Geoderma 330 (2018) 254–263

Contents lists available at ScienceDirect

Geoderma
journal homepage: www.elsevier.com/locate/geoderma

Laboratory-based hyperspectral image analysis for predicting soil carbon, T

nitrogen and their isotopic compositions
⁎ ⁎
Iman Tahmasbiana, , Zhihong Xua, , Sue Boyda, Jun Zhoub, Roya Esmaeilanic, Rongxiao Chea,
Shahla Hosseini Baid
a
Environmental Futures Research Institute, School of Environment and Science, Griffith University, Nathan, Brisbane, Queensland 4111, Australia
b
School of Information and Communication Technology, Griffith University, Nathan, Brisbane, Queensland 4111, Australia
c
Faculty of Computing, University Technology Malaysia 81300, Skudai, Johor, Malaysia
d
Faculty of Science, Health, Education and Engineering, University of the Sunshine Coast, Maroochydore DC, QLD, 4558, Australia

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

Handling Editor: A.B. McBratney The common methods of determining soil carbon (C), nitrogen (N) and their isotopic compositions (δ13C and
Keywords: δ15N) are expensive and time-consuming. Therefore, alternative low-cost and rapid methods are sought to ad-
Native forest dress this issue. This study aimed to investigate the potential of hyperspectral image analysis to predict soil total
Chemometrics carbon (TC), total nitrogen (TN), δ13C and δ15N. Hyperspectral images were captured from 96 ground soil
Multivariate analysis samples using a laboratory-based visible to near-infrared (VNIR) hyperspectral camera in the spectral range of
PLSR, VNIR, δ13C, δ15N 400–1000 nm. Partial least squares regression (PLSR) models were developed to correlate the values of TC, TN,
δ13C and δ15N, obtained from isotope ratio mass spectrometry method, with their spectral reflectance. The
developed models provided acceptable predictions with high coefficient of determination (R2c) and low root
mean square error (RMSEc) of calibration set for TC (R2c = 0.82; RMSEc = 1.08%), TN (R2c = 0.87;
RMSEc = 0.02%), δ13C (R2c = 0.82; RMSEc = 0.27‰) and δ15N (R2c = 0.90; RMSEc = 0.29‰). The prediction
abilities of the models were then evaluated using the spectra of an external test set (24 samples). The models
provided excellent predictions with high R2t and ratio of performance to deviation (RPD) of test set for TC
(R2t = 0.76; RPD = 2.02), TN (R2t = 0.86; RPD = 2.08), δ13C (R2t = 0.80; RPD = 2.00) and δ15N (R2t = 0.81;
RPD = 1.94). The results indicated that the laboratory-based hyperspectral image analysis has the potential to
predict soil TC, TN, δ13C and δ15N.

1. Introduction
The biogeochemical cycles of carbon (C) and nitrogen (N) in ter-
restrial ecosystems have received considerable attention because of
their significant role in climate change (Fu et al., 2010; Zaehle et al.,
2010). The C and N pools are key indicators of soil quality and it is
therefore important that they are monitored in terrestrial ecosystems
(Che et al., 2018; Darby et al., 2016; Hosseini Bai et al., 2015a; Nguyen
et al., 2017). The soil natural isotopic abundance of C and N (δ13C and
δ15N) have also been widely used as powerful tools to gain insight into
C and N cycling in the terrestrial ecosystems (Bai et al., 2017; Che et al.,
2017; Fujiyoshi et al., 2017; Hosseini Bai et al., 2017; Hosseini Bai
et al., 2015b; Wang et al., 2015). The isotopic compositions, vary over
the time due to different processes such as microbial activities, plant
uptake and leaching (Farquhar et al., 1989; Högberg, 1997), and need
to be analyse frequently. Conventional methods, such as isotope ratio
mass spectroscopy, used for analysing soil total C (TC), total N (TN),
δ13C and δ15N, however, are time-consuming and expensive when ap-
plied to a large number of samples (Jiang et al., 2017). Hence, the al-
ternative inexpensive and rapid methods are sought to predict soil TC,
TN, δ13C and δ15N.
The visible to near infrared (VNIR) airborne/laboratory-based hy-
perspectral imaging technique (400–2500 nm), which uses the ad-
vantages of both conventional imaging and VNIR spectroscopy, is a
physical, rapid, reproducible and low-cost alternative to conventional
methods of soil chemical analysis (Gmur et al., 2012; Jung et al., 2015;
Stevens et al., 2008; Vaudour et al., 2016). More importantly, hyper-
spectral imaging permits a quantitative assessment of different soil
properties from a single measurement (Ben-Dor and Banin, 1995a; Jia
et al., 2016; Viscarra Rossel et al., 2006; Xu et al., 2016). The spatial
dimension provided by airborne/laboratory-based hyperspectral ima-
ging technology gives this method an advantage over the VNIR spec-
troscopy (Manley, 2014). For example, the spatial dimensions are used
to obtain the chemical image/map for identification and localisation of

⁎
Corresponding authors at: Environmental Future Research Institute, Griffith University, Nathan, Queensland 4111, Australia.
E-mail addresses: iman.tahmasbian@griffithuni.edu.au (I. Tahmasbian), zhihong.xu@griffith.edu.au (Z. Xu).

