Infrared Physics and Technology 98 (2019) 62–68

Contents lists available at ScienceDirect

Infrared Physics & Technology

journal homepage: www.elsevier.com/locate/infrared

Regular article

Determination of viability of Retinispora (Hinoki cypress) seeds using FT- T

NIR spectroscopy
Perez Mukasaa, Collins Wakholia, Changyeun Mob, Mirae Ohc, Hye-Joon Jood, Hyun Kwon Suha,
⁎
Byoung-Kwan Choa,
a
Department of Biosystems Machinery Engineering, College of Agricultural and Life Science, Chungnam National University, 99 Daehak-ro, Yuseong-gu, Daejeon 34134,
Republic of Korea
b
National Academy of Agricultural Science, Rural Development Administration, 310 Nonsaengmyeong-ro, Wansan-gu, Jeollabuk-do 54875, Republic of Korea
c
USDA-ARS Environmental Microbial and Food Safety Laboratory, Henry A. Wallace Beltsville Agricultural Research Center, Beltsville, MD 20705, USA
d
National Forest Seed & Variety Center, Korea Forest Service, 72 Suhoeri-ro, Suanbo-myeon, Chungju-si, Chungcheongbuk-do 27495, Republic of Korea

A R T I C LE I N FO A B S T R A C T

Keywords: The use of Hinoki cypress for the formation of healing forest is gradually increased in South Korea, but the
FT-NIR spectroscopy germination rate of these seeds is low, and viability determination by conventional methods is destructive and
Hinoki cypress time-consuming. Therefore, the aim of this study was to investigate the potential of Fourier transform infrared
Seed viability (FT-NIR) spectroscopy in determining the viable seeds of Hinoki cypress nondestructively. FT-NIR reflectance
Seed quality
spectra for single seeds were collected in the range of 4000–10,000 cm−1 (1000–2500 nm), and a germination
PLS-DA
test was carried out to determine viability. To differentiate between viable and nonviable seeds, a multivariate
classification with partial least square discriminant analysis (PLS-DA) was developed. The best PLS-DA model
assigned the seeds to their respective classes, with 97.7–99.2% and 94.4–95.4% accuracy in the calibration and
validation sets, respectively. The PLS-DA Beta coefficient revealed the important wavelength to differentiate
viable from nonviable seeds, which was attributed to changes in the chemical composition of the seeds, such as
lipids and proteins, which might be responsible for the germination ability of the seeds. Variable importance of
projection (VIP) was applied on the spectral data which reduced original variables from 1557 to 27. The de-
veloped VIP-PLS-DA model resulted into classification accuracy of 97.7% in calibration and 91.7% in the vali-
dation set, with maximum normalization data preprocessing method. In conclusion, the results demonstrate the
potential of FT-NIR spectroscopy as a powerful nondestructive method for determination of viable Hinoki cy-
press seeds, which could be applied in the development of an online sorting technique for seed companies and
nurseries.

1. Introduction frequent formation of anatomically unsound seeds that are empty and
underdeveloped seeds [5]. The empty seeds are totally devoid of
Chamaecyparis obtusa is a species of native cypress (Hinoki cypress) megagametophyte (storage organ) and embryo [6], thus contributing to
and one of the most important conifer species for wood in the southern the nonviable seed lot. Successful cultivation of Hinoki cypress thus
part of Korea, mainly Jeju [1]. Hinoki wood has gained usage as a demands a steady and persistent supply of high-quality seeds for the
structural material because of its superior mechanical properties, ex- production of seedlings in nurseries or for direct sowing in the field.
cellent durability and the beautiful color of its heartwood [2]. This tree Recent studies have revealed that most of these nonviable Hinoki
species is one of the representative trees for forest bathing, which is a cypress seeds cannot be separated from viable seeds by visual ob-
common health practice in East Asia, and essential oil from this tree is servation with the naked eyes [5]. Thus, seed sorting has been a
widely used in commercial products of air purifiers or deodorants [3]. common practice in the seed handling routine to upgrade the seed lot
Seeds from this species are the major regeneration material for quality by removing viable seeds from nonviable seeds. To upgrade the
seedlings and reforestation purposes; hence, their demand has in- seed lot quality, commercial seed companies [7] have deployed special
creased [1]. The seeds have low seed germination rates [4], and studies equipment to screen the nontree seeds based on some attributes such as
have shown that the causes of low germination are attributed to the color, size, viability, vigor, genetic purity, seedling performance [8],

