Mirchooli, F. Et Al. 2020

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Ecological Indicators 116 (2020) 106473

Contents lists available at ScienceDirect

Ecological Indicators
journal homepage: www.elsevier.com/locate/ecolind

Spatial distribution dependency of soil organic carbon content to important T


environmental variables
Fahimeh Mirchoolia, Mahboobeh Kiani-Harchegania, Abdulvahed Khaledi Darvishana,

Samereh Falahatkarb, Seyed Hamidreza Sadeghia,
a
Department of Watershed Management Engineering, Faculty of Natural Resources, Tarbiat Modares University, Noor, Iran
b
Department of Environmental Sciences, Faculty of Natural Resources Tarbiat Modares University, Noor, Iran

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

Keywords: Soil organic carbon (SOC) has a vital role in the physical, chemical and biological behavior of the soil, and
Land degradation therefore prediction of the amount and mapping spatial distribution of SOC is necessary for sustainable soil
Landsat images management. However, the relationship between SOC and remotely sensed and easily accessible variables have
Soil quality been rarely reported. The main objective of the present study is, therefore, to estimate SOC using the remote
Carbon spatial distribution
sensing of satellite images as well as some field variables for the Shazand Watershed, Iran. Towards that, 140 soil
Semi-arid region
samples were taken from the top 30-cm of the soil from homogeneous units representing an area > 1 km2 to
measure SOC. The potential relationship between SOC and some remote sensing-based indices including
Enhanced Vegetation Index (EVI), Difference vegetation index (DVI), Optimized Soil Adjusted Vegetation Index
(OSAVI), Soil Adjusted Vegetation Index (SAVI), Modified Soil Adjusted Vegetation Index (MSAVI), Normalized
Difference Vegetation Index (NDVI), Coloration Index (CI), Normalized Difference Moisture Index (NDMI), and
some topographical and soil texture factors viz. elevation, slope, aspect, Topographical Wetness Index (TWI),
clay and silt content of soil were assessed for the Shazand Watershed. In this regard, the relationships between
SOC and environmental variables were analyzed using Ordinary Least Square (OLS), Geographically Weighted
Regression (GWR), and Random Forest (RF) analyses. The results showed that none of the models indicated a
good predictive performance due to low R2. However, CI was found as a better predictor (R2 = 0.258 for GWR
and R2 = 0.040 for OLS; p-value < 0.05) for SOC in comparison with other indices. Accordingly, it was con-
cluded that the remote sensing-based and environmental indices did not have enough potential for estimating
SOC to generate managerial plans for the study watershed. The low SOC content (0.63 ± 0.29%) and low
variation across the Shazand Watershed were found as determinant factors on the low fitness of the study
relationships. Further studies are therefore needed to find out reliable predictive models for SOC for the con-
ditions under considerations.

1. Introduction concentration of SOC improve soil structure and inhibit soil erosion
which leads to improvement of water storage capacity, regulation of
Soil as a vital part of the ecosystem is the net source or sink of soil nutrient and stabilization of soil aggregates (Miller et al., 2015; Wang
organic carbon (SOC), since about two or three times more SOC lie in et al., 2019), soil productivity (Parshotam et al., 2001; Oueslati et al.,
soil rather than the atmosphere (Polyakov and Lal, 2004; Wang et al., 2013; Castaldi et al., 2019) and food production (Bhunia et al., 2017).
2003; Dai et al., 2014; Muchana, 2017; Zissimos et al., 2019). Soil can In other words, SOC is a critical component to these ecosystem services
sequestrate CO2 existing in the atmosphere as a result of both natural (Grinand et al., 2017) that are important in watershed health and
and anthropogenic activities. As SOC increases, agriculture and en- sustainability, and decreases land degradation (Davudirad et al., 2016;
vironment are benefited, and atmospheric carbon reduction is effec- Sadeghi and Hazbavi, 2017; Kiani-Harchegani and Sadeghi, 2020). The
tively leading to climate change mitigation (Wang et al., 2003; Knadel spatial and temporal distribution of SOC is variable due to different
et al., 2015; Somarathna et al., 2016). In this way, the higher land uses/covers, climates, soil management approaches, topographic


Corresponding author.
E-mail addresses: m.kiani@modares.ac.ir (M. Kiani-Harchegani), a.khaledi@modares.ac.ir (A. Khaledi Darvishan),
Samereh.falahatkar@modares.ac.ir (S. Falahatkar), Sadeghi@modares.ac.ir (S.H. Sadeghi).

https://doi.org/10.1016/j.ecolind.2020.106473
Received 4 March 2019; Received in revised form 22 April 2020; Accepted 28 April 2020
1470-160X/ © 2020 Published by Elsevier Ltd.
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