https://doi.org/10.1016/j.geoderma.2018.06.008
Received 22 September 2017; Received in revised form 20 April 2018; Accepted 11 June 2018
0016-7061/ © 2018 Elsevier B.V. All rights reserved.
I. Tahmasbian et al.
chemical compounds in the samples with non-homogenous surfaces
(Manley, 2014). However, for homogenised samples such as ground
soils, the spatial information are used by averaging the several pixels in
the spatial dimensions of hyperspectral images (Manley, 2014). Using
the average of the pixels increases the signal to noise ratio (Manley,
2014), and therefore improves the accuracy and reproducibility of
predictions compared with the single points used in VNIR spectroscopy.
One of the main disadvantages of hyperspectral imaging compared with
VNIR spectroscopy include the high cost of hyperspectral sensors,
especially when wavelengths between 1000 nm to 2500 nm are re-
quired (Manley, 2014). Fast computers and substantial data storage
capacity are also required due to the large size of hyperspectral images
(Manley, 2014). Despite the higher accuracy and capability of labora-
tory-based sensors to measure both surface and deep soils, they can
cover smaller areas compared with airborne hyperspectral systems.
The hyperspectral imaging technique utilises the reflectance mea-
sured on different materials using airborne or laboratory-based hyper-
spectral sensors (Bai et al., 2018; Ben-Dor et al., 1997; Gama et al.,
2018; Grahn and Geladi, 2007; Manley, 2014). Multivariate analysis is
then used to correlate the spectral reflectance data with their corre-
sponding reference values of the targeted variables measured using the
reference chemical methods e.g., isotope ratio mass spectrometry
(Manley, 2014; Wijewardane et al., 2016). However, different pre-
processing techniques such as wavelength selection and spectral data
transformation are usually applied to reduce the collinearity of the
spectral data and the impact of artefacts, such as random noise and light
scattering (Tahmasbian et al., 2018). The data pre-processing usually
increase the signal to noise ratio in the spectral data prior to performing
the multivariate analyses (Manley, 2014; Wei et al., 2017). First deri-
vative, second derivative, multiplicative scatter correction (MSC), or-
thogonal signal correction (OSC), standard normal variate (SNV) and
detrending (DT) are the most common transformation techniques ap-
plied to hyperspectral reflectance data in wide range of studies (Fearn,
2000; Kamruzzaman et al., 2016b; Lin et al., 2016; Rinnan et al., 2009;
Siripatrawan et al., 2011). The best-transformed data are then used to
develop the multivariate models using the selected important wave-
lengths. This step employs an appropriate chemometric algorithm such
as partial least squares regression (PLSR).
PLSR is the most commonly used multivariate linear model for
chemometrics (Coûteaux et al., 2003; Jiang et al., 2017; Mouazen et al.,
2007). PLSR enables us to analyse small data sets with multiple vari-
ables, which are strongly correlated and noisy (Höskuldsson, 1988;
Wold et al., 1984; Wold et al., 2001). PLSR finds a few linear combi-
nations (latent variables) of the original X-values and uses only those
linear combinations in the regression equation. This leads to discarding
the irrelevant and unstable information to use the most relevant X-
variation for the regression analysis (Höskuldsson, 1988; Wold et al.,
1984; Wold et al., 2001). The developed PLSR models are then used to
predict the target variables in new samples.
The VNIR airborne hyperspectral imaging and VNIR spectroscopy
analyses have been used to analyse soil C and N in a diverse type of soils
(Barthès et al., 2006; Brunet et al., 2007; Denis et al., 2014; Stevens
et al., 2008; Udelhoven et al., 2003; Wei et al., 2017). For example, soil
organic and total C, as well as soil TN, have been analysed using air-
borne hyperspectral imaging and VNIR spectroscopy in the spectral
region of 350–2500 nm; these studies provided results with a wide
range of accuracies, R2 from 0.35 to over 0.90 (Cozzolino et al., 2013;
Gmur et al., 2012; Gopal et al., 2015; Jiang et al., 2017; Stevens et al.,
2008). Limited information, however, is available on using laboratory-
based hyperspectral images for the prediction of TC and TN in soil
samples. To the best of our knowledge, there is also no report on using
the laboratory-based hyperspectral imaging sensors to predict soil δ13C
and δ15N. The main objective of this study was to assess the potential of
using the laboratory-based hyperspectral imaging systems to predict
TC, TN, δ13C and δ15N in soil samples. The models developed in this
study will be used to predict the long-term dynamics of soil TC, TN,
255
Geoderma 330 (2018) 254–263
δ13C and δ15N in real-time after prescribed burning in Toohey Forest,
East Queensland, Australia.
The techniques used in airborne hyperspectral imaging and VNIR
spectroscopy follow the same principles as those of laboratory-based
hyperspectral imaging. Therefore, it was hypothesised that the con-
centration of TC and TN in soil samples could be predicted using the
spectra extracted from laboratory-based hyperspectral images in the
spectral region of 400–1000 nm. Furthermore, the featureless (without
a specific spectral response) properties in soil samples can also be
predicted using their internal correlation with other properties which
possess specific spectral responses (Ben-Dor and Banin, 1995b; Ben-Dor
et al., 2002). Soil TC, TN and their isotopic compositions, δ13C and δ15N
have shown to be correlated (Abdullah, 2016; Hobbie et al., 2000; West
et al., 2009). Therefore, it was also hypothesised that δ13C and δ15N can
be predicted using laboratory-based hyperspectral reflectance in the
visible to near infrared regions (400–1000 nm) due to their correlation
with TC and TN, respectively. We additionally hypothesised that im-
portant wavelengths selection would improve the accuracy of the de-
veloped models by reducing the collinearity of the spectral data.
2. Materials and methods
2.1. Study site description
Soil samples used in this experiment were collected from Toohey
Forest, a suburban native forest located in south-eastern Queensland,
Australia (27°30′S, 153°E) (Hosseini Bai et al., 2014a). Toohey Forest is
located in a subtropical region, with average minimum temperature
(1939–2016) of 14.4 °C, average maximum temperature (1939–2016)
of 26.3 °C and mean annual precipitation (1969–2016) of 1163.5 mm.
Geological units of Toohey Forest were diverse, and include argil-
lites, shales, sandstones and shales of the Woogaroo subgroup; soft se-
dimentary rocks of the Tingalpa formation (Ipswich coal measures);
cherts and arsenite of the Neranleigh-Fernvale beds; and sandstones of
the Sunnybank formation (Catterall et al., 2001a). The most forest soils
were thin lithosols, with a thickness varying between 60 cm and 1 m
(Abdullah, 2016). The lithosols contained both stony components and
loam to silty loam, which may be rich in humus (Catterall et al., 2001a).
The soil pH varied between 4.4 and 5.5. The forest was subjected to
prescribed burning since 1993 with different intervals for different sites
(Butler et al., 2016; Catterall et al., 2001b).
2.2. Site establishment, soil sample collection and chemical analyses
Two study sites were established randomly in different sectors of
Toohey Forest, included Site 1 (27° 32′ 39.42̋ ″ S, 153° 03' 6.81″ E) and
Site 2 (27° 32′ 45.33″ S, 153° 03′ 14.72″ E). The two study sites covered
