean white noise.

The linear solution of Wiener filter is 𝑓̂ = 𝑚𝑓 + Γ𝑓𝑔 Γ𝑔𝑔−1

(𝑔 − 𝑚𝑔 ) where Γ𝑓𝑔 is the covariance of
f and g, and Γ𝑔𝑔 is the variance-covariance matrix of g. From the received image we could estimate
Γ𝑔𝑔 . But as the noise and the signal are uncorrelated, Γ𝑔𝑔 = Γ𝑓𝑓 + Γ𝑣𝑣 and Γ𝑓𝑔 = Γ𝑓𝑓

Since the noise is zero-mean, mf=mg and the equation becomes

𝑓̂ = 𝑚𝑔 + 𝐻(𝑔 − 𝑚𝑔 ) and 𝐻 = (Γ𝑔𝑔 − Γ𝑣𝑣 ) Γ𝑔𝑔

−1

Using the local region model, Γ𝑔𝑔 is estimated and mg is updated at each pixel.

3.3.2.3. Multiway filtering

Multiway filtering is another reformulation of the Wiener filter based on modelling the
hyperspectral image by a third order Tensor.
The collected hyperspectral image R is modeled as the sum of the desired original image X and the
additive white and Gaussian noise N
𝑅 =𝑋+𝑁

40
The goal is to estimate the original image by applying multidimensional filtering on the received
data

𝑋̂ = 𝑅1 𝐻1 2 𝐻2 3 𝐻3
Where n represents the n-mode product. The n-mode product between a data tensor R and matrix
Hn represents the consecutive matrix product between matrix Hn and the In-dimensional vectors
obtained from R by varying index in and keeping the other indexes fixed [66].
In order to determine the optimal n-mode filters H1, H2 and H3, the criterion used is the
minimization of the mean squared error between the estimated signal 𝑋̂ and the original one 𝑋.
e(H1,H2,H3)= E[||X − 𝑅1 𝐻1 2 𝐻2 3 𝐻3||2]
To estimate Hn, an Alternative Least Square algorithm is used, consisting of the following
steps [66]:
1. Initialization k = 0: R0 = R ⇔ H0n = IIn for all n = 1 to N (=3 in this case).
2. ALS loop: while ||X − Rk||2 > thr, with thr > 0 fixed a priori.
(a) for n = 1 to N:
i. Rkn = R ×1 H1k · · · ×n−1 Hn-1k ×n+1 Hn+1k . . . ×N HNk.
ii. Hnk+1=argmin ||X−Rkn×nQn||2 subject to 𝑄𝑛 = H1𝑇 𝐻1 ⨂ … H𝑛−1
𝑇 𝑇
𝐻n−1 ⨂H𝑛+1 𝑇
𝐻𝑛+1 ⨂. . . H𝑁 𝐻N
In×In
Qn∈ R .

(b) Rk+1 = R ×1 H1 k+1 · · · ×N HN k+1, k ← k + 1.

3. Output: 𝑋̂ = 𝑅1 𝐻1 2 𝐻2 3 𝐻3

Step (2)(a)(ii) of the ALS algorithm can be decomposed into the following sub-steps:
1.n-mode unfold R kn into R nk =Rn(H1k⊗…Hnk−1⊗Hnk+1...⊗HNk),and R
into Rn;

2. Compute γRRn = E(RnkRnT), perform its eigenvector decomposition (EVD) and place the
eigenvalues in λγk, for k = 1 to In;
3. Estimate Kn using Akaike Information Criterion or Minimum Description Length criterion.
1 𝐼𝑛 𝛾
4. Estimate σ γ (n)2 by computing 𝐼 −𝐾 ∑𝑘=𝐾 𝜆 and estimate βi by computing λγi- σ γ (n)2 for i=
𝑛 +1 𝑘
𝑛 𝑛
1 to Kn;
5. compute ΓRR (n) = E(Rn kRn kT ), perform its EVD, keep in matrix Vsn the Kn eigenvectors
associated with the Kn largest eigenvalues of ΓRR (n), and keep the Rn largest eigenvalues λnΓk for
k = 1 to Kn;

41
6. Compute the (k + 1)th iteration of n-mode Wiener filter Hn k+1 using the expression of n-mode
Wiener filter.
This method has been tested in [66] on different images and proved its efficiency by increasing
the SNR by about 3dB. However, one of the main drawbacks is an increased complexity and
computational time.

3.3.3. Segmentation

In the remote sensing community, segmentation is defined as the process of searching for
homogenous regions in an image, that is later followed by the classification of these regions [67].
In image processing, there are many methods used for segmentation, however not all of them are
applicable to multispectral and hyperspectral images. Some methods like watershed algorithms
have been upgraded in order to segment hyperspectral images. Globally, segmentation algorithms
are divided into two categories: Boundary-based and Region-based. Boundary based methods
detect the boundary using the discontinuity property. In region-based algorithm, pixels in a region
are grouped using the similarity property. In the following, we present the main methods used in
hyperspectral image segmentation.
3.3.3.1. Watershed Algorithm

The watershed algorithm is a powerful tool usually used for mathematical morphology
segmentation. In [68] the authors proposed to use spatial gradients and spectral markers for
segmentation. The algorithm works as follows:
First, to avoid obtaining a large number of minima while flooding the watershed using the gradient
function (over-segmentation), they determine markers for each region of interest using Clara
Clustering algorithm [69]. Then, the Factor Correspondence Analysis FCA [70] data reduction
method is applied to remove the redundancy of channels and filter the image. Next, a chi-squared
distance based gradient is performed on the filtered image, then watershed segmentation is
computed. This approach works well and proves that an adapted data reduction is necessary for
multivariate gradient segmentation.
3.3.3.2. Hierarchical segmentation

In 1989, Beaulieu and Goldberg [71] proposed a hierarchical process to segment images based on
hierarchical step-wise optimization. Hierarchical segmentation is defined as a set of segmentations
of the same image at different levels of detail in which the segmentations at coarser levels can be
produced from a simple merging of regions at finer levels [71]. First, each pixel is assigned to a
region label. Then, spatially adjacent regions with small dissimilarity value are merged. The
dissimilarity between new spatially adjacent regions are calculated and the pairs with smallest
value are merged. The process is repeated until the number of regions needed is obtained or all

42
values of dissimilarity are below a predefined threshold. The drawback of this method is the long
computational time while dealing with large data.
Tilton in 1998 [72] proposed a new hierarchical segmentation method called HSEG. The main
improvement of this method is that non-adjacent regions could be merged together and the
dissimilarity function is selectable. Another recursive version of this algorithm called RHEG was
proposed in [73] to overcome the problem of long computational time of HSEG. These algorithms
are registered patents for US government.

3.3.4. Feature extraction

Feature extraction consists in transforming the data from a high dimensional space to a lower
dimensional space chosen in such a way as to conserve as much as possible the information of
interest in the data. Feature extraction is used in hyperspectral image analysis to overcome the
problem of a low number of data training samples in comparison to the high spectral resolution of
the image and to reduce the computational time. There are many feature extraction algorithms that
are introduced; some are linear while others are nonlinear. While working on landmine or target
detection, not all feature extraction algorithms are useful, because the targets of interest are
generally sparse and the feature extraction may remove the key features of the target. In the
following, we are going to list some of these algorithms, their implementation and their
advantages.
3.3.4.1. Principal Component Transformation (PCT)

Principal Component Transformation, also called principal component analysis, Hotelling

transformation or Karhunen-Loeve transformation is a dimensionality reduction method based on
the minimization of the representation error. The idea is to choose the most representing bands
with the help of the eigenvalue decomposition of the covariance matrix of the hyperspectral
image [74]. The first step of PCT is the calculation of the covariance matrix of the image matrix.
Then, the eigenvalues of the covariance matrix are calculated and the eigenvectors are extracted.
Finally, the image matrix is projected onto the new subspace formed by the k orthogonal
eigenvectors corresponding to the highest eigenvalues. Y=WT x where x is a d x1 -dimensional
vector representing one image pixel, y is the transformed k x1-dimensional sample in the new
subspace and W is the transformation matrix of k orthogonal eigenvectors.
Note that while computing the PCT algorithm, the variance of the projections along the principal
components is equal to the eigenvalues of the principal components. In theory, PCT transformation
affects the classification of hyperspectral images. However, the overall effect on classification
does not change the general class patterns and, therefore, the dominating classification result
remains correct.

43
 3.3.4.2. Linear Discriminant Analysis (LDA)

Linear discriminant analysis is a statistical based method often used for feature extraction and
dimensionality reduction. It is also named Discriminant Analysis Feature Extraction (DAFE). It is
an extension of the well-known Fisher discriminant analysis, which is limited to binary class
decomposition. LDA computes an optimal transformation by minimizing the within-class distance
and maximizing the between-class distance simultaneously, thus achieving maximum class
discrimination [75]. Therefore, the first step is to calculate the within-class, between-class and
total scatter matrices. A transformation matrix is then defined and computed by applying the
eigenvector decomposition on the scatter matrix [76]. The main disadvantage of this method is
that it requires that the scatter matrix of the data be nonsingular. This method has also other
drawbacks: the maximum number of features extracted is equal to the number of classes minus
one. The number of training samples should be large enough to estimate the between-class and
within-class scatter matrix reliably. The between-class will be biased toward the class that has very
different mean value. Also, it is very time consuming compared to other methods. In addition, it
requires more training samples for hyperspectral images to calculate the class statistical parameters
at full dimension [77]. Many LDA extensions have been proposed to deal with the singularity
problem like PCA+LDA, regularized LDA (RLDA) , null space LDA (NLDA) , orthogonal
centroid method (OCM) , uncorrelated LDA (ULDA) , orthogonal LDA (OLDA), LDA/GSVD,
etc. [78].

In addition to the main methods we described above for feature extraction of hyperspectral
images, many other techniques exist like matched pursuit [77], neighborhood embedding [79],
Sammon’s mapping [80] and nonparametric weighted feature extraction [81].

3.3.5. Classification

It is the most important step in landmine and target detection. The performance of the algorithms
used in each of the previous steps and in the classification phase are evaluated by the study of the
classification results. The classification phase in an image based target detection process could be
defined as the step in which the pixels are discerned between target and non-target. Globally, the
classification algorithms are divided into two main classes: Supervised and unsupervised.
Supervised classification methods are based on the knowledge of the target and the use of training
samples. Unsupervised classification methods consist of grouping pixels that have similar
properties without the knowledge of target properties. Considering the way the classifier computes
the information in the pixels, classification algorithms are divided into per pixel classifiers,
subpixel classifiers, per-field classifiers, knowledge based classifiers, contextual and multiple
classifiers [82]. In landmine detection, unsupervised classification techniques are used when there
is no information on the type of mine present in the field or when there is the possibility that a
particular type of mine is deployed but its reflectance spectrum is not in the library of known

44
spectra. However, unsupervised classification methods do not work well in every possible
condition and suffer from high false alarm rate due to the generally low number of target pixels
compared to background pixels. While the use of unsupervised methods could help in detecting
unknown types of landmines, the use of supervised classification methods is necessary for the
identification of mines. In the following, we are going to mention the major classification methods
used in landmine detection:

3.3.5.1. Support vector machine (SVM)

Support vector machine is a powerful non-parametrical supervised classification method. Firstly,

it was proposed for binary classification and regression [83]. Then, it has been used in the
classification of hyperspectral images [84]. SVM consists in finding the best separation between
two classes based on the separation of representative training samples called support vectors. In
addition, SVM does not suffer from Hughes effect and may perform separation of classes having
very close means even with a very small number of training samples [85]. First, we start with a
couple of training samples (xi,yi) where yi is a class label equal to ±1 which indicates the class of
the pixel and xi is a d-dimensional vector which represents the spectrum of the pixel in d
wavelengths in the case of hyperspectral images. If the classes are linearly separable by a
hyperplane, the SVM classifier is represented by the function f(x)=w.x+b where w is a vector ∈
Rd and b is a real bias ∈ R that could separate the classes without errors. The decision is made
according to the sign of f. The SVM approach consists in finding the separating hyperplane that
has the largest distance from the closest training samples. This distance is expressed as 1/||w||. The
margin is defined as 2/||w||. So to calculate W and b, the following optimization must be calculated:
min{1/2 ||w||2} with yi(w.x+b)≥1, for all samples. By introducing the Lagrangian formalism, the
problem is transformed to the dual problem:
1
Max of: ∑𝑁 𝑁 𝑁 𝑁
𝑖=1 𝛼 𝑖 − 2 ∑𝑖=1 ∑𝑗=1 𝛼𝑖 𝛼𝑗 𝑦𝑖 𝑦𝑗 (𝑥𝑖 . 𝑥𝑗 ) with the condition ∑𝑖=1 𝛼𝑖 𝑦𝑖 = 0 𝑎𝑛𝑑 𝛼𝑖 ≥ 0.
Where 𝛼𝑖 are Lagrange multiplier that can be estimated using quadratic programming.
If the samples are not linearly separable, suitable kernel functions are used to project the data into
a higher dimensional feature space in which the data could be linearly classified. Profiting from
this transformation, the inner product in the maximization (𝑥𝑖 . 𝑥𝑗 )is replaced with the function
𝑘(𝑥𝑖 . 𝑥𝑗 ).

There are many types of kernel functions, including: polynomial: K(xi, xj) = (1 + xi.xj)q ;Gaussian
radial basis K(xi, xj) = exp(−||xi−xj||2/(2σ2)) ; Laplacian radial basis K(xi, xj) = exp(−||xi−xj||/(2σ2))
; Sigmoidal K(xi, xj) = tanh(α0(xi.xj) + σ2). In the case of multiclass classification, two approaches
could be used: One against all, where each class is discriminated using the samples of all classes.
One against one, where a larger number of classifiers are computed using each time the training
samples of two different classes.

45
 3.3.5.2. K means clustering

K means clustering is one of the most used clustering methods for hyperspectral images. In k
means clustering, the pixels of the image are grouped into classes based on spectral similarity.
First, k random centroids are assigned. Then each pixel is assigned to the closest centroid. The
norm used to calculate the distance between the pixel and the centroid could be the Euclidian
distance, Manhattan distance, max distance, or linear combination of the above distances. After
that, new centroids are found by calculating the mean value of each cluster. Then, the clusters are
reformulated. This process is repeated until the total number of iterations is achieved or the total
distance between classes is minimized [86].
3.3.5.3. Orthogonal subspace projection (OSP)

Orthogonal subspace projection is a supervised classification method used to detect targets in

hyperspectral images at subpixel level. This method is based on the theory of spectral unmixing
which consists in subdividing the reflectance spectra of each pixel into endmembers spectra. This
method was proposed by Harsanyi and Chang in 1994 [87] in order to exploit a priori knowledge
of the target and facilitate the target detection. Suppose the image pixel is modeled by the equation:
x=ta +Bα+ξ where:
x = spectral vector characterizing the pixel
t = spectral vector associated with the target
a = unknown fractional abundance of the target within the pixel
B = matrix of vectors of the scene endmembers (materials found in the scene background)
α = unknown fractional abundance of each basis vector
ξ = residual error associated with this model.
After the background suppression, OSP uses the matched filter to determine if the target spectrum
is a part of the pixel spectra by calculating its abundance. This is done using the OSP operator
δOSP(x)= tTPB┴x where PB┴=I-BB# is the orthogonal background operator, and I is the identity
matrix. The fractional abundance of the target within the pixel can be computed as follows:
𝑎̂=Tosp(x)= (tTPB┴t)-1 δOSP(x) [87].

