You are on page 1of 11

Geoderma 337 (2019) 1309–1319

Contents lists available at ScienceDirect

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

Estimating soil salinity from remote sensing and terrain data in southern T
Xinjiang Province, China

Jie Penga,b, Asim Biswasc, Qingsong Jianga,d, Ruiying Zhaoa, Jie Hua, Bifeng Hua, Zhou Shia,
a
Institute of Applied Remote Sensing and Information Technology, College of Environmental and Resource Sciences, Zhejiang University, Hangzhou 310058, China
b
College of Plant Science, Tarim University, Alar 843300, China
c
School of Environmental Sciences, University of Guelph, 50 Stone Road East, Guelph, ON N1G 2W1, Canada
d
College of Information Engineering, Tarim University, Alar 843300, China

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

Keywords: Soil salinization is one of the main reasons for soil health and ecosystem deterioration in most degraded arid and
Soil salinization semiarid areas. To monitor its spatial variation as precise as possible over a large area, we collected 225 samples
Cubist using traditional field experiment and laboratory analysis method from the southern part of the Xinjiang
Partial least squares regression Province, China, affected by soil salinity under strong arid climate. Then, we constructed both Cubist and partial
Remote sensing
least square regression (PLSR) models on electrical conductivity (EC) (150 ground-based measurements as ca-
Digital soil mapping
libration set) using various related covariates (e.g. terrain attributes, remotely sensed spectral indices of vege-
tation and salinity from landsat8 OLI satellite) that are at the same time period corresponding to soil sampling.
Two models were validated using remaining 75 independent ground based measurements and were then used to
map the soil salinity over the study area. Finally, the validation results of two models were compared under
different intervals of EC, soil moisture content and vegetation coverage. The results indicated that Cubist model
could predict EC value with better accuracy and stability under variable environment than PLSR. The R2, RMSE,
MAE and RPD of the Cubist model were 0.91, 5.18 dS m−1, 3.76 dS m−1 and 3.15 while corresponding values of
the PLSR model were 0.66, 10.46 dS m−1, 8.21 dS m−1 and 1.56 in validation dataset, respectively. Additionally,
the map derived from Cubist model revealed more detailed variation information of the spatial distribution of EC
than that from PLSR model across the study area. Thus, Cubist model was recommended for mapping soil salinity
using indices derived from satellite and terrain in other arid areas.

1. Introduction agricultural operations and the need for irrigation water increased
sharply. On the other hand, high demand of water for reclamation
As a current global issue, soil salinization is critically affecting our process increased the water scarcity and affected the ecosystem func-
limited soil resource and deteriorating the ecosystem health (Li et al., tioning (Deng et al., 2013). This also poses a serious threat to the sus-
2015). Meanwhile, it leads to soil desertification and land degradation, tainable development of regional agriculture and economy (Wang et al.,
and results in sharp decrease in soil productivity, vegetation cover and 2006). Therefore, only dynamic monitoring of soil salinization can
biodiversity (Dehaan and Taylor, 2003; Metternicht and Zinck, 2003; provide much more quantitative information for land reclamation.
Farifteh et al., 2006; Zhang et al., 2015; Gorji et al., 2017). Almost 3% Traditional laboratory analysis of characterizing soil is time-con-
of the global soil resource is salt-affected (Gorji et al., 2015) and this suming. Owing to the great temporal and spatial variation, it is very
area is expanding at a speed of 2.00 × 106 ha per year (Abbas et al., difficult to obtain the dynamic information (Allbed et al., 2014;
2013). In semi-arid and arid regions, the soil salinization is especially Barbouchi et al., 2015; Harti et al., 2016) and to reveal the evolution
prominent due to the scarcity of rainfall (Ma et al., 2018), the great process and trend of soil salinization. Compared with traditional
amount of evaporation, the high water table and the high water soluble methods, satellite remote sensing technology has great advantages in
salt content. Due to increasing demand for food and living materials observing ground at large spatial scales and high temporal resolutions.
from the rapidly growing population put an extraordinary pressure on It can also provide spectral information of soil salinization repeatedly
already degraded soil resources, saline soils have been reclaimed for with short temporal intervals. At present, there are many satellite


Corresponding author.
E-mail address: shizhou@zju.edu.cn (Z. Shi).