parameters and soil types. Some studies have focused on the spatial to analyze spatial heterogeneity in the relationships. In fact, GWR is a
pattern of SOC due to different effective parameters such as land use localized regression that examines spatially non-stationary parameters,
dynamic (Dorji et al., 2014; Stumpf et al., 2018), vegetation (Bhunia so that the model performance will be improved by reducing spatial
et al., 2017), topographic factors (Tsui et al., 2013; Oueslati et al., autocorrelations. In recent years, GWR has been widely employed in
2013), soil type (Wang et al, 2003), soil properties (Were et al., 2015), different field such as groundwater quantity (Taghipour Javi et al.,
climate (Luo et al., 2017), landslide (Restrepo and Laquay, 2018) and 2014; Wu et al., 2017; Almeida Salles et al., 2018), rainfall and some
period of cultivation (Dalal and Mayer, 1986). So that, the information environmental indices (Georganos et al., 2017; Ahmadi et al., 2018;
of spatial and temporal changes of SOC is hence important for policy- Salimi et al., 2018; Zhou et al., 2020), land surface temperature (Kalota,
makers for proper decision on agro-environment climate change and 2017; Zhao et al., 2018), urban and regional analysis (Dadashpoor
assessment the short- and long-term carbon consequences as a result of et al., 2019; Duncan et al., 2019) and ecology and even human geo-
alternative strategies managements (Bartholomeus et al., 2008; Grinand graphy (Tu, 2011) and its good performances were confirmed. This
et al., 2017; Jebari et al., 2018). approach has also been applied in soil and land use management and
Laboratory measurement of SOC is accurate but time-consuming, aided in SOC prediction and assessment. In this regard, Mishra et al.
expensive, and labor intensive. It also requires abundant soil samples. (2010) compared three models of multiple linear regression (MLR),
Therefore, the measurement of SOC can only be applied in small areas, regression kriging (RK) (Gao and Li, 2011) and GWR for SOC predic-
while managers need to estimate soil properties at watershed or re- tion, in the Midwest of the United States. In this study, GWR had lower
gional scales. Hence, discovering potential relationships between SOC root mean square errors (RMSE) of 22% and 2% over MLR and RK.
and easily accessible variables may help decision makers and practi- Similarly, Shi et al. (2014) reported the good performance of GWR for
tioners economically designate appropriate and applied strategies. To spatial prediction of SOC based on spatial climate and socio-economic
this end, remote sensing is a suitable technology that can be applied for predictors in China. Song et al. (2016) compared GWR with MLR,
rapid monitoring of soil properties in broader scales with a frequent and geographically weighted ridge regression (GWRR), kriging with an
repetitive time basis (Muchana, 2017; Yu et al., 2018; Gholizadeh et al., external drift (KED), and GWR plus ordinary kriging of model residuals
2018; Lin et al., 2020). (GWRSK) for predicting the spatial distribution of SOC in the Heihe
Moreover, the potential of some different remote sensing-based in- basin, China. They concluded the improved prediction accuracy of
dices such as Normalized Difference Vegetation Index (NDVI), Soil GWR in the spatial variability of SOC. Wang et al. (2018) indicated the
Adjusted Vegetation Index (SAVI), Enhanced vegetation index (EVI), good performance of GWR model for predicting the influence of land
Bare Soil Index (BSI), and Compound Topographic Index (CTI) has been use on SOC at the watershed scale in coastal agro-ecosystems of China.
reported to estimate and predict SOC using band- rationing algorithm Furthermore, some machine learning techniques such as random
(Takata, 2010; Kumar et al., 2016; Mondal et al., 2017). In this vein, forest (RF) was proved to be a robust method which has the potential to
Ladoni et al. (2010) quantitatively estimated SOC using the digital enhance the prediction of SOC distribution (Grimm et al., 2008). This
value of soil reflectance in satellite bands and three models of simple method performs data from different sources and various measurement
regression models (SRM), principal component analysis (PCA), and the scales and uses no statistical assumptions (Rahmati et al., 2016). The
soil line approach. The results revealed that SRM provided the most helpful applicability of this method has been also confirmed by
accurate SOC predictions. Kumar et al. (2018) assessed the relationship Rodriguez-Galiano et al. (2014), Naghibi et al. (2017), and Wang et al.
between SOC and leaf area index (LAI) measured through satellite data (2018).
and regression analysis and concluded the efficient role of LAI for It is seen from the literature that many studies worked on the re-
predicting SOC with R2 = 0.95 (p-value < 0.05). Kumar et al. (2018) lationships between soil organic carbon content and remotely sensed
proved NDVI was the most effective variable among eight study remote data. However, the spatial variations of these relationships have been
sensing indices for predicting SOC in India. Wang et al. (2019) esti- less studied across the world at watershed scale. This makes the present
mated and mapped SOC content using environmental variables and endeavor different from very valuable former researches. Accordingly,
three methods of inverse distance weighting (IDW), regression kriging GWR was applied to determine the spatial heterogeneity in the re-
(RK) and multi-factor weighted regression model (MWRM) in grass- lationships between study variables, which is supposed as the novelty of
lands of China. They concluded MWRM was the best method and NDVI the current research. Therefore, the present study tried to collabora-
was among the critical variable for predicting SOC. Recently, Zhou tively apply satellite images indices, field and laboratory data, en-
et al. (2020) assessed the ability of multi-temporal Sentinel-1A data in vironmental and biophysical variables to map SOC variability and also
SOC prediction in the Heihe River Basin in northwestern China using s to explore the SOC content- environmental variables relationships using
set of 162 topsoil (0–20 cm) samples and with the help of 15 en- OLS, GWR and RF methods. The study was formulated for the Shazand
vironmental variables. The results were compared to those obtained watershed due to existing background researches (Davudirad et al.,
using the random forest (RF) and boosted regression tree (BRT) models. 2016; Sadeghi and Hazbavi, 2017), easy access and complete data bank
They reported that the combination of all environmental variables with for further scientific endeavors.
further emphasis on precipitation, elevation, and temperature had the
best prediction performance explaining 75% of SOC variation. The 2. Materials and methods
multi-temporal Sentinel-1A data in RF and BRT models explained 9%
and 7%, respectively. The applicability of fused Sentinel-2 and Sentinel- 2.1. Study area
3 remote sensing images for SOM estimation using 107 field samples
was successfully investigated by Lin et al. (2020) for Donghai County, The Shazand watershed located in Markazi Province is one of the
China. They also reported that the modeling accuracy was higher sub-watersheds of the Namak Lake with a total area of 1740 km2 half of
(R2 > 0.75) for the SOM content more than 24 g kg−1, and lower which includes highlands and hard formations, and another half con-
(R2 > 0.35) when the SOM content was lower. tains alluvial sediments and sub-mountain screes. A general view of the
The majority of studies have been investigated to estimate SOC study watershed has been shown in Fig. 1.
using environmental indices and traditional statistical analyses, which A complex topography with elevation ranges from 1810 m to more
produced average parameter estimates. Therefore, the spatial variation than 3343 m (above mean sea level), results in steep gradients of
of relationships among non-stationary independent predictors and SOC rainfall on both spatial and temporal scales. More than 90% of geolo-
has been ignored. As a result, proposing a sophisticated technique to gical formations of the watershed belong to second and third geological
overcome previously mentioned limitations was found crucial. The era (i.e., Quaternary, Neogene, and Cretaceous). The climate is semi-
geographically weighted regression (GWR) was accordingly developed arid with a mean annual temperature of 12 °C and an average annual

2
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Fig. 1. Location of study area and spatial distribution of soil sampling points in the Shazand Watershed, Markazi Province, Iran.

precipitation of 420 mm, mostly falling in winter, autumn and spring. low-vegetation canopy areas using background adjustment factor (Jin
Agricultural lands are located mostly in 5–15% slopes and mainly are et al., 2014). Although, NDVI is one of the most widely used vegetation
utilized for rainfed wheat. The soils have low organic matter (about indices, however, it is susceptible to some errors such as variable at-
1%). The soil of the study watershed is mainly clay loam and calcareous mospheric and canopy background conditions (Matsushita et al., 2007).
with CaCO3 contents from 5 to 55% (Kiani Harchegani et al., 2019a). DVI is another vegetation index that does not account for the difference
between reflectance and radiance caused by atmospheric effects or
2.2. Soil sampling shadows. The coordinates of sampling points were transferred to ArcGIS
10.3 software and corresponding environmental variables were ex-
Soil samples were collected from the center of four corners of each tracted.
quadrant by driving a core sampler up to 30 cm in depth (Kiani-
Harchegani et al. 2019a; Kiani-Harchegani and Sadeghi, 2020). Some 2.4. Analysis methods
140 points were determined based on homogeneous land units resulted
from overlying soil types, land use/cover and slope layers of the Sha- 2.4.1. OLS and GWR models
zand watershed. Three soil samples were collected in each point and The OLS and GWR were used to investigate the relationships be-
then mixed into composite samples and taken to the laboratory of tween SOC and some environmental variables. GWR is a local tech-
Tarbiat Modares University of Iran for the analyses. Soil samples were nique, which resolves the limitations of non-stationary in OLS. It im-
purified from stones, granules, leaves, and roots. The samples were then proves OLS results using spatial information and explores the
air-dried, crushed, mixed thoroughly, and sifted with a 2-mm sieve relationship among spatial variables as detailed is given by
(Islam et al., 2016; Kiani-Harchegani et al., 2019b). Ultimately, SOC Fotheringham et al. (2002). OLS as one of the global spatial models
contents of soil samples were analyzed using Walkey–Black procedure examined the average state of SOC over the space and expressed in Eq.
(Nelson and Sommers, 1996) and appropriately stored in the data bank (1):
developed in Excel 2013 software package. n
y = β0 + ∑ βj xi + ε
2.3. Landsat data and environmental predictors j=1 (1)