3.3.5.4. Matched Filter (MF)

This technique is based on the statistical approach. The problem is posed as a hypothesis testing
problem between the two hypotheses:
H0: Mine absent (Background material)

46
H1: Mine present
In the statistical model, the background and the target mine are usually considered to be following
a Multivariate normal distribution (MVN) as follows [88]:
𝐻0 : 𝑥~𝑁(𝜇𝑏 , 𝜎 2 Σ𝒃 )
𝐻1 : 𝑥~𝑁(𝑎𝑆𝑡 , 𝜎 2 Σ𝒕 )
Where 𝜇𝑏 is the background mean vector, 𝑆𝑡 is the target spectrum, Σ𝑏 and Σ𝑡 are the variance-
covariance matrices of background and target respectively, a and 𝜎 are scaling factors.
The detectors should satisfy the Generalized Likelihood Ratio Test (GLRT)
𝑓(𝑥⁄𝐻 )
𝐿(𝑥) = 𝑓(𝑥⁄𝐻1 )
0

where 𝑓(𝑥⁄𝐻𝑖 ), 𝑖 = 0,1 is the conditional probability density function (PDF) of the input x given
hypothesis Hi. The unknown parameters of the PDFs are replaced with their maximum likelihood
estimates.
In order to simplify the model, we consider that the target and the background have the same
covariance matrix. In addition, we remove the mean of the background from all pixels and from
the target. The new hypotheses are as follows [89]:
𝐻0 : 𝑥~𝑁(0, 𝜎 2 Σ𝒃 )
𝐻1 : 𝑥~𝑁(𝑎𝑆, 𝜎 2 Σ𝒃 )
Where S=St-µb.
The conditional density functions are:
1 1 𝑇 −1
𝑓(𝑥⁄𝐻0 ) = 𝑝 1 𝑝 exp (− 𝑥 𝛴𝑏 𝑥)
2𝜎 2
(2𝜋)2 |Σ𝑏 |2 (𝜎 2 )2
1 1
𝑓(𝑥⁄𝐻1 ) = 𝑝 1 𝑝 exp(− 2𝜎2 (𝑥 − 𝑎𝑆)𝑇 Σ𝑏−1 (𝑥 − 𝑎𝑆))
(2𝜋)2 |Σ𝑏 |2 (𝜎2 ) 2

To simplify the calculation of the maximum likelihood ratio, a monotonic function (usually
logarithmic) is applied without affecting the performance of the matched filter [90]. The derivative
of the logarithmic of the likelihood ratio is equal to
1 1
𝐿′ (𝑥) = − 2𝜎2 (𝑥 − 𝑎𝑆)𝑇 Σ𝑏−1 (𝑥 − 𝑎𝑆) + 2𝜎2 𝑥 𝑇 Σ𝑏−1 𝑥

expanding L’(x) we will arrive to the MF detector

𝑦 −1
𝑆 𝑇 Σ𝑏 𝑥
𝑀𝐹=
√𝑆 𝑇 Σ−1
𝑏 𝑆

In the literature, we may find the MF defined as a finite impulse response (FIR) filter y=hTx where

47
 𝑇
𝑆 𝑇 Σ𝑏−1
h =
√𝑆 𝑇 Σ𝑏−1 𝑆

3.3.5.5. Constrained energy minimization (CEM)

This algorithm is derived from a different point of view than the MF but in the end, it provides a
very similar solution. The main difference is that CEM uses the correlation matrix instead of the
covariance matrix. Therefore, there is no need in CEM to subtract the mean of image scene from
all pixels.
The objective of CEM is to design a FIR linear filter W=(w1,w2…,wL)T that maximizes the
response for a given target d while minimizing the output power [91].
𝑑𝑇 𝑤 = 1
The output of the filter is given by:
𝑦 = 𝑤 𝑇 𝑥 = ∑𝐿𝑖=1 𝑤𝑖 𝑥𝑖
The average output power is equal to:
1 1 1
∑𝐿𝑖=1 𝑦𝑖2 = ∑𝐿𝑖=1(𝑟𝑖𝑇 𝑤)𝑇 𝑟𝑖𝑇 𝑤 = 𝑤 𝑇 ( ∑𝐿𝑖=1 𝑟𝑖 𝑟𝑖𝑇 )𝑤 = 𝑤 𝑇 𝑅𝐿𝑥𝐿 𝑤
𝑁 𝑁 𝑁

where RLxL is the autocorrelation matrix of the samples.

The solution is found using the Lagrange multipliers methods as shown in [92].
The optimal filter coefficients are:
−1
𝑅𝐿𝑥𝐿 𝑑
𝑤𝑜𝑝𝑡 = 𝑇 −1
𝑑 𝑅𝐿𝑥𝐿 𝑑

3.3.5.6. Multiple Target CEM (MTCEM)

This is an extended version of the CEM algorithm that supports the detection of multiple targets.
Suppose we have a matrix D=[S1,S2,…Sp] that contains the signature of p targets. The objective
now is to minimize the output energy with the constraint DT. w=1. Where 1 is a px1 column vector
of ones.
The solution is given in [93]by:
−1 −1
𝑦𝑀𝑇𝐶𝐸𝑀 = (𝑅𝐿𝑥𝐿 𝐷(𝐷𝑇 𝑅𝐿𝑥𝐿 𝐷)−1 𝟏)𝑇 . 𝑥

48
 3.3.5.7. Winner take all CEM (WTACEM) and Sum CEM (SCEM)

Another two techniques for multitarget detection based on CEM are Winner take all CEM and
Sum CEM [93]. These algorithms necessitate running the CEM algorithm each time for each
target. Then the results of the detectors are summed up in SCEM or we take the maximum among
other detectors in case of WTACEM. One advantage of WTACEM over SCEM is that if the results
are noisy, the noise is not summed up in the final detector.

3.3.5.8. Adaptive Coherent/Cosine estimator (ACE)

In the derivation of the Adaptive Coherence Estimator, we will use the same procedure we used
in the case of MF. However, in deriving the ACE we will assume that the covariance matrix in the
two hypotheses is scaled by different factors 𝜎02 𝑎𝑛𝑑 𝜎12 . Therefore, the hypotheses are drawn as
follows [88]:

𝐻0 : 𝑥~𝑁(0, 𝜎02 Σ𝒃 )
𝐻1 : 𝑥~𝑁(𝑎𝑆, 𝜎12 Σ𝒃 )
The likelihood ratio will be:
−𝑝/2
𝜎2 1 1
𝐿(𝑥) = (𝜎12 ) 𝑒𝑥𝑝 {− 2𝜎2 (𝑥 − 𝑎𝑆)𝑇 Σ𝑏−1 (𝑥 − 𝑎𝑆) + 2𝜎2 𝑥 𝑇 Σ𝑏−1 𝑥}
0 1 0

The MLE of the scaling factors 𝜎02 𝑎𝑛𝑑 𝜎12 are obtained by differentiating f(x|H0) and f(x|H1) with
respect to 𝜎 2 .
The results are:
1
𝜎12 = (𝑥 − 𝑎𝑆)𝑇 Σ𝑏−1 (𝑥 − 𝑎𝑆)
𝑝
1 𝑇 −1
𝜎02 = 𝑥 Σ𝑏 𝑥
𝑝
The MLE of a is given by:
𝑥 𝑇 Σ−1 𝑆
𝑏
𝑎 = 𝑆𝑇 Σ−1
𝑏𝑆

By replacing these estimates in the likelihood ratio equation, we can arrive to the ACE detector
given by the equation:

49
 𝑥 𝑇 Σ𝑏−1 𝑆(𝑆 𝑇 Σ𝑏−1 𝑆)−1 𝑆 𝑇 Σ𝑏−1 𝑥
𝑦𝐴𝐶𝐸 =
𝑥 𝑇 Σ𝑏−1 𝑥

ACE could be considered as a little extension of the MF algorithm where the result is normalized
by the length of the whitened input pixel √𝑥 𝑇 Σ𝑏−1 𝑥 [133].
3.3.5.9. Fully constrained least square (FCLS)

This is another method based on the linear mixing approach. As its name suggests, the abundances
in this method are calculated so as to respect all abundances’ constraints: the non-negativity and
the sum-to-one constraints. This makes the method useful for material quantification in
hyperspectral imagery [94].
Returning to the linear mixing model, each pixel is written as:
𝑟 = 𝑀𝛼 + 𝑛
Where r represents the spectrum of the pixel, M is a concatenation of target and backgrounds
signatures, 𝛼 is the abundance factor and n represents the noise.
The least square cost function is given by:
𝑝
1
𝐽 = (𝑟 − 𝑀𝛼)(𝑟 − 𝑀𝛼)𝑇 − 𝜆(∑ 𝛼𝑗 − 1)
2
𝑗=1

𝜆 is Lagrange multiplier. Differentiating the cost function with respect to α and making it equal
to zero we obtain: (𝑀𝑇 𝑀)−1 𝑀𝑇 𝑟 − 𝜆=0
We obtain 𝜆 = (1 − 𝟏𝑇 𝛼̂)/(𝟏𝑇 𝑠) with s=(𝑀𝑇 𝑀)−1 𝟏 and 𝛼̂ = (𝑀𝑇 𝑀)−1 𝑀𝑇 𝑟. The solution of
FCLS is found using the following procedure [94]:
 Calculate 𝛼̂
 Compute 𝜆 and set 𝛼̂
𝐹𝐶𝐿𝑆 = 𝛼
̂ − 𝜆𝑠
 If all components of 𝛼̂𝐹𝐶𝐿𝑆 are positive, the algorithm stops.
 If not, divide each negative value of 𝛼̂𝐹𝐶𝐿𝑆 by its corresponding component in the vector s,
set the maximum absolute fraction to zero and remove its corresponding endmember
signature. Return to step 1.
3.3.5.10. Adaptive Matched Subspace Detector (AMSD)

This technique is based on Generalized Likelihood Ratio Test between the two hypotheses. The
noise is supposed to be a zero-mean normal distribution with covariance matrix 𝜎 2 𝐼 [88].
𝐻0 : 𝑥~𝑁(𝑈𝛼𝑈 , 𝜎02 𝑰)
50
 𝐻1 : 𝑥~𝑁(𝑆𝛼𝑆 + 𝑈𝛼𝑈 , 𝜎12 𝑰)
By replacing the unknown with their MLE in the likelihood ratio we arrive to the AMSD detector
given by:
˔
𝑥 𝑇 (𝑃𝑈 −𝑃𝑍˔ )𝑥
𝐷𝐴𝑀𝑆𝐷 (𝑥) = 𝑥 𝑇 𝑃𝑍˔ 𝑥

𝑃𝑈˔ = 𝐼 − 𝑈(𝑈 𝑇 𝑈)−1 𝑈 𝑇 𝑎𝑛𝑑 𝑃𝑍˔ = 𝐼 − 𝐸(𝐸 𝑇 𝐸)−1 𝐸 𝑇

E is defined as the concatenation of the background and target signatures.

3.3.5.11. Hybrid Unstructured Detector (HUD)

HUD is based on mixing both statistical and physical models. The first step is to calculate the
abundances using non-negative least square or FCLS. By this, we use the physical model
information. The obtained abundances are then used as inputs of statistical based detector like
ACE. The detector is written as follows [88]:
𝑥 𝑇 Σ−1 ̂
𝑏 𝑆𝛼
𝐷𝐻𝑈𝐷 (𝑥) = 𝑆 𝑇 Σ−1
𝑏 𝑆

Where 𝛼̂ is the abundance estimate obtained using constrained least squares.

3.3.5.12. Spectral Angular Mapper (SAM)

Like Euclidian distance, spectral angular mapper is a measure of similarity between two vectors.
For hyperspectral target detection, SAM is used to calculate the angle between the target
reflectance spectra and pixel reflectance spectra treated as vectors. The smaller the angle, the more
similar the pixel is to the target. The angle is calculated using the following equation [93]:
𝑃⃗ .𝑇
⃗
𝐷𝑆𝐴𝑀 = cos −1 (‖𝑃⃗‖‖𝑇⃗‖ )

where 𝑃⃗ is the pixel vector and 𝑇

⃗ is the target vector.

3.3.5.13. Spectral Information divergence (SID)

Another way to detect the presence of a target in the hyperspectral image is to calculate the
similarity between each pixel and the target using spectral information divergence. This technique
is inspired from information theory where the degree of similarity is calculated using the entropy
formula [95]:

51
 𝑆𝐼𝐷(𝑥, 𝑦) = 𝐷(𝑥||𝑦) + 𝐷(𝑦||𝑥)
𝑝 𝑥𝑙
Where 𝐷(𝑥||𝑦) = ∑𝐿𝑙=1 𝑝𝑙 log(𝑞𝑙) and 𝑝𝑙 = ∑𝐿
𝑙 𝑙=1 𝑥𝑙

As in the SAM case, the result of the detection depends on the precision of the target reflectance
spectra and is sensitive to the spectral variability.

3.4. Recent developments in target detection using

hyperspectral images.

In recent years, researchers proposed various new algorithms to detect targets in a hyperspectral
image. Although the different approaches are devoted to generic target detection, they represent
promising candidates for improving the performance of current landmine detection techniques. As
a matter of fact, landmines constitute a special type of targets, since they are usually rare and sparse
in the scene, and they have different shapes, colors and reflectance spectra. For example, various
approaches to model a hyperspectral image, in addition to a comparison between supervised
Matched filter and unsupervised Reed-Xioli target detection algorithms, are presented in [89]. A
nonlinear version of the algorithm Target Constrained Interference Minimized Filter based on
kernels is recently proposed in [96]. In [97], the authors propose a new endmember extraction
process to detect anomalies in a hyperspectral image. Some researchers proposed new models to
interpret the hyperspectral data in order to simplify the target detection process. Here we mention:
Forward modelling working in radiance space [98], Sparse Representation Based Binary
Hypothesis Model (SRBBH) [99], Sparsity and Compressed sensing based models [100] and
spatio-spectral Gaussian random field modeling [101].

52
Chapter Four
4.Experiments and Results
4.1. Preliminary test of applicability of hyperspectral images
for landmine detection

In this chapter, we will show the first tests done in order to verify the applicability of
hyperspectral imaging for the detection of landmines. Here we did a test on simulated data in
order to precise the spectral information in the hypercube that helps in distinguishing
landmines from background materials. In addition, we tested both Supervised and
unsupervised classifiers in order to highlight the pros and cons of each type of approach. The
results were mainly presented in [102].

4.1.1. Detection Using VNIR, SWIR AND TIR

Even if sometimes mines are just laid on the surface or very close to it, they are still hard to
be detected. New mine casings are made similar to the background and hard to be visible by
naked eye at visible wavelengths. Because of that, visible wavelengths are not sufficient to
detect landmines, especially the buried ones, so we will use infrared bands. While camouflage
matches mine coating reflectivity to that of the background in an average sense, exact matches
only occur at a few points across the visible and near infrared spectrum (Fig.10)[103][32] . It
is difficult to match a coating to a background over a wide range of the spectrum. Quite narrow
bands may have large differences in reflectivity between mines and background. Such subtle
mismatches between mine and background spectra in VNIR range can be discerned if the
spectral range is finely divided [104].

53
 figure 10: VNIR reflectance spectra of mines and background materials [32].

The differences between surface-laid mines and background are more important in the SWIR
spectral region. Earlier studies have shown that the reflectivity of many mine coatings and
background materials is significantly lower in the VNIR than in the SWIR region (Fig.11) [26].
For example, the reflectivity of the AP landmine is very close to that of a healthy leaf in the visible
spectrum, but different in most of the infrared region. In order to achieve better detection, we may
focus on the SWIR spectrum as in this range the contrast between background materials and man-
made objects is much larger. Plastic mines and painted unexploded ordnance have special
pigments in their reflectance spectrum that allow simple classifiers to distinguish synthetic objects
from natural features such as vegetation and soil.

Figure 11: VNIR and SWIR reflectance spectra of mines and

background materials [26]

Thermal infrared hyperspectral (TIR, 8000 to 12000 nm wavelength) have the potential to detect
buried mines in certain types of recently disturbed soils. The most common mineral constituent of
sand in the Earth's continental crust is quartz (SiO4 silicon–oxygen tetrahedra); Soil disturbance

54
has measurable impact on the quartz reflectance spectrum, presumably due to mixing of different
soil particle sizes [103]. The presence of other materials, such as carbonates, may also cause similar
reflectivity changes in the TIR region, which also may be suitable for detection of soil disturbance.
However, the use of TIR images for the detection of disturbance depends on results of other
unfinished researches that should precise the effect of weather changes and time on the disturbance
detection in different types of soil.

4.1.2. Supervised and unsupervised classification

An approach to analyze a hyperspectral image is to attempt to match each pixel spectrum

individually to one of the reference reflectance spectra in a spectral library. This is the supervised
classification method. Correlation-based classifiers work well with multipixel-sized mines,
whereas spectral unmixing methods can detect subpixel-sized mines [105]. Some supervised
algorithms are shown in section 3.3.5.
Unsupervised classification, or learning, is a term for grouping objects with similar properties
together, without any foreknowledge of those properties. Clustering can be used for classification
on multi-dimensional images. The image clustering result is an assignment of each spatial position
to a spectral class based on the values of the different points in the image bands. The results of
clustering can be used to determine the location and number of classes present. A supervised
connection can later be applied to the results with available spectral reference data.