https://doi.org/10.1016/j.geoderma.2018.08.006
Received 16 November 2017; Received in revised form 20 April 2018; Accepted 6 August 2018
Available online 22 August 2018
0016-7061/ © 2018 Elsevier B.V. All rights reserved.
J. Peng et al. Geoderma 337 (2019) 1309–1319

remote sensing data with moderate or high spatial and temporal re- between ecological water and irrigation water. Meanwhile, it also en-
solutions, which provide new opportunities for monitoring spatial dis- dangers the sustainable development of agriculture and ecological en-
tribution characteristics of soil salinization using digital soil mapping vironment in this area.
techniques (Farifteh et al., 2006; Mehrjardi et al., 2014). In addition,
only a small amount of soil salinization survey data (ground-based) is 2.2. Sampling campaigns and laboratory measurement
needed for constructing predictive model between remotely sensed soil
salinity information and ground based monitoring. Therefore, satellite On the 24th and 25th July 2016, a total of 225 topsoil mixed
remote sensing technology shows high promise in monitoring and samples were collected at a depth of 0–20 cm. In regions where the
mapping of soil salinization at high spatial and temporal resolutions. vehicle can reach, 250 quadrats were selected with the same spatial
The theoretical basis for remote sensing monitoring is building upon resolutions (30 m) of the Landsat8 OLI. Applying Quincunx sampling
the spectral characteristics of saline soils. The soil reflectance increases method (Cheng et al., 2007), five topsoil samples were collected in each
with the increase of soil salinization in visible-near infrared-shortwave quadrat using a 5.72 cm diameter soil drill. The latitude, longitude and
infrared bands (Bannari et al., 2008; Sidike et al., 2014; Harti et al., altitude of each sampling point were recorded using a handheld GPS.
2016). However, in the natural environment, it is very difficult to ob- Soil samples were sealed in plastic bags (Ziploc bag) to prevent the loss
tain pure spectral information of saline soils using remote sensing due of water through evaporation and brought back to the laboratory for
to the interferences from other factors, such as soil moisture, vegetation further analysis. At each quadrat, five samples were gathered and the
cover and data acquisition time. In general, vegetation coverage shows locations of the samples have also been recorded. If at least 3 of 5
a decreasing trend with increasing soil salinity. Overall, Monitoring samples were located in the same pixel of Landsat8 OLI, which meant
accuracy in the dry season is usually higher than that in the wet season this quadrat is valid, or invalid. In this way, there were 225 valid
(Ding and Yu, 2014). Additionally, most linear models have large errors quadrats among the 250. Using quartering method, all the valid sam-
in the lower and higher values over regions where spatial variability of ples at each valid quadrat were fully mixed and subsampled to retain
salinity is greater (Scudiero et al., 2015; Sidike et al., 2014). 300 g of soil. After removing the stones, plant roots and litter, the oven-
How to reduce interference of soil moisture, vegetation and other drying method was applied to a part of soil sample to determine the
factors on the remote sensing monitoring of soil salinization? How can moisture content. Each of the remaining soil samples was air-dried,
we improve monitoring accuracy, especially in diverse types of land- ground and sieved through a 2 mm sieve. To measure soil electrical
scapes? These questions became unavoidable in remote sensing mon- conductivity, 250 ml distilled water was added into every 50 g pro-
itoring of soil salinization and were the objectives of the current study. cessed soil sample and mixed properly using a shaker for 30 mins (Peng
Firstly, field experiments were conduct for ground validations in the et al., 2016). The leachate was extracted to measure the electrical
southern part of Xinjiang province, China, with typical alluvial fan and conductivity using a LeiCi DDS-307 (ShengKe, ShangHai, China) con-
diverse types of land use. Hereafter, piecewise linear tree model (Ma ductivity meter (Peng et al., 2016). Soil pH was measured in a 1:2 soil/
et al., 2017) and partial least squares regression model were used for distilled water suspension using a pre-calibrated glass electrode
quantitative inversion and digital mapping on EC. Further, the Cubist (Masoud, 2014) of HORIBA D-53 pH meter at a resolution of 0.01. The
model was recommended for achieving high-accuracy prediction on soil solutions of pH 7 and 10 were used for calibration before every mea-
salinization. surement.