where xi and y are respectively the independent and dependent vari-


The Landsat 8 Operational Land Imager (OLI) was used to calculate ables; n is the number of independent variables; β0 and βi represent
environmental indices (Liou et al., 2015). The necessary geo-refer- intercept and coefficient respectively; ε represents error.
enced, radiometric and FLAASH atmospheric corrections were then GWR was calculated using Eq. (2) as an extended form of Eq. (1)
performed (Zhang et al., 2017). Considering estimation of SOC using through considering local parameters (Almeida Salles et al., 2018):
remote sensing is mostly related to decreasing the reflectance in visible
n
and near infrared bands (Bartholomeus et al., 2008), some remote
sensing-based variables including band 4, band 5, band 6 of Landsat
yi = β0 (ui , vi ) + ∑ βj (ui, vi ) xij + εi
j=1 (2)
OLI, SAVI, MSAVI, OSAVI, NDVI, EVI, DVI, CI and BI, and topo-
graphical variables of elevation, slope, aspect and TWI, and also clay where yi is the value of the dependent variable at location i, (ui, vi) is the
and silt contents were calculated based on procedures shown in Table 1. spatial location of i-th data, β0 (uj, vj) acts as intercept and βj (ui. vi) is
SAVI, MSAVI, and OSAVI are kinds of vegetation indices that suited the value of the j-th parameter at location of i, εi is the random error for
for low vegetated areas because of enhancing vegetation brightness in location i (Georganos et al., 2017).

3
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Table 1
Methods and formulae used to calculate environmental and topographic variables.
Variables Index Abbreviation Source or Formula Reference

Remote sensing-based Landsat band (1, 2, 3, 4, 5, 6) B 1 to B 6 – –


Brightness Index BI Red2 + NIR2 Asfaw et al. (2016)
Enhanced Vegetation Index EVI G.
NIR − R Liu and Huete, (1995)
L + NIR + (C1 . Red) + (C2 . B)
Optimized Soil Adjusted Vegetation Index OSAVI NIR − R Mokarram et al. (2015)
NIR + R + 0.16
Soil Adjusted Vegetation Index SAVI (NIR − Red)
(1 + L) Huete (1988); Mokarram et al. (2015)
(NIR + Red + L)
Modified Soil Adjusted Vegetation Index MSAVI 2NIR + 1 − (2NIR + 1)2 − 8(NIR − R) Qi et al. (1994)
2
Normalized Difference Vegetation NDVI NIR − R Robinson et al. (2017)
NIR + R
Coloration Index CI R−G Ray et al. (2004)
R+G
Normalized Difference Moisture Index NDMI NIR − SWIR Rokni et al. (2014)
NIR + SWIR
Difference vegetation index DVI NIR − R Jin et al. (2014)

Topographic Elevation ELE Digital Elevation Model (DEM) –


Slope SLP DEM –
Aspect ASP DEM –
Topographic Wetness Index TWI a ⎞
Song et al. (2016)
ln ⎛⎜ ⎟
tanβ
⎝ ⎠
Clay and silt content Clay Lab analyses –
Silt

R = Red band, NIR = Near Infrared, B = Blue band, G = Green band, G = 2.5, C1 = 6, L = 0.5, C2 = 7.5, L = 1.

The weight of data was measured by considering its distance to 3. Results and discussion
location i. Therefore, the weighting factor varied with i. The matrix
form of parameter estimation for i was then written as Eq. (3): 3.1. Analysis of remote sensing-based indices

β (ui , vi ) = [X T W (ui , vi ) X ]−1X T W (ui , vi ) Y (3) The descriptive statistics of measured different variables and vege-
tation indices (VIs) of SAVI, MSAVI, OSAVI, NDVI, EVI, and DVI for the
where W (ui, vi) is a spatial weighting matrix, X is an independent data
Shazand Watershed have been summarized in Table 2. The spatial
matrix and Y is dependent data vector (Fotheringham et al. 2002; Gao
variations of the study variables have been mapped as shown in Fig. 2.
and Li, 2011). Among different methods for calculating the spatial
The results of Table 2 showed that SOC distribution was different
weighting function, the most common method of Gaussian function was
from 0.00 to 1.30% over the watershed. In the central, northern and
used as expressed in Eq. (4):
southern watershed had the lowest SOC (Bednář and Šarapatka, 2018)
wij = e−0.5(dij h )2
(4) where arable lands existed. SAVI varied from 0.83 to −0.39 with a
mean of 0.10. NDVI also ranged from −0.99 to 0.99 with a mean of
where Wij is the weight of observed data at location j for estimating the 0.14. MSAVI developed to further minimize soil influences (Liu and
dependent variable at location i, and h is referred as a bandwidth. In Huete, 1995) ranged from −0.34 to 0.99. OSAVI and EVI with re-
this function, the weight of location j decreases with its distance from spective mean values of 0.05 ± 0.04 and 0.10 ± 0.06 have been
location i which is being regressed (Wang et al., 2005). developed to overcome the limitations of NDVI. The NDVI results also
Finally, the accuracy and performance of these methods were showed vigorous vegetation in the south and central parts of the area
measured using R2 and Akaike Information Criterion (AIC) in which the while the southeastern parts were characterized by lower NDVIs. DVIs,
higher value of R2 and lower value of AIC represented the better per- which is simpler than NDVI, were different between −0.28 and 0.64. CI
formance of the model. developed to evaluate the soil color (Mandal, 2016) varied from −0.99
to 0.99. Furthermore, BI as an indicator for average reflectance mag-
2.4.2. RF model nitude ranged from 0.00 to 1.34. Ultimately, NDMI, which detected the
The random forest package in R environment (www.r-project.org) vegetation water liquid differed between −0.99 and 1.00.
was used for spatial prediction of SOC in the Shazand Watershed. RF as
a nonparametric technique and an extension of regression tree model Table 2
was used to generate classifications and regression trees to improve Descriptive statistics of measured variables for the Shazand Watershed, Iran.
prediction accuracy and reduces model over fitting (Grimm et al., 2008;
Variables Minimum Maximum Average Standard Coefficient of
Dharumarajan et al., 2017). In the RF algorithm, a random subset of
Deviation variation (%)
predictor factors was used to determine each split of the tree through
bootstrapping techniques. A random selection of training data was used SOC (%) 0.00 1.30 0.63 0.29 46.00
for model building and the others were known as out-of-bag (OOB). SAVI −0.39 0.83 0.10 0.06 60.00
MSAVI −0.34 0.99 0.09 0.06 66.00
Each tree was contributed to minimizing error. Therefore, the final
OSAVI −0.39 0.79 0.10 0.06 60.00
output of the model was the average results of all trees. Furthermore, in EVI −0.11 0.96 0.05 0.04 80.00
order to minimize error, two parameters of the number of tree (ntree), NDVI −0.99 0.99 0.14 0.09 64.00
and number of predictors were optimized for fitting the tree (mtry). RF DVI −0.28 0.64 0.06 0.04 66.00
algorithm determined the importance of each predictor variable CI −0.99 0.99 0.15 0.04 26.00
BI 0.00 1.34 0.34 0.08 23.00
through assessing the prediction error increasing as OOB for that
NDMI −0.99 1.00 −0.02 0.06 −300.00
variable changed and others were unchanged (Rahmati et al., 2016; B6 −3.36 0.99 0.28 0.06 21.00
Zabihi et al., 2016).

4
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Fig. 2. Spatial distribution of environmental variables in the Shazand Watershed, Markazi Province, Iran.