4.1.3. Experiments

To compare between the two types of algorithms we did a test in order to detect the spectrum of
landmine in a hyperspectral scene. The experiments were done on the image SalinasA (which
could be found on the website [106]. This scene has 224-band over Salinas Valley, California, and
is characterized by high spatial resolution. It includes vegetables and bare soils. It comprises 50*50
pixels and includes six classes. In this preliminary experiment, we substitute the spectrum of one
pixel with the spectrum of the landmine. The reflectance spectra of the landmines inserted and
bare soil are shown in figure 13:

55
 Figure 12: The reflectivity spectrum of one pixel of Salinas
ground, mine1, mine2.

The application of two supervised classification methods Normalized Cross correlation and
orthogonal subspace projection are shown in fig 14. Note that Red circles designate pixels
classified as mine type1 and magenta stars are mine type2.

Figure 13: Detection performance of supervised methods: Normalized Cross Correlation (left) and Orthogonal
Subspace projection (right)

56
 Same procedure is done but using the Kmeans and Fuzzy Cmeans unsupervised classification
techniques and the results are shown in the fig.12. Red circles are for pixels classified as mine
type1 and mine type2. A supervised connection can later be applied to the results with available
spectral reference data to discriminate between type1 and type2.

Figure 14: Detection performance of unsupervised methods: Kmeans (left) and Fuzzy Cmeans (right)

All the mines are detected with a 100% probability of detection and no false alarm rate. The
discrimination is accomplished easily in our work because the spectral data of the implemented
mines are spectral data of two types of plastic mines. Also the implemented spectra are not covered
with any ambiguities as for the real case where mines are covered with dirt or vegetation or even
camouflaged to match surrounding.
So the pixels that contain the reflectance of the mines show great difference in the reflectivity
spectrum than the surrounding (Fig. 12).
There is a need for real images of minefields to investigate accurately the classification and
clustering methods.
The tested processing methods showed the potential for a high probability of detection, although
further investigation is required for detection in more difficult scenarios. The performed
experiment shows that mines possess spectral features that allow them to be distinguished from
other materials. Successful surface landmine detection in the VNIR has been shown using spectral
signatures. However, the graphs showed that there is more distinguishing spectral characteristics
in the SWIR than in the VNIR. More spectral characteristics may increase the detection and
identification rates and lower the false alarms. Although reliable detection is not obtained yet, TIR
HS imagers suggests promise for buried landmine detection.
In another scenario, the planted spectrum was mixed with the reflectance spectrum of background
material. We mixed the reflectance spectrum of landmines by a portion of 0.3 background and 0.7

57
mine. We applied the two supervised method (NCC and OSP) and the two unsupervised clustering
methods (Kmeans and Fuzzy Cmeans) on the new scene. Using the two supervised methods, we
were still able to detect landmines with 0 FAR. However, using the unsupervised method, many
FA showed up (see Fig.15). In case of k means, if we run it several times, few false alarms at the
borders shows up (Fig. 15 left). But if we run the FCM several times, same large number of false
alarms will show up every time. A possible explanation is that by implanting the mixed spectrum
(landmine + background), the contrast between the implanted target and surrounding background
is reduced in such a way that the implanted spectrum could not be distinguished in a specific
cluster. In addition, other rare events that exist in the scene and were not planted are marked as
targets.

Figure 15: Kmeans clustering after several run (left) and FCM clustering (right) in case of subpixel target

4.2. Full pixel and subpixel mine detection

In this section, we describe the simulation tests done in order to emulate a hyperspectral scene of
a minefield and evaluate the performance of different types of supervised detection algorithms. In
the first part, we present the image used in addition to the methodology that we followed to
simulate a minefield. In the second section, we present the results of this simulation followed by a
discussion and conclusions deducted out of this experiment.

4.2.1. Data description

In this experiment, we use a part of an AVIRIS image scene named f100902t01p00r03 available
on the following website (http://aviris.jpl.nasa.gov). This image was acquired by the airborne
AVIRIS sensor that acquire hyperspectral images in 224 bands ranged between 395 nm and
58
2500nm with 10nm spectral resolution. The spatial resolution of the image is 0.8m. The original
size of the image was 995 samples and 8716 lines. A part of the original scene containing grass
and sand in the background was taken. The size of the chosen area is 148x123 pixels (118.4 x
98.4m= 11651 m2). In order to emulate a minefield, the spectrum of a landmine is inserted in
different locations of the hyperspectral image scene.
In the first step, as the image is given in radiance unit, it is converted to reflectance domain. This
process is called atmospheric correction. Usually working in the reflectance domain is preferred
because the reflectance value is independent of the illumination and weather conditions. Then we
spatially upsample the image to arrive to a pixel size equal to the size of the landmine to be inserted.
For this, we use the bicubic interpolation. By upsampling the image, we increase the number of
pixels per unit area and the spectrum of the added pixels are interpolated according to the spectra
of the surrounding pixels. So we obtain an image with lower Ground Sample Distance (GSD)
where each pixel represents smaller area but have the same characteristics of the original one (same
components).
On the other hand, the information acquired in some bands are too noisy as these bands correspond
to the water absorption bands. We may remove these bands in order to reduce the noise and reduce
the size of the image at the same time.
After that, depending on the surrounding background material, we replace some pixels in the image
with the reflectance of the landmine. By this, we implant the mines in the scene. Now we can apply
the classification algorithms on this image to detect the full-pixel target.
In order to simulate the case of subpixel target, the same classification algorithms are applied on
downsampled versions of the image. The downsampling was done by grouping the neighbor pixels
using bicubic interpolation after using an anti-aliasing low pass filter. The downsampled images
have the same area size of the original one but the pixel size is larger. Therefore, the reflectance
spectrum of the mined area now is the reflectance of the mine mixed with surrounding background
spectra.
Finally, to study the effect of downsampling, we apply the same algorithms on images that have
the same size of downsampled images but have pixel size equal to the size of the mine. In this
case, we try to detect full-pixel targets in smaller images. However, the area covered by these
images is smaller.
Each time we apply a classification algorithm, we obtain a metric for each pixel that represents the
degree of similarity between the pixel and the target that we are searching for. After that, we choose
a threshold to classify the suspected regions from clean ones. Changing this threshold will change
the probability of detection and the False Alarm Rate (FAR). In this study, we chose the threshold
in a way to detect all targets and then we registered the FAR obtained when using each algorithm.
To note that by FAR here we mean the number of wrongly detected landmines per square meter.
The following chart resumes the characteristics of the images on which we tested the classification
algorithms:

59
 Original image (809340 pixels,
GSD=12cm, Total area=11654
m2)

Downsampled image 2 Downsampled image 2 times smaller 4 times smaller

times (202540 pixels, 4 times (50635 image(202540pixels, image (50635pixels,
GSD=24cm, pixels, GSD=48cm, GSD=12cm, area= GSD=12cm, area=
Area=11654 m2) Area=11654 m2) 2916.6m2) 1458 m2)

Due to water absorption, here we use 189 of 224 bands. The deleted bands are between 1353 nm
& 1443nm and between 1812 & 1958 nm.

4.2.2. Classification Results

In this section, we show the results obtained after we applied the following classification
algorithms:
 SAM: Spectral Angular Mapper
 OSP: Orthogonal Subspace Projection
 ACE: Adaptive Coherence Estimation
 CEM: Constrained Energy Minimization
 SID: Spectral Information Divergence
 FCLS: Fully Constrained Least Square
 AMSD: Adaptive Matched Subspace Detector
 MF: Matched Filter
 HUD: Hybrid Unstructured Detector

Table 2 contains the FAR obtained after applying different algorithms on the 5 images. The
computation times in each case are registered in Table 3.

60
 Table 3: FAR (nb of false alarms/m2)

Original D2 D4 S1 S2
CEM 0 0 0.0049 0 0
MF 0 0 0.0058 0 0
ACE 0 0 0.0003 0 0.0048
OSP 0.001 2.3236 2.4261 0.0014 0
SID 0 3.6867 2.7836 0 0
AMSD 0 16.5982 4.151 0 0
FCLS 0.443 12.0172 3.6171 0.2859 0.6076
HUD 0.3782 0.9234 3.738 0.5119 0.4786
SAM 0 4.8432 3.1551 0 0

Table 4: Computation time in seconds

Original D2 D4 S1 S2
CEM 19.8 4.62 2.22 4.8 2.17
MF 23 5.12 2.37 5.7 2.1
ACE 28.8 6.63 2.65 7 2.5
OSP 19 4.54 2.1 5 2
SID 27.2 6.5 2.9 7 2.5
AMSD 27.3 6.83 2.72 6.8 2.4
FCLS 177 19.66 6.43 42 11.2
HUD 152 38 11.6 39 10.4
SAM 23 5.15 2.3 5 1.68

As we can notice from the tables, ACE, MF and CEM show the best performance for detecting
the landmines in the hyperspectral images because even when the image was spatially

61
downsampled by a factor of 2, the FAR remained zero at full detection. When the image was
downsampled by a factor of 4, few false alarms appeared whereas for the other algorithms the FAR
is too high to consider the detection as useful. Moreover, the computation times of these algorithms
are acceptable and are lower than other algorithms.
As we have seen in section 2, the coefficients of the filter in CEM and MF methods are very
similar. They differ in using the correlation matrix or the covariance matrix. This explains why the
computation time and the FAR obtained when applying CEM and MF algorithms are very close.
In addition, we can see that the OSP algorithm, which is based on linear unmixing model, is
sensitive to the target abundance. The detection was good in case of full-pixel target but the FAR
increased significantly in case of subpixel target in D2 and D4 images. AMSD performance shows
the same behavior, but in addition to its sensitivity to target abundance, the computation time is a
bit higher.
Spectral information divergence (SID) and spectral angular mapper (SAM) belongs to the same
family of detectors as both measure the difference between the target and the pixel. Globally, SAM
has higher FAR than SID but has a lower computation time. A comparison between both
techniques could be found in [95].
Fully constrained least square (FCLS) algorithm is used to calculate the abundances of the
background and the target at each pixel. Therefore, it takes the reflectance spectra of the
background materials and of the target as input and calculates the abundance of each component
in every pixel of the image taking into consideration the non-negative and sum-to-one constraints
of the abundances. It is a complex process, which explains the long computation time. The high
FAR obtained using this method demonstrates that the estimated abundance of the target could not
be used alone as a decision metric of the presence of the target. As the background and target
spectra are used in FCLS processing, the detection results depend on the quality of the input spectra
and the number of background materials used in the input.
Finally, The two-step detection process of the hybrid unstructured detector (HUD) explains the
high computation time. The high false alarm rates may be due to errors in estimating the
abundances as in the case of FCLS.

4.2.3. Discussion

Several approaches have been proposed for target detection using hyperspectral imaging. Some
of these approaches are based on linear mixing model where the reflectance of each pixel is made
of mixing the endmembers’ reflectance spectra in different abundances with additional white
noise. OSP, FCLS are algorithms based on this approach. However, the detection performance of
these algorithms is too sensitive to the choice of the endmembers. If the number or the type of the
endmembers was wrongly chosen, the detection will be difficult.
Another approach simulates the spectral variability of the targets and background materials using
statistical models like MF and ACE. This approach proves its efficiency in detecting the mines at
subpixel level in an acceptable computation time.

62
MF and CEM methods do not require other information than the reflectance spectrum of the target.
This makes the detection of the target faster and simpler but makes it dependent on the precision
of the spectrum used in the search. When the spectrum of the target in the acquired scene is
different from the spectrum that we are looking for, the detection performance get worse. This may
occur due to differences in weather and illumination conditions at the moment of registering the
spectral response of the landmine (maybe taken in lab conditions) and at the moment of image
acquisition.

4.2.4. Conclusions

In this study, we arrived to the preliminary result that the CEM, MF and ACE algorithms are three
of the best algorithms to be used when trying to detect landmines using hyperspectral imagery.
This result is in agreement with the results obtained in the hyperspectral target detection tests [107]
that proofs the effectiveness of ACE in target detection. Linear mixing model based algorithms
depend on the definition of the endmembers and the fill fraction of the target. The definition of the
endmember may differ between images and between users changing the detection results. In the
future, we will try to improve the detection of the algorithms in the case of a multi-target scenario
as we will see in chapter 4.6.

4.3. Effect of PCA Feature Selection Prior To Detection

In another experiment, we tested the effect of dimensionality reduction prior to detection on the
classification performance. Here we used the Principal component analysis (PCA) to choose the
most representative bands out the 224 bands of the image and then we applied different detection
methods. Our goal is to evaluate the consistency of the detection algorithms if less information is
used in the detection.

4.3.1. Data Description

Here we use a part of the same AVIRIS image scene named f100902t01p00r03 used in the previous
test. However, in this experiment we chose another part of the scene where the main background
material is sand. The size of the chosen part is 588 samples and 1430 lines (238.4x262.4 m). Here
we chose a larger area in order the see the effect in computation time. As in the previous test, the
image is pretreated before implanting the mines. First, atmospheric correction is applied to
transform it from radiance into reflectance domain. Then, the image is upsampled in order to obtain
pixel size equal to the size of landmine. After that, the bands corresponding to water absorption
bands are discarded. In the resultant image composed of 189 bands, we implant the spectral
reflectance of landmines. Then, we made another copy of the image reduced to 100 bands with the
use of Principal Component Analysis algorithm. By applying the PCA algorithm, we are taking
the bands that have the largest variability. Thus, these bands contains most of the information in
the hypercube.

63
The same algorithms used in the previous chapter were applied in this test. We applied them on
the reduced and normal image. Also in this test, we chose the threshold to discriminate between
target ad background so as to detect all planted mines, and then we registered the False alarms and
computation time. The results are shown in the next section

4.3.2. Results

In this section we present the results obtained when the classification algorithms are applied on
the original and reduced image using PCA. the tests were done on Windows server with the
following characteristics: CPU quad Core 2.9 GHz, 32 GB RAM and 1TB Memory.
The variation of FAR when using each algorithm is shown in fig 14 and the Computation time in
Fig 16:

Effect of the use of Principal Component Transform

on False Alarm Rate
(nb of pixels=6255616)
350000
300000 300000
300000
243138
250000
200000
150000
100000
46781
50000
0 0 0 0 0 0 2853 0 11617 0 0 0
0
OSP ACE CEM SID FCLS AMSD MF HUD
224 bands 100 bands chosen using PCT

Figure 16: Effect of PCA on FAR

64
 Effect of the use of Principal Component
Transform on computation time(s) (nb of
pixels=6255616)
5000 4449
3879
4000 3694
3208
3000
2000
1000 699 482301
244294 360 240 305 257371 428359
0
OSP ACE CEM SID FCLS AMSD MF HUD
224 bands 100 bands chosen using PCT

Figure 17: Effect of PCA on computation time

As we see in the charts, after dimensionality reduction using PCA, the performance of ACE, MF
and CEM did not change in terms of FAR. Zero FAR rate is obtained in both cases. However, in
case of OSP, AMSD and HUD too many false alarms show up after the size of the image is
reduced. In case of SID and FCLS, the number of false alarms became too high to consider the
detection as effective. So in these algorithms, we lost the information useful to distinguish the
targets from background. To note that the false alarm here is higher than the 300000 limit that we
considered as maximum value to consider the detection as useful. 300000 is almost half of number
of pixels of the image. An algorithm is not effective if half of the field is considered as landmine,
because it is not a realistic result. Usually landmines are rare in the scene. In addition, in the case
of landmine detection, each false alarm will require about one hour of work to take the necessary
precautions before starting the deactivation process. Therefore, if we have a high FAR, the wasted
time is too long making the detection using other preliminary techniques more effective.
Talking about the computation time, here we register the computation time needed to perform both
the dimensionality reduction and the classification. We see that the computation time is reduced
about 48.5% in case of ACE and is reduced about 37.5% in case of MF. In both cases, the FAR
remained zero after the use of PCA. However, in case of CEM, the total time to perform PCA and
apply CEM on reduced image is higher than the time needed to compute the detection on the
original image. This phenomena is repeated with OSP, SID and FCLS however using these
algorithms, the FAR became higher after using PCA. So reducing the size of the data prior to use
these algorithms is not beneficial in terms of both FAR and computation time.

65
 4.3.3. Conclusions

As in the previous test, ACE and MF showed the best performance. Even if the image was reduced
to 100 bands we were still able to detect all landmines with 0 FAR. In addition we gained some
time in the computation time. This performance improvement do not apply to CEM: even though
the FAR obtained is zero in case of reduced image, the computation time in the two steps
(reduction+classification) algorithm is higher than when applying the detection on all image.
Spectral unmixing based methods are too sensitive to the dimension of data used. As we see here,
when we applied the OSP AMSD and HUD on reduced image, the FAR obtained was too high.
So the use of these algorithms on reduced images won’t be effective.
If we would like to use a dimensionality reduction method prior to classification in a real target
detection scenario, as on board of a quadrotor while acquiring the hyperspectral image, we should
verify if the use of reduction method is useful. Because using some methods, we may have
increased computational time as in case of OSP and we may have large FAR.