2. Material and methods 2.3. Remote sensing data and processing

2.1. Study area Two Landsat 8 OLI images acquired on July 24th, 2016 were used in
this study. The path/row was 146/31 and 146/32, respectively for the
There is 4.67 × 106 ha of saline soil, accounting for 22.36% of the images and cloud cover was lower than 10% for both images. Seven
available land area in Southern Xinjiang and 55.1% of the total area spectral bands in visible and infrared wavelength (B1 (Costal,
with saline soils in Xinjiang (Turhong et al., 2008). The study area is a 0.43–0.45 μm), B2 (Blue, 0.45–0.51 μm), B3 (Green, 0.53–0.59 μm), B4
typical alluvial fan and is located in Wensu county of southern Xinjiang (Red, 0.64–0.67 μm), B5 (NIR, 0.85–0.88 μm), B6(SWIR1,
province in northwestern China (40°47′ 40′–41°34′ N, 80°43′–81°19′ E) 1.57–1.65 μm) and B7 (SWIR2, 2.11–2.29 μm)) with a grid resolution of
(Fig. 1). The area extends from east to west for 61 km and from north to approximately 30 m, and one thermal infrared band (B10(TIRS1,
south for 93 km. The elevation ranges from 960 to 1460 m with an 10.60–11.19 μm) with a grid resolution of approximately 100 m were
increasing gradient from the southwest to the northeast. The climate is used in this study. To decrease the effect of atmospheric scattering on
typical Continental warm temperate arid climate. The mean annual the quality of the remote sensing images, radiation correction and at-
precipitation is 60 mm while evaporation is 1800 mm with a ratio of mospheric correction using FLAASH model were performed. Principal
1:30. The study area owes abundant light and heat resources with an Component Analysis (PCA) and Tasseled Cap (TC) transformation based
average annual total solar radiation of 544–590 kJ cm−2, an annual on the correlation matrix were carried out with the corrected images to
total sunshine time of 2855–2967 h and a frost-free period of reduce the total number of data layers and for better discrimination
205–219 days (Shi et al., 2006). In the direction from north to south, between saline and non-saline soil. In addition, a number of vegetation
there is a cemented road, which traverses the study area along with spectral indices and salinity spectral indices were computed and they
several tractor-ploughing regions. The types of land use include farm- are listed in Table 1.
land, newly cultivated land and desert. The main crops are cotton,
wheat, corn and apples. Most of the natural vegetation is halophytes, 2.4. Terrain attributes
such as Tamarix, Halocnemum strobilaceum, halostachyscaspica, reed
and alhagi sparsifolia shap. The vegetation cover showed a low–- Terrain data is the most commonly used surface parameter in digital
high–low trend from south to north. The degree of soil salinization soil mapping studies (Moore et al., 1991; Abdel-Kader, 2011; Gallant
shows an increasing and then decreasing trend in the direction from and Dowling, 2003). Terrain conditions play important roles in salt
south to north, and an increasing–decreasing–increasing trend in the distribution and redistribution. While high steep landscape benefits
direction from west to east. In recent years, due to the rapid increase of salinity migration, the low-lying landscape promotes salt accumulation
population, various lands with severe or extremely saline soil has been (Fan et al., 2016). The influence of terrain data on soil salinization has
reclaimed into cultivated land in the south of the study area, which lead been considered in this study (Mehrjardi et al., 2014). In this study, 15
to sharp increase of water demand and intensified the contradiction terrain attributes were calculated using SAGA GIS from a 30 m digital

1310
J. Peng et al. Geoderma 337 (2019) 1309–1319

Fig. 1. Study area and sampling points within the Xinjiang Province. A total of 225 points were measured for soil EC in two directions and showed as spatial series.