an area of 4000 m2. Each site had four randomly established circular
plots (radius of 12.62 m, area of 500 m2). Each plot was divided into
four sampling areas (ca. 125 m2) for soil collection. The soils samples
were collected separately from the depths of 0–5 and 5–10 cm of the
sampling areas to increase the variability of data and make a wide-
range data set. The soil samples were collected from different locations
within the sampling areas at each sampling time.
A total of 120 soil samples were collected from the study sites. Soil
samples from Site 1 (56 samples) were collected seasonally (every three
months), within an 18-month period, starting in November 2014. Soil
samples from Site 2 (64 samples) were collected monthly in the first six
months and then seasonally (every three months) for the next six
months, within a 12-month period, starting in Jun 2015.
The collected soil samples were carried to the laboratory in separate
plastic zipper bags and were air-dried in the Griffith University Soil
Laboratory. The air-dried soil samples were sieved with a 2-mm sieve
and ground into fine powder. Approximately 20 mg of ground soil
samples were transferred into tin capsules for TC, TN, δ13C and δ15N
analyses using an isotope ratio mass spectrometer (Hosseini Bai et al.,
I. Tahmasbian et al.
2014b). The values for δ13C and δ15N are defined in Eq. (1):
Rsample − Rstandard
δ (‰) = × 1000
Rstandard
where δ is the C or N isotopic composition, Rsample is the C/ C
13 15 13
ratio of a sample for δ13C or 15N/14N for δ15N, and Rstandard is the
13
C/12C ratio of the international PeeDee Belemnite standard for δ13C
or atmospheric N2 for δ15N (Prasolova et al., 2000; Shearer and Kohl,
1986).
2.3. Hyperspectral imaging system and image acquisition
A visible-near infrared (VNIR) hyperspectral imaging system, lo-
cated in Ecology Laboratory, University of the Sunshine Coast, was used
in this experiment to capture images from soil samples. The system
consisted of a 12-bit line scanner camera (Pika XC2, USA), operated in
the spectral range of 400 nm to 1000 nm, with a spectral resolution of
1.3 nm producing a total 462 spectral band; a lens with 23 mm of focal
length; four current-controlled wide spectrum quartz halogen lights; a
linear moving stage operated by a stepper motor; and data acquisition
software (SpectrononPro 2.94, Resonon, USA). The system is illustrated
in Fig. 1.
Approximately 10 g of the ground soil samples were placed in
plastic weigh-trays (2 cm × 2 cm) and positioned on the moving stage
and moved into the camera's field of view. The speed of moving stage
and the exposure time were set at 147 pps and 27.31 ms, respectively.
Samples were shaken after capturing each image for homogenisation
and imaging was repeated until reproducible spectral signatures were
obtained for consecutive images (Tahmasbian et al., 2017).
2.4. Spectral profile extraction and data calibration
Sample images were cropped from the background. Approximately
100,000 (317 × 317) pixels were selected from the cropped images for
data extraction. The averages of the corrected reflectance (R) of the
selected pixels were extracted using the SpectrononPro software
package (version 2.94, Resonon, USA) (Supplementary material Fig.
S1). Image correction was conducted by dark and white calibration
during the image acquisition using the SpectrononPro software. Dark
calibration was applied by capturing an image in a dark room with the
lens cap on, prior to capturing the images from samples. White cali-
bration was applied by capturing the reflectance of a white Teflon sheet
(Resonon, USA) with the reflectance of approximately 99%. The R was
then calculated from the uncorrected spectral reflectance (R0) using Eq.
(2) (Ariana et al., 2006):
R = (R 0 − D)/(W − D)
Fig. 1. Hyperspectral imaging system used in the current experiment.
Geoderma 330 (2018) 254–263
where D is dark current and W is the reflectance of white Teflon sheet.
Fig. 2 illustrates the whole procedure of hyperspectral image analysis.
(1)
2.5. Outliers detection and data transformation
12
Spectral outliers can influence the robustness of the developed
models and therefore should be detected and removed from the datasets
(Kamruzzaman et al., 2016a; Morellos et al., 2016). Principal-compo-
nent-analysis (PCA) was performed to identify the spectral outliers
using a Hotelling's T2 test within 99% of confidence level in the current
study (Jiang et al., 2016; Li and Qu, 2016; Morellos et al., 2016).
In order to have an independent dataset to test the precision of final
models, the full dataset was then randomly divided into calibration sets
(80% of the data) and test sets (20% of the data) for each target variable
(TC, TN, δ13C and δ15N) (Kohavi, 1995). The test dataset was kept se-
parately to examine the accuracy of models developed using calibration
dataset. A t-test was carried out to compare both sets, to confirm con-
sistent coverage of the whole range of values for target variables
(Barbin et al., 2012).
A wide range of transformation techniques were performed on the
spectra obtained from calibration samples, to remove the influence of
undesired effects and improve the signal to noise ratio (Rinnan et al.,
2009). The transformation techniques included first derivative, second
derivative, orthogonal signal correction (OSC), multiplicative scatter
correction (MSC), reciprocal logarithm (RL) and standard normal
variate (SNV) (Fearn, 2000; Gao et al., 2014; Lin et al., 2016; Rinnan
et al., 2009; Wei et al., 2017). Outlier detection and data transformation
were performed using the Unscrambler® X software package (ver-
sion10.4, CAMO Software Inc., Trondheim, Norway).
2.6. Initial model development and evaluation
Initial PLSR models were developed using raw and transformed
calibration data to correlate the values of TC, TN, δ13C and δ15N in soil
samples with the corresponding hyperspectral reflectance ranging from
400 nm to 1000 nm.
In general, PLSR models find a few orthogonal variables in the X
matrix, which are estimations of latent variables (LV). The LVs are then
used to estimate the dependent variable Y. The algorithm can be
summarised in Eq. (3).
Y = Xβ + F (3)
where β is the regression coefficient matrix obtained from X and Y-
weight matrices and F is the matrix of Y-residuals (Wold et al., 2001).
The optimal number of LVs were selected at the minimum predicted
(2) residual error sum of squares (PRESS) obtained from the cross-valida-
tion sets (Abdi, 2010). PRESS was calculated as shown in Eq. (4) (Abdi,
2010; Li et al., 2002).
n
PRESS =  2
∑ (Y–Y)
i i
i=1 (4)
where Yi and Ŷi are the reference and predicted values of each target
variable in the i th sample, respectively.
The cross-validation was a k-fold (k = 10) method (Kohavi, 1995);
in general data are split into k sets for a k-fold cross-validation. In this
case, the models were developed iteratively using k–1 sets, with one set
left out for validation on each iteration. The process was repeated until
all the sets were left out of the models once (Afandi et al., 2016; Kohavi,
1995; Parsons et al., 2011).
The best-fitted models were selected by choosing the models with
the highest coefficient of determination in the calibration sets (R2c) and
cross-validation sets (R2cv) and the lowest root mean square error in the
calibration sets (RMSEc) and cross-validation sets (RMSEcv). R2 and
RMSE are defined in Eqs. (5) and (6), respectively (Barrett, 1974;
Bellon-Maurel et al., 2010). The biases of the predictions in the
256
I. Tahmasbian et al. Geoderma 330 (2018) 254–263