4.4. Effect of spectral variability on landmine detection

In this part, we will test the same algorithms but on images in which the mines were planted with
different errors. Here we would like to test the possibility to detect landmines if their spectral
signature in the hyperspectral image is not exactly the same signature that we have in a library.
This very usual case occurs when the mine is covered by another background material like sand,
soil or vegetation, or happens in the case of low spatial resolution image. So the reflectance
spectrum in the pixel where the mine exists is a mixture of the signature spectrum of landmine and
other background material.

4.4.1. Data description

In this test, we use a small part of AVIRIS image named f100902t01p00r03. The chosen part
contains mainly vegetation and soil. The size of the chosen area is 494x410 pixels (59.28 x 49.2
m). In different locations of the image, we planted the spectrum of the PMN landmine mixed the
spectrum of green leaf ( vegetation) that is the background material most dominant in the scene.
We mixed the target and the background material in different proportions: 0.5 PMN+0.5 leaf; 0.6
PMN+0.4 leaf; 0.8 PMN+0.2 leaf. In the following section we show the results.

4.4.2. Results

In this paragraph, we will show the results obtained when we tried to detect the PMN signature in
the hyperspectral image in which the PMN was planted in different proportions. As in the previous
tests, the threshold was set in such a way as to detect all landmines and then we registered the
number of false alarms and the computation time.

66
In this test, the graph of computation time do not have any differences between images of different
abundance factor because the images have the same number of pixels. But it helps to compare
between different detection algorithms.
In consistency with previous tests, even in this test, CEM ACE and MF show the best performance
as using these algorithms, we are still able to detect landmines with 0 FAR even when the
abundance factor of target is 0.5. In addition, AMSD show the same performance in this test.

67
In OSP case, even when the abundance factor of landmine is 0.8, we did not detect the landmine
without false alarms. The number of false alarms increases if we are searching for mines with
lower abundance. The same could be deduced in case of HUD algorithm. This may be due to
linear unmixing step in which the abundance of targets was wrongly estimated.
By using the FCLS, the computation time is too high by comparison with other algorithms. In
addition the FAR is high. Therefore this method is not preferred for landmine detection case.
SID and SAM methods are similar as they both compare between target signature and pixel
signature based on similarity measure. SAM uses the angular distance between the spectra while
SID measures the mutual entropy between them. However, these methods are sensitive to target
abundance. Because of that, we see 0 FAR in case of pure target or target with abundance factor
of 0.8, but the detection become harder and too many false alarms will show up in case we try to
detect the landmine with abundance factor of 0.6 or 0.5.

4.4.3. Conclusions

In this test, we arrived to almost same result of previous tests that ACE CEM and MF are some
of the best algorithms to detect landmines with different abundance factors.
The linear unmixing based methods like OSP do detect the landmines but a right definition of
background endmembers is necessary to reduce the false alarm rate or to detect landmines with
low abundance factor. We can say the same conclusion for the hybrid detector (HUD) where the
first step of its computation is based on linear unmixing model.
SAM, SID or other similarity or distance calculation methods are not reliable for landmine
detection, because if the abundance of target is 0.6 or less, the detection of targets is possible but
we will have several false alarms. This is not practical especially in case of landmines where each
false alarm will take at least one hour of precautions and land preparations.

4.5. MLP Neural network for landmine detection using

hyperspectral imaging

In this chapter, we present the tests done in order to detect landmines in hyperspectral images
using MLP neural networks. In order to have good performance (reduce the FAR at full detection),
we did several tests. In each test, we change some factors that affects the results. These factors are:
 Training data set
 Number of neurons in the hidden layer
 activation function of the neurons
 Error minimization strategy used in the training phase.

68
In the following, we will present a brief introduction about the multi-layer perceptron used in this
experiment. Next to it, we will show the results when applied on 17 images and finally the
conclusions are provided.

4.5.1. Multi-Layer Perceptron (MLP) Neural networks

The general structure of MLP NN is presented in section 11.1. It is a kind of feed-forward neural
networks constituted of at least three layers: Input layer, Hidden layer and output layer. Feed-
forward neural networks provide a general framework for representing non-linear functional
mappings between a set of input variables and a set of output variables [126]. The network may
have an arbitrary number of hidden layers, which in turn may have an arbitrary number of
perceptrons [127].

Figure 18: Example of multilayer perceptron NN

The activation function of the hidden layer could be linear, hyperbolic tangent, sigmoid or other.
Usually sigmoid activation function is preferred as the output of the perceptron are limited between
zero and one so they can be considered as probabilistic values. To train a MLP NN, usually
supervised learning is used. The error function to be minimized by backpropagation in the learning
process tested here are the sum-of-square errors and cross-entropy. In general, we got a better
performance in case we used cross entropy error function.

4.5.2. MLP training and application

In order to train the MLP NN, we used a part of AVIRIS hyperspectral image composed of 280
rows, 150 columns that contain green leaf and sand in the scene. We use the most useful 189 bands
out of 224 bands after excluding the water absorption bands. In addition, we test the trained NN
on the same 17 images used in chapter 4 section 6.

69
 Here in follow, to give a nomenclature of the results, the training image used in this test will be
named TI. The 17 images on which the resultant neural networks are tested are named according
to the number of the field number in our data set. The 17 images are named respectively: field2,
field3, field4, field51, field52, field53, field54, field55, field56, field61, field62, field63, field64,
field71, field72, field73, field10. We used 17 images that have different background materials in
order to test the performance of the neural networks in different case studies.
In this experiment, the simulation are performed using MATLAB simulation tool. I referred to
NETLAB toolbox [128]. It contains a group of functions prepared to simplify setting up the
network parameters.
In the first test, we trained a NN using data from TI. The input is composed of 189 neurons
corresponding to the reflectance bands. The size of the output is 3: 100 for PMN, 010 for M20 and
001 for background. The hidden layer is composed of 115 neurons. The training dataset is
composed of 200 entry referring to PMN reflectance spectrum, 100 entry referring to M20 and 200
background pixels from TI. The network is composed of 2 hidden layers. This NN is named
“netTI”. When applied on TI, all mines in the training image are detected with 0 FAR. When we
used this network to detect landmines in the 17 images, all mines in all images were detected.
However, in some images some FA appeared. The average FAR obtained is 0.121/m2 the detection
was done in 3.64seconds.
In the second test, I used the data from images field3 field4 field51. The dataset were composed
of all pixels of these images including the pixels were the landmines are inserted. This NN is
named “netfield3451”. When applied on all fields, the obtained Pd was 0.85 with FAR 1.385/m2.
the computational time is on average 6.5 s.
In the third test, in an attempt to increase the Pd and decrease the FAR, I used for training the
images field2 field3 field4 field51. The input for training was the reflectance spectrum of each
pixel in these images. This NN is named “netfield23451”. By comparison with the previous test,
the probability of detection has increased to 0.94, the FAR has also increased to 7.3074/m2. the
computation time was 6.65s.
In a fourth test, I trained a NN using the data of images field4 field71 field72 field73 field9 field10.
The number of neurons in the hidden layer is fixed to 115. This network is named
“net4717273910withoutpmn”. When this NN is applied on all fields, the average Pd obtained is
near 1 (0.99) but the FAR was High (29.31 / m2).
Then I build another NN in an attempt to reduce the FAR using the pixels of field4 field71 field72
field73 field9 field10 in the training phase. In addition, I introduced in the training phase 500
replica of the spectrum of PMN. The number of neurons in the hidden layer is 115. The trained
NN is named “net4717273910”. This strategy didn’t work well. The probability of detection was
near zero and the FAR so. So introducing the spectra of the targets several times will not improve
the detection. After that, I trained another NN with the same data of the same fields with addition
replica of PMN landmines spectra with abundance factor 0.9 mixed with other background
material. This NN is named “net4717273910with9”. The results are similar to the results of
previous neural networks where the Pd and FAR obtained is near zero.

70
 In another test, I trained the data of the fields that usually had high Pd and low FAR in other tests.
So in this test I trained a NN using the data of field2 field51 field52 field53 field56 field62. The
NN is named “net25152535662”. When applied on all fields, the average Pd and FAR were close
to zero.
As I found that training data of TI give better results than training few images from the test images,
I choose to use it in the training. This time I trained a NN named “n1” using 20 spectra of PMN
mine, 10 spectra of M20 mine, and 20000 spectra of background from TI image. The number of
hidden neurons was 2. Using this NN, we were able to detect all landmines in all fields. But in
some images we got some false alarms. the average FAR obtained is 0.464. The detection is done
in an average of 3.8 seconds.
To reduce the FAR obtained when using “n1” NN, I used the same data used in the previous test.
The training set is composed of 20 spectra of PMN mine, 10 spectra of M20 mine, 20000 spectra
of background materials in addition to 150 pixel spectra of background road from image field3
usually marked as FA this time is included in the training. To note that the new samples included
in the training of “n2” and successive NN are not used for testing the NN performance. When “n2”
is used to detect the landmines in all fields, we detected them all landmines. The FAR is reduced
to 0.142/m2. The average computational time is 3.36. So including some FA in training may help
to reduce the FAR. The same NN named “n2” was applied on all images but this time the fields
contains mines with different abundances. The results are showed under name n2_2. We got the
same FAR as in case of n2 but the Pd dropped to 0.35.
To improve the results of “n2”, I took the pixels that were marked as false alarms using n2 and
add them to the training sample. So the training set of “n3” is composed of 20 spectra of PMN, 10
spectra of m20, 20000 background spectra from TI and pixels marked as FA when using “n2”.
When I applied this NN on all fields, all mines are detected in 4.18 seconds. The average FAR
obtained is 0.609. So the FAR did not decrease as expected.
After that, instead of using all false alarms found in case of n2, I used in the training of “n4” the
data used to train “n2” in addition to FA obtained with fields4 52 53 54 and 55. In this case, I got
Pd=1with FAR=0.397/m2 in 3.673s. So in this case, the FA is reduced by comparison with the
results of “n2”.
Finally, I used in the training of new NN named “n5” the same data used to train “n2” in addition
to pixels marked as FA in case of when “n2” was applied of field54 field55 field72 field9. when
“n5” is applied on all fields, all mines are detected in 3.68s with slightly high FAR of 3.797/ m2.
In a new series of experiments, we first applied a feature extraction method to reduce the size of
the image prior to train the MLP NN. Here I used the Net analyte signal presented in section 11.1
to choose the best bands that represent the landmines. In the first test, I chose 20 bands using this
method. then I trained a MLP NN names “net_nas20” using the data in the image field3. This NN
when applied on all fields we got low Pd=0.12 with FAR of 0.209 in 0.58 seconds. So using this
method, the computational time is reduced but the detection performance has decreased.

71
Then I trained another NN this time using 50 bands chosen using Net analyte signal. I used the
data in field2 in the training. The resultant NN is named “net_nas50_f2”. This network when
applied on all fields, we detected all mines but with very high FAR of 65.74. To decrease the FAR,
I used the same 50 bandswidth a training data set composed of the pixels of the fields field2 field3
field4. When I applied this NN named “net_nas50_f234” on all fields, all mines were detected
with FAR of 6.214.
The charts that resume all results obtained are shown in the following figures:

Figure 19: Average Probability of detection

72
Figure 20: Average False Alarm Rate

Figure 21: Average computational time

73
 4.5.3. Conclusions

Referring to all tests done in this experiment, we can deduce the following:
 Repeating the spectra of the targets in the training sample is not necessary as we obtained
similar performance when using n2 and netTI.
 The training should be rich and included almost all cases in which a landmine could be in
the hyperspectral scenes in order to be detected using MLPNN.
 The use of few neurons in the hidden layer with more representative data is more effective
than using too many neurons with intensive Training samples. this is concluded when
comparing between 'net4717273910withoutpmn' that have 115 neurons and ‘n2’.
 When the landmines in the image have different abundance factors, the training sample
used to train the MLP NN must include several samples of targets with low abundance
factor. Without this, the Probability of detection won’t be sufficient as happened in case of
n2_2.

4.6. Multi Target Detection Using Neural Networks

In this chapter, we evaluate different classification algorithms used for multitarget detection using
hyperspectral imaging. We take into consideration different scenarios of landmine detection in
which we compare the performance of each method in various cases. In addition, we introduce the
detection of targets using artificial intelligence based methods in order to increase the probability
of detection, to reduce the false alarm rate and to foster the detection. These algorithms were tested
on simulated data where the spectra of landmines is planted in different proportions with respect
to the pixel size in a hyperspectral image scene. We retested these algorithms on real image with
real targets. The results show that we can use a well-trained radial basis function (RBF) neural
network in order to detect targets using hyperspectral imagery.
Several algorithms have been proposed for target detection in hyperspectral imagery. Some of
them are mentioned in the fifth section of chapter 5. Most of them do not support multitarget
detection unless we run them several times each run for a specific target. However, this will be a
time-consuming process especially if the number of targets is high. Some algorithms were
extended for multitarget case e.g the Constrained Energy minimization (CEM) algorithm
originally made to give an estimation of the abundance of the target, has several extension to fit
the multi target detection: multiCEM, SumCEM, Winner-take-all CEM (WTACEM) and
others [108]. Other unsupervised algorithms may be used to detect targets without referring to
their reflectance spectrum [109]. But it has been proved that this type of algorithms usually have
high False alarm rates as some inert low frequency pixels may be marked as targets while they are
not.

74
In this chapter, we test different supervised classification algorithms used for multitarget detection
in a landmine detection scenario and show the possibility of detecting targets using artificial
intelligence based techniques. A comparison of the results will be discussed. The types of tests
will be carried out: the first one uses images where the targets have been spectrally added to an
AVIRIS image while the second one uses real images containing manmade targets.

4.6.1. Neural Networks based Target detection

In this section, we will introduce the use of artificial intelligence in order to detect targets in
hyperspectral imagery. Specifically, we will work on neural networks (NN). This approach is
adopted due to several reasons:
First, to construct a neural network classifier, two phases are needed: training and classification.
The training phase could be done offline and then the detection is achieved online during image
acquisition. Therefore, this method may be optimized for real time detection as most of the
workload is done offline in the training phase. Secondly, we can customize the detector to detect
a large number of targets in one scan, which means that the detection of several targets is fast and
requires only one scan. In addition, the network could be customized for different types of
backgrounds i.e. we could have several trained neural networks, each for different types of scenes
(background, water, sand or forest); therefore, we reduce the FAR by taking the combination of
results of several NN.
Artificial Neural Network (ANN) is a computational model used for various machine learning and
computer vision tasks. It is designed to work in the same way as the neural networks of the human
brain work [110]. It is composed of a network of connected units called “neurons” where each
connection has a weight. The neurons are grouped into layers. In addition to the weights, each
layer has a bias that plays a crucial role in the detection [111]. A basic NN is composed of two
layers: input layer and output layer. This type of NN is called Single layer NN. Other type of NN
may have additional hidden layers between the input and output layers. In this category, we can
find the Multi Layer Perceptron. This kind of ANN has the ability to solve nonlinear complex
problems that the single layer NN will not be able to solve [111].
Another type of neural networks is the Radial basis functions neural networks. It has the same
structure of layers as the MLP. However, in the hidden layer, the activation function is a kernel
function (usually Gaussian) [112]. Usually, MLP NN are faster than RBF NN as their computation
do not necessitate the use of kernels and therefore is simpler. However, in case of high dimensional
data, as in our case where the pixel is of 189 band dimensions, the RBF performs better. RBF
showed better performance in our case and thus we will adopt this method in the comparison.
Here, we used two-layers RBF neural networks. The activation function of the first layer (hidden
layer) is Gaussian. The activation function of the output layer is linear. The number of neurons is
empirically estimated to minimize the global problem.