elevation model that was corrected and filled with depressions. The grid better prediction accuracy than regression trees (Henderson et al.,
size of 30 m was chosen as it was proven to be the most suitable for soil- 2005; Bui et al., 2009; Ma et al., 2017). The size of the decision tree can
landscape analyses (Park et al., 2009), and it can also match the spatial also be controlled by two parameters in Cubist, an optional constraint
resolution of Landsat8 OLI. on the minimum number of observations upon which to base a rule and
a brevity factor, which uses heuristics to control the complexity of the
2.5. Modeling method and performance assessment model (Henderson et al., 2005). Over the past decades, the Cubist
model played a significant role in the prediction and mapping of many
The quantitative inversion model of soil electrical conductivity was soil properties, and it could significantly improve the prediction accu-
constructed using Partial Least Squares Regression (PLSR) and Cubist. racy (Viscarra Rossel et al., 2014; Henderson et al., 2005; Bui et al.,
PLSR is the most common way to build a multi-independent model, and 2009). But with respect to soil salinity, few studies were reported.
an advantage of PLSR is that it can handle data with strong collinearity All 225 samples were sorted in ascending order according to their
and noise, as well as situations in which the number of variables con- EC values, after that, for every 3 adjacent samples in the resulting se-
siderably exceeds the number of available samples (Peng et al., 2016). quence, 2 of them were selected for model calibration, and the re-
PLSR has been widely used in hyperspectral remote sensing monitoring maining samples were used for validation.
of soil salinization, (Farifteh et al., 2007; Weng et al., 2008; Weng et al., The coefficients of determination (R2), the root mean square error
2010), and is more suitable for prediction of soil salt content than the (RMSE), the mean absolute error (MAE), and the ratio of performance
stepwise multiple regression method (Weng et al., 2008; Sidike et al., to deviation (RPD) were used to evaluate and compare the performance
2014). In the study of multi-spectral remote sensing monitoring of soil of the above models. A predictive model is accurate if R2 and RPD
salinization, some researchers have attempted to construct a soil salt values are higher than 0.91 and 2.5, respectively. R2 between 0.82 and
quantitative inversion model using the PLSR method (Fan et al., 2015; 0.9 and RPD higher than 2 indicate a good prediction, whereas the R2 of
Fan et al., 2016; Sidike et al., 2014). They also obtained satisfactory an approximate prediction is considered to lie between 0.66 and 0.81
results, and the determination coefficient reported between 0.51 and with an RPD of higher than 1.5. R2 between 0.5 and 0.65 indicates a
0.99. poor relationship (Farifteh et al., 2007). Generally, a model that per-
Cubist is a piecewise linear tree model and uses a recursive parti- forms well would have high R2 and RPD values, and a low RMSE and
tioning of the predictor variable space (Ma et al., 2017). It takes a di- MAE.
vide-and-conquer strategy and seeks to minimize the intra subset var-
iation at each node. Cubist models take the form: if [conditions] then 3. Results and discussion
[linear model] approach. If the predictor variables associated with an
observation satisfy a set of conditions, the linear model is used to 3.1. Descriptive statistics for soil properties
predict the response. The advantage of the condition set in each rule is
that they enable interactions to be handled automatically by allowing The descriptive statistics for the whole, calibration and validation
different linear models to capture the local linearity in various parts of subsets are shown in Table 2. There was no significant difference in the
the predictor variable space. This can often lead to smaller trees and characteristic statistics (the maximum value, the minimum value and

1311
J. Peng et al. Geoderma 337 (2019) 1309–1319

Table 1
Land surface parameters (under major category of terrain attributes, remote sensing data vegetation spectral indices and salinity spectral indices) used to predict soil
EC in this study along with their abbreviations, calculation formula and the reference of calculation.
Auxiliary data Land surface parameters Abbrev Formulations References

Terrain attributes Elevation DEM Mehrjardi et al.