⁎
Corresponding author.
E-mail address: [email protected] (B.-K. Cho).

https://doi.org/10.1016/j.infrared.2019.02.008
Received 16 October 2018; Received in revised form 11 February 2019; Accepted 13 February 2019
Available online 14 February 2019
1350-4495/ © 2019 Elsevier B.V. All rights reserved.
P. Mukasa, et al. Infrared Physics and Technology 98 (2019) 62–68

specific gravity, shape [9], and surface texture [10], since good and bad 2.2. FT-NIR spectra acquisition
seeds differ basically according to these traits [6]. However, these
methods have limitations such as subjectivity, invasiveness, low spe- Spectra for a single seed were acquired using a FT-NIR spectrometer
cificity, long time computation and low accuracy [11]. Therefore, to (Antaris II FT-NIR Analyzer, Thermo Scientific Co., MA, USA), with a
overcome these limitations, there have been interests in the search of an spectral range of 4000–10,000 cm−1 (1000–2500 nm) and a resolution
efficient and nondestructive sorting system that can be applied across of 4 cm−1. The reflectance spectra of a single seed were acquired by
this species of tree seeds for viability determination. Thus Fourier placing individual seeds at the center of the scanning glass window and
transform infrared (FT-NIR) spectroscopy as a robust seed sorting covering with the instrument lid that had a black background. Each
system has been a subject of interest. seed was scanned 32 times, and the mean spectra were obtained.
Over the years, near-infrared spectroscopy has been proven useful
for identification of viable and nonviable tree seeds [12–14]. NIR 2.3. Germination test
spectroscopy in combination with multivariate analysis methods can
measure moisture and chemical composition of biological materials After spectra acquisition, the seeds were soaked in running water for
based on the absorption of near-infrared radiation by bonds between 48 h and then placed in a petri-dish for top paper germination ac-
light atoms, such as C − H , O − H and N − H in biological samples cording to the standards of International Seed Testing Association [19].
that result in overtones and combination bands detectable in the The germination conditions were as follows: exposure to light for a
780–2500 nm wavelength region [15]. period of 8 h and 16 h of darkness, temperature of 20–24 °C and relative
Spectroscopy has recently gained success in viability determination humidity of 60–80%. The germination rate was calculated after 7 days,
of artificially aged seeds, for example, [16,17], used near-infrared with everyday inspection for a period of 21 days. Seeds with a germi-
spectroscopy to successfully discriminate artificially aged (nonviable) nation length of 1 cm were counted as viable.
corn and wheat respectively from the viable ones.
Like any other plant seed, Hinoki has suffered from seed perfor- 2.4. Data preprocessing and multivariate analysis
mance decline as storage time increases. This is termed as aging or loss
of vigor which is demonstrated by delayed and fall in germination, thus We analyzed the spectral data using MATLAB software (Version
seed aging is a problem in agriculture that needs to be handled so as to 2015a, The Mathwork, Natik, MA, USA). It should be noted that FT-NIR
have high quality and yielding seeds. Artificial seed aging is normally spectroscopy data are usually not amendable for direct analysis due to
done based on the considerable increase in seed deterioration at high light scattering, baseline shift, instrumental drift and path length dif-
temperatures and relative humidity, and so seeds with lower quality ferences [20]. Such systematic noise should be removed from the raw
will deteriorate faster under these conditions than the more vigorous spectral signals to prevent them from dominating the chemical signals.
ones. Studies made by [18], showed some appreciable difference be- We used four types of mathematical preprocessing techniques for
tween the two aging types where artificial aging increased the level of data correction namely; normalization methods (minimum, maximum
antioxidant (α-tocopherol) in soybeans than in natural aging. Anti- and range), standard normal variate transformation (SNV), and mul-
oxidants inhibit the process of germination thus leading to poor seed tiple scatter correlation (MSC) and Savitzky-Golay smoothing (1st and
performance. 2nd derivatives). Normalization preprocessing served to fit spectral
More still, given the fact that seed companies store seeds in bulk and data to similar range (0–1) to compensate for differences due to sample
at a considerable period of time, there is a need to determine the via- quantity and path length. SNV was used to correct shifts in the spectral
bility of naturally aged seeds so as to have high-quality grad seeds. Up data caused by scattering and differences in the particle size [21]. MSC
to date, there is no single method that can differentiate naturally aged was used to correct the scatter level of the spectra [22]. Savitzky-Golay
seeds from its counterpart nondestructively without seed pretreatment smoothing method was used to remove overlapping peaks and baseline
and anatomical observation. correction of the data [23].
Therefore, for a seed to be tested for its germinability, the viability Principal component analysis (PCA) and partial least squares dis-
at the time of sowing is needed to be evaluated in a nondestructive criminant analysis (PLS-DA) techniques were applied to develop clas-
manner, which needs to be established for Hinoki and Retinispora seeds sification models for the viable and nonviable Hinoki cypress seeds. We
[5]. used PCA for visualization of data distribution with regard to distinc-
This study aims at the potential of FT-NIR spectroscopy to classify tion between nonviable and viable seed samples.
viable (naturally aged) from nonviable seeds nondestructively. The The PLS-DA classification model in Eq. (1) was used to analyze the
study is also conducted to determine an optimum multivariate classi- data for a reliable prediction accuracy as follows:
fication model (s) to distinguish between viable and nonviable Hinoki
Y = X ·B + E (1)
cypress seeds.
where Y is the matrix of response binary variables that relates to the
2. Materials and methods sample classes; X is the matrix that holds the predictor (spectral) values
for each class; B is the regression coefficient matrix for the predictor
2.1. Seed samples values; and E is the matrix of error terms (residuals). To find the re-
lationship (between Y and X variables) the X and Y matrix are de-
Hinoki cypress seeds were obtained from National Forest Seed and composed by latent variables which is given as;
Variety Center in South Korea. The seed batches used in the study were (3)
X = TPT + E
labeled Retinispora_C69_2016, Retinispora_6P′_2017 and
Retinispora_80P′_2017. The first batch of seeds was used for viability Y= UQT +E (4)
model development, while the second and third seed batches were used
for the blind test. The samples were first stored at 4 °C until use. A total U=β∗T (5)
of 1000 seeds from Retinispora_C69_2016, were randomly selected and
Yi = Xi ·β + β0 (6)
used for viability model development. A total 400 seeds were also
randomly picked from Retinispora_6P′_2017 and Retinispora_80P′_2017 where T and U are the score matrices andP and Q are the loading
for the blind test measurement. It should be noted that no seed pre- matrices. β0 and β are the beta coefficients of the PLS model. Yi is the
treatment or anatomical observation analysis was carried out on the predicted responses from a new data matrix Xi .
seeds in this study. We used calibration data to build the PLS model, Beta/regression