75
 Input layer Hidden layer Output layer

Bias Bias
1 1
O1
w0
w1
I1 w2 O2

On
Ip

Figure 22: Multi-layer RBF Neural network

Several studies introduced the use of deep learning neural networks for target detection using
hyperspectral imaging [113]. However, this will not be our work in this study. Here we focus on
ordinary neural networks due to several reasons: First, the main objective of our test is the detection
of landmines at subpixel level, therefore extracting some features from a window of pixels as in
the preliminary step of convolutional neural networks will make the detection harder. Secondly,
we are dealing in this paper with hyperspectral images where each pixel is composed of hundreds
of bands. So by mixing several bands we loose some spectral information that are necessary in the
detection. In addition we are searching for a simple solution to make the detection faster.
In order to reduce the size of the neural network, our first step will be the feature mapping. In this
stage, some key features of the hypercube will be chosen in such a way to reduce the size of the
input image and rely on useful information. For this objective, there are several methods that could
be used: Principal Component Analysis (PCA), matched pursuit [114], neighborhood embedding
Error! Reference source not found., Sammon’s mapping [115], multicriteria method [117],
nonparametric weighted feature extraction [118] linear discriminant analysis (LDA) and
others [119]. In this experiment, we use the concept of Net analyte signal introduced by Lorber
1986 in order to specify the unique part of an analyte signal in chemical spectrum analysis. The
idea is to find the part of the signal that belongs to the orthogonal plane of all materials other than
the target. By this, we choose the most representative bands of the target. These bands will be used
as input to the neural networks in order to detect the targets instead of detecting the complete signal
spectrum.
The chosen bands are calculated as follows [120]

nj =( I -S-j (ST-j S-j )-1 ST-j ) sj

where sj is the target spectrum, S-j is a matrix of background analyte spectra and nj is the portion
of sj that is orthogonal to S-j. First we estimate all endmembers spectra of the image using

76
 Automatic Target Generation Process (ATGP) algorithm [121], S-j is obtained after removing the
endmembers corresponding to the targets.

4.6.2. Experiment on simulated data

4.6.2.1. Data description

In the first scenario, we tested the target detection algorithms on 17 hyperspectral images taken
using Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) of JPL NASA Laboratory. These
scenes are available online on the site [122]. The spatial resolution of the images depends on the

Figure 24: Reflectance spectrum of the pmn mine (target) Figure 23: Reflectance spectrum of the vs-2.2 mine (target)
inserted in the image inserted in the image

altitude of the airplane during image acquisition. We can find different scenes of different spatial
resolutions. In the chosen scenes, we introduced in different locations the spectrum of manmade
targets that will be PMN landmine (Fig.24) and VS-2.2 mine (Fig.25). The reflectance spectra of
the landmines were taken in our Lab using Field Spec 4 Hi-Res spectroradiometer. This device is
able to acquire the reflectance spectrum between 350 and 2500 nm with spectral resolution of 1nm.
We took the spectral signature in different conditions: in lab where specific source of light is used,
in grass field and in soil field during a sunny day. Here we plant the spectral reflectance taken
when thin layer of grass covered the landmine in the AVIRIS scenes. The insertion was done after
several image-preprocessing steps: firstly, atmospheric correction is done to convert the image
from radiance domain that depends on the illumination and weather conditions into unified
reflectance domain scaled between 0 and 1. Some bands characterized by low SNR due to vapor
absorption are discarded. Then the image is up sampled in order to increase the spatial resolution
of the image to arrive to pixel size equivalent to the size of the mine.

77
 In order to test the full pixel and subpixel cases, the targets were planted in different proportions
in the images. The signatures of the targets were mixed with the neighbor pixel signatures in
different fill fractions: 𝑃𝑆 = 𝛼 ∗ 𝑇 + (1 − 𝛼)𝐵 where PS represents the planted spectrum in the
image, T is a vector containing the target reflectance spectrum, B the background reflectance
spectrum and 𝛼 the target fill fraction varying between 0.6 and 0.9. In the 17 images, the total
number of pixels with landmine abundance factor of 𝛼 = 0.6 (PMN& VS-2.2) is 136, 170 have
landmine abundance factor 𝛼 = 0.7, 102 have landmine abundance factor 𝛼 = 0.8 and 110 pixels
have landmine abundance factor 𝛼 = 0.9 . By this, we evaluate the ability of a target detection
technique to detect subpixel targets.
As the target in this scenario is a landmine, the risk of missing a target is much dangerous than
having a FAR. Therefore, the decision threshold to discriminate between target and background
material is set such a way to detect all targets (Pd=1) and then the FAR is registered. Therefore, a
technique is said to be more efficient if it has lower FAR giving that all targets have been detected.
4.6.2.2. Results

In this part, we show the results obtained when applying the detection techniques on all images
that contain targets in different abundances. We show the average false alarm rate and computation
time at full target detection. The tests where done on Dell server with 64 cores, 128 GB RAM and
1TB Memory.
The tested algorithms are the following:
 SAM: Spectral Angular Mapper
 OSP: Orthogonal Subspace Projection
 ACE: Adaptive Coherence Estimation
 CEM: Constrained Energy Minimization
 MTCEM: Multiple target CEM
 WTACEM & SCEM:Winner take all CEM and Sum CEM
 SID: Spectral Information Divergence
 MF: Matched Filter
 RBF NN: Radial basis function Neural Network
In case of neural networks, the best NN in terms of false alarm rate was individuated after several
tests where we took into consideration different training samples and spread values. First, we
randomly divided the 17 images between training and testing data where we used some images in
order to train the NN and the other images to evaluate the performance. Using this strategy, the
training was very intensive process, took a long time, necessitates large number of neurons to
consider all possible cases and we did not arrive to zero FAR. To make sure that we are training
the useful data without repetition, we decided to use another strategy. We found that training few
pixels that represent the image endmembers is sufficient to obtain a NN able to estimate the
abundance of targets and background in each pixel. The input training dataset is the background

78
reflectance spectra automatically estimated using Automatic Target Generation Process (ATGP)
algorithm [121]. The PMN reflectance spectra and the VS-2.2 reflectance spectra. The training
data is composed as follows: 377 spectra represents various background materials, 5 spectra of
PMN landmine and 5 spectra of VS-2.2 landmine. The corresponding output are respectively:
001,100 & 010. To note that the reflectance spectrum of the targets exist with different fill fraction
in the scene (0.6 0.7 0.8 &0.9). However, in the training phase, the pure reflectance spectrum of
the target is introduced. Using this training strategy, the output for each pixel will be abundance
fraction of PMN, VS-2.2 or general background.

The following charts represent the results of the adopted methods:

Figure 25: Average computational time /algorithm

79
Figure 26: average FAR/ algorithm

In figure 25, we see the average time needed for each algorithm to detect all the targets in the 17
images. As we see in the figure, to detect both types of targets using ACE we needed about 349
seconds while the computation time of MF is 250.77 s and is 238 s for CEM. About the same time
is needed using SCEM and WTACEM as these algorithms are based on running the same detector
2 times each run to detect one target. They differ in the decision making step as follows: in case
of CEM we set a threshold for each target; in case of SCEM, we add the outputs of the detectors
and set one threshold for the sum; or we take the maximum of the outputs and set the threshold
accordingly as in WTACEM. SAM and SID are faster than other algorithms but they have very
high FAR as we see in Fig 26.
Talking about the FAR, we see that almost all algorithms could detect the targets with very low
FAR except for SID and SAM that have high FAR. Both algorithms are based on comparing the
spectrum of pixels with the target’s spectrum both treated as vectors. Thus, they depend on how
similar the pixel is to the target. In case of target with abundance 0.6, they are too much different
causing this high false alarm rate.
The other algorithms show very good performance even in case of small abundance factor where
few FA shows up when trying to detect low abundance targets using MF and CEM. ACE algorithm
gives the ability to detect all targets with 0 FAR. This confirms the previous tests used for target
detection [107],[123].
It is worth to note that when applying CEM 2 times and setting a threshold for each target, we got
false alarm of 0.00027. Most of the false alarms refer to VS-2.2 targets. However, when we took
the sum of the results or their maximum as in SCEM and WTACEM, no more false alarms are
obtained. This is due to the increased contrast between targets and background in case of Sum
CEM or by ignoring noise effect while taking most valuable results in case of WTACEM.
On the other hand, MTCEM has better performance as all targets are detected without any false
alarms with lower computation time. It should be pointed out that, using ACE MF CEM SID SAM

80
and OSP, we have an additional advantage when identifying the targets since using these
algorithms we are able to distinguish between PMN and VS-2.2 targets. While using the other
algorithms, we can know the presence of a target without knowing its type. This type of
information is crucial in some target detection tasks, especially in case of landmine detection in
order to determine the best strategy to isolate the landmine according to its blast and fuse type.
However, this ability comes at an additional cost in terms of time and/or computational resources.
As we see in the charts, using the adopted training strategy, we got an RBF NN able to detect the
landmines without any false alarm. By setting a large value of spread while training the NN, the
output was less sensitive to the spectral variability of the input pixel and able to distinguish the
presence of target even with low abundance factor.
On the other side, the computation time needed to get this result is lower than ACE that has also 0
FAR, but is higher than other multitarget detection algorithms MTCEM, SCEM and WTACEM.
However, using RBFNN, we are able to distinguish between targets whereas in these algorithms
we are not.

4.6.3. Real target experiment

In this section, we show the results obtained when we applied previously mentioned algorithms in
order to detect targets in real hyperspectral images. This is done in order to prove the applicability
of these algorithms in real case scenarios. Even if it has been proven in [124]that the target implant
method does provide accurate relative predictions in terms of both target difficulty and detector
performance, but reliably predicting the actual number of false alarms for a given target at a given
fill fraction is difficult or impossible [124].
4.6.3.1. Test Image

Target detection algorithms tested in previous sections will be applied on the hyperspectral data
collected over Viareggio city, Tuscany, Italy by Centro Interforze Studi e Applicazioni Military
(CISAM) in collaboration with university of Pisa [125]. The image that we worked on is named
D1_F12_H1 and contains 5 different targets (2 panels and 3 vehicles). The data has spectral
resolution of about 1.2 nm between 400 nm and 1000 nm with 0.6 m spatial resolution. The targets
are as follows: two green colored panels made of carton named P1 and P2, Ford fiesta car named
V1, a FIAT DUCATO mini commercial vehicle named V3 and a Ford Focus car named V4. These
targets are located in different positions of the image scene. The positons are given in a ROI file
for performance evaluation. The spectra of the targets are also given. The first step we did is to
convert the image from radiance into reflectance. Then, some bands are chosen to reduce the error.
In case of RBF NN, we used in the training phase, the endmembers of the image automatically
extracted through the ATGP algorithm, knowing that the spectra of the targets are given in the file.
The total training sample is composed of 396 background pixels and 5 spectra corresponding to
the targets.

81
In the next paragraph, we will show the results of FAR and the computation time.
4.6.3.2. Results

Here we will show the results obtained when we applied the previously mentioned algorithms on
the image containing real targets. Note that as in the previous case, the threshold for classifying
targets is chosen in such a way in order to detect all the targets and then we compare the false
alarms obtained in each method.
Table 5: Nb of false alarms and computation time obtained when applying each algorithm

Algorithm/Target P1 P2 V1 V3 V4 Total Time (s)

ACE 0 0 0 3 0 3 63.7
MF 6 3 3 7 0 19 35.8
CEM 8 3 3 7 0 21 28
OSP 544 0 0 24 0 568 25.8
SID 43608 11309 10 12 30 54969 27
SAM 19498 111 1 46 0 19656 5.37
MTCEM NA NA NA NA NA 44 29
SumCEM NA NA NA NA NA 24 28
WTACEM NA NA NA NA NA 40 28
RBFNN 0 0 0 0 0 0 37

As we see in Table 5, similarly to the previous section, SID and SAM have a high false alarm rate
in comparison with other techniques. These algorithms are too sensitive to the spectral signature
of the targets and; therefore, will not be able to distinguish it in case of mixture with other spectra.
Thought the computation time of SAM is minimal; the high false alarm rate makes this algorithm
useless for this task.
ACE algorithm is one of the best algorithms to use in order to detect the targets because it shows
very low FAR (just 3 pixels marked as V3), but its computation time is high (63.7s) in comparison
with other techniques. The performance of CEM and MF in terms of FAR and computation time
is similar. Both have an acceptable false alarm rate with a small advantage for CEM over MF in
terms of computation time and vice versa (for MF over CEM) in terms of FA. This is due to the
similarity in the model in which they only differ in using the correlation matrix in case of CEM,
while the covariance matrix is used in case of MF.
In addition, in case of OSP, we found out that several FA appeared especially while searching for
P1 target. This is because there are three panels in the scene, so we lowered the threshold in order
to detect the third panel that exists in the scene with low abundance fraction. This caused this high
number of false alarms; however, we didn’t notice this huge change when using other algorithms.
When using the multi target versions of CEM MTCEM, SCEM and WTACEM, we couldn’t
identify the target using this type of algorithms because we were setting one threshold for the
mixture of all detectors. For this reason, the false alarms under these algorithms are marked as Non

82
Available (NA) in the table. However, an additional similarity test may help us in identifying their
type. In comparison with the results obtained in the simulated image test, we were setting one
threshold to discriminate the two types of landmines from background. Here in this image we have
5 different types of targets so we are setting the same threshold for the 5 detectors corresponding
to each target. This causes the appearance of more FA as the number of targets is higher.
When applying the proposed type of NN, we were able to detect all targets without any false alarm
for all of them. Even though the computation time is a bit high (37 s) in comparison with other
multitarget detection algorithms, however detecting all targets without false alarms is more
important for this kind of application. It is worth noting that when training the RBFNN to detect
the targets in this image, we used Gaussian activation function in the hidden layer with high spread
value to overcome the spectral variability of the targets in the scene. The results prove the
advantage of using RBFNN with the above-mentioned training strategy in order to detect several
targets using hyperspectral imagery in one scan.

4.6.4. Conclusions

In this experiment, we tested some supervised multitarget detection algorithms using hyperspectral
imagery. We tested simulated data where the reflectance spectrum of the target was planted in the
scene in different proportions and, at the same time, when applied on real hyperspectral image
with real targets.
Some of the tested algorithms (ACE, MF, CEM, OSP, SAM, SID) are designed for the detection
of one target. They were applied several times each run for each target. This may be a time
consuming process, especially if the number of targets is high. This is why we addressed the
multitarget detection in this paper. When using some of the multitarget detection process, we lose
the privilege of identifying the type of the target. However, this can be recovered by an additional
similarity test to classify each target.
SAM and SID are similarity measures between pixel signatures and target reflectance spectrum.
The former calculates the angle between the target spectrum and the pixel spectrum, while the
latter calculates the entropy between them. When the pixel spectrum is a linear mixture of target
and other background material, the similarity measure will differ according to the abundance
fraction of the target making the detection process harder. This is the reason of the high false alarm
rate that appeared in the simulated data and real image tests.
ACE is one of the best algorithms for target detection. It showed few false alarms by comparison
with other algorithms in this test and previously done tests [[107],[123]]. However, its computation
time is high. This limits the use of this algorithm in such situation where the detection should be
fast like the case of real-time detection.
Using RBFNN, we are able to detect, identify the targets and to estimate the abundance fraction
without any problem. The proposed strategy for training an RBF NN has reduced the size of the
used NN making also possible to estimate the abundance fraction of the targets. In both tests shown

83
in this test, we got a full detection rate without any false alarm rate. This was not achieved by any
of other algorithms, which proves the advantage of using NN for target detection in the proposed.
On the other hand, the computation time is a bit higher than other techniques, but it can be reduced
if the size of the NN is reduced.

4.7. Created Spectra method

In this test, we propose a new algorithm for multi target detection using hyperspectral imaging.
The objective is to detect the presence of two signatures referring to 2 types of landmines in one
scan. The idea here is instead of running the classification algorithm to detect the spectrum of the
target two times each run for each target, we run the algorithm one time for both. In this only one
run, we search for a synthetic new spectrum that represents both targets. The details about how
this spectrum is created in addition to results of tests are shown in the next paragraph.