Fig. 2. The procedure of hyperspectral image analysis and model development.

calibration and cross-validation sets were also calculated using Eq. (7). 2.8. Identification of the most important/principal wavelengths in the final
n 2
models
∑i = 1 (Yi − Ŷ)
i
R2 = 1 − n
∑i = 1 (Yi − Ȳ)2 (5) After the final models were developed using informative wave-
lengths, the most important/principal wavelengths in the final models
n
∑i = 1 (Yi − Ŷ)
i
2 were selected, using both the β coefficient and variable importance in
RMSE = projection (VIP) as defined in Eq. (8). This was done in an effort to
N (6)
simplify/highlight the important features of the overall results obtained
n
i − Y)
(Y i
after modelling (Chong and Jun, 2005).
Bias = ∑ N
i=1 (7) F
∑ f = 1 w jf2 . SSYf . J
VIPj =
where Yi and Ŷi are the reference and predicted values of target vari- SSYt . F (8)
ables in the i th sample, respectively. The Ȳi is the mean of reference
th
values, while N is the number of samples. where, VIPj was the importance of j wavelength in the model with F
number of components, Wjf was the loading weight of the corre-
2.7. Uninformative wavelengths elimination and final model development sponding wavelength in the f th component, SSYf was the explained sum
of squares of the target variable in the f th component, SSYt was the total
The VNIR hyperspectral imaging sensor provides large sets of sum of squares of the target variable and J was to total number of
strongly collinear wavelengths, and these may adversely affect the wavelengths.
success of soil TC, TN, δ13C and δ15N quantification. Although the PLSR The wavelengths with a β coefficient larger than their standard
latent variables partly compensate for these effects, due to the projec- deviation, and the corresponding VIP values > 1, were identified as
tion of the spectral data into a low-dimensional space, further im- principal wavelengths and selected for further discussion (Li et al.,
provements in the accuracy of the PLSR model, can be achieved by 2006).
eliminating uninformative and/or noisy variables (Kamruzzaman et al.,
2012; Vohland et al., 2016; Wold et al., 2001). 2.9. Evaluating model performance using external test sets
In the current study, the wavelengths having low β coefficient
(uninformative wavelengths) were removed from the initial models and The developed models were then tested using the external test set.
the final PLSR models were redeveloped using the remaining (in- The accuracies of the models were evaluated by the R2 and RMSE of the
formative) wavelengths (Chong and Jun, 2005). The removed/ne- test set (R2t and RMSEt). The ratio of performance to deviation (RPD)
glected wavelengths were returned to the model in the case where R2cv was also used to assess the prediction ability of the developed models.
of the final model was less than the R2cv of the initial model The RPD is defined in Eq. (8) (Morellos et al., 2016).
(Tahmasbian et al., 2017). The process was repeated until the highest
R2cv was achieved. RPD = SDt /RMSEt (9)

257
I. Tahmasbian et al. Geoderma 330 (2018) 254–263

3.2.2. Uninformative wavelengths elimination and final model

Fig. 3. Hotelling's T2 test conducted for identification of outliers in the spectra.
The circle shows the 99% of confidence level. The points out of the circuit were
considered as outliers.
where SDt is the standard deviation of the predicted values in the test
set.
The PLSR models were developed using the Unscrambler® X soft-
ware package (version10.4, CAMO Software Inc., Trondheim, Norway).
The RPD was calculated using Microsoft Excel 365 ProPlus.
3. Results
redevelopment
Removing the uninformative wavelengths from the models reduced
the number of wavelengths used in the final models without reducing
the accuracies.
Uninformative wavelengths (low β coefficient) were eliminated
from the transformed spectra and the final models were then developed
using the total number of 438, 377, 378 and 44 informative wave-
lengths remained in the models for TC, TN and δ13C δ15N, respectively
(Supplementary material Fig. S3a-d).The transformations used for de-
veloping the final models were not changed from the initial models. The
final models, developed using only informative wavelengths, were also
fitted well to the values of TC, TN, δ13C and δ15N in the corresponding
soil samples with the R2c of 0.82, 0.87, 0.82 and 0.90; R2cv of 0.71, 0.81,
0.70 and 0.85; RMSEc of 1.08%, 0.02%, 0.27‰ and 0.29‰ and RMSEcv
of 1.21%, 0.03%, 0.35‰ and 0.36‰ respectively. The bias values of all
the developed models were insignificant (absolute bias < 0.001) for
both calibration and cross-validation sets (Fig. 4a-d).
The results showed that the accuracies of the final models with only
informative wavelengths were slightly improved for all TC, TN and δ13C
δ15N variables, comparing to the initial models with the full number of
wavelengths (Table 3 vs. Fig. 4a-d).
The number of LVs selected for TC, TN, δ13C and δ15N were 7, 2, 12
and 3, respectively. These were not changed from initial to final models
and hence there were no changes in the dimensionality of the initial and
final models.