3.2. Development of OLS and GWR models being composite in term of Variance Inflation Factor (VIF). VIF mea-
sures redundancy and multi-collinearity among explanatory variables.
In this research, single and different combinations of environmental Since, it was not possible to consider all variables into a composite
variables were considered to predict SOC using OLS and GWR in the model with high value of VIF (VIF > 7.5) among some variables in-
Shazand watershed. Descriptive details of independent variables of cluding SAVI, MSAVI, OSAVI, EVI and DVI, model 18 was developed
each model have been given in Table 3. Model 1–14 consider NDVI, using variables without internal correlation. The detailed statistical
NDMI, CI, B6, SAVI, OSAVI, MSAVI, EVI, DVI, elevation, slope, aspect, results of VIF have also been summarized in Table 4.
TWI, and percentage of clay and silt as independent variables in dif- In single factor analysis in OLS model (Table 3), CI has the highest
ferent models, separately. Model 15–17 set up with remote sensing and influence with the lowest AIC and highest R2 and adjusted R2. In GWR
topographical variables. Model 18 contains all variables, which allowed model, CI is also the most important factor in SOC prediction with the

5
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Fig. 2. (continued)

Table 3 Table 4
Performance of OLS and GWR models with different independent variables for The results of VIF analysis for the study variables for prediction of SOC for the
estimating SOC for the Shazand Watershed, Iran. Sahzand Watershed, Iran.
Model Independent variables AIC R2 Parameter Coefficient Standard Error VIF

OLS GWR OLS GWR Intercept 0.752 0.174 0.000


NDVI 0.202 0.213 2.560
1 NDVI 48.990 46.276 0.013 0.145 NDMI −0.007 0.220 2.704
2 NDMI 48.865 48.054 0.014 0.134 CI −0.366 0.145 1.140
3 CI 44.941 35.914 0.040 0.258 Band 6 −0.079 0.160 1.093
4 B6 49.979 45.565 0.006 0.162 Elevation 0.173 0.119 1.292
5 SAVI 50.917 39.999 0.000 0.236 Slope −0.002 0.166 1.514
6 OSAVI 50.921 39.822 0.000 0.237 Aspect −0.038 0.092 1.061
7 EVI 50.906 38.990 0.000 0.239 TWI 0.101 0.154 1.280
8 DVI 50.907 40.517 0.000 0.235 Silt −0.001 0.002 1.093
9 Elevation 50.207 48.026 0.004 0.119
10 Slope 50.918 49.187 0.000 0.123
11 Aspect 50.898 50.539 0.000 0.110
However, none of these models was satisfied with SOC prediction.
12 TWI 50.025 48.901 0.006 0.098
13 Clay 50.687 50.004 0.001 0.127 Generally, GWR models performed better than OLS models did.
14 Silt 50.553 48.205 0.002 0.150 The outputs of GWR, including local parameter estimates, provided
15 NDVI, NDMI, CI 47.782 44.134 0.050 0.287 spatial explanation to the independent variables in explaining SOC
16 Elevation, Slope, Aspect, TWI 57.597 55.684 0.011 0.051 across the Shazand Watershed. The results of regression coefficients
17 Clay, Elevation, Slope, Aspect, TWI 57.667 60.372 0.013 0.058
between SOC values and each independent variable in model 18, and
18 NDVI, NDMI, CI, B 6, Elevation, Slope, 61.164 58.953 0.041 0.142
Aspect, TWI, Clay the contribution of variables to the SOC at each location were shown
using boxplot in Fig. 3. The regression coefficients are interpreted as the
dependent variable with respect to the influences of various in-
lowest AIC of 35.914 and the highest R2 of 0.04. In multi-factor ana- dependent variables (Wu et al., 2017). The coefficients of each in-
lysis, model 15, which is a combination of NDVI, NDMI, and CI have the dependent variable were different at various locations, which revealed
best performance of AIC, R2 and adjusted R2 among other models. the influence of the variables in the model was different across the

6
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Fig. 3. Ranges of local parameter estimates for model 18 in GWR.

watershed. Among different independent variables in model 18, some potential of remote sensing-based and environmental variables due to
topographic variables and NDVI had generally positive impact on SOC low level of SOC in arid ecosystems and watersheds has already been
in most areas of watershed, which conformed to Wang et al. (2018). In verified by Muchana (2017). A similar finding has been reported by Lin
the same vein, elevation had a positive impact on SOC distribution et al. (2020) for the SOM content < 24 g kg−1 for Donghai County,
because it had potential to create microclimate and influence on the China. In addition, Wang et al. (2018) found the weak correlation be-
microbial activity leading to change of SOC content. Many researches tween vegetation index (NDVI) and SOC rather than other variables.
were also investigated elevation impact on SOC distribution (Martin
et al., 2014; Sun et al., 2015; Wang et al., 2018; Zhou et al., 2020). TWI,
which distinguishes potential wet areas, was used to quantitatively 3.3. Development of RF model
calculate soil moisture conditions. TWI coefficient of GWR model
showed substantial variation across the study area with a variable po- In the RF model, two parameters, number of predictors and number
sitive to almost zero (0.4). It was also proved existence of vegetation of tree should be optimized. In this way, the best performance of RF
increased the carbon input into the soil potentially as implied by (Yang occurs when the optimal values of these parameters were 6 and 1000
et al., 2015). It could be easy to found a positive relationship between for mtry and ntree, respectively. The performance of RF model was
NDVI and SOC with range of 0.07 to 0.26 in different parts of the evaluated by calculating uncertainty indicators viz. R2 and RMSE.
watershed. In contrary to other variables, clay content showed no ob- Results indicated that RF predicted SOC with R2 and RMSE values of
vious impact on SOC distribution all over the watershed. It also in- 0.021 and 0.3, respectively. The poor performance of RF may be related
dicated there was a negative relationship between CI and SOC across to the low amounts of SOC and respective low variations as well. These
the watershed, which is consistent with Mandal. (2016). results agree Vancampenhout et al. (2006) and Lin et al. (2020) who
In this context, the adaptive Kernel density estimation was used to concluded that the low performance of models in soils with low OC
provide the geographic weighting in the model as the observations were compared to soils having high OC. So that, in the present study, the
not regularly positioned in the study area. Spatial correlations between performance of models for prediction of SOC is lower than some pre-
independent variables and SOC in model 18 have been determined by vious studies such as Taghizadeh-Mehrjardi et al. (2016), Grinand et al.
local R2 values. As indicated in Fig. 4, the explanatory power of GWR is (2017), Dharumarajan et al. (2017) and Wang et al. (2018).
a little spatially different in term of local R2. It obviously verified that Furthermore, RF model evaluates the important variables through
the mentioned remote sensing-based indices had low ability in esti- removing one or more variables and assess how prediction would be
mating SOC in all parts of the Sahzand watershed. changed (Fox et al., 2017). The importance of variables for RF was
The regression quality variables were compared to other related evaluated by the percentage increase in the mean-squared error (%In-
regional-scaled researches. In case of GWR, Wang et al. (2012) and cMSE). The larger value of %IncMSE for a variable means the influence
Wang et al. (2018) evaluated spatial distribution of SOC using GWR and of variable on the accuracy of model prediction is larger. IncNodePurity
regression kriging (RK) and concluded the better performance of GWR. represents the total decrease in node impurity from splitting on a pre-
In addition, Kumar et al. (2013) referred that GWR performed better dictor in the tree construction process, averaged over all trees.
rather than another used model in capturing SOC in Ohio, USA. The As depicted in Fig. 5, some explanatory variables namely CI and
small R2 in this study may be attributed to the low SOC values (< 2%) NDMI are more important rather than other variables for SOC predic-
with non-significant spatial variability in various parts of the watershed tion in the watershed. In decreasing order of importance, the other
led to incomprehensible bias in the analyses. Furthermore, the low variables included in the RF model were elevation, NDVI, slope, TWI, B
6. Aspect had very little impact on SOC prediction, shuffling lead to a