(2014)
Analytical hill shading AH SAGA GIS
Aspect AS SAGA GIS
Catchment area CA SAGA GIS
Channel network base level CNBL SAGA GIS
Closed depressions CD SAGA GIS
Convergence index CI SAGA GIS
Cross-sectional curvature CSC SAGA GIS
Longitudinal curvature LC SAGA GIS
LS factor LSF SAGA GIS
Relative slope position RSP SAGA GIS
Slope S SAGA GIS
Topographic wetness index TWI SAGA GIS
Valley depth VD SAGA GIS
Vertical distance to channel network VDCN SAGA GIS
Remote sensing data Principal components of seven landsat8 OLI PC1, PC2,
bands PC3
Tasseled cap transformation of seven landsat8 TC1, TC2,
OLI bands TC3
Brightness index BI [(B4)2 + (B5)2]0.5 Khan et al. (2005)
Band10 B10 10.6–11.2 μm
Vegetation spectral Normalized difference vegetation index NDVI (B5 − B4) / (B5 + B4) Rouse et al. (1973)
indices Enhanced vegetation index EVI g × (B5 − B4) / (B5 + C1 × B4 − C2 × B2 + L) Huete et al. (2002)
Generalized difference vegetation index GDVI (B52 − B42) / (B52 + B42) Wu et al. (2014)
Non-linear vegetation index NLI (B52 − B4) / (B52 + B4) Goel and Qin (1994)
Green atmospherically resistant vegetation GARI {B5 − [B3 + γ × (B2 − B4)]} / {B5 + [B3 + γ × (B2 − B4)]} Gitelson et al. (1996)
index
Canopy response salinity index CRSI {[(B5 × B4) − (B3 × B2)] / [(B5 × B4) + (B3 × B2)]}0.5 Scudiero et al. (2014)
Soil-adjusted vegetation index SAVI [(B5 − B4) × (1 + L)] / (B5 + B4 + L) Huete (1988)
Salinity spectral indices Salinity index SI (B4 × B2)0.5 Khan et al. (2005)
Salinity index 1 SI1 (B4 × B3)0.5 Douaoui et al. (2006)
Salinity index 2 SI2 [(B5)2 + (B4)2 × (B3)2]0.5 Douaoui et al. (2006)
Salinity index 3 SI3 [(B4)2 + (B3)2]0.5 Douaoui et al. (2006)
Salinity index I S1 B2/B4 Abbas and Khan
(2007)
Salinity index II S2 (B2 − B4) / (B2 + B4) Abbas and Khan
(2007)
Salinity index III S3 B3 × B4 / B2 Abbas and Khan
(2007)
Salinity index IV S4 B2 × B4 / B3 Abbas and Khan
(2007)
Salinity index V S5 B4 × B5 / B3 Abbas and Khan
(2007)
Salinity index VI S6 B6 / B7 Bannari et al. (2008)
Salinity index VII S7 (B6 − B7) / (B6 + B7) Bannari et al. (2008)
Salinity index VII S8 (B3 + B4) / 2 Douaoui et al. (2006)
Salinity index IX S9 (B3 + B4 + B5) / 2 Douaoui et al. (2006)

Note:where the aerosol and soil correcting parameters g, C1, C2, L and γ are set to 2.5, 6, −7.5, 1, and 0.9, respectively.

mean value) of the EC, SWC, and pH between calibration and validation and median were 20.65 dS m−1 and 17.43 dS m−1, respectively with
dataset. standard deviation (SD) of 16.35 and coefficient of variation (CV) of
There was a high variation of EC in the whole study area. The EC for 79.19%. Spatially, EC showed a trend of increase-decrease-increase
the whole dataset varied between 0.12 and 67.70 dS m−1. The mean from the west to east of the study area (Fig. 1). The samples with

Table 2
Summary statistics of soil properties in calibration, validation and total datasets.
Datasets Soil properties Maximum Minimum Mean Median SD CV (%)

−1
Calibration EC(dS m ) 67.70 0.12 20.65 17.31 16.36 79.24
SMC(g kg−1) 325.15 0.03 153.06 162.27 62.84 41.06
pH 9.25 7.96 8.51 8.52 0.25 2.94
Validation EC(dS m−1) 62.05 0.14 20.65 17.46 16.33 79.09
SMC(g kg−1) 243.46 0.46 143.34 149.48 56.54 39.44
pH 9.17 8.01 8.52 8.52 0.25 2.93
Total EC(dS m−1) 67.70 0.12 20.65 17.43 16.35 79.19
SMC(g kg−1) 325.15 0.03 149.82 158.56 60.98 40.71
pH 9.25 7.96 8.52 8.52 0.25 2.96

Note: Where the SMC denote soil moisture content.