63
P. Mukasa, et al. Infrared Physics and Technology 98 (2019) 62–68

values of the PLS model were then multiplied with the test data to
determine new data responses. These data responses were then used to
determine the accuracy of the PLS model in relation to original re-
sponses. Calibration and prediction accuracy of the models were cal-
culated based on the number of samples correctly identified using the
various preprocessing methods. Those samples with the highest percent
accuracy were considered superior as a reliable and robust classification
model for viable and nonviable Hinoki cypress seeds.
After the germination test, the FT-NIR seed spectra acquired was
divided into 2 groups, viable and nonviable. The final PLS-DA model
was developed using the spectra range of 1000–2500 nm. The calibra-
tion set was developed with 284 seeds (142 seeds for both viable and
nonviable seeds). The validation set was evaluated using 122 seeds (61
seeds for both viable and nonviable seeds). The remaining 594 seeds
were treated as test data set. Calibration and validation accuracy of the
model were calculated based on the number of samples correctly
identified using the various preprocessing methods.

2.5. Variable selection

To analyze absorption bands that accounted for the discrimination

of viable and nonviable Hinoki cypress seeds, a parameter called vari-
able importance on projection (VIP) was computed as follows in Eq. (2)
as follows:
A
⎛ K ⎞
⎜ ∑ (w ak X (SSYa − 1 − SSYa )) X (SSY − SSY ) ⎟
VIPAK = 2
0 A (2)
⎝ a=1 ⎠
where VIP for A components and K variables is a weighted sum of
squares of the PLS weights (w) for a given component a and k variables,
taking into account the amount of explained Y-variance (SSY) of a
component, and SSY0 and SSYA are the sum of squares of the response
variable Y before and after extracting A, the number of components,
respectively. The advantage is that there will be only one VIP-vector
summarizing Y-variables, thereby enabling absorption bands that in-
fluence the discriminant models to be identified. As a rule, predictors
with a VIP value greater than 1.0 have a strong influence on the model,
but a cut-off approximately 0.7–0.8 has been suggested to discriminate
between relevant and irrelevant predictors [24].