4.7.1. Spectrum creation

As said in the introduction of this section, in this experiment we create a new spectrum out of the
reflectance spectrum of the landmines that we are searching for in order to detect several targets
in one scan. Our goal in this step is to foster the detection process to arrive in the future to real
target detection at the same time during image acquisition.
In the same AVIRIS scene used in the previous simulated data tests, here we implanted the
spectrum of 2 types of landmines: PMN anti-personnel mine and M20 anti-tank mine. The scene
is composed of 987x 820 pixels with 0.12 m spatial resolution. The total number of planted mines
30: 15 PMN mines and 15 M20 mines.
In this experiment, we will test the difference in probability of detection, false alarm rate and
computation time when applying the CEM, ACE and MF algorithms 2 times each time for each
target and when applying the detection algorithms once searching for the created spectra.
Having the image and the reflectance spectrum of landmines in 189 bands between 395 and
2500nm, to create the representative spectrum of the two target, firstly we calculate the distance
between the two vectors PMN-M20 and we sort it in increasing order. We take 5 bands where the
distance is maximum to represent the PMN mine, we take the 5 bands where the distance is
minimum to represent the m20 and we take 70 bands where the distance in near zero. The new
spectrum is composed of 80 bands. An image of the created spectra is shown in the next figure:

84
Figure 27: Created reflectance spectrum

Then we applied the classification algorithm ACE, CEM and MF in addition to MultiCEM on
reduced image to 80 bands. The results are shown in the following chart:

Comparison of performance parameters of

120 different methods
100 100 100 100 100 100 100
100

80

60 52

40 36.4
30 28
19.2 15.4
20 13.42
0 0 0 0 0 0 0
0
ACE applied ACE using CEM CEM using multiCEM Matched MF using
2 times created applied 2 the created filter created
spectra times spectra applied 2 spectra
times
Pd FAR Computation time (s)

Figure 28: Created spectrum performance

85
 As we see in the chart, in the three algorithms, the created spectra method has the same
performance in terms of Pd and FAR as if we applied the algorithms two times. However, we have
a gain of more than 57% in computation time. Using this method, we would be able to detect both
landmines in less time. To note that in case of CEM, the method of created spectrum is even faster
than the MultiCEM that is specified to detect multiple targets.
We applied the same procedure on another image 870x1330 pixel that have different background
materials. We got similar results:

Comparison of performance parameters of different

methods
120
100 96 100 100 100 100 100
100

80

60 53
42
40 28
20 22
20
0 0 0 0 0 0
0
ACE applied 2 ACE using CEM applied 2 CEM using the Matched filter MF using
times created times created applied 2 created
spectra spectra times spectra
Pd FAR Computation time (s)

Figure 29: Performance of created spectrum method when applied on another image

However, when we applied these algorithms on downsampled images by a factor of two, so the
spectrum of landmines in the image is corrupted by neighbor pixels, the detection of landmines
became harder. Too many false alarms show up making the use of this method in case of subpixel
targets inefficient.

4.7.2. Conclusions

Using the new method, the computational time has decreased up to 60% while conserving 100%
Pd and FAR=0 when applied on original image with full pixel targets. This method does not
perform well when applied on downsampled image or on in case of subpixel targets because we

86
are taking part of the information to distinguish the landmines and this part is corrupted while
downsampling.
The number of bands taken to represent each mine affect the Probability of detection. So we may
take more bands to represent each landmine in the resultant spectrum. The number of common
bands also affects the detection probability. If we use more bands to produce the created spectrum,
we may have better detection but with lower false alarms. This depends on the type of the spectra
that we are trying to mix and how much they are similar.
A study to apply this method in case of three or more targets and different types of targets must be
conducted in order to optimize the bands selection and spectra creation.

4.8. Field Experiment

In order to study the spectral characteristics of landmines and how they may change according to
the environment conditions and the type of background where they are planted, it was necessary
to collect mines, plant them and acquire their reflectance spectra using a spectrometer in different
conditions. For this purpose, we contacted the Lebanese army to get some samples of landmines
already found in the Lebanese territory that have been deactivated. After we got the necessary
permissions, the Lebanese army gave us 6 samples of 6 types of landmines. 4 are anti-personnel
mines and 2 are anti-tank mines. (see Fig. 30 )

TM-46

PMN

VS-2.2

M411 VS 50

PMD-6

Figure 30: Samples of landmines used for acquiring their reflectance spectra

87
To collect the reflectance spectra, we used a high-resolution spectroradiometer named FieldSpec
4 HI-RES. This device is made by ASD Inc. Company. It has the capability to detect the spectrum
in the range between 350 and 2500 nm. The specifications of the spectroradiometer are given in
table 5.
The device is composed of several components:
 The main radiometer in which is connected the probe via 1.5m fiber connected wire.
 Power bank battery to supply the radiometer and very useful in case of field experiment
where no power source is near.
 Computer with specialized software to control the instrument and register the data
 White board used to calibrate the device before registering the data.
There are other optional accessories. To note that the connection between the computer and the
spectrometer is wireless. This made simpler the data collection in the field.
Table 6: FieldSpec 4 Hi-res spectroradiometer specifications

Spectral Range 350-2500 nm

Spectral Resolution 3 nm @ 700 nm
8 nm @ 1400/2100 nm
Spectral sampling 1.4 nm @ 350-1000 nm
(bandwidth) 1.1 nm @ 1001-2500 nm
Scanning Time 100 milliseconds
Stray light specification VNIR 0.02%, SWIR 1 & 2 0.01%
Wavelength 0.1 nm
reproducibility
Wavelength accuracy 0.5 nm
Maximum radiance VNIR 2X Solar, SWIR 10X Solar
Channels 2151
Detectors VNIR detector (350-1000 nm): 512 element silicon array
SWIR 1 detector (1001-1800 nm): Graded Index InGaAs Photodiode,
Two Stage TE Cooled
SWIR 2 detector (1801-2500 nm): Graded Index InGaAs Photodiode,
Two Stage TE Cooled
Input 1.5 m fiber optic (25° field of view). Optional narrower field of view
fiber optics available.
Noise Equivalent VNIR 1.0 X10-9 W/cm2/nm/sr @700 nm
Radiance (NEdL) SWIR 1 1.4 X10-9 W/cm2/nm/sr @ 1400 nm
SWIR 2 2.2 X10-9 W/cm2/nm/sr @ 2100 nm

Weight 5.44 kg (12 lbs)

88
Calibrations Wavelength, absolute reflectance, radiance*, irradiance*. All
calibrations are NIST traceable. (*radiometric calibrations are
optional)
Computer Windows® 7 64-bit laptop (instrument controller)

The Remote Sensing department of the National Research Council in Lebanon owns this device.
They helped us in the acquisition of the spectra of landmines.
To test the variation of the reflectance spectra in different conditions, we acquired the reflectance
spectra in Lab, in grass field and in soil field as we will show in the next sections.

4.8.1. Reflectance spectra of landmines acquired in the lab

In a first step, in order to register the reflectance spectrum of the landmines that we have, we
acquired the spectrum in Lab conditions. By this, we mean that in a lab room, we used a specific
source of light that produces light of different wavelengths between 350 and 2500nm of the same
intensity. It is designed to produce stable output with a smooth spectral curve into the SWIR range
and to minimize backscatter and any change of lamp energy output over time [131]. The lamp
produces a well-defined beam to maximize the amount of light energy on a sample area while
minimizing stray light from surrounding surfaces [131].
The Illuminator benefits as stated by the manufacturer are [131]:
 The 70 watt quartz-tungsten-halogen light source with integrated reflector produces stable
illumination over the 350 to 2500 nm range
 Stable output yields accurate and dependable reflectance measurements
 Well-defined beam maximizes light energy on sample area
 Precise voltage regulation for high stability light output
 Multiple mount options for lab stands or tripods
The source of light is used in order to avoid the noise and other artifacts that we may face in the
field. In addition, the sunlight when reaches the soil, it will not have the same intensity at all
wavelengths due to CO2 ,water vapor and other pollutants that absorb radiations of specific
wavelengths.

This experiment was done in Scientic research center in Engineering at Lebanese University
faculty of Engineering.

89
Figure 31: Acquisition of
the reflectance spectrum
of TM-46 landmine

90
Figure 32: trying different incident angle

91
As we see in the image, to acquire the reflectance spectrum of the landmines, we used the source
of light shown in the figures. First, the detector is calibrated on the white board. The calibration is
necessary so after it is achieved, the registered spectra are automatically converted from radiance
into reflectance value according to the reflection of light on the white board.
We registered the reflectance spectra of the 6 mines. The results are as follows:

Figure 33: PMN reflectance spectrum taken in LAB

Figure 34: VS-50 reflectance spectrum taken in lab

92
Figure 35: PMD-6 reflectance spectrum taken in lab

figure 36: M411 reflectance spectrum taken in lab

93
Figure 37: VS 2.2 reflectance spectrum taken in lab

Figure 38: TM 46-reflectance spectrum taken in lab

As we can see in the images, each landmine has a specific reflectance spectrum. This proves the
utility of hyperspectral imaging technique in landmine detection because by profiting from the
spectral information in the hyperspectral images, we are able to detect the landmines and in
addition to distinguish their type.
Another thing we can notice is that metal cased landmines like the TM46 antitank mine have an
increasing reflectance value as the wavelength increases in the VNIR and SWIR domains.
However, other plastic cased mines, like VS50 M411 and PMD6 have a decreasing value in the

94
 VNIR and SWIR ranges. In the next section, we will see how these reflectance spectra will change
when landmines are planted in grass.

4.8.2. Reflectance spectra of landmines acquired in grass Field

In this part, we show the spectra of reflectance of the landmines when they were planted in grass
field. This experiment was done at Lebanese university campus-Hadath in a sunny day. The
weather was clear and sunny without clouds in the sky. Here we planted all mines in the field. In
addition, we acquired the reflectance spectrum of the grass just to compare the signature with and
without landmines.

Figure 39: AP mines planted in grass

The reflectance spectrum of the grass is as follows:

95
Figure 40: Grass reflectance spectrum

In this test, we removed the bands that corresponds to water absorption bands. In these bands, the
data collected is too noisy. Just for example, the complete reflectance spectrum of the grass
including the water absorption bands is as follows:

Figure 41: Grass reflectance spectrum including water absorption bands

It is clear that the reflectance spectrum contains some erroneous registrations as we see that the
reflectance value in some bands is higher than one. This is not possible as at maximum an object
can reflect the entire incident light (reflectance value =1). These erroneous values are registered in

96
the wavelengths that correspond to water absorption bands. So the humidity in the air causes this
error.
The reflectance spectrum of the landmines in the grass field is as follows:

Figure 42: PMN reflectance spectrum when covered by grass

Figure 43: VS50 reflectance spectrum when covered by grass

97
Figure 44: PMD 6 reflectance spectrum when covered by grass

Figure 45: M411 reflectance spectrum when covered by grass

98
Figure 46: VS 2.2 reflectance spectrum when covered by grass

Figure 47: TM-46 reflectance spectrum when covered by grass

99
As we notice in the figures, when landmines are covered with grass, the reflectance spectrum
changes and became more similar to grass. The maximum reflectance value sensed in case of pure
grass is 0.58. However, in case of landmine presence, the maximum achieved is lower than 0.5.
In the case of plastic mines, we notice that the reflectance values at high wavelength value (1500
nm- 2500 nm) are lower than the case of grass material and have similar shape to landmine
(especially in case of PMN). However, in case of metallic antitank mine (TM-46), the reflectance
values are a bit higher.
These are some key points that we noticed in this test. Many other details would help us to estimate
the abundance of landmine in the sensed pixel.

Figure 48: Four AP mines exist in this scene. Could you localize them all?

4.8.3. Reflectance spectra of landmines acquired in soil Field

During the same day when we acquired the reflectance spectra in the grass field, we acquired the
reflectance spectra of landmines in another field made of soil only and surrounded by trees of
pines. Also this test was done at Lebanese University Campus in the same weather conditions.

100
Figure 49: Holding the device on my back, we acquired the spectra of the landmines after burying them in the soil

101
In this test, we acquired the reflectance spectrum of four anti-personnel mines. In the following
some figure of the landmines buried in the soil:

Figure 50: Landmines buried in soil

The spectrum of soil without landmines is as follows:

102
Figure 51: Untouched soil reflectance spectrum

The reflectance spectra of the landmines buried in soil are as follows:

Figure 52: PMN reflectance spectrum when buried in soil

103
Figure 53: VS 50 reflectance spectrum when buried in soil

Figure 54: PMD 6 reflectance spectrum when buried in soil

104
Figure 55: M411 reflectance spectrum when buried in soil

We see in these spectra, that once the mines are covered by soil, the reflectance spectrum will have
an aspect very similar to the spectrum of inert soil without landmine. Just the small details in the
spectrum will help to distinguish the presence of landmine as untouched soil will have a spectral
response different than returned or excavated soil.
In the visible domain, the spectral response is almost identical for all types of mines and is similar
to the response of the soil. That is expected, as all mines look like soil to naked eye.
In case of M411, VS 50 and PMD 6 we arrived to maximum reflectance spectra of higher than 0.3
whilst in case of bare soil, the maximum is less than 0.3. This gives a sign of presence of another
material other than soil and led us to suspect the presence of landmines.

4.8.4. Conclusions

In this field experiment, we collected the spectral reflectance of 6 different types of landmines in
various background conditions. Some of these spectra are used in the experiment shown in chap.
4.6.

105
In real experiments, we are relying on the sun to acquire data. Sun do not produce light of same
intensity at all wavelengths. Also, there are other artifacts that may affect the detection mainly the
water absorption and CO2 absorption bands. These artifacts cause the difference in spectral
signatures of landmines when acquired in Lab or in another fields.
When the landmine is covered by another material like grass or soil, the spectral signature will
change also. The changes will depend on the proportion of background material covering the mine.
However, it has been shown that we still have some spectral characteristics that help in the
detection of landmine.
The spectral response of landmines with plastic case have different shape than the reflectance of
background material. This gives an advantage of hyperspectral imaging technique over the well-
known metal detectors in the detection of plastic mines as nowadays most landmines are made of
plastic that the metal detector is not able to detect.

106
Chapter Five
5. Conclusions and Future Work
The main contribution of this thesis is that we proved the applicability of hyperspectral imaging
to detect surface laid mines. Landmines buried by soil or in grass are still detectable but proper
classifiers must be used.
The problem of landmines is expanding worldwide. Although it is necessary to ban the use of
landmines immediately, there is a need to find a new solution able to detect landmines of different
types and shapes and is at the same time safe, fast and reliable. Hyperspectral imaging technology
is a good candidate for this purpose. The mostly used technique until now is the metal detector
thanks to its low cost. However, most of landmines nowadays are made of plastic, which made
their detection using the metal detector harder. One of the advantage of hyperspectral imaging
technique is that it detects the presence of landmines whatever the type of the case is. The
hyperspectral imagers are too expensive, but when mounted on a UAV to scan minefields on large
scale, its efficiency will be comparable to other techniques especially if the time of detection is
considered in the comparison. It is expected that the price of hyperspectral cameras decreases as
more companies fabricates this type of devices in addition to finding new technologies that makes
the fabrication of optical devices cheaper.
Every material has its special spectral signature. Therefore, knowing the mine spectral curve, by
comparison between the mine spectrum and the pixel spectrum, we can decide on the presence or
the absence of the mine at that specific position.
It has been shown in the previous tests that using VNIR band, recently buried landmines could be
detected. Also, the fusion of VNIR and SWIR could give better results. Landmine burying changes
the thermal properties of the upper level of some type of soils. It also changes its surface reflectivity
and stresses vegetation. Hence, buried landmines can be detected by measuring the change of
reflectivity both between manipulated soil and background and between stressed and unstressed
vegetation. Consequently, as anti-tank mine deployment is done by digging up a larger area of
surface (soil and/or vegetation) and a larger volume of soil is disturbed, the possibility of detecting
them is higher than with anti-personnel mines. MWIR and LWIR bands are also used to detect
buried landmines. Even if SWIR and VNIR alone could detect soil disturbances due to buried

107
mines, MWIR and LWIR can reduce the false alarm rate. However, the use of SWIR bands is more
common since the majority of manufactured imagers operates in the VNIR and SWIR bands. After
testing several hyperspectral imagers of different bands, it was found that imagers in LWIR bands
have the potential to detect buried landmines with the use of proper algorithms. The algorithms
could be supervised or unsupervised based on the data availability. Note that this does not eliminate
the possibility to detect landmines with the use of other bands. However, proper algorithms and
thresholds should be used for each case.
If we consider high spatial resolution images, which means the image has ground sample distance
close to the size of landmine, the possibility to detect a landmine is higher as the reflectance
spectrum of the pixel will result only from the reflectance of the mine, or at least the reflectance
of the landmine will be present with a high abundance. In addition, military target detection could
be achieved at subpixel level using hyperspectral images. This means that by acquiring images
from high altitude, using UAV or aircrafts, fast target detection is possible even if the target
constitutes a small part of the pixel.
In order to attain quasi real-time detection, all the processes involved, starting from geocorrection
until classification, must be studied and organized so as to reduce the computational time. Since
the detection performance will be possibly affected by some optimizations, a tradeoff between
computational time and detection performance has to be achieved.
Several factors affect the reflectance signature obtained by the imager. Wind and rain are the main
factors, but the effect of rain is the dominant one. In the case of buried landmines, rainfall decreases
the reflected portion of the thermal energy and therefore the reflectance signal received. However,
the shape of the signature remains the same. More rainfall will result in more reduction and
therefore the reflected signal will be more and more similar to the background.
The use of PCA or other feature extraction method prior to classification do not always reduce the
total computational time. Depending on the target detect algorithm used in the following step,
reducing the dimension of the data may be effective or not. For example, the computational time
have been reduced and the good detection performance have been preserved if we used PCA with
ACE algorithm. However, the performance became worse when using similarity-based detection
methods like SAM and SID.
The use of supervised detection methods is preferred over unsupervised detection techniques
because usually higher FAR is obtained in case of unsupervised techniques as low frequency
elements in the scene are marked as targets while they are not. However, supervised detection
techniques necessitates knowing the target reflectance spectra and sometimes the reflectance
spectrum of the background materials prior to detection. This kind of information is not always
present.
There are different types of supervised detection techniques. Some may be less tolerant to the
spectral variability or the abundance fraction of the target as in case of ACE, CEM and MF. Other
techniques will not be able to detect the target if the spectral signature is slightly different from
the reference spectra as in the case of SAM detector.