1312
J. Peng et al. Geoderma 337 (2019) 1309–1319

EC < 10 dS m−1 were mainly distributed in the west corner of the study
area, while those larger than 40 dS m−1 were mostly distributed in the
zone with a distance of 6 - 20 km. From south to north, EC showed a
trend of increase and then decrease. The EC in the zone with a distance
of 60 - 90 km was nearly 10 dS m−1, while samples larger than 40 dS
m−1 were concentrated in the zone with distance of 30 - 60 km (as
shown in Fig. 1). The locations with smaller EC were mainly distributed
in the farmlands and the Gobi Desert (a barren plateau in southern
Mongolia and northern China) at the top of alluvial fans. In the farm-
lands, due to the leaching from artificial irrigation, most of the soluble
salts were leached below 0 - 20 cm soil layer. Generally, the top of the
fan mainly consisted of sand with low absorbability and most of the
salts were leached out. Samples with high EC were mainly distributed in
the uncultivated soil. Due to the strong evaporation and low pre-
cipitation, it was difficult to leach the surface soil salt to the soil layer
below 20 cm, thus, salinity often accumulated in the natural soil sur-
face. The SMC ranged from 0.03 to 325.15 g kg−1, and the mean,
median, SD and CV values were 149.82 g kg−1, 158.56 g kg−1, 60.98
and 40.71%, respectively. The samples with low SMC content were Fig. 2. Reflectance spectra curves of different soil electrical conductivity in the
mainly distributed in the fan top in the northern part of the study area, study area.
while samples with high SMC content were distributed in the farmland.
Soil pH varied from 7.96 to 9.25 within the study area, and the values which is abundant in Epsomite (MgSO4∗7H2O), Bischofite
of mean, median, SD and CV were 8.52, 8.52, 0.25 and 2.96%, re- (MgCl2∗6H2O) and Halite (NaCl) was statistically significantly dif-
spectively. The variation of pH was small in the whole study area. ferent. Farifteh et al. (2007) reported that the coefficient of determi-
Further statistical analysis showed that there are only two samples with nation between EC and the reflectance soil that is rich MgSO4, and
pH < 8, 100 samples with the pH between 8.0 and 8.5, 114 samples MgCl2 was larger than that with soils rich in NaCl and KCl.
with pH from 8.5 to 9.0, and 9 samples with pH over 9.0, accounting for
0.89%, 44.44%, 50.67%, and 4% of the whole samples, respectively.
3.3. Correlation between land surface parameters and measured EC