7
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Fig.4. Local R2 values estimates for model 18, GWR in the Shazand Watershed, Iran.

watershed. Overall, the findings of present study in the Shazand Wa-


tershed with arid and semi-arid ecosystems and calcareous soils dis-
agree Kumar et al. (2016) because of difference in the ecosystem under
consideration (i.e., forest ecosystem). They conclude that optical re-
mote sensing data have potential for predicting SOC. Similarly, Siewert
(2018) reported the high potential of remote sensing for prediction of
SOC in tropical and forest ecosystems.

4. Conclusions

By considering the importance of SOC in regional carbon cycling


and environmental management, a precise spatial mapping of SOC will
potentially help policymakers take reasonable land use and manage-
ment measures for ecological restoration and rehabilitation. In this
way, exploring the ability of remote sensing oriented techniques and
methods is crucial due to capability of overcoming the governing con-
straints of field surveys. Therefore, the primary aim of the current study
was to estimate SOC in the study area using environmental variables as
measures of vegetation and soil conditions to estimate SOC in the
Fig. 5. Rankings variable importance of RF model for prediction of SOC (% Shazand Watershed, Iran. The relationship between study variables and
IncMSE = percent increase in Mean Square Error and IncNodPurity = decrease SOC was described by GWR, OLS, and RF. Since some 98% of the study
in node impurity). watershed is covered by soils with very low SOC content of below 1%,
none of studied variables was significantly related to SOC, while if the
slight increase in accuracy, which make small negative importance correlation is high at the appropriate level of significance and corre-
score and it can be considered as zero (Richardson, 2015). These sponding errors are low, the information on the spatial mapping and
variables were relatively weak predictors in comparison with first order dynamic of SOC, and the identification of the determinant factors on
ones. SOC distribution, would be sufficiently reliable. In other words, this
Sun et al. (2015) reported that land use and topography are the study showed the low ability of Landsat images and environmental and
most important variables for prediction of SOC density in China. Liu remote sensing-based indices for estimating SOC in the study arid wa-
et al. (2015) also concluded the vegetation biomass, climatic variables tershed. Considering the importance of accurate predicting and map-
and soil moisture all influenced distribution characteristics of SOC. ping of SOC, it is recommended soil samples to be taken from the top
Bhunia et al. (2017) found significant relationships between vegetation 50 cm of the soil profile with specified intervals. In addition, further
indices and SOC. Zhou et al. (2020) reported that the combination of research should be explored to consider human-induced variables for
precipitation, elevation, and temperature had the best prediction per- estimating spatial distribution of SOC. Besides, conducting comparative
formance explaining 75% of SOC variation. Nevertheless, in present studies using some processed-based models such as DNDC and TECO-R
study, due to lesser SOC variations across the watershed, the variables models is recommended to compare the outputs obtained by the remote
did not emerge as important factors in predicting SOC in the study sensing-based indices in agricultural and rangeland ecosystems, re-
spectively, to improve the results in future research. These models were