3.1. Descriptive analysis
No outlier was identified in the spectral dataset as analysed using
Hotelling's T2 test (Fig. 3). The dataset was randomly divided into ca-
libration and test sets using 96 samples (80% of the data) and 24
samples (20% of the data), respectively. The spatial distribution of the
calibration and test samples were presented in Table 1. The averages of
TC, TN, δ13C and δ15N in the calibration and the test datasets were
5.99% and 5.86%, 0.19% and 0.18%, − 26.59 and – 26.73, and – 0.19
and – 0.48, respectively. There were no significant differences between
the averages in the calibration and the test datasets using t-test
(p < 0.05). The descriptive analysis of the calibration and the test
datasets were reported in Table 2.
3.2. Attributes of the models developed using calibration datasets
3.2.1. Transformation technique selection and initial model development
Different transformation techniques were investigated to increase
the signal to noise ratio in the spectral data. The initial models were
developed using differently transformed data as well as raw data. The
transformations leading to highest R2c and R2cv and lowest RMSEc and
RMSEcv were selected as the best transformation techniques.
The best initial models, with the highest R2c and R2cv were developed
using the second derivative, SNV-OSC and SNV-DT-OSC transformation
techniques for TC, TN and δ15N, respectively (Table 3, Supplementary
material Fig. S2b-d). However, the best initial model for δ13C was de-
veloped using the raw spectral data (Table 3, Supplementary material
Fig. S2a).
3.2.3. Identification of most important/principal wavelengths in the final
models
In our study, TC, TN, δ13C and δ15N had 21, 29, 10 and 6 most
important/principal wavelengths respectively (Table 4; Fig. S3a-d). The
principal wavelengths were selected from informative wavelengths
when the wavelengths had β coefficient larger than their corresponding
standard deviation and indicated VIP > 1 (Jiang et al., 2017).
3.3. Evaluating model performance using external test sets
The developed models successfully predicted all TC, TN, δ13C and
δ15N values of the corresponding test datasets (Fig. 5a-d). The results
showed an “excellent” prediction of TC using the PLSR model with the
R2t of 0.76, RMSEt of 1.27%, RPD of 2.02 and biast of – 0.15% (Fig. 5a).
A regression model is considered “excellent” when the RPD calculated
for prediction is above 2 (Bellon-Maurel et al., 2010). The values of TN
in the test set were also predicted “excellently” using the developed
PLSR with the R2t of 0.86, RMSEt of 0.04%, RPD of 2.08 and biast of –
0.006% (Fig. 5b). An “excellent” prediction of δ13C in the test dataset
was also achieved with the R2t of 0.80, RMSEt of 0.33‰, RPD of 2.00
and biast of – 0.12‰ (Fig. 5c). The prediction of δ15N in the test set
showed the R2t of 0.81, RMSEt of 0.50‰, RPD of 1.94 and biast of –
0.24‰. The RPD calculated for δ15N was between 1.4 and 2, which is
considered as “good” (Fig. 5d) (Bellon-Maurel et al., 2010). The dy-
namics of TC, TN, δ13C and δ15N during the time, predicted using the
developed models in the test sets, were illustrated in supplementary
material Fig. S4.
4. Discussion

Table 1 The final PLSR models, developed using the selected informative
Distribution of the samples collected from different plots (P) of the Sites 1 and 2 wavelengths, provided acceptable correlations for soil TC, TN, δ13C and
in the calibration and test sets. δ15N with their corresponding values obtained from isotope ratio mass
Calibration set Test set spectrometry analysis in the calibration and test datasets.

P1 P2 P3 P4 Total P1 P2 P3 P4 Total 4.1. Attributes of the models developed using calibration datasets
Site 1 11 12 9 11 43 3 3 4 3 13
Site 2 15 13 14 11 53 2 3 4 2 11 Developing the models using the selected informative wavelength
slightly increased the accuracies of the final models, developed to

258
I. Tahmasbian et al. Geoderma 330 (2018) 254–263

Table 2
Descriptive analysis of total carbon (TC), total nitrogen (TN), carbon isotopic composition (δ13C) and nitrogen isotopic composition (δ15N) in the calibration and test
sets.
Calibration (96 sample) Test set (24 sample)

TC (%) TN (%) δ13C (‰) δ15N (‰) TC (%) TN (%) δ13C (‰) δ15N (‰)

Average 5.99 0.19 −26.59 −0.19 5.86 0.18 −26.73 −0.48

Std 2.39 0.08 0.67 0.91 3.14 0.10 0.54 1.05
CV 0.40 0.41 −0.03 −4.91 0.54 0.57 −0.02 −2.16
Skewness 0.98 1.03 −0.66 −0.44 0.93 0.93 0.59 −0.40
Min 2.33 0.07 −29.10 −2.29 2.46 0.07 −27.50 −2.76
Max 14.27 0.47 −25.29 1.49 11.97 0.40 −25.58 1.04

Std: standard deviation; CV: Coefficient of variations.

Table 3 near 403, 470, 687, and 846 nm have been found to be principal for
Characteristics of the initial PLSR model developed for total carbon (TC), total predicting TN in soil (Gmur et al., 2012). The reflectance in 538 nm,
nitrogen (TN), carbon isotopic composition (δ13C) and nitrogen isotopic com- 617, 760, 905, 1094, 1239, 1442, 1477 and 1500 have been also re-
position (δ15N). ported as the principal wavelengths for predicting TN and nitrate-N in
TC (%) TN (%) δ13C (‰) δ15N (‰) soil samples (Gopal et al., 2015; Wei et al., 2017). The inconsistency in
the principal wavelengths reported for predicting TN in different stu-
TF SDr SNV + OSC – SNV + DT + OSC dies might be associated with indirect estimation of soil N in different
compounds (Blackmer et al., 1994; Li et al., 2014).
TF details Polynomial – – DT polynomial
order:2, Left order:2 There is no specific report on N reflectance in soil samples measured
points:12, Right using laboratory-based hyperspectral imaging systems. In plant tissues,
points:12 however, the reflectance of N in the visible (VIS) region (400–700 nm)
has been mainly attributed to the reflectance of chlorophyll (Curran,
WL No. 462 462 462 462
LV No. 7 2 12 3 1989; Ferwerda et al., 2005). Since we used a forest soil in this study,
R2c 0.82 0.86 0.82 0.88 the presence of plant residues containing chlorophyll in the soil samples
R2cv 0.67 0.79 0.68 0.81 was not unlikely. Therefore, the principal wavelengths in VIS region
RMSEc 1.08 0.02 0.27 0.31 might be associated with the presence of chlorophyll in the soil sam-
RMSEcv 1.49 0.03 0.37 0.37
ples. Reflectance in NIR region has been attributed to the weak over-
TF: transformation method; SDr: Second derivative; SNV: Standard normal tones and combination of fundamental vibrations caused by stretching
variate; OSC: Orthogonal signal correction; DT: Detrending; WL: wavelength; and bending of NeH, OeH and CeH groups, which may reflects the
LV: latent variable; R2c: coefficient of determination of calibration set; R2cv: presence of protein–N, starch or cellulose in the analysed soil samples
coefficient of determination of cross-validation set; RMSEc: root mean squared (Curran, 1989; Ferwerda et al., 2005; Wetzel, 1983). In addition, using
error of calibration set; RMSEcv: root mean squared error of cross-validation VNIR spectroscopy, overlaps were observed between the principal
set. wavelengths for predicting TC and TN. Overlaps between these prin-
cipal wavelengths have been attributed to the inherent links between C
predict TC, TN, δ13C and δ15N, compared to those of the initial models, and N (Jiang et al., 2017). We also observed high correlation (r = 0.97)
which was consistent with our hypothesis. Such results are also con- between TC and TN in the soil samples used in this study and this may
sistent with those described in the literature, which indicated an in- further explain the overlaps observed in our study for the principal
crease in the accuracy of models after elimination of uninformative wavelength of soil TC and TN.
wavelengths in PLSR modelling (Kamruzzaman et al., 2012; Wold et al., Most of the principal wavelengths for predicting δ13C were located
1996). This might be due to removing the inter-correlation from the in the 400–410 nm and 960–1000 nm regions, while those of δ15N were
wavelengths and facilitating the computation (Kamruzzaman et al., only located in the 950–1000 nm region. Although there is limited in-
2012; Tahmasbian et al., 2018; Wold et al., 1996). Therefore, it might formation on predicting δ13C and δ15N in soil samples, the possibility of
be concluded that eliminating the uninformative wavelengths reduces predicting δ13C and δ15N in canopy and leaf litterfall samples have been
the model's size without reducing the accuracy of the models. examined using airborne and laboratory-based hyperspectral imaging,
The most important/principal wavelengths selected for predicting as well as VNIR spectroscopy (Elmore and Craine, 2011; Lorentz, 2013;
TC were mostly observed to be in the spectral regions of 740–800 nm Tahmasbian et al., 2017; Wang et al., 2007). For example, in our
and 930–1000 nm. The importance of reflectance in the spectral region complementary study, we successfully predicted δ15N in leaf litterfall
of 700–1000 nm for predicting soil C content using airborne hyper- samples using a laboratory-based hyperspectral imaging system with a
spectral imaging and VNIR spectroscopy techniques has also been re- lower camera resolution than the present study (10 nm vs. 1.3 nm) but
ported (Cozzolino and Morón, 2006; Gmur et al., 2012; Ingleby and with sampling over a wider range of wavelengths (400 nm to 1700 nm).
Crowe, 2000; Mouazen et al., 2007; Yang and Li, 2013). The reflectance However, we were unable to predict δ13C in that study because of (a)
in the region between 700 mm and 800 nm has been suggested to be the small variation in δ13C in the samples studied and (b) the lower
associated with humic compounds, pigments derived from chlorophyll camera resolution (Tahmasbian et al., 2017). The suggested principal
and phenolic compounds during decomposition of organic matter and wavelengths for predicting δ15N in leaf litterfall samples are 590, 610,
plant residues in soil (Dalal and Henry, 1986; Daughtry, 2001; Fidêncio 650, 900 and 1270 nm (Tahmasbian et al., 2017). The δ15N has also
et al., 2002). been predicted successfully in fresh leaves, as well as dried and ground
The principal wavelengths for predicting soil TN were mainly ob- leaves, using VNIR spectroscopy in the spectral range of 350–2500 nm,
served in the 900–1000 nm region with a few principal wavelengths in with the highest correlation between reflectance spectra and δ15N lo-
400–410 nm. Various wavelengths have been reported as principal cated in 400–600 nm and 1500–2000 nm regions (Elmore and Craine,
wavelengths for predicting TN, which in some cases are in the same 2011).
regions as those of this study. For example, the wavelengths centred Hobbie and co-workers and West and co-workers have all suggested