108
To detect several targets in one scan, we may run single target detection algorithm several times,
each run to detect one target. However, this may be time consuming and not effective for real time
detection. Other target detection algorithms supports the detection of several targets
simultaneously.
After several experiments, we proved the advantage of using neural networks in landmine
detection using hyperspectral imaging. Even if the abundance of landmines was about 0.6, an RBF
neural network trained with few background endmember data and target spectra was able to detect
the landmines, identify them and estimate their abundance. Also MLP neural networks were
examined to detect the spectra of landmines in hyperspectral scenes. The results are obtained in a
very fast computational time using this type of NN, but we were not able to detect all landmines
with 0 FAR. So MLP NN have a good potential to be used in real-time detection but proper
preparation of training data and network parameters calculation must be conducted first in a future
work.
In the field experiment, we collected valuable data of landmines reflectance spectrum in different
environment. We acquired the reflectance spectrum of 4 types of AP mines and 2 types of AT
mines. The experiment shows how much the reflectance spectrum change when taken in lab
conditions and in field situation where too many factors affects the registration. The main factors
that changes are the sun emission spectrum that is not uniform in all wavelengths, the water vapor
and CO2 in the air that absorb light in some specific bands, in addition to other artifacts and noise.
In a future work, a scan of real minefield using hyperspectral imager mounted on an UAV must
be conducted in order to validate the algorithms proposed and developed here in real case scenario.
In addition, MLP training must be optimized to obtain 0 FAR.
Also, we would collect data in new background scenarios to study the possibility to detect
landmines in situations other than we can find in Lebanon like in desert.

109
References

[1] Landmine monitor report 2015 http://www.the-

monitor.org/media/2152583/Landmine-Monitor-2015_finalpdf.pdf
[2] UNICEF, “Children and Landmines: A Deadly Legacy”, New York Sep 2013
https://www.unicef.org/french/protection/files/Landmines_Factsheet_04_LTR_HD.p
df
[3] N. Cumming-Bruce, K.Harrison, M.Loddo, L. Pinches, S. Casey-Maslen,
“CLEARING CLUSTER MUNITION REMNANTS”, Geneve 5-7 Septembet 2016.
http://www.mineactionreview.org/assets/downloads/Clearing_Cluster_Munition_Rem
nants_2016_WEB.pdf
[4] B. Ghali, “The landmine crisis: a humanitarian disaster. Foreign Affairs”, Oct 1994
https://www.foreignaffairs.com/articles/1994-09-01/land-mine-crisis-humanitarian-
disaster
[5] Landmine Monitor Core Group,” Towards a Mine Free World: Executive Summary”,
Feb 2016
[6] A.A Berhe, "The contribution of landmines to land degradation." Land Degradation
& Development 18, no. 1: 1-15. California Aug 2006
http://eps.berkeley.edu/~aaberhe/Berhe%202007-%20LM%20and%20LD.pdf
[7] N. Anderson, S. Paredes, “Social cost of land mines in four countries: Afghanistan,
Bosnia, Cambodia, and Mozambique“, British Medical Journal ,British Sep 2016
[8] A.A. Berhe, "Landmines and land degradation: a regional political ecology perspective
on the impacts of landmines on environment and development in the developing
world." Michigan State University: East Lansing, MI (2000).
[9] M.A. Buenker, “Landmines: A threat to wildlife and sustainability”, World
Conservation. 1: 19–20, Jan 2000
[10] Global Security Fact Sheets, “Landmines”, Jul 7, 2011
[11] R.Keeley, “Understanding Landmines and Mine Action. MIT”, Massachusetts
Institute of Technology (US), Feb 2003
[12] United Nations Mine Action Service (UNMAS), “Glossary of mine action terms,
definitions and abbreviations “, New York Jan 2003
[13] J. MacDonald, J.R. Lockwood, J. McFee, T. ALtshuler, T. Broach, L. Carin,R.
Harmon, C. Rappaport, W. Scott,R. Weaver,” Alternatives for Landmine Detection”,
Pittsburg, RAND Science and Technology Policy Institute, Online book 2003, page:6-
47.

110
[14] M.G.Kale, V. R. Ratnaparkhe, A.S.Bhalchandra. “Sensors For Landmine Detection
And Techniques: A Review”, International Journal of Engineering Research &
Technology (IJERT)Vol. 2 Issue 1, January- 2013 p:2-4.
[15] C. P. Gooneratne, S. C. Mukhopahyay and G. Sen Gupta. “A Review of Sensing
Technologies for Landmine Detection: Unmanned Vehicle Based Approach” , Institute
of Information Sciences and Technology, Massey University, Palmerston North, New
Zealand December 2004. p:401-407
[16] M. Carrasco, D. Mery, L. Robledo. “A survey of land mine detection technology”
International Journal of Remote Sensing Vol. 30, No. 9, 10 May 2009, p:2400-2404.
[17] I. Makki, R. Younes, C. Francis, T. Bianchi, & M. Zucchetti (2017). “A survey of
landmine detection using hyperspectral imaging”, ISPRS Journal of Photogrammetry
and Remote Sensing, 124, 40-53.
[18] J. E. McFee ; K. L. Russell and M. R. Ito, "Detection of surface-laid minefields using
a hierarchical image processing algorithm", Proc. SPIE 1567, Applications of Digital
Image Processing XIV, 42 (December 1, 1991); doi:10.1117/12.50802
[19] J. E. McFee, H. T. Ripley, R. Buxton, A. M. Thriscutt, “Preliminary study of detection
of buried landmines using a programmable hyperspectral imager”, Proc. SPIE 2765,
Detection and Remediation Technologies for Mines and Minelike Targets, 476 (May
31, 1996); doi:10.1117/12.241250
[20] J. E. McFee and H. T. Ripley, "Detection of buried land mines using a CASI
hyperspectral imager", Proc. SPIE 3079, Detection and Remediation Technologies for
Mines and Minelike Targets II, 738 (July 22, 1997)
[21] S. B. Achal ; J. E. McFee and C. D. Anger, "Identification of surface-laid mines by
classification of compact airborne spectrographic imager (CASI) reflectance spectra",
Proc. SPIE 2496, Detection Technologies for Mines and Minelike Targets, 324 (June
20, 1995);
[22] T. Skauli, and I. Kåsen. "The effect of spatial resolution on hyperspectral target
detection performance." In European Symposium on Optics and Photonics for Defence
and Security, pp. 59870V-59870V. International Society for Optics and Photonics,
2005.
[23] S. B. Achal; C. D. Anger; J. E. McFee and R. W. Herring "Detection of surface-laid
mine fields in VNIR hyperspectral high-spatial-resolution data", Proc. SPIE 3710,
Detection and Remediation Technologies for Mines and Minelike Targets IV, 808
(August 2, 1999)
[24] T. Ivanco, S. B. Achal, C. D. Anger, and J. E. McFee, “Casi Real-Time Surface-Laid
Mine Detection System”, Proc. SPIE 4394, Detection and Remediation Technologies
for Mines and Minelike Targets VI, 365 (October 18, 2001)

111
[25] T. Ivanco, S. Achal, J. E. McFee, C. Anger, J. Young, “Real-time airborne
hyperspectral imaging of landmines”, Proc. SPIE 6553, Detection and Remediation
Technologies for Mines and Minelike Targets XII, 655315 (April 26, 2007);
doi:10.1117/12.720442
[26] J. E. McFee, S. Achal, T. Ivanco and C. Anger, "A short wave infrared hyperspectral
imager for landmine detection", Proc. SPIE 5794, Detection and Remediation
Technologies for Mines and Minelike Targets X, 56 (July 08, 2005);
[27] S. Achal, J. E. McFee, T. Ivanco and C. Anger, "A thermal infrared hyperspectral
imager (tasi) for buried landmine detection", Proc. SPIE Conference on Detection and
Remediation Technologies for Mines and Mine-like Targets XII, Vol. 6553, Orlando,
FL, USA, 9-13 April, 2007, pp.655316-1-655316-11.
[28] J. E. McFee, S. B. Achal, A. U. Diaz and A. A. Faust, "Comparison of broad-band
and hyperspectral thermal infrared imaging of buried threat objects", Proc. SPIE
Conference on Detection and Sensing of Mines, Explosive Objects and Obscured
Targets XVIII, Volume 8709, Baltimore, MD, USA, 29 April - 03 May 2013.
[29] A. A. Faust, J. E. Mcfee, R. H. Chesney, K. L. Russell, Y. Das, “Canadian
Teleoperated landmine detection systems part I”, International Journal Of Systems
Science 36(9):511-528 · JULY 2005
[30] J. E. McFee, K.L. Russell, R. H. Chesney, A. A. Faust, Y. Das, “The Canadian forces
ILDS- A military fielded multi-sensor, vehicle mounted , teleoperated landmine
detection system”, Detection and Remediation Technologies for Mines and Minelike
Targets XI, Proceeding of Spie (2006)
[31] A.A. Faust, R.H. Chesney, Y. Das, J.E. McFee, K.l Russell, “Canadian teleoperated
landmine detection systems Part II: Anti personal landmine detection”, International
journal of Systems Science, vol.36, no.9, 15 July 2005, p 529-543.
[32] J. E. McFee ; C. Anger ; S. Achal and T. Ivanco, "Landmine detection using passive
hyperspectral imaging", Proc. SPIE 6554, Chemical and Biological Sensing VIII,
655404 (April 26, 2007)
[33] J. E. McFee and S. Achal, "Infrared and hyperspectral systems'', in Subsurface
Sensing, Section 7.6, Editors: A.S.Turk, A.K.Hocaoglu, A.A.Vertiy, ISBN 978-0-470-
13388-0, John Wiley and Sons, Wiley Series in Microwave and Optical Engineering,
New York, Series Volume 197, August 2011, pp.465-483.
[34] J. E. McFee, S. Achal, A. A. Faust, E. Puckrin, A. House, D. Reynolds, W. McDougall
and A. Asquini, "Detection and dispersal of explosives by ants", Proc. SPIE Conference
on Detection and Sensing of Mines, Explosive Objects and Obscured Targets XIV, Vol.
7303, Orlando, FL, USA, 13-17 April, pp.730302, 2009.
[35] S. Achal, J. E. McFee and J. Howse, "Gradual dispersal of explosives by ants and its
possible implication for future landmine production", Proc. 7th International
Symposium on Humanitarian Demining 2010, Croatian Mine Action Center
(CROMAC), Sibenik, Croatia, 27-29 April 2010, paper available on line at
http://www.ctro.hr/universalis/148/dokument/bookofpapers\_373441161.pdf, page 60.
[36] J. E. McFee, A.A. Faust, Y.Das, K.L. Russell, “Final report Shield ARP 12 rl” –
Optical imaging of explosive threats, August 2010.

112
[37] J.R. Simard, P. Mathieu, G. R. Fournier, V. Larochelle, and Stephen K. Babey.
"Range-gated intensified spectrographic imager: an instrument for active hyperspectral
imaging." In AeroSense 2000, pp. 180-191. International Society for Optics and
Photonics, 2000.
[38] J.-P Ardouin, J. Levesque and T.A. Rea, “Demonstration of Hyperspectral Image
Exploitation for Military Applications”, Information Fusion, 2007 10th International
Conference on 9-12 July 2007
[39] L. B. Wolff ; D. A. Socolinsky ; C. K. Eveland ; J. I. Yalcin and J. H. Holloway, Jr.
"Image fusion of shortwave infrared (SWIR) and visible for detection of mines,
obstacles, and camouflage", Proc. SPIE 5089, Detection and Remediation
Technologies for Mines and Minelike Targets VIII, 1298 (September 15, 2003).
[40] E.M. Winter, M. J. Schlangen, A.P. Bowman, M. R. Carter, C. L. Bennett, D. J. Fields,
W. D. Aimonetti et al. "Experiments to support the development of techniques for
hyperspectral mine detection." In Aerospace/Defense Sensing and Controls, pp. 139-
148. International Society for Optics and Photonics, 1996.
[41] A. P. Bowman, E. M. Winter, A. D. Stocker, and P. G. Lucey. "Hyperspectral infrared
techniques for buried landmine detection." In Detection of Abandoned Land Mines,
1998. Second International Conference on the (Conf. Publ. No. 458), pp. 129-133. IET,
1998.
[42] A. M. Smith, A.C. Kenton, R. Horvath, L. S. Nooden, J. Michael, J. A. Wright, J. L.
Mars et al. "Hyperspectral mine detection phenomenology program." In AeroSense'99,
pp. 819-829. International Society for Optics and Photonics, 1999.
[43] A.C. Kenton, C. R. Schwartz, R. Horvath, J. N. Cederquist, L. S. Nooden, D. R.
Twede, J. A. Nunez, J. A. Wright, J. W. Salisbury, and K. Montavon. "Detection of
land mines with hyperspectral data." In AeroSense'99, pp. 917-928. International
Society for Optics and Photonics, 1999.
[44] M. T. Eismann, C. R. Schwartz, J. N. Cederquist, J.A. Hackwell, and R. J. Huppi.
"Comparison of infrared imaging hyperspectral sensors for military target detection
applications." In SPIE's 1996 International Symposium on Optical Science,
Engineering, and Instrumentation, pp. 91-101. International Society for Optics and
Photonics, 1996.
[45] E. M. Winter, M. A. Miller, C. G. Simi, A. B. Hill, T. J. Williams, D. Hampton, M.
Wood, J. Zadnick, and M. D. Sviland. "Mine detection experiments using hyperspectral
sensors." In Proceedings of SPIE, vol. 5415, pp. 1035-1041. 2004.
[46] A. M Thomas., and J. Michael Cathcart. "Applications of grid pattern matching to the
detection of buried landmines." IEEE Transactions on Geoscience and Remote Sensing
48, no. 9 (2010): 3465-3470.
[47] N. Playle, "Detection of landmines using hyperspectral imaging." In Defense and
Security Symposium, pp. 62170A-62170A. International Society for Optics and
Photonics, 2006.
[48] G. Suganthi, R. Korah, “Discrimination of Mine-Like Objects in Infrared Images
Using Artificial Neural Network”, Indian Journal Of Applied Research, Volume : 4
Issue : 12 , India, p 206-208, Dec 2014