3. Results and discussion

3.1. Germination and spectral profile

The germination results showed 20.3% viability of the seed with 42

seeds mold infected. The mold growth could have been due to un-
developed and juvenile seeds that may have been chosen randomly
during the experiment which have a low shelf life.
we plotted a raw data spectral graph (Fig. 1a), which showed no
characteristic difference between the viable seeds and the counterparts.
The spectra described a uniform characteristic for all seeds, and were
widely dispersed and contained noise. At this stage we couldn’t dis- Fig. 1. FT-NIR spectral plots: (a) Raw spectral plot, (b) SNV preprocessed
tinguish between viable and nonviable seeds by color of their spectral spectral data and (c) SNV mean preprocessed spectra.
plot. The spectral scattering effect was then corrected with SNV data
preprocessing method, as shown in (Fig. 1b), where there were still no patterns, could have resulted from differences in the chemical compo-
clear direct differences between the viable and nonviable spectra. nents of the seeds.
(Fig. 1c) shows the SNV preprocessed mean spectral plot con-
structed. Significant differences in the two spectral profile of the viable
and nonviable seeds were identified and marked as shown in the plot. 3.2. Principal Component Analysis (PCA)
These differences in the average absorbance spectral profile in the 2
groups, were evident in the spectral range of 1700–1800 and PCA was used for dimensionality reduction and visualization of data
2200–2390 nm. distribution with regard to the distinction between nonviable and vi-
More still, it was observed that the nonviable seeds had higher ab- able seed samples as shown in the 3-D (Fig. 2a). Here, dimensionality
sorbance values in the 1000–1500 nm and 2300–2500 nm ranges while reduction was performed by transforming the wavelength variables into
the viable seeds had more absorbance values in the 1700–1910 nm principal components (PCs). The first four PCs accounted for a cumu-
wavelength region. These differences observed in the mean spectral lative total variation of 99.93% in the seed viability and the remaining

64
P. Mukasa, et al. Infrared Physics and Technology 98 (2019) 62–68

peaks at described wavelength formed the basis for further analysis to

build classification models using PLS-DA.

3.3. PLS-DA model results

The values of viable and nonviable seeds in a y-vector were set as 1

and 0, respectively. For modeling development, 0.5 was considered as
the baseline and any value above the baseline was classified as viable
and below as nonviable. The model required 5–7 latent variables to
describe a 97.7–99.2% calibration and 94.4–95.4% validation accuracy
by the different data preprocessing methods. The classification results
of the PLS-DA model with each preprocessing method in the different
class sets are shown in (Table 1). The accuracy results in the table
clearly demonstrate that FT-NIR spectral data contained much in-
formation to differentiate between viable and nonviable Hinoki cypress
seeds. These results confirm that there are differences between the vi-
able and nonviable seeds because of the different chemical composi-
tions [15].