8
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

not applied in the current study due to the need for a wide range of vegetation and surface temperature at the city scale. Sci. Total Environ. 656,
input data not comprehensively existing for the country. 118–128. https://doi.org/10.1016/j.scitotenv.2018.11.223.
Fotheringham, A.S., Brunsdon, C., Charlton, M., 2002. Geographically Weighted
Regression: The Analysis of Spatially Varying Relationships. Wiley, Chichester, pp.
CRediT authorship contribution statement 269.
Fox, E.W., Hill, R.A., Leibowitz, S.G., Olsen, A.R., Thornbrugh, D.J., Weber, M.H., 2017.
Assessing the accuracy and stability of variable selection methods for random forest
Fahimeh Mirchooli: Conceptualization, Methodology, modeling in ecology. Environ. Monit. Assess. 189 (7), 316. https://doi.org/10.1007/
Investigation, Formal analysis, Resources, Data curation, Software, s10661-017-6025-0.
Writing - original draft. Mahboobeh Kiani-Harchegani: Gao, J., Li, S., 2011. Detecting spatially non-stationary and scale-dependent relationships
between urban landscape fragmentation and related factors using geographically
Conceptualization, Methodology, Investigation, Resources, Data cura- weighted regression. Appl. Geogr. 31 (1), 292–302. https://doi.org/10.1016/j.
tion. Abdulvahed Khaledi Darvishan: Methodology, Investigation, apgeog.2010.06.003.
Software. Samereh Falahatkar: Methodology, Investigation, Software. Georganos, S., Abdi, A.M., Tenenbaum, D.E., Kalogirou, S., 2017. Examining the NDVI-
rainfall relationship in the semi-arid Sahel using geographically weighted regression.
Seyed Hamidreza Sadeghi: Conceptualization, Methodology,
J. Arid Environ. 146, 64–74. https://doi.org/10.1016/j.jaridenv.2017.06.004.
Investigation, Validation, Resources, Data curation, Software, Writing - Gholizadeh, A., Žižala, D., Saberioon, M., Borůvka, L., 2018. Soil organic carbon and
review & editing, Supervision, Project administration, Funding acqui- texture retrieving and mapping using proximal, airborne and Sentinel-2 spectral
sition. imaging. Remote Sens. Environ. 218, 89–103. https://doi.org/10.1016/j.rse.2018.
09.015.
Grimm, R., Behrens, T., Märker, M., Elsenbeer, H., 2008. Soil organic carbon con-
Declaration of Competing Interest centrations and stocks on Barro Colorado Island—digital soil mapping using random
forests analysis. Geoderma 146 (1–2), 102–113. https://doi.org/10.1016/j.
geoderma.2008.05.008.
The authors declare that they have no known competing financial Grinand, C., Le Maire, G., Vieilledent, G., Razakamanarivo, H., Razafimbelo, T., Bernoux,
interests or personal relationships that could have appeared to influ- M., 2017. Estimating temporal changes in soil carbon stocks at ecoregional scale in
Madagascar using remote sensing. Int. J. Appl. Earth Obs. Geoinf. 54, 1–14. https://
ence the work reported in this paper. doi.org/10.1016/j.jag.2016.09.002.
Islam, N., Hossen, S., Baten, A., 2016. Soil carbon and nitrogen dynamics in agricultural
Acknowledgements soils of Mymensingh, Bangladesh. Int. J. Agric. Biosyst. Eng. 1(1), 1–8. https://doi:
10.2134/jeq2018.10.0384.
Jebari, A., del Prado, A., Pardo, G., Rodríguez Martín, J.A., Álvaro-Fuentes, J., 2018.
The authors wish to thank the National Elites Foundation, Islamic Modeling regional effects of climate change on soil organic carbon in Spain. J.
Republic of Iran for it’s partly supports of this research (Project No. Environ. Qual. 47 (4), 644–653. https://doi.org/10.2134/jeq2017.07.0294.
Jin, Y., Yang, X., Qiu, J., Li, J., Gao, T., Wu, Q., Zhao, F., Ma, H., Haida, Y., Xu, B., 2014.
15.90186). The partial support of the Agrohydrology Research Group of
Remote sensing-based biomass estimation and its spatio-temporal variations in
Tarbiat Modares University (Grant No. IG-39713), Iran, concerning the temperate grassland, Northern China. Remote Sens. 6 (2), 1496–1513. https://doi.
corresponding author is also acknowledged. org/10.3390/rs6021496.
Kalota, D., 2017. Exploring relation of land surface temperature with selected variables
using geographically weighted regression and ordinary least square methods in
References Manipur State, India. Geocarto Int. 32 (10), 1105–1119. https://doi.org/10.1080/
10106049.2016.1195883.
Ahmadi, M., Kashki, A., Roudbari, A.D., 2018. Spatial modeling of seasonal pre- Kiani Harchegani, M., Sadeghi, S.H.R., Boor, S., 2019a. a. Comparative analysis of soil
cipitation–elevation in Iran based on aphrodite database. Model. Earth Syst. Environ. variables in different land uses of the Shazand watershed, Iran. J. Agric. For. 65 (1),
4 (2), 619–633. https://doi.org/10.1007/s40808-018-0444-y. 81–87.
Almeida Salles, L., Lima, J.E.F.W., Roig, H.L., Malaquias, J.V., 2018. Environmental Kiani-Harchegani, M., Sadeghi, S.H.R., Falahatkar, S., 2019b. Comparative analysis of soil
factors and groundwater behavior in an agricultural experimental basin of the erodibility factor in Shazand Watershed, Iranian. J Ecohydrology 6 (1), 153–163.
Brazilian central plateau. Appl. Geogr. 94, 272–281. https://doi.org/10.1016/j. Kiani Harchegani, M., Sadeghi, S.H.R., 2020. Practicing land degradation neutrality
apgeog.2018.02.007. (LDN) approach in the Shazand Watershed, Iran. Sci. Total Environ. 698 (1). https://
Bartholomeus, H.M., Schaepman, M.E., Kooistra, L., Stevens, A., Hoogmoed, W.B., doi.org/10.1016/j.scitotenv.2019.134319.
Spaargaren, O.S.P., 2008. Spectral reflectance based indices for soil organic carbon Knadel, M., Thomsen, A., Schelde, K., Greve, M.H., 2015. Soil organic carbon and particle
quantification. Geoderma 145 (1–2), 28–36. https://doi.org/10.1016/j.geoderma. sizes mapping using vis–NIR, EC and temperature mobile sensor platform. Comput.
2008.01.010. Electron. Agric. 114, 134–144. https://doi.org/10.1016/j.compag.2015.03.013.
Bednář, M., Šarapatka, B., 2018. Relationships between physical–geographical factors Kumar, S., Lal, R., Liu, D., Rafiq, R., 2013. Estimating the spatial distribution of organic
and soil degradation on agricultural land. Environ. Res. 164, 660–668. https://doi. carbon density for the soils of Ohio, USA. J. Geogr. Sci. 23 (2), 280–296. https://doi.
org/10.1016/j.envres.2018.03.042. org/10.1007/s11442-013-1010-1.
Bhunia, G.S., Kumar Shit, P., Pourghasemi, H.R., 2017. Soil organic carbon mapping Kumar, P., Pandey, P.C., Singh, B.K., Katiyar, S., Mandal, V.P., Rani, M., Rani, M., Tomar,
using remote sensing techniques and multivariate regression model. Geocarto Int. V., Patairiya, S., 2016. Estimation of accumulated soil organic carbon stock in tro-
1–12. https://doi.org/10.1080/10106049.2017.1381179. pical forest using geospatial strategy, Egypt. J. Remote Sens. Sp. Sci. 19 (1), 109–123.
Castaldi, F., Hueni, A., Chabrillat, S., Ward, K., Buttafuoco, G., Bomans, B., Vreys, K., https://doi.org/10.1016/j.ejrs.2015.12.003.
Brell, M., van Wesemael, B., 2019. Evaluating the capability of the Sentinel 2 data for Kumar, P., Sajjad, H., Tripathy, B.R., Ahmed, R., Mandal, V.P., 2018a. Prediction of
soil organic carbon prediction in croplands. ISPRS J. Photogramm. Remote Sens. 147, spatial soil organic carbon distribution using Sentinel-2A and field inventory data in
267–282. https://doi.org/10.1016/j.isprsjprs.2018.11.026. Sariska Tiger Reserve. Nat. Hazards 90 (2), 693–704. https://doi.org/10.1007/
Dadashpoor, H., Azizi, P., Moghadasi, M., 2019. Land use change, urbanization, and s11069-017-3062-5.
change in landscape pattern in a metropolitan area. Sci. Total Environ. 655, 707–719. Kumar, N., Velmurugan, A., Hamm, N.A., Dadhwal, V.K., 2018b. Geospatial mapping of
https://doi.org/10.1016/j.scitotenv.2018.11.267. soil organic carbon using regression Kriging and remote sensing. J. Indian Soc.
Dalal, R.C., Mayer, R.J., 1986. Long term trends in fertility of soils under continuous Remote Sens. 46 (5), 705–716. https://doi.org/10.1007/s12524-017-0738-y.
cultivation and cereal cropping in southern Queensland. I. Overall changes in soil Ladoni, M., Alavipanah, S.K., Bahrami, H.A., Noroozi, A.A., 2010. Remote sensing of soil
properties and trends in winter cereal yields. Soil Res. 24 (2), 265–279. https://doi. organic carbon in semi-arid region of Iran. Arid Land Res. Manag. 24 (4), 271–281.
org/10.1071/SR9860293. https://doi.org/10.1080/15324982.2010.502917.
Dai, F., Zhou, Q., Lv, Z., Wang, X., Liu, G., 2014. Spatial prediction of soil organic matter Lin, C., Zhu, A.Z., Wang, X., Ma, R., 2020. The refined spatiotemporal representation of
content integrating artificial neural network and ordinary kriging in Tibetan Plateau. soil organic matter based on remote images fusion of Sentinel-2 and Sentinel-3. Int. J.
Ecol. Indic. 45, 184–194. https://doi.org/10.1016/j.ecolind.2014.04.003. Appl. Earth Obs. Geoinf. 89, 102094. https://doi.org/10.1016/j.jag.2020.102094.
Davudirad, A.A., Sadeghi, S.H.R., Sadoddin, A., 2016. The impact of development plans Liou, Y.A., Nguyen, K.A., Li, M.H., Lin, C.Y., 2015. Landsat 8 operational land imager-
on hydrological changes in the Shazand Watershed. Iran. L. Degrad. Dev. 27 (4), derived variables for environmental risk assessment in Taoyuan. IEEE Int. Geosci.
1236–1244. https://doi.org/10.1002/ldr.2523. Remote Sens. Symp., pp. 874–877. https://doi.org/10.1109/IGARSS.2015.7325904.
Dharumarajan, S., Hegde, R., Singh, S.K., 2017. Spatial prediction of major soil properties Liu, H.Q., Huete, A., 1995. A feedback based modification of the NDVI to minimize ca-
using Random Forest techniques-A case study in semi-arid tropics of South India. nopy background and atmospheric noise. IEEE Trans. Geosci. Remote Sens. 33 (2),
Geoderma Reg. 10, 154–162. https://doi.org/10.1016/j.geodrs.2017.07.005. 457–465. https://doi.org/10.1109/36.377946.
Dorji, T., Odeh, I.O., Field, D.J., Baillie, I.C., 2014. Digital soil mapping of soil organic Liu, S., An, N., Yang, J., Dong, S., Wang, C., Yin, Y., 2015. Prediction of soil organic
carbon stocks under different land use and land cover types in montane ecosystems, matter variability associated with different land use types in mountainous landscape
Eastern Himalayas. Forest Ecol. Manag. 318, 91–102. https://doi.org/10.1016/j. in southwestern Yunnan province, China. Catena 133, 137–144. https://doi.org/10.
foreco.2014.01.003. 1016/j.catena.2015.05.010.
Duncan, J.M.A., Boruff, B., Saunders, A., Sun, Q., Hurley, J., Amati, M., 2019. Turning Luo, Z., Feng, W., Luo, Y., Baldock, J., Wang, E., 2017. Soil organic carbon dynamics
down the heat: an enhanced understanding of the relationship between urban jointly controlled by climate, carbon inputs, soil properties and soil carbon fractions.
Glob. Change Biol. 23 (10), 4430–4439. https://doi.org/10.1111/gcb.13767.