259
I. Tahmasbian et al.
Fig. 4. Reference vs. predicted values of soil (a) total carbon (TC), (b) total nitrogen (TN), (c) carbon isotopic composition (δ13C) and (d) nitrogen isotopic com-
position (δ15N) in the calibration sets. The dotted lines are regression lines for calibration sets.
Table 4
The most important/principal wavelengths selected based on variable im-
portance in projection (VIP) and β coefficient for predicting TC, TN, δ13C and
δ15N. The threshold for VIP was set to 1, and the threshold for β coefficient was
their standard deviation.
Principal wavelengths (nm)
TC 407.21, 743.86, 745.2, 761.32, 762.66, 764.01, 792.26, 939.84, 941.2,
942.57, 957.55, 971.18, 972.55, 973.91, 975.28, 976.64, 979.37, 980.74,
984.83, 987.56, 988.93
TN 401.97, 403.28, 407.21, 874.66, 909.93, 912.65, 914.01, 945.29, 948.01,
950.74, 956.19, 957.55, 963, 964.37, 967.09, 968.46, 971.18, 972.55,
973.91, 976.64, 979.37, 980.74, 982.1, 984.83, 987.56, 994.39, 995.76,
997.13, 999.86
δ13C 400.67, 405.9, 407.21, 778.8, 961.64, 973.91, 976.64, 978.01, 984.83,
994.39
δ15N 953.46, 967.09, 973.91, 975.28, 986.2, 998.49
that it is the correlation between TN and δ15N that enables the quan-
tification of δ15N using reflectance data (Hobbie et al., 2000; West
et al., 2009). However, in our case, we found no significant correlation
between TN and δ15N (r = 0.02). The soils studied in this experiment
were course-grained shallow acidic soils, located in a subtropical region
with hot summers and heavy showers in wet seasons. These conditions
might have influenced N cycling through increasing leaching or vola-
tilization which alter the soil δ15N and may explain the lack of re-
lationship between TN and δ15N. Other studies have also reported that
tropical forests have more open N cycles with greater losses via frac-
tionation pathways than temperate forests, resulting in significant
variations in soil δ15N (Martinelli et al., 1999; Piccolo et al., 1996). We
are unsure of why we were able to predict both TN and δ15N in-
dependently, however, other authors have mentioned that successful
quantification of δ15N may be affected by absorption features related to
the wave motion of molecular bonds involving N, and this might be
altered in samples with considerable values of 15N (Elmore and Craine,
2011; Lorentz, 2013). A significant correlation (r = − 0.50) between
TC and δ13C in our soil samples was also observed. Therefore, we think
that the successful prediction of soil δ13C using hyperspectral imaging
260
Geoderma 330 (2018) 254–263
technique in the current study might be related to the well-established
relationship between reflectance data and C content in soil (Gmur et al.,
2012; Yang and Li, 2013).
The number of principal wavelengths used for predicting δ13C and
δ15N were considerably lower than those of TC and TN. This is because
quantification of TC and TN by hyperspectral imaging is conducted
using the reflectance of all compounds containing C and N. However,
13
C and 15N may be enriched in only some types of compounds because
the variations of 13C and 15N, and therefore δ13C and δ15N, are influ-
enced by different processing in soil, such as transformation, fractio-
nation and leaching (Bai et al., 2017; Högberg, 1997; Hosseini Bai et al.,
2017; Hosseini Bai et al., 2015a). Hence, 13C and 15N are abundant
within fewer compounds, and therefore fewer principal wavelengths
are used for estimating and predicting δ13C and δ15N.
4.2. Evaluating model performance using external test sets
The models developed using the selected informative wavelengths
provided high accuracy predictions of soil TC, TN, δ13C and δ15N in the
test datasets. We were only able to reasonably compare the accuracies
of prediction obtained from studies which used airborne hyperspectral
imaging and VNIR spectroscopy (Gopal et al., 2015; Jiang et al., 2017;
Sorenson et al., 2017) to our results because of the limited available
literature on predicting TC, TN, δ13C and δ15N in soil samples using
laboratory-based hyperspectral image analysis. The accuracies of the
models, developed for predicting TC, TN, δ13C and δ15N, in the current
study were in the range of other studies which have used airborne
hyperspectral imaging and VNIR spectroscopy techniques for predicting
TC and TN in soil samples (Gmur et al., 2012; Gopal et al., 2015; Jiang
et al., 2017; Sorenson et al., 2017; Zhang et al., 2016). Overall, la-
boratory-based hyperspectral imaging technique could successfully
predict TC, TN, δ13C and δ15N in soil samples.
In our future studies, the developed models need to be validated and
improved using a wider range of soils. The mechanisms of the predic-
tion of δ15N need to be understood possibly by finding the correlations
between δ15N and soil chromophore properties. Further optimisation
I. Tahmasbian et al.
Fig. 5. Reference vs. predicted values of soil (a) total carbon (TC), (b) total nitrogen (TN), (c) carbon isotopic composition (δ13C) and (d) nitrogen isotopic com-
position (δ15N) in the test sets. The dotted lines are regression lines for test sets.
and programming are needed for using the method in real-time.
Additionally, the potential of outdoor proximate hyperspectral imaging
needs to be investigated.
5. Conclusion
The potential of a VNIR (400–1000 nm) laboratory-based hyper-
spectral imaging sensor for the predicting of TC, TN, δ13C and δ15N in
the soil samples was examined. PLSR models were developed using the
spectra obtained from 96 soil samples. The developed PLSR models
were successfully predicted TC, TN, δ13C and δ15N in the 24 external
test samples. The results indicated that removing the uninformative
wavelengths from the models slightly increased the prediction ac-
curacies, possibly due to removing the collinearity existed between
wavelengths. We think that the correlation observed between soil TC
and δ13C was the possible mechanism through which the prediction of
soil δ13C was possible. Despite the high-accuracy prediction of δ15N, no
correlation was found among δ15N, TC and TN. Overall, laboratory-
based hyperspectral image analysis can be used to predict soil TC, TN,
δ13C and δ15N.
Acknowledgement
This study has been funded by the Griffith University (grant number
NSC 1010). The Authors acknowledge Professor Helen M Wallace and
Ecology Laboratory at University of the Sunshine Coast for providing
access to hyperspectral imaging camera. The authors would also like to
acknowledge Mr. Geoffrey Lambert and Dr. Thi Thu Nhan Nguyen for
their fieldwork assistance.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.geoderma.2018.06.008.
261
Geoderma 330 (2018) 254–263
References
Abdi, H., 2010. Partial least squares regression and projection on latent structure re-
gression (PLS regression). Wiley Interdiscip. Rev. Comput. Stat. 2 (1), 97–106.
Abdullah, K.M., 2016. Carbon and Nitrogen Dynamics Following Prescribed Burning in a
Suburban Native Forest of South-east Queensland. Griffith University.
Afandi, S.D., Herdiyeni, Y., Prasetyo, L.B., Hasbi, W., Arai, K., Okumura, H., 2016.
Nitrogen content estimation of rice crop based on near infrared (NIR) reflectance
using artificial neural network (ANN). Procedia Environ Sci 33, 63–69.
Ariana, D.P., Lu, R., Guyer, D.E., 2006. Near-infrared hyperspectral reflectance imaging
for detection of bruises on pickling cucumbers. Comput. Electron. Agric. 53 (1),
60–70.
Bai, S.H., Trueman, S.J., Nevenimo, T., Hannet, G., Bapiwai, P., Poienou, M., Wallace,
H.M., 2017. Effects of shade-tree species and spacing on soil and leaf nutrient con-
centrations in cocoa plantations at 8 years after establishment. Agric. Ecosyst.
Environ. 246, 134–143.
Bai, S.H., Tahmasbian, I., Zhou, J., Nevenimo, T., Hanet, G., Walton, D., Randall, B.,
Gama, T., Wallace, H.M., 2018. A non-destructive determination of peroxide values,
total nitrogen and mineral nutrients in an edible tree nut using hyperspectral ima-
ging. Comput. Electron. Agric (In press).
Barbin, D.F., Elmasry, G., Sun, D.-W., Allen, P., 2012. Predicting quality and sensory
attributes of pork using near-infrared hyperspectral imaging. Anal. Chim. Acta 719,
30–42.
Barrett, J.P., 1974. The coefficient of determination—some limitations. Am. Stat. 28 (1),
19–20.
Barthès, B.G., Brunet, D., Ferrer, H., Chotte, J.-L., Feller, C., 2006. Determination of total
carbon and nitrogen content in a range of tropical soils using near infrared spectro-
scopy: influence of replication and sample grinding and drying. J. Near Infrared
Spectrosc. 14 (5), 341–348.
Bellon-Maurel, V., Fernandez-Ahumada, E., Palagos, B., Roger, J.-M., Mcbratney, A.,
2010. Critical review of chemometric indicators commonly used for assessing the