3.3.1. Interpretation of the beta coefficient

In multivariate data analysis, beta coefficients are important since
they measure how many units of standard deviation the criterion
variables in some multiple regression equation changes with a change
of one standard deviation unit in a predictor variable, while controlling
the other predictor variables. Beta coefficient was used to interpret the
direction of the relationship between NIR spectra as predictor variables
and viability as the criterion variables with PLS-DA. The Beta coeffi-
cient plot of PLS-DA model in (Fig. 3) shows that the difference in the
energy absorption between viable and nonviable seeds described in
(Fig. 1c), was due to the differences in chemical compositions. The plot
indicated the absorption bands that contributed to the classification of
viable from nonviable Hinoki cypress seeds.
The Beta coefficient plot showed four peaks in the wavelength re-
gion of 1000–1400 nm centered at 1167, 1209, 1330, and 1392 nm. The
wavelengths of 1167 and 1208 nm correspond to the C − H second
overtone stretching vibration which represents the presence of CH3
functional groups. 1330–1392 nm waveband region represented the
second combination C − H stretching bond reflecting the presence of
lipids.
The PLS regression showed seven important peaks in 1400–2000 nm
wavelength region, centered at 1520, 1637, 1685, 1714, 1740, 1798
and 1944 nm. The 1520 nm waveband represented the first overtone
N − H and O − H stretching bond of H2 O and protein moieties [15].
1637–1798 nm represented C − H first stretching bond representing
lipids. 1944 nm waveband was likely attributed to the presence of water
due to the O − H first overtone and combination bands [25].
In the waveband regions of 2000–2400 nm, ten wavebands were
showed which were peaked at 2035, 2111, 2146, 2188, 2231, 2284,
2301, 2317, 2346 and 2379 nm. Those wavebands at 2035 and
2111 nm showed the combination N − H and O − H stretching, re-
Fig. 2. PCA score plot for raw data (a), 4th and 2nd PCs score plot (b) and PCA
presenting several compounds of proteins and water[26]. The region of
loading plots (c).
2200–2460 nm absorption bands with a 2284 nm peak were found to
characterize the CH2 stretch bend combinations[15]. Several oil crops
PCs corresponded to 0.07%. PC1, PC2, PC3 and PC4 had 98.15%, have shown a positive correlation to this waveband region, describing
1.59%, 0.1% and 0.09% percentage separation, respectively. We ob- several fatty acids, notably polyunsaturated fatty acids [6].
served that PC2 and PC4 were the most relevant components for via- Therefore, FT-NIR spectroscopy was able to detect differences in the
bility separation and thus used for the visual exploration in identifying chemical compounds between the viable and nonviable seeds because
viable seeds from nonviable one as shown in 2D (Fig. 2b). of the differences in the reserve compounds of lipids, proteins and
We further developed PC loading plot for the first four loading moisture.
components (PC1, PC2, PC3 and PC4), which described the correlation
weights of the variables. These loading profiles revealed some im- 3.3.2. Wavelength selection that accounted for classification of viable and
portant information as shown (Fig. 2c). From the profiles described, non-viable seeds
four peaks centered at 1440 nm, 1713 nm, 1926 nm and 2305 nm, were Variable selection was carried out using variable importance on
observed to have been responsible for the viability separation between projection and the developed VIP PLS-DA model. The VIP model was
the viable seeds from nonviable. The distinction between viable and developed with different cut off values that ranged from 0 to 3 and the
nonviable Hinoki cypress seeds by the PCA and the presence of spectral cut-off threshold hold value determined as shown in (Fig. 4a).

65
P. Mukasa, et al. Infrared Physics and Technology 98 (2019) 62–68

Table 1
PLS-DA model summary on different preprocessing methods.
Preprocessing method Calibration (284 samples) Validation (122 samples) Latent variables

Nonviable Viable Overall Accuracy Nonviable Viable Overall Accuracy

1
Normalization Mean 96.2 99.2 97.7 98 93.1 95.4 5
Max2 97.7 100 98.8 98 93.1 95.4 7
Range3 98.5 100 99.2 98 91.4 94.4 7
MSC4 97.7 99.2 98.4 98 93.1 95.4 6
SNV5 97.7 100 98.8 100 91.4 95.4 6

Derivatives SG16 96.2 99.2 97.7 98 93.1 95.4 5

SG27 96.2 99.2 97.7 98 91.4 94.4 5
Raw8 96.6 99.2 98.1 98 93.1 95.4 6

1
Mean normalization.
2
Maximum normalization.
3
Range normalization.
4
Multiplicative scatter correction.
5
Standard normal variance.
6
Savitzky-Golay first derivative.
7
Savitzky-Golay second derivative.
8
Raw data.

Fig. 3. SNV preprocessed data, PLS-DA model Beta coefficients plot.