9
F. Mirchooli, et al. Ecological Indicators 116 (2020) 106473

Mandal, U.K., 2016. Spectral color indices based geospatial modeling of soil organic Stumpf, F., Keller, A., Schmidt, K., Mayr, A., Gubler, A., Schaepman, M., 2018. Spatio-
matter in Chitwan district. Nepal. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. temporal land use dynamics and soil organic carbon in Swiss agroecosystems. Agric.
- ISPRS Arch. 41. https://doi.org/10.5194/isprsarchives-XLI-B2-43-2016. Ecosyst. Environ. 258, 129–142. https://doi.org/10.1016/j.agee.2018.02.012.
Martin, M.P., Orton, T.G., Lacarce, E., Meersmans, J., Saby, N.P.A., Paroissien, J.B., Sun, W., Zhu, H., Guo, S., 2015. Soil organic carbon as a function of land use and to-
Arrouays, D., 2014. Evaluation of modelling approaches for predicting the spatial pography on the Loess Plateau of China. Ecol. Eng. 83, 249–257. https://doi.org/10.
distribution of soil organic carbon stocks at the national scale. Geoderma 223, 1016/j.ecoleng.2015.06.030.
97–107. https://doi.org/10.1016/j.geoderma.2014.01.005. Taghipour Javi, S., Malekmohammadi, B., Mokhtari, H., 2014. Application of geo-
Matsushita, B., Yang, W., Chen, J., Onda, Y., Qiu, G., 2007. Sensitivity of the enhanced graphically weighted regression model to analysis of spatiotemporal varying re-
vegetation index (EVI) and normalized difference vegetation index (NDVI) to topo- lationships between groundwater quantity and land use changes (case study:
graphic effects: a case study in high-density cypress forest. Sensors 7 (11), Khanmirza Plain, Iran). Environ. Monit. Assess. 186 (5), 3123–3138. https://doi.org/
2636–2651. https://doi.org/10.3390/s7112636. 10.1007/s10661-013-3605-5.
Mishra, U., Lal, R., Liu, D., Van Meirvenne, M., 2010. Predicting the spatial variation of Taghizadeh-Mehrjardi, R., Nabiollahi, K., Kerry, R., 2016. Digital mapping of soil organic
the soil organic carbon pool at a regional scale. Soil Sci. Soc. Am. J. 74 (3), 906–914. carbon at multiple depths using different data mining techniques in Baneh region,
https://doi.org/10.2136/sssaj2009.0158. Iran. Geoderma 266, 98–110. https://doi.org/10.1016/j.geoderma.2015.12.003.
Miller, B.A., Koszinski, S., Wehrhan, M., Sommer, M., 2015. Comparison of spatial as- Takata, Y., 2010. Analysis of spatial and temporal variation of soil organic carbon budget
sociation approaches for landscape mapping of soil organic carbon stocks. Soil 1 (1), in northern Kazakhstan. JPN Agric. Res. 44 (4), 335–342. https://doi.org/10.1016/
217. https://doi.org/10.5194/soil-1-217-2015. S0160-4120(01)00070-8.
Mondal, A., Khare, D., Kundu, S., Mondal, S., Mukherjee, S., Mukhopadhyay, A., 2017. Tsui, C.C., Tsai, C.C., Chen, Z.S., 2013. Soil organic carbon stocks in relation to elevation
Spatial soil organic carbon (SOC) prediction by regression kriging using remote gradients in volcanic ash soils of Taiwan. Geoderma 209, 119–127. https://doi.org/
sensing data, The Egypt. J. Rem. Sens. Space Sci. 20 (1), 61–70. https://doi.org/10. 10.1016/j.geoderma.2013.06.013.
1016/j.ejrs.2016.06.004. Tu, J., 2011. Spatially varying relationships between land use and water quality across an
Muchana, R., 2017. Estimating soil carbon stocks in a dry Miombo ecosystem using re- urbanization gradient explored by geographically weighted regression. Appl. Geogr.
mote sensing. Forest Res. 6 (1), 1–6. https://doi.org/10.4172/2168- 9776.1000198. 31 (1), 376–392. https://doi.org/10.1016/j.apgeog.2010.08.001.
Naghibi, S.A., Ahmadi, K., Daneshi, A., 2017. Application of support vector machine, Vancampenhout, K., Nyssen, J., Gebremichael, D., Deckers, J., Poesen, J., Haile, M.,
random forest, and genetic algorithm optimized random forest models in ground- Moeyersons, J., 2006. Stone bunds for soil conservation in the northern Ethiopian
water potential mapping. Water Resour. Manag. 31 (9), 2761–2775. https://doi.org/ highlands: Impacts on soil fertility and crop yield. Soil Tillage Res. 90 (1–2), 1–15.
10.1007/s11269-017-1660-3. https://doi.org/10.1016/j.still.2005.08.004.
Nelson, D.W., Sommers, L.E. 1996. Total carbon, organic carbon and organic matter, in: Wang, S., Tian, H., Liu, J., Pan, S., 2003. Pattern and change of soil organic carbon
Sparks, D.L. (Ed.), Methods of Soil Analysis. Part 3. Chemical Methods, SSSA Book storage in China: 1960s–1980s. Tellus B Chem. Phys. Meteorol. 55 (2), 416–427.
Ser. 5. 3., Madison, USA, Soil Sci. Soc. Am. J., 961-1010. https://doi.org/10.1034/j.1600-0889.2003.00039.x.
Oueslati, I., Allamano, P., Bonifacio, E., Claps, P., 2013. Vegetation and topographic Wang, Q., Ni, J., Tenhunen, J., 2005. Application of a geographically-weighted regression
control on spatial variability of soil organic carbon. Pedosphere 23 (1), 48–58. analysis to estimate net primary production of Chinese forest ecosystems. Glob. Ecol.
https://doi.org/10.1016/S1002-0160(12)60079-4. Biogeogr. 14 (4), 379–393. https://doi.org/10.1111/j.1466-822X.2005.00153.x.
Parshotam, A., Saggar, S., Tate, K., Parfitt, R., 2001. Modelling organic matter dynamics Wang, K., Zhang, C., Li, W., 2012. Comparison of geographically weighted regression and
in New Zealand soils. Environ. Int. 27 (2–3), 111–119. https://doi.org/10.1016/ regression kriging for estimating the spatial distribution of soil organic matter.
S0160-4120(01)00070-8. GIScience Remote Sens. 49 (6), 915–932. https://doi.org/10.2747/1548-1603.49.6.
Polyakov, V., Lal, R., 2004. Modeling soil organic matter dynamics as affected by soil 915.
water erosion. Environ. Int. 30 (4), 547–556. https://doi.org/10.1016/j.envint.2003. Wang, S., Adhikari, K., Wang, Q., Jin, X., Li, H., 2018. Role of environmental variables in