113
[49] J.S. Groot, Y.H.L. Janssen, “Remote Land Mine(Field) Detection, an overview of
techniques”, TNO Physics and Electronics Laboratory ,September 7, 1994.
[50] N. Milisavljević and I. Bloch. "How can data fusion help humanitarian mine action?."
International Journal of Image and Data Fusion 1, no. 2 (2010): 177-191.
[51] D. Letalick, I. Renhorn, O. Steinvall, “Multi-Optical Mine detection System (MOMS)
final report”, FOI Swedish Defence Research Agency, ISSN 1650-1942, December
2009.
[52] M.-A. Gagnon, P. Lagueux, J.-P. Gagnon, S. Savary, P. Tremblay, V. Farley, É. Guyot
and M. Chamberland. "Airborne thermal infrared hyperspectral imaging of buried
objects." In SPIE Security+ Defence, pp. 96490T-96490T. International Society for
Optics and Photonics, 2015.
[53] J. M Bioucas-Dias, A. Plaza, G. Camps-Valls, P. Scheunders, N. Nasrabadi, and J.
Chanussot. "Hyperspectral remote sensing data analysis and future challenges." IEEE
Geoscience and Remote Sensing Magazine 1, no. 2 (2013): 6-36.
[54] V. A. Kotkar , S. S. Gharde, “Review Of Various Image Contrast Enhancement
Techniques, International Journal of Innovative Research in Science, Engineering and
Technology Vol. 2, Issue 7, July 2013.
[55] C. P.Lee, "Mine detection techniques using multiple sensors." The Project in Lieu of
Thesis, Electrical and Computer Engineering The University of Tennessee at Knoxville
(2000).
[56] N. Sengee, A. Sengee and H.K. Choi, “Image Contrast Enhancement using Bi-
Histogram Equalization with Neighborhood Metrics” IEEE Transactions on Consumer
Electronics, Vol 56, No.4. November 2010.
[57] Y. Wang, Q. Chen, and B. Zhang. "Image enhancement based on equal area dualistic
sub-image histogram equalization method." IEEE Transactions on Consumer
Electronics 45, no. 1 (1999): 68-75.
[58] S. Der Chen and R. Ramli, “Minimum Mean Brightness Error Bi-Histogram
Equalization in Contrast Enhancement”, IEEE Transactions on Consumer Electronics,
Vol. 49, No. 4, pp.1310-1319, 2003.
[59] D. Chen and R. Ramli, “Contrast Enhancement Using Recursive Mean-Separate
Histogram Equalization for Scalable Brightness Preservation”, IEEE Transactions on
Consumer Electronics, Vol. 49, No. 4, pp.1301-1309, 2003.
[60] D. Menotti, L. Najman, J. Facon and A. Araujo, “Multi-Histogram Equalization
Methods for Contrast Enhancement and Brightness Preserving”, IEEE Transactions on
Consumer Electronics, Vol. 53, No. 3, pp.1186-1194, 2007.
[61] H. Ibrahim and N. S. P. Kong. "Brightness preserving dynamic histogram equalization
for image contrast enhancement." Consumer Electronics, IEEE Transactions on 53, no.
4 (2007): 1752-1758.
[62] M. Kim and M. G. Chung. "Recursively separated and weighted histogram
equalization for brightness preservation and contrast enhancement." Consumer
Electronics, IEEE Transactions on 54, no. 3 (2008): 1389-1397.
[63] U.K. Sharma, K. Umesh and K. Kumawat. "Review Of Histogram Based Image
Contrast Enhancement Techniques. International Journal of Research in Engineering

114
 & Technology ISSN(E): 2321-8843; ISSN(P): 2347-4599 Vol. 3, Issue 2, Feb 2015,
65-76
[64] J-M Gaucel, M. Guillaume and S. Bourennane. "Adaptive-3D-Wiener for
hyperspectral image restoration: Influence on detection strategy." Signal Processing
Conference, 2006 14th European. IEEE, 2006.
[65] Filtrage adaptatif: théorie et algorithmes. Hermes Science, 2005.
[66] N. Renard, S. Bourennane and J. Blanc-Talon. "Multiway filtering applied on
hyperspectral images." Advanced Concepts for Intelligent Vision Systems. Springer
Berlin Heidelberg, 2006.
[67] P.M. Mather and Magaly Koch. “Computer processing of remotely-sensed images: an
introduction”. John Wiley & Sons, 2011.
[68] G. Noyel, J. Angulo, and D. Jeulin. "Morphological segmentation of hyperspectral
images." Image Anal. Stereol 26, no. 3 (2007): 101-109.
[69] L. Kaufman, and P. J. Rousseeuw. Finding groups in data: an introduction to cluster
analysis. Vol. 344. John Wiley & Sons, 2009.
[70] J. P. Benzecri, 1973." L'analyse de données 2, Dunod (1973).
[71] J.M. Beaulieu, and M. Goldberg. "Hierarchy in picture segmentation: A stepwise
optimization approach." Pattern Analysis and Machine Intelligence, IEEE Transactions
on 11.2 (1989): 150-163.
[72] J. C Tilton. "Image segmentation by region growing and spectral clustering with
natural convergence criterion." International geoscience and remote sensing
symposium. Vol. 4. Institute Of Electrical & Electronicsengineers, Inc (Ieee), 1998.
[73] J.C. Tilton, "Split-remerge method for eliminating processing window artifacts in
recursive hierarchical segmentation." U.S. Patent No. 7,697,759. 13 Apr. 2010.
[74] C. Rodarmel and J. Shan. "Principal component analysis for hyperspectral image
classification." Surveying and Land Information Science 62.2 (2002): 115.
[75] J. Ye and S. Ji. "Discriminant analysis for dimensionality reduction: An overview of
recent developments." Biometrics: Theory, Methods, and Applications. Wiley-IEEE
Press, New York (2010).
[76] K. Fukunaga. “Introduction to statistical pattern recognition.” Academic Press
Professional, Inc., San Diego, CA, USA, 2nd edition, 1990.
[77] P.-H. Hsu, "Feature extraction of hyperspectral images using wavelet and matching
pursuit." ISPRS Journal of Photogrammetry and Remote Sensing 62, no. 2 (2007): 78-
92.
[78] P. Howland, M. Jeon, and H. Park. Structure preserving dimension reduction for
clustered text data based on the generalized singular value decomposition. SIAM
Journal on Matrix Analysis and Applications, 25(1):165–179, 2003.
[79] X. He, D.Cai, S. Yan and H.J. Zhang, "Neighborhood preserving embedding."
Computer Vision, 2005. ICCV 2005. Tenth IEEE International Conference on. Vol. 2.
IEEE, 2005.
[80] J. W. Sammon, "A nonlinear mapping for data structure analysis." IEEE Transactions
on computers 5 (1969): 401-409.

115
[81] B.-C. Kuo, and D. A. Landgrebe. "Nonparametric weighted feature extraction for
classification." Geoscience and Remote Sensing, IEEE Transactions on 42, no. 5
(2004): 1096-1105.
[82] R. Ablin, and C. H. Sulochana. "A survey of hyperspectral image classification in
remote sensing." International Journal of Advanced Research in Computer and
Communication Engineering 2.8 (2013): 2986-3000.
[83] V. N. Vapnick, Statistical Learning Theory. John Wiley and Sons Inc.,1998.
[84] J. A. Gualtieri and R. F. Cromp, “Support vector machines for hyperspectral remote
sensing classification,” in Proceedings of the SPIE, vol.3584, 1999, pp. 221–232.].
[85] G. Mercier and M. Lennon. "Support vector machines for hyperspectral image
classification with spectral-based kernels." Geoscience and Remote Sensing
Symposium, 2003. IGARSS'03. Proceedings. 2003 IEEE International. Vol. 1. IEEE,
2003.
[86] P.-N. Tan, M. Steinbach and V. Kumar, “Cluster Analysis: Basic Concepts and
Algorithms”, Chapter 8 in Introduction to Data Mining, Addison-Wesley, 2006, 487-
567.
[87] J.C. Harsanyi and C.-I. Chang. "Hyperspectral image classification and
dimensionality reduction: an orthogonal subspace projection approach." Geoscience
and Remote Sensing, IEEE Transactions on 32.4 (1994): 779-785.
[88] J. Broadwater, and R. Chellappa. "Hybrid detectors for subpixel targets." IEEE
transactions on pattern analysis and machine intelligence 29, no. 11 (2007)
[89] D. Manolakis, E. Truslow, M. Pieper, T. Cooley, and M. Brueggeman. "Detection
algorithms in hyperspectral imaging systems: An overview of practical algorithms."
IEEE Signal Processing Magazine 31, no. 1 (2014): 24-33.
[90] R.O. Duda, P. E. Hart and D. G. Stork. Pattern classification. John Wiley & Sons,
2012.
[91] H. Ren, Q. Du, C.-I. Chang, and J. O. Jensen. "Comparison between constrained
energy minimization based approaches for hyperspectral imagery." In Advances in
Techniques for Analysis of Remotely Sensed Data, 2003 IEEE Workshop on, pp. 244-
248. IEEE, 2003.
[92] O.L. Frost, "An algorithm for linearly constrained adaptive array processing."
Proceedings of the IEEE 60, no. 8 (1972): 926-935.
[93] L. Wang, and C. Zhao. Hyperspectral Image Processing. Springer, 2015, p236.
[94] D.C. Heinz, "Fully constrained least squares linear spectral mixture analysis method
for material quantification in hyperspectral imagery." IEEE transactions on geoscience
and remote sensing 39, no. 3 (2001): 529-545.
[95] C.-I Chang,. "Spectral information divergence for hyperspectral image analysis." In
Geoscience and Remote Sensing Symposium, 1999. IGARSS'99 Proceedings. IEEE
1999 International, vol. 1, pp. 509-511. IEEE, 1999.
[96] T. Wang, Bo Du, and L. Zhang. "A kernel-based target-constrained interference-
minimized filter for hyperspectral sub-pixel target detection." IEEE journal of selected
topics in applied earth observations and remote sensing 6, no. 2 (2013): 626-637.

116
[97] A. Ertürk, D. Çeşmeci, M. K. Güllü, D. Gerçek, and S. Ertürk. "Endmember extraction
guided by anomalies and homogeneous regions for hyperspectral images." IEEE
Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7, no.
8 (2014): 3630-3639.
[98] M. Axelsson, O. Friman, T.V. Haavardsholm, and I. Renhorn. "Target detection in
hyperspectral imagery using forward modeling and in-scene information." ISPRS
Journal of Photogrammetry and Remote Sensing 119 (2016): 124-134.
[99] Y. Zhang, Bo Du, and L. Zhang. "A sparse representation-based binary hypothesis
model for target detection in hyperspectral images." IEEE Transactions on Geoscience
and Remote Sensing 53.3 (2015): 1346-1354.
[100] R.M. Willett, M. F. Duarte, M. A. Davenport, and R. G. Baraniuk. "Sparsity and
structure in hyperspectral imaging: Sensing, reconstruction, and target detection."
IEEE signal processing magazine 31, no. 1 (2014): 116-126.
[101] O. Ahmad, C. Collet, and F. Salzenstein. "Spatio-spectral Gaussian random field
modeling approach for target detection on hyperspectral data obtained in very low
SNR." 2015 IEEE International Conference in Image Processing (ICIP), pp. 2090-
2094. IEEE, 2015.
[102] I. Makki, R. Younes, C. Francis and M. Zucchetti, “Mathematical Methods for
Hyperspectral Imaging in Landmine Detection”, Transactions of the American Nuclear
Society, Vol. 112, San Antonio, Texas, June 7–11, 2015.
[103] J.B. ZARE, P. GADER and M. SCHATTEN, ”Vegetation Mapping for Landmine
Detection Using LongWave Hyperspectral Imagery”, IEEE Transactions On
Geoscience And Remote Sensing, 46,1, 172 (2008)
[104] J. KHODER, R. YOUNES. “Proposal for Preservation Criteria to Rare Event.
application on Multispectral/Hyperspectral Images”. 25th IEEE International
Conference on Microelectronics (2013)
[105] N. PLAYLE, “Detection of landmines using hyperspectral imaging”. Proc. SPIE
6217, Detection and Remediation Technologies for Mines and Minelike Targets XI,
62170A (2006).
[106] www.ehu.es/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes
[107] D. Manolakis, R. Lockwood, T. Cooley, and J. Jacobson. "Is there a best
hyperspectral detection algorithm?." In SPIE Defense, Security, and Sensing, pp.
733402-733402. International Society for Optics and Photonics, 2009
[108] J. Yin, Y. Wang, Y. Wang, and Z. Zhao. "A modified algorithm for multi-target
detection in hyperspectral image." In Informatics in Control, Automation and Robotics
(CAR), 2010 2nd International Asia Conference on, vol. 3, pp. 105-108. IEEE, 2010.
[109] A. Plaza, P. Martínez, R. Pérez, and J. Plaza. "A new method for target detection in
hyperspectral imagery based on extended morphological profiles." In Geoscience and
Remote Sensing Symposium, 2003. IGARSS'03. Proceedings. 2003 IEEE
International, vol. 6, pp. 3772-3774. IEEE, 2003.
[110] W. McCulloch, W. Pitts (1943). "A Logical Calculus of Ideas Immanent in Nervous
Activity". Bulletin of Mathematical Biophysics. 5 (4): 115–133.
doi:10.1007/BF02478259

117
[111] C.M. Bishop, “Pattern recognition and machine learning.” springer, 2006 p228
[112] M.JL. Orr, "Introduction to radial basis function networks." (1996).
http://twyu2.synology.me/htdocs/class_2008_1/nn/Slides/Introduction%20to%20Radi
al%20Basis%20Function%20Networks%20(1996).pdf
[113] Y. Chen, H. Jiang, C. Li, X. Jia, and P. Ghamisi. "Deep feature extraction and
classification of hyperspectral images based on convolutional neural networks." IEEE
Transactions on Geoscience and Remote Sensing 54, no. 10 (2016): 6232-6251.
[114] P.-H. Hsu, "Feature extraction of hyperspectral images using wavelet and matching
pursuit." ISPRS Journal of Photogrammetry and Remote Sensing 62, no. 2 (2007): 78-
92.
[115] X. He, D.Cai, S. Yan and H.J. Zhang, "Neighborhood preserving embedding."
Computer Vision, 2005. ICCV 2005. Tenth IEEE International Conference on. Vol. 2.
IEEE, 2005.
[116] J.W. Sammon, "A nonlinear mapping for data structure analysis." IEEE Transactions
on computers 5 (1969): 401-409.
[117] M. Khoder, S. Kashana, J. Khoder, and R. Younes. "Multicriteria classification
method for dimensionality reduction adapted to hyperspectral images." Journal of
Applied Remote Sensing 11, no. 2 (2017): 025001-025001.
[118] B.-C. Kuo, and D. A. Landgrebe. "Nonparametric weighted feature extraction for
classification." Geoscience and Remote Sensing, IEEE Transactions on 42, no. 5
(2004): 1096-1105.
[119] J. Khodr, and R. Younes. "Dimensionality reduction on hyperspectral images: A
comparative review based on artificial datas." In Image and Signal Processing (CISP),
2011 4th International Congress on, vol. 4, pp. 1875-1883. IEEE, 2011.
[120] H. Grahn, and P. Geladi, eds. “Techniques and applications of hyperspectral image
analysis”. John Wiley & Sons, 2007 p185.
[121] H. Ren and C.-I. Chang, “Automatic spectral target recognition in hyperspectral
imagery,” IEEE Transactions on Aerospace and Electronic Systems, vol. 39, no. 4, pp.
1232–1249, October 2003.
[122] http://aviris.jpl.nasa.gov
[123] I. Makki, R. Younes, C. Francis, T. Bianchi, and M. Zucchetti. "Classification
algorithms for landmine detection using hyperspectral imaging." In Landmine:
Detection, Clearance and Legislations (LDCL), 2017 First International Conference
on, pp. 1-6. IEEE, 2017.
[124] W.F. Nanceb, E. Basenera, and J. Kerekesa. "The target implant method for
predicting target difficulty and detector performance in hyperspectral imagery." In
Proc. of SPIE Vol, vol. 8048, pp. 80481H-1. 2011.
[125] rsipg.dii.unipi.it
[126] C. M. Bishop, 1996, Neural Networks for Pattern Recognition. Oxford Univ. Press,
Oxford
[127] O. Laurino, R. D’Abrusco, G. Longo, G. Riccio; “Astroinformatics of galaxies and
quasars: a new general method for photometric redshifts estimation”, Monthly Notices

118
 of the Royal Astronomical Society, Volume 418, Issue 4, 21 December 2011, Pages
2165–2195, https://doi.org/10.1111/j.1365-2966.2011.19416.x
[128] I. Nabney, “NETLAB: algorithms for pattern recognition.”, Springer Science &
Business Media, 2002.
[129] M. Zucchetti, M. Khoder, I. Makki, R. Younes, C. Francis and T. Bianchi,
"Landmines: Crisis, legacy, international and local action," 2017 First International
Conference on Landmine: Detection, Clearance and Legislations (LDCL), Beirut,
2017, pp. 1-6.doi: 10.1109/LDCL.2017.7976954
[130] http://www.icbl.org/en-gb/the-treaty/treaty-status.aspx
[131] https://www.asdi.com/products-and-services/accessories/illuminator-reflectance-
lamp
[132] Fundamentals of remote sensing, Canada center for remote sensing chap 2.8.
http://pages.csam.montclair.edu/~chopping/rs/CCRS/chapter2/chapter2_8_e.html
[133] E. Truslow, “Performance evaluation of the adaptive cosine estimator detector for
hyperspectral imaging applications.”, Diss. NORTHEASTERN UNIVERSITY, 2012.

119