Additionally, the number of variables (wavelength) selected by the VIP

decreased from 1557 in the original spectral data with an increase in
the cut off values. The threshold VIP cut off value obtained was 2.43
with 27 selected variables as shown in (Fig. 4b). A new PLS-DA model
with VIP threshold selection was developed and compared with the
original PLS-DA model as shown in (Table 2), with the maximum and
mean normalization data preprocessing methods having the best ac-
curacy of classification. The VIP score plot is shown in (Fig. 5).
The VIP score plot in (Fig. 5) shows that the major peak in the
waveband region of 1600–1800 nm centered at 1713 nm had a strong
influence on the discrimination of viable and nonviable seeds. This
region represents C − H first stretching bond, probably representing
the presence of lipids as the major chemical composition. It was ob-
served from (Fig. 5) that shifting the dotted red line upwards, would
eliminate the 1713 nm wavelength. This would lower the prediction
accuracy of the model; thus, any VIP cut off point beyond the threshold
value will eliminate this important wavelength and therefore the results
Fig. 4. Variation of the model accuracy (a) and selected variables (b) with VIP
will be of low accuracy as shown in (Fig. 4a). cut-off values.
For a better comparison on the test data set that contained 594
seeds, SNV data preprocessing method was used for both the developed
PLS-DA model and VIP to determine the accuracy of the prediction as factors that deter seed viability. More still, 42 seeds were mold infected
shown in (Table 3). Originally, all samples (594) were considered during germination and never germinated. This could have been
nonviable per the germination results but the model results showed probably due to seeds with underdeveloped embryo that may have been
some percentage of viable seeds, which was more in the VIP PLS-DA randomly selected from the seed batch.
than in the originally developed PLS-DA model. The predicted viable The model developed was further tested on Hinoki seeds from two
seeds in both models may not have been seen in the germination results, seed batches, namely, Retinispora_6p′_2017 and 80p′_2017. The dif-
probably because of germination conditions, seed dormancy and other ferent model parameters and accuracy are shown in (Table 4). Two

66
P. Mukasa, et al. Infrared Physics and Technology 98 (2019) 62–68

Table 2
Original PLS-DA model versus developed threshold VIP PLS-DA model.
Preprocessing method PLS-DA (1556 variables) VIP PLS-DA (27 variables)

Calibration Validation Calibration Validation

Normalization Mean 97.7 95.4 97.3 91.7

Max 98.8 95.4 97.7 91.7
Range 99.2 94.4 96.9 90.7
MSC 98.4 95.4 97.7 88.9
SNV 98.8 95.4 97.7 88.9

Derivatives SG1 97.7 95.4 93.4 85.2

SG2 97.7 94.4 93.4 88.0
Raw 98.1 95.4 97.3 90.7

seed batch 80P′ _2017 than in 6P′_ 2017, because these were harvested
from different location and years and that seed batch 80P’ _2017 had
more viable seeds.

4. Conclusion

The discrimination of the viability of Hinoki cypress seeds using FT-

NIR spectroscopy was successful. The results showed the potential of
FT-NIR spectroscopy as a nondestructive method to differentiate viable
Hinoki cypress seeds from the nonviable ones. The PLS-DA model had
97.7–99.2% accuracy in the calibration set and 94.4–95.4% in the va-
lidation set with different preprocessing methods in classifying viable
and nonviable seeds. The Beta coefficient of the PLS-DA model provided
a better way of observing the important wavelength region that best
differentiated between viable and nonviable seeds. Variable selection
method reduced the size of the wavelengths from 1557 to 27 wave-
Fig. 5. Plot of VIP scores and selected wavebands.
lengths, describing a 98.3% reduction in wavelength. This process
would reduce the computation time and cost of system design for cases
Table 3 of automation machine design. Therefore, the FT-NIR spectroscopy
Developed PLS-DA and VIP threshold PLS-DA prediction on 594 test data set. technique is rapid and requires no pretreatment or sample preparation.
Percentage Germination PLS-DA model VIP PLS-DA model FT-NIR measurement technique of seeds can be addressed on the la-
Predicted Groups results (1557 variables) (27 variables) boratory scale at present because of the measurement of a single seed at
time which may not be feasible for large scale production sorting. Thus,
Non-Viable (%) 100 90.74 86.36
Viable (%) 0 9.26 13.64 there is need to investigate other nondestructive methods that can
allow bulk sorting of Hinoki cypress seeds by their viability. The study
focused on Hinoki Cypress seeds as one of the member for Retinispora
Table 4 family, more research is needed to construct a global model that can
Model blind test on two selected seed batches. predict the viability of all the Retinispora seeds.
Model parameters Retinispora 6P′_ 2017 (n = 200) 80P′_2017 (n = 200)
Acknowledgements
Specificity 97.87 98.94
Sensitivity 83.33 91.67 This research was partially supported by Export Strategy
Precision 71.43 84.61
Technology Development Program, Ministry of Agriculture, Food and
Accuracy 98 98.5
Rural Affairs (MAFRA), Republic of Korea and by National Forest Seed
and Variety Center, Korea Forest Service, Republic of Korea. Authors
hundred seeds from each seed batch were tested with the model for would like to appreciate National Forest Seed & Variety Center, Korea
viability determination. Failure of the predicated viable seeds to ger- Forest Service, Republic of Korea for providing seed samples.
minate within the test period could be contributed to dormancy, sub-
optimal culture conditions and loss of viability. This outcome could be Conflicts of interest
due to the long-term storage of the seeds.
Sensitivity described the percentage of viable seeds expected in the The authors declare no conflict of interest.
seed sample while precision described the percentage of the actual vi-
able seeds predicted by the model. From the table, it’s observed that the Appendix A. Supplementary material
model had both high sensitivity and precision performance on the two
seed batches. High sensitivity and precision means that most of the Supplementary data to this article can be found online at https://
viable seeds (positives) are correctly recognized and that there are a lot doi.org/10.1016/j.infrared.2019.02.008.
of true or actual viable seeds (positives) predicted by the model.
Specification described the percentage of the actual predicted nonvi- References
able seeds by the model while accuracy described the overall model
accuracy in separating the viable seeds from nonviable. [1] Y.M. Kim, Y.S. Shin, J.H. Jeong, Development and Characterization of
Microsatellite Primers for Chamaecyparis obtusa (Cupressaceae), Appl. Plant Sci. 4
The model performance showed high sensitivity and precision in (2016) 1500136, https://doi.org/10.3732/apps.1500136.