3.2. Spectral reflectance of different soil salinity


Out of 43 parameters, the correlation between 14 parameters and
EC were significant with p < 0.01 (Table 3). The correlation between
The saline soil has obvious absorption characteristic in the visible
another 4 parameters (RSP, VDCN, S6 and S7) and EC was significant at
(400–770 nm), near-infrared (900–1030 nm, 1270–1700 nm,1900–
p = 0.05, while the remaining 25 parameters had no significant cor-
2150 nm, 2150–2310 nm, 2320–2400 nm) (Sidike et al., 2014; Nawar
relation with EC (Table 3). There were 24 parameters that showed a
et al., 2014). Characteristics spectral absorptions were associated with
negative correlation with EC (e.g. DEM, AH, AS). The correlation be-
the degree of salinization and had been used in the quantitative mon-
tween TC2 and EC was the strongest, with a correlation coefficient of
itoring of soil salinization by remote sensing. However, the spectral
−0.61. While, other 19 parameters (e.g. CD, TWI, PC1) and EC showed
characteristics were largely affected by the presence of vegetation and
a positive correlation. The strongest positive correlation was between
soil crust that changed the soil surface roughness and caused a strong
B10 and EC, with a correlation coefficient of 0.59 (Table 3).
reflective effect in visible and near-infrared electromagnetic waves and
Of all the 15 topographic factors, only DEM and CNBL showed a
higher reflection than non-saline or moderately saline soil. The reflec-
significant correlation with EC (p < 0.01). The correlation of RSP and
tion increased with the amount of the soil salinity (Bouaziz et al.,
VDCN with EC was significant at p = 0.05, which may be mainly due to
2011). This high reflection characteristics of strongly saline soil (pre-
the flat terrain of the study area. In addition, limited surface runoff
sence of thick crust) were easily captured by satellite remote sensing
from rainfall in this area greatly weakened the influence of topographic
and played a vital role in the remote monitoring of soil salinization in
factors on redistribution of soil salinity (Akramkhanov et al. 2011;
large area (Abbas et al., 2013).
Mehrjardi et al. 2014).
As shown in Fig. 2, Soil samples with different electrical con-
Among the 8 remote sensing factors, PC2, TC2, BI and B10 had a
ductivity showed a very similar shape of the spectral reflectance curves
significant correlation with EC (p < 0.01), while other four factors
which was sharply increased from visible band to infrared band, espe-
cially, from blue light to near infrared band (B2–B5) and reaching the
Table 3
maximum value at the short-wave infrared1 (SWIR1). From SWIR1 to Correlation coefficients between land surface parameters and measured EC
SWIR2, the reflectance decreased. The reflectance of soil with different values.
EC increased with the increase of EC during the Costal to SWIR1 (B1B6)
Factors R Factors R Factors R Factors R
band, with a smaller difference in Costal-Blue (B1–B2) than that in
Blue-SWIR1 (B2–B6). While, there was no obvious regularity in the DEM −0.43⁎⁎ S −0.06 B10 0.59⁎⁎ SI3 0.06
relationship of reflectance and different EC levels at the SWIR2 band. AH −0.08 TWI 0.03 NDVI −0.29⁎⁎ S1 −0.06
This finding is consistent with that of Harti et al. (2016) who observed AS −0.06 VD −0.08 EVI −0.30⁎⁎ S2 −0.05
CA −0.01 VDCN −0.18⁎ GDVI −0.42⁎⁎ S3 0.11
similar spectral curves from slightly to strongly saline soils in the
CNBL −0.43⁎⁎ PC1 0.04 NLI −0.33⁎⁎ S4 0.01
SWIR2 band. The reflectivity of soil at 2200 nm band was mainly in- CD 0.01 PC2 0.33⁎⁎ GARI −0.28⁎⁎ S5 0.29⁎⁎
fluenced by the content of soil organic molecules (i.e. CH2, CH3, and CI −0.11 PC3 0.01 CRSI −0.25⁎⁎ S6 0.14⁎
NH3), SieOH bonds, cation-OH bonds in phyllosilicate minerals (i.e. CSC −0.05 TC1 0.07 SAVI −0.30⁎⁎ S7 0.15⁎
kaolinite, montmorillonite), CO32– ion in carbonate minerals (i.e. cal- LC −0.04 TC2 −0.61⁎⁎ SI 0.10 S8 0.13
LSF −0.08 TC3 −0.01 SI1 0.10 S9 0.11
cite), or any combination of these components (Peng et al., 2016). In RSP −0.16⁎ BI 0.23⁎⁎ SI2 0.10
addition, the difference of salt composition is also an important factor
affecting the spectral characteristics of soil in the SWIR2 band (Farifteh ⁎⁎
Significant at the 0.01 probability level.

et al., 2006). The relationship of EC value and reflectance of the soil Significant at the 0.05 probability level.

1313
J. Peng et al. Geoderma 337 (2019) 1309–1319

were not significant at all. According to remote sensing images, high

5181.408506 − 22,079 ∗ TC2 − 25634 ∗ PC2 − 13892 ∗ TC1 − 7525 ∗ PC3 − 2627 ∗ TC3 + 683 ∗ GDVI − 401 ∗ PC1 − 537 ∗ BI − 147.8 ∗ NLI − 342 ∗ S6 − 138 ∗ S3 − 119 ∗ S1 −-
2500.606488 − 10,045 ∗ TC2 − 11987 ∗ PC2 − 6966 ∗ TC1 − 4388 ∗ PC3 − 2222 ∗ TC3 − 110.3 ∗ NLI + 236 ∗ GDVI − 139 ∗ PC1 − 186 ∗ BI − 206 ∗ S6 − 89 ∗ S1 − 0.058 ∗ CNBL-