10.011. the spatial distribution of soil carbon (C), nitrogen (N), and C: N ratio from the
Rahmati, O., Pourghasemi, H.R., Melesse, A.M., 2016. Application of GIS-based data northeastern coastal agroecosystems in China. Ecol. Indic. 84, 263–272. https://doi.
driven random forest and maximum entropy models for groundwater potential org/10.1016/j.ecolind.2017.08.046.
mapping: a case study at Mehran Region, Iran. Catena 137, 360–372. https://doi.org/ Wang, S., Fan, J., Zhong, H., Li, Y., Zhu, H., Qiao, Y., Zhang, H., 2019. A multi-factor
10.1016/j.catena.2015.10.010. weighted regression approach for estimating the spatial distribution of soil organic
Restrepo, C., LaQuay, G., 2018, April. Variation in carbon dynamics mediated by land- carbon in grasslands. Catena 174, 248–258. https://doi.org/10.1016/j.catena.2018.
slide activity in a Central American mountainscape. In EGU General Assembly 10.050.
Conference Abstracts, vol. 20, p. 17184. Were, K., Bui, D.T., Dick, Ø.B., Singh, B.R., 2015. A comparative assessment of support
Richardson, H.J., 2015. A comparison of random forests and linear stepwise regressions to vector regression, artificial neural networks, and random forests for predicting and
model and map soil carbon in south-central British Columbia Grasslands using mapping soil organic carbon stocks across an Afromontane landscape. Ecol. Indic. 52,
Normalized Difference Vegetation Index based models (Doctoral Dissertation). 394–403. https://doi.org/10.1016/j.ecolind.2014.12.028.
Thompson Rivers University. Wu, S.S., Yang, H., Guo, F., Han, R.M., 2017. Spatial patterns and origins of heavy metals
Rodriguez-Galiano, V., Mendes, M.P., Garcia-Soldado, M.J., Chica-Olmo, M., Ribeiro, L., in Sheyang River catchment in Jiangsu, China based on geographically weighted
2014. Predictive modeling of groundwater nitrate pollution using Random Forest and regression. Sci. Total Environ. 580, 1518–1529. https://doi.org/10.1016/j.scitotenv.
multisource variables related to intrinsic and specific vulnerability: a case study in an 2016.12.137.
agricultural setting (Southern Spain). Sci. Total Environ. 476, 189–206. https://doi. Yang, R., Rossiter, D.G., Liu, F., Lu, Y., Yang, F., Yang, F., Zhao, Y., Li, D., Zhang, G., 2015.
org/10.1016/j.scitotenv.2014.01.001. Predictive mapping of topsoil organic carbon in an alpine environment aided by
Sadeghi, S.H.R., Hazbavi, Z., 2017. Spatiotemporal variation of watershed health pro- Landsat TM. PLoS One 10 (10), 1–20. https://doi.org/10.1371/journal.pone.
pensity through reliability-resilience-vulnerability based drought index (case study: 0139042.
Shazand Watershed in Iran). Sci. Total Environ. 587, 168–176. https://doi.org/10. Yu, H., Kong, B., Wang, G., Du, R., Qie, G., 2018. Prediction of soil properties using a
1016/j.scitotenv.2017.02.098 0048-9697. hyperspectral remote sensing method. Arch. Agron. Soil Sci. 64 (4), 546–559.
Salimi, S., Balyani, S., Hosseini, S.A., Momenpour, S.E., 2018. The prediction of spatial https://doi.org/10.1080/03650340.2017.1359416.
and temporal distribution of precipitation regime in Iran: the case of Fars province. Zabihi, M., Pourghasemi, H.R., Pourtaghi, Z.S., Behzadfar, M., 2016. GIS-based multi-
Model. Earth Syst. Environ. 4 (2), 565–577. https://doi.org/10.1007/s40808-018- variate adaptive regression spline and random forest models for groundwater po-
0451-z. tential mapping in Iran. Environ. Earth Sci. 75 (8), 665. https://doi.org/10.1007/
Shi, S., Cao, Q., Yao, Y., Tang, H., L.Z., 2014. Influence of climate and socio-economic s12665-016-5424-9.
factors on the spatio-temporal variability of soil organic matter: a case study of Zhang, X., Estoque, R.C., Murayama, Y., 2017. An urban heat island study in Nanchang
central Heilongjiang Province, China, J. Int. Agric, 13(7), 1486–1500. https://doi. City, China based on land surface temperature and social-ecological variables.
org/10.1016/S2095-3119(14)60815-7. Sustain. Cities Soc. 32, 557–568. https://doi.org/10.1016/j.scs.2017.05.005.
Siewert, M.B., 2018. High-resolution digital mapping of soil organic carbon in permafrost Zhao, C., Jensen, J., Weng, Q., Weaver, R., 2018. A geographically weighted regression
terrain using machine learning: a case study in a sub-Arctic peatland environment. analysis of the underlying factors related to the surface urban heat Island phenom-
Biogeosci. Discuss 15 (6), 1663–1682. https://doi.org/10.5194/bg-2017-323. enon. Remote Sens. 10 (9), 1428. https://doi.org/10.3390/rs10091428.
Somarathna, P.D.S.N., Malone, B.P., Minasny, B., 2016. Mapping soil organic carbon Zissimos, A.M., Christoforou, I.C., Cohen, D.R., Mooney, S.D., Rutherford, N.F., 2019.
content over New South Wales, Australia using local regression kriging. Geoderma Spatial distribution and controls on organic and inorganic carbon in the soils of
Reg. 7 (1), 38–48. https://doi.org/10.1016/j.geodrs.2015.12.002. Cyprus. J. Geochem. Explor. 196, 95–104. https://doi.org/10.1016/j.gexplo.2018.
Song, X.D., Brus, D.J., Liu, F., Li, D.C., Zhao, Y.G., Yang, J.L., Zhang, G.L., 2016. Mapping 10.005.
soil organic carbon content by geographically weighted regression: a case study in the Zhou, T., Geng, Y., Chen, J., Liu, M., Haase, D., Lausch, A., 2020. Mapping soil organic
Heihe River Basin, China. Geoderma 261, 11–22. https://doi.org/10.1016/j. carbon content using multi-source remote sensing variables in the Heihe River Basin
geoderma.2015.06.024 0016-7061. in China. Ecol. Indic. 114, 106288. https://doi.org/10.1016/j.ecolind.2020.106288.

10

You might also like