67
P. Mukasa, et al. Infrared Physics and Technology 98 (2019) 62–68

[2] Y. Kijidani, N. Sakai, K. Kimura, Y. Fujisawa, Y. Hiraoka, J. Matsumura, S. Koga, seeds in response to stratification and smoke treatments, and detection of insect-
Termite resistance and color of heartwood of hinoki (Chamaecyparis obtusa) trees damaged seeds with VIS + NIR spectroscopy, New For. 33 (2007) 155–169,
in 5 half-sib families in a progeny test stand in Kyushu, Japan, J. Wood Sci. 58 https://doi.org/10.1007/s11056-006-9020-9.
(2012) 471–478, https://doi.org/10.1007/s10086-012-1285-3. [15] L. Workman, J. Weyer, Practical Guide and Spectral Atlas for Interpretive Near-
[3] S.H. Lee, H.S. Do, K.J. Min, Effects of essential oil from Hinoki Cypress, Infrared Spectroscopy, second ed., CRC Press, 2012.
Chamaecyparis obtusa, on physiology and behavior of flies, PLoS One 10 (2015) [16] L. Esteve Agelet, D.D. Ellis, S. Duvick, A.S. Goggi, C.R. Hurburgh, C.A. Gardner,
1–14, https://doi.org/10.1371/journal.pone.0143450. Feasibility of near infrared spectroscopy for analyzing corn kernel damage and
[4] J.M. Baskin, The woody plant seed manual, Nativ. Plants J. 10 (2009) 300–301, viability of soybean and corn kernels, J. Cereal Sci. 55 (2012) 160–165, https://doi.
https://doi.org/10.2979/NPJ.2009.10.3.300. org/10.1016/J.JCS.2011.11.002.
[5] O. Matsuda, M. Hara, H. Tobita, K. Yazaki, T. Nakagawa, K. Shimizu, A. Uemura, [17] D.D. Wang, F.E. Dowell, D.S. Chung, Assessment of heat-damaged wheat kernels
H. Utsugi, Determination of seed soundness in conifers cryptomeria japonica and using near-infrared spectroscopy, p. 1, 2001 Sacramento, CA July 29–August 1,
chamaecyparis obtusa using narrow-multiband spectral imaging in the short-wa- 2001, American Society of Agricultural and Biological Engineers, St. Joseph, MI,
velength infrared range, PLoS One 10 (2015) 1–21, https://doi.org/10.1371/ 2001https://doi.org/10.13031/2013.7466.
journal.pone.0128358. [18] M.I.K. Giurizatto, O. Ferrarese-Filho, M.De.L.L. Ferrarese, A.D. Robaina,
[6] M. Farhadi, M. Tigabu, L.G. Stener, P.C. Odén, Feasibility of visible + near infrared M.C. Gonçalves, C.A.L. Cardoso, α-Tocopherol levels in natural and artificial aging
spectroscopy for non-destructive verification of European × Japanese larch hybrid of soybean seeds, Acta Sci. Agron. 34 (2012) 339–343, https://doi.org/10.1590/
seeds, New For. 47 (2016) 271–285, https://doi.org/10.1007/s11056-015-9514-4. s1807-86212012000300015.
[7] F.Y. McDonald, M.B. Kwong, Flower Seeds: Biology and Technology, CABI [19] IST International Seed Testing Association, Seed science and technology, Seed Sci.
Publishing, Wallingford, 2005https://doi.org/10.1079/9780851999067.0000. Technol. 13 (1985) 299–513.