quality of the prediction of soil attributes by NIR spectroscopy. TrAC Trends Anal.
Chem. 29 (9), 1073–1081.
Ben-Dor, E., Banin, A., 1995a. Near-infrared analysis as a rapid method to simultaneously
evaluate several soil properties. Soil Sci. Soc. Am. J. 59 (2), 364–372.
Ben-Dor, E., Banin, A., 1995b. Near infrared analysis (NIRA) as a method to simulta-
neously evaluate spectral featureless constituents in soils. Soil Sci. 159 (4), 259–270.
Ben-Dor, E., Inbar, Y., Chen, Y., 1997. The reflectance spectra of organic matter in the
visible near-infrared and short wave infrared region (400–2500 nm) during a con-
trolled decomposition process. Remote Sens. Environ. 61 (1), 1–15.
Ben-Dor, E., Patkin, K., Banin, A., Karnieli, A., 2002. Mapping of several soil properties
using DAIS-7915 hyperspectral scanner data - a case study over clayey soils in Israel.
Int. J. Remote Sens. 23 (6), 1043–1062.
Blackmer, T.M., Schepers, J.S., Varvel, G.E., 1994. Light reflectance compared with other
nitrogen stress measurements in corn leaves. Agron. J. 86 (6), 934–938.
Brunet, D., Barthès, B.G., Chotte, J.-L., Feller, C., 2007. Determination of carbon and
nitrogen contents in Alfisols, Oxisols and Ultisols from Africa and Brazil using NIRS
I. Tahmasbian et al.
analysis: effects of sample grinding and set heterogeneity. Geoderma 139 (1–2),
106–117.
Butler, O.M., Lewis, T., Chen, C., 2016. Prescribed fire alters foliar stoichiometry and
nutrient resorption in the understorey of a subtropical eucalypt forest. Plant Soil 410
(1), 181–191.
Catterall, C., Piper, S., Bunn, S.E., Arthur, J.M., 2001a. Flora and fauna assemblages vary
with local topography in a subtropical eucalypt forest. Austral Ecol. 26 (1), 56–69.
Catterall, C.P., Piper, S.D., Bunn, S.E., Arthur, J.M., 2001b. Flora and fauna assemblages
vary with local topography in a subtropical eucalypt forest. Austral Ecology. 26 (1),
56–69.
Che, R., Wang, F., Wang, W., Zhang, J., Zhao, X., Rui, Y., Xu, Z., Wang, Y., Hao, Y., Cui,
X., 2017. Increase in ammonia-oxidizing microbe abundance during degradation of
alpine meadows may lead to greater soil nitrogen loss. Biogeochemistry 136 (3),
341–352.
Che, R., Qin, J., Tahmasbian, I., Wang, F., Zhou, S., Xu, Z., Cui, X., 2018. Litter amend-
ment rather than phosphorus can dramatically change inorganic nitrogen pools in a
degraded grassland soil by affecting nitrogen-cycling microbes. Soil Biol. Biochem.
120, 145–152.
Chong, I.-G., Jun, C.-H., 2005. Performance of some variable selection methods when
multicollinearity is present. Chemom. Intell. Lab. Syst. 78 (1), 103–112.
Coûteaux, M.-M., Berg, B., Rovira, P., 2003. Near infrared reflectance spectroscopy for
determination of organic matter fractions including microbial biomass in coniferous
forest soils. Soil Biol. Biochem. 35 (12), 1587–1600.
Cozzolino, D., Morón, A., 2006. Potential of near-infrared reflectance spectroscopy and
chemometrics to predict soil organic carbon fractions. Soil Tillage Res. 85 (1), 78–85.
Cozzolino, D., Cynkar, W.U., Dambergs, R.G., Shah, N., Smith, P., 2013. In situ mea-
surement of soil chemical composition by near-infrared spectroscopy: a tool toward
sustainable vineyard management. Commun. Soil Sci. Plant Anal. 44 (10),
1610–1619.
Curran, P.J., 1989. Remote sensing of foliar chemistry. Remote Sens. Environ. 30 (3),
271–278.
Dalal, R., Henry, R., 1986. Simultaneous determination of moisture, organic carbon, and
total nitrogen by near infrared reflectance spectrophotometry. Soil Sci. Soc. Am. J. 50
(1), 120–123.
Darby, I., Xu, C.-Y., Wallace, H.M., Joseph, S., Pace, B., Hosseini Bai, S., 2016. Short-term
dynamics of carbon and nitrogen using compost, compost-biochar mixture and or-
gano-mineral biochar. Environ. Sci. Pollut. R. 23 (11), 11267–11278.
Daughtry, C.S., 2001. Discriminating crop residues from soil by shortwave infrared re-
flectance. Agron. J. 93 (1), 125–131.
Denis, A., Stevens, A., Van Wesemael, B., Udelhoven, T., Tychon, B., 2014. Soil organic
carbon assessment by field and airborne spectrometry in bare croplands: accounting
for soil surface roughness. Geoderma 226, 94–102.
Elmore, A.J., Craine, J.M., 2011. Spectroscopic analysis of canopy nitrogen and nitrogen
isotopes in managed pastures and hay land. IEEE T Geosci. Remote. 49 (7),
2491–2498.
Farquhar, G.D., Ehleringer, J.R., Hubick, K.T., 1989. Carbon isotope discrimination and
photosynthesis. Annu. Rev. Plant Biol. 40 (1), 503–537.
Fearn, T., 2000. On orthogonal signal correction. Chemom. Intell. Lab. Syst. 50 (1),
47–52.
Ferwerda, J.G., Skidmore, A.K., Mutanga, O., 2005. Nitrogen detection with hyperspec-
tral normalized ratio indices across multiple plant species. Int. J. Remote Sens. 26
(18), 4083–4095.
Fidêncio, P.H., Poppi, R.J., de Andrade, J.C., Cantarella, H., 2002. Determination of or-
ganic matter in soil using near-infrared spectroscopy and partial least squares re-
gression. Commun. Soil Sci. Plant Anal. 33 (9–10), 1607–1615.
Fu, X., Shao, M., Wei, X., Horton, R., 2010. Soil organic carbon and total nitrogen as
affected by vegetation types in northern loess plateau of China. Geoderma 155 (1),
31–35.
Fujiyoshi, L., Sugimoto, A., Tsukuura, A., Kitayama, A., Lopez Caceres, M.L., Mijidsuren,
B., Saraadanbazar, A., Tsujimura, M., 2017. Spatial variations in larch needle and soil
δ15N at a forest–grassland boundary in northern Mongolia. Isot. Environ. Health
Stud. 53 (1), 54–69.
Gama, T., Wallace, H.M., Trueman, S.J., Tahmasbian, I., Bai, S.H., 2018. Hyperspectral
imaging for non-destructive prediction of total nitrogen concentration in almond
kernels. Acta Hortic (In press).
Gao, X., Yang, Y., Zhang, W., Jia, W., Li, J., Tian, C., Zhang, Y., He, L., 2014. Visible-near
Infrared Reflectance Spectroscopy for Estimating Soil Total Nitrogen Contents in the
Sanjiang Yuan Regions, China: A Case Study of Yushu County and Maduo County,
Qinghai Province. SPIE Asia-Pacific Remote Sensing, Beijing, China.
Gmur, S., Vogt, D., Zabowski, D., Moskal, L.M., 2012. Hyperspectral analysis of soil ni-
trogen, carbon, carbonate, and organic matter using regression trees. Sensors 12 (8),
10639–10658.
Gopal, B., Shetty, A., Ramya, B.J., 2015. Prediction of the presence of topsoil nitrogen
from spaceborne hyperspectral data. Geochem. Int. 30 (1), 82–92.
Grahn, H., Geladi, P., 2007. Techniques and Applications of Hyperspectral Image
Analysis. John Wiley & Sons.
Hobbie, E.A., Macko, S.A., Williams, M., 2000. Correlations between foliar δ15N and
nitrogen concentrations may indicate plant-mycorrhizal interactions. Oecologia 122
(2), 273–283.
Högberg, P., 1997. 15N natural abundance in soil-plant systems. New Phytol. 137 (2),
179–203.
Höskuldsson, A., 1988. PLS regression methods. J. Chemom. 2 (3), 211–228.
Hosseini Bai, S., Blumfield, T.J., Xu, Z., 2014a. Physiological traits of Acacia concurrens
and Eucalyptus crebra with respect to radical site preparation practices in a re-
vegetation trial, south East Queensland, Australia. J. Soils Sediments 14 (6),
1107–1115.
262
Geoderma 330 (2018) 254–263
Hosseini Bai, S., Blumfield, T.J., Xu, Z.H., 2014b. Survival, growth and physiological
status of Acacia disparrima and Eucalyptus crebra seedlings with respect to site
management practices in Central Queensland, Australia. Eur. J. For. Res. 133 (1),
165–175.
Hosseini Bai, S., Reverchon, F., Xu, C.-Y., Xu, Z., Blumfield, T.J., Zhao, H., Van Zwieten,
L., Wallace, H.M., 2015a. Wood biochar increases nitrogen retention in field settings
mainly through abiotic processes. Soil Biol. Biochem. 90, 232–240.
Hosseini Bai, S., Xu, Z., Blumfield, T.J., Reverchon, F., 2015b. Human footprints in urban
forests: implication of nitrogen deposition for nitrogen and carbon storage. J. Soils
Sediments 15 (9), 1927–1936.
Hosseini Bai, S., Dempsey, R., Reverchon, F., Blumfield, T.J., Ryan, S., Cernusak, L.A.,
2017. Effects of forest thinning on soil-plant carbon and nitrogen dynamics. Plant Soil
411 (1), 437–449.
Ingleby, H., Crowe, T., 2000. Reflectance models for predicting organic carbon in
Saskatchewan soils. Can. Agric. Eng. 42 (2), 57–64.
Jia, S., Li, H., Wang, Y., Tong, R., Li, Q., 2016. Recursive variable selection to update
near-infrared spectroscopy model for the determination of soil nitrogen and organic
carbon. Geoderma 268, 92–99.
Jiang, Q., Chen, Y., Guo, L., Fei, T., Qi, K., 2016. Estimating soil srganic carbon of
cropland soil at different levels of soil moisture using VIS-NIR spectroscopy. Remote
Sens. 8 (9), 755.
Jiang, Q., Li, Q., Wang, X., Wu, Y., Yang, X., Liu, F., 2017. Estimation of soil organic
carbon and total nitrogen in different soil layers using VNIR spectroscopy: effects of
spiking on model applicability. Geoderma 293, 54–63.