[8] L.C. Deleuran, M.H. Olesen, N. Shetty, R. Gislum, B. Boelt, Importance of seed [20] T. Naes, T. Isaksson, T. Fearn, T. Davies, Introduction a User-Friendly Guide to
quality for the fresh cut chain, Acta Hortic. (2011) 35–40. Multivariate Calibration and Classification, NIR Publications, 2002.
[9] J. Harmond, N. Brandenburg, L. Klein, Mechanical seed cleaning and handling, [21] A. Candolfi, R. De Maesschalck, D. Jouan-Rimbaud, P.A. Hailey, D.L. Massart, The
agriculture handbook no. 354, Mech. Seed Clean. Handl. 1968, p. 58. influence of data pre-processing in the pattern recognition of excipients near-in-
[10] R.P. Karrfalt, Producing the Target Seed: Seed Collection, Treatment, and Storage frared spectra, J. Pharm. Biomed. Anal. 21 (1999) 115–132, https://doi.org/10.
(Technical Coord. Natl. Proc. For. Conserv. Nurs. Assoc.), in: L.E. Riley, D.L. Haase, 1080/13554790490495140.
J.R. Pinto (Eds.), Proc. RMRS-P-65, USDA, Rocky Mt. Res. Station, 2011, pp. 67–73. [22] Å. Rinnan, F. van den Berg, S.B. Engelsen, Review of the most common pre-pro-
[11] A. Ambrose, S. Lohumi, W.H. Lee, B.K. Cho, Comparative nondestructive mea- cessing techniques for near-infrared spectra, TrAC – Trends Anal. Chem. 28 (2009)
surement of corn seed viability using Fourier transform near-infrared (FT-NIR) and 1201–1222, https://doi.org/10.1016/j.trac.2009.07.007.
Raman spectroscopy, Sens. Actuators, B Chem. 224 (2016) 500–506, https://doi. [23] B.G. Osborne, T. Fearn, P.T. Hindle, Practical NIR Spectroscopy with Applications in
org/10.1016/j.snb.2015.10.082. Food and Beverage Analysis, Longman Scientific and Technical, Harlow, Essex,
[12] M. Tigabu, P.E.R.C. Oden, Simultaneous detection of filled, empty and insect- in- England, 1993.
fested seeds of three Larix species with single seed near-infrared transmittance [24] L. Eriksson, T. Byrne, E. Johansson, J. Trygg, C. Vikström, Multi-and Megavariate
spectroscopy, New For. 27 (2004) 39–53, https://doi.org/10.1023/ Data Analysis Basic Principles and Applications, Umetrics Academy, 2013.
A:1025037408922. [25] M. Tigabu, P.C. Odén, Discrimination of viable and empty seeds of Pinus patula
[13] A. Soltani, T.A. Lestander, M. Tigabu, P.C. Odén, Prediction of viability of oriental Schiede & Deppe with near-infrared spectroscopy, New For. 25 (2003) 163–176,
beechnuts, Fagus orientalis, using near infrared spectroscopy and partial least https://doi.org/10.1023/A:1022916615477.
squares regression, J. Near Infrared Spectrosc. 11 (2003) 357–364, https://doi.org/ [26] K.S. Kim, S.H. Park, M.G. Choung, Y.S. Jang, Use of near-infrared spectroscopy for
10.1255/jnirs.387. estimating fatty acid composition in intact seeds of rapeseed, J. Crop. Sci.
[14] M. Tigabu, J. Fjellström, P.C. Odén, D. Teketay, Germination of Juniperus procera Biotechnol. 10 (2007) 15–20.

68