Jung, A., Vohland, M., Thiele-Bruhn, S., 2015. Use of a portable camera for proximal soil
sensing with hyperspectral image data. Remote Sens. 7 (9), 11434–11448.
Kamruzzaman, M., Elmasry, G., Sun, D.-W., Allen, P., 2012. Prediction of some quality
attributes of lamb meat using near-infrared hyperspectral imaging and multivariate
analysis. Anal. Chim. Acta 714, 57–67.
Kamruzzaman, M., Makino, Y., Oshita, S., 2016a. Parsimonious model development for
real-time monitoring of moisture in red meat using hyperspectral imaging. Food
Chem. 196, 1084–1091.
Kamruzzaman, M., Makino, Y., Oshita, S., 2016b. Rapid and non-destructive detection of
chicken adulteration in minced beef using visible near-infrared hyperspectral ima-
ging and machine learning. J. Food Eng. 170, 8–15.
Kohavi, R., 1995. A study of cross-validation and bootstrap for accuracy estimation and
model selection. In: International Joint Conference on Artificial Intelligence (Ijcai),
Canada, pp. 1137–1145.
Li, W., Qu, H., 2016. Determination of total organic carbon and soluble solids contents in
Tanreqing injection intermediates with NIR spectroscopy and chemometrics.
Chemom. Intell. Lab. Syst. 152, 140–145.
Li, B., Morris, J., Martin, E.B., 2002. Model selection for partial least squares regression.
Chemom. Intell. Lab. Syst. 64 (1), 79–89.
Li, B., Liew, O.W., Asundi, A.K., 2006. Pre-visual detection of iron and phosphorus de-
ficiency by transformed reflectance spectra. J. Photochem. Photobiol. B Biol. 85 (2),
131–139.
Li, F., Mistele, B., Hu, Y., Chen, X., Schmidhalter, U., 2014. Reflectance estimation of
canopy nitrogen content in winter wheat using optimised hyperspectral spectral in-
dices and partial least squares regression. Eur. J. Agron. 52, 198–209 Part B.
Lin, L., Wang, Y., Teng, J., Wang, X., 2016. Hyperspectral analysis of soil organic matter
in coal mining regions using wavelets, correlations, and partial least squares re-
gression. Environ. Monit. Assess. 188 (2), 1–11.
Lorentz, L.J., 2013. Hyperspectral Reflectance and Stable Isotopic Nitrogen: Tools to
Assess Forest Ecosystem Nitrogen Cycling. Virginia Tech.
Manley, M., 2014. Near-infrared spectroscopy and hyperspectral imaging: non-destruc-
tive analysis of biological materials. Chem. Soc. Rev. 43 (24), 8200–8214.
Martinelli, L., Piccolo, M., Townsend, A., Vitousek, P., Cuevas, E., McDowell, W.,
Robertson, G., Santos, O., Treseder, K., 1999. Nitrogen stable isotopic composition of
leaves and soil: tropical versus temperate forests. Biogeochemistry 46 (1–3), 45–65.
Morellos, A., Pantazi, X.-E., Moshou, D., Alexandridis, T., Whetton, R., Tziotzios, G.,
Wiebensohn, J., Bill, R., Mouazen, A.M., 2016. Machine learning based prediction of
soil total nitrogen, organic carbon and moisture content by using VIS-NIR spectro-
scopy. Biosyst. Eng. 152, 104–116.
Mouazen, A., Maleki, M., De Baerdemaeker, J., Ramon, H., 2007. On-line measurement of
some selected soil properties using a VIS–NIR sensor. Soil Tillage Res. 93 (1), 13–27.
Nguyen, T.T.N., Xu, C.-Y., Tahmasbian, I., Che, R., Xu, Z., Zhou, X., Wallace, H.M., Bai,
S.H., 2017. Effects of biochar on soil available inorganic nitrogen: a review and meta-
analysis. Geoderma 288, 79–96.
Parsons, S.A., Lawler, I.R., Congdon, R.A., Williams, S.E., 2011. Rainforest litter quality
and chemical controls on leaf decomposition with near-infrared spectrometry. J.
Plant Nutr. Soil Sci. 174 (5), 710–720.
Piccolo, M.C., Neill, C., Melillo, J.M., Cerri, C.C., Steudler, P.A., 1996. 15N natural
abundance in forest and pasture soils of the Brazilian Amazon Basin. Plant Soil 182
(2), 249–258.
Prasolova, N., Xu, Z.H., Farquhar, G.D., Saffigna, P.G., Dieters, M.J., 2000. Variation in
branchlet δ13C in relation to branchlet nitrogen concentration and growth in 8-year-
old hoop pine families (Araucaria cunninghamii) in subtropical Australia. Tree
Physiol. 20 (15), 1049–1055.
Rinnan, Å., Berg, F.v.d., Engelsen, S.B., 2009. Review of the most common pre-processing
techniques for near-infrared spectra. TrAC Trends Anal. Chem. 28 (10), 1201–1222.
Shearer, G., Kohl, D.H., 1986. N2-fixation in field settings: estimations based on natural
15
N abundance. Funct. Plant Biol. 13 (6), 699–756.
Siripatrawan, U., Makino, Y., Kawagoe, Y., Oshita, S., 2011. Rapid detection of
Escherichia coli contamination in packaged fresh spinach using hyperspectral ima-
ging. Talanta 85 (1), 276–281.
Sorenson, P., Small, C., Tappert, M., Quideau, S.A., Drozdowski, B., Underwood, A., Janz,
I. Tahmasbian et al.
A., 2017. Monitoring organic carbon, total nitrogen, and pH for reclaimed soils using
field reflectance spectroscopy. Can. J. Soil Sci. 97 (2), 241–248.
Stevens, A., van Wesemael, B., Bartholomeus, H., Rosillon, D., Tychon, B., Ben-Dor, E.,
2008. Laboratory, field and airborne spectroscopy for monitoring organic carbon
content in agricultural soils. Geoderma 144 (1), 395–404.
Tahmasbian, I., Xu, Z., Abdullah, K., Zhou, J., Esmaeilani, R., Nguyen, T.T.N., Hosseini
Bai, S., 2017. The potential of hyperspectral images and partial least square regres-
sion for predicting total carbon, total nitrogen and their isotope composition in forest
litterfall samples. J. Soils Sediments 17 (8), 2091–2103.
Tahmasbian, I., Hosseini Bai, S., Wang, Y., Boyd, S., Zhou, J., Esmaeilani, R., Xu, Z., 2018.
Using laboratory-based hyperspectral imaging method to determine carbon func-
tional group distributions in decomposing forest litterfall. Catena 167, 18–27.
Udelhoven, T., Emmerling, C., Jarmer, T., 2003. Quantitative analysis of soil chemical
properties with diffuse reflectance spectrometry and partial least-square regression: a
feasibility study. Plant Soil 251 (2), 319–329.
Vaudour, E., Gilliot, J.-M., Bel, L., Lefevre, J., Chehdi, K., 2016. Regional prediction of
soil organic carbon content over temperate croplands using visible near-infrared
airborne hyperspectral imagery and synchronous field spectra. Int. J. Appl. Earth
Obs. Geoinf. 49, 24–38.
Viscarra Rossel, R., Walvoort, D., Mcbratney, A., Janik, L.J., Skjemstad, J., 2006. Visible,
near infrared, mid infrared or combined diffuse reflectance spectroscopy for si-
multaneous assessment of various soil properties. Geoderma 131 (1), 59–75.
Vohland, M., Ludwig, M., Harbich, M., Emmerling, C., Thiele-Bruhn, S., 2016. Using
variable selection and wavelets to exploit the full potential of visible–near infrared
spectra for predicting soil properties. J. Near Infrared Spectrosc. 24 (3), 255–269.
Wang, L., Okin, G.S., Wang, J., Epstein, H., Macko, S.A., 2007. Predicting leaf and canopy
15N compositions from reflectance spectra. Geophys. Res. Lett. 34 (2).
Wang, Y., Xu, Z., Zheng, J., Abdullah, K.M., Zhou, Q., 2015. δ 15N of soil nitrogen pools
and their dynamics under decomposing leaf litters in a suburban native forest subject
263
Geoderma 330 (2018) 254–263
to repeated prescribed burning in Southeast Queensland, Australia. J. Soils Sediments
15 (5), 1063–1074.
Wei, Y., Zhu, X., Li, C., Cheng, L., Wang, L., Zhao, G., Jiang, Y., 2017. Monitoring soil
nitrate nitrogen based on hyperspectral data in the apple orchards. Science 8, 21–32.
West, J.B., Bowen, G.J., Dawson, T.E., Tu, K.P., 2009. Isoscapes: Understanding
Movement, Pattern, and Process on Earth through Isotope Mapping. Springer.
Wetzel, D.L., 1983. Near-infrared reflectance analysis. Anal. Chem. 55 (12),
1165A–1176A.
Wijewardane, N.K., Ge, Y., Morgan, C.L.S., 2016. Moisture insensitive prediction of soil
properties from VNIR reflectance spectra based on external parameter orthogonali-
zation. Geoderma 267, 92–101.
Wold, S., Ruhe, A., Wold, H., Dunn III, W.J., 1984. The collinearity problem in linear
regression. The partial least squares (PLS) approach to generalized inverses. SIAM. J.
Sci. Stat. Comp. 5 (3), 735–743.
Wold, J.P., Jakobsen, T., Krane, L., 1996. Atlantic salmon average fat content estimated
by near-infrared transmittance spectroscopy. J. Food Sci. 61 (1), 74–77.
Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics.
Chemom. Intell. Lab. Syst. 58 (2), 109–130.
Xu, C., Zeng, W., Huang, J., Wu, J., Van Leeuwen, W.J., 2016. Prediction of soil moisture
content and soil salt concentration from hyperspectral laboratory and field data.
Remote Sens. 8 (1), 42.
Yang, H., Li, J., 2013. Predictions of soil organic carbon using laboratory-based hyper-
spectral data in the northern Tianshan mountains, China. Environ. Monit. Assess. 185
(5), 3897–3908.
Zaehle, S., Friedlingstein, P., Friend, A.D., 2010. Terrestrial nitrogen feedbacks may ac-
celerate future climate change. Geophys. Res. Lett. 37 (1), 1–5.
Zhang, Y., Li, M., Zheng, L., Zhao, Y., Pei, X., 2016. Soil nitrogen content forecasting
based on real-time NIR spectroscopy. Comput. Electron. Agric. 124, 29–36.