2784.676212 − 12,537 ∗ TC2 − 13967 ∗ PC2 − 7136 ∗ TC1 − 3327 ∗ PC3 + 614 ∗ GDVI − 361 ∗ PC1 − 483 ∗ BI − 700 ∗ TC3 − 124 ∗ S3 − 0.068 ∗ CNBL + 10.4 ∗ B10 − 101 ∗ S2-
370.460059 − 2508 ∗ TC2 − 2709 ∗ PC2 − 1342 ∗ TC1 − 1029 ∗ GDVI + 1044 ∗ BI − 477 ∗ PC1 − 568 ∗ S3 + 124.1 ∗ NLI − 306 ∗ PC3 − 206 ∗ S2 − 151 ∗ TC3 − 0.017 ∗ DEM +-
values of PC2 were mainly distributed in high brightness salt spots, and
the areas of low values were mainly farmland or wetland (Gutierrez and
Johnson, 2010). As TC2 carries the green information of an image, the
zone of high TC2 values corresponded to the green vegetation coverage
area and zone of low TC2 values corresponded to high EC values. As BI
carries the brightness information of an image, the areas of high BI
corresponded to the high value of EC within the image. The result of
Allbed et al. (2014) showed that BI was positively related to EC, but the
intensity of correlation was different in different areas. Therefore, the
correlation between BI and EC was not only related to salt content, but
also may have influenced by vegetation cover, soil moisture content,
soil texture and other factors. B10 belongs to the thermal infrared band
and has been widely used in the study of soil salinity and soil water
(Goossens and Van Ranst, 1998). The B10 value represents the surface
temperature and the high value associated with the high surface tem-

726.239824 + 2132 ∗ TC2 − 2486 ∗ PC2 − 1364 ∗ TC1 − 518 ∗ PC3 − 0.149 ∗ CNBL − 0.091 ∗ VD − 91 ∗ TC3 − 26 ∗ CRSI
perature. The land surface temperature was mainly controlled by soil

176.178798 + 1015 ∗ TC2 − 665 ∗ GDVI + 528 ∗ PC2 + 317 ∗ TC1 − 0.111 ∗ CNBL + 152 ∗ PC3 + 42 ∗ TC3 − 19 ∗ S6
moisture. The areas with high EC were usually the zones with low soil
moisture content, where the surface was usually covered with white salt
crust. There was a negative correlation between the soil moisture and

−158.455582 − 1474 ∗ GDVI + 1495 ∗ BI − 473 ∗ PC1 − 678 ∗ S3 + 157.7 ∗ NLI + 474 ∗ PCA3 − 226 ∗ S2
EC with the correlation coefficient of −0.42.
All vegetation spectral indices had a significant correlation with EC
at p = 0.01. Among the salinity indices, S5 was significant at p = 0.01,
but S6 and S7 were significant at p = 0.05. The response of vegetation
spectral index and salinity spectral index to EC was affected by many
factors including vegetation coverage, salt tolerance, soil moisture, and
soil type (Metternicht and Zinck, 2003). The result may vary greatly
under different environmental conditions with a better result for ve-

0.086 ∗ CNBL + 11.5 ∗ B10 − 113 ∗ S2 − 0.068 ∗ DEM + 0.1 ∗ VD − 6 ∗ CRSI


getation spectral index in areas with high vegetation cover while
spectral indices opposite of this (Allbed et al., 2014; Alhammadi and
Glenn, 2008; Zhang et al., 2011). Therefore, although the vegetation
spectral index and salinity spectral index showed satisfactory results in
monitoring salinity all over the world, it is worth noting that there is no
− 48 ∗ S3 + 4 ∗ B10 − 39 ∗ S2 + 0.023 ∗ DEM + 0.035 ∗ VD

universal spectral index that can show a satisfactory result in any en-
vironmental conditions (Allbed et al., 2014). Strategies to solve this

− 0.061 ∗ DEM − 113 ∗ S6 + 0.09 ∗ VD − 32 ∗ CRSI


problem are: (1) to choose the existing spectral indices suitable for local
environmental conditions in different study areas; (2) to revise the
existing spectral indices or construct new spectral indices based on local
environmental conditions; and (3) to choose suitable sensitive spectral
index based on vegetation coverage. For example, in low vegetation
2.8 ∗ B10 − 31 ∗ S6 + 0.025 ∗ VD

coverage area, the salinity spectral index should be given priority while
the vegetation spectral index should be considered for areas with high
vegetation coverage.

3.4. Classification of the study area and the dominant factors based on
Cubist model
Linear models

Table 4 lists the rules for partitioning regions using the Cubist model
and the corresponding linear model for each sub region. According to
the model output, when Cubist model divided the region, the higher
contribution factors were CNBL, GDVI, B10, and TC2 with values of
GDVI ≤ 0.05259998 & CNBL > 1039.14
B10 > 10.7678 & TC2 > −0.114027 &

TC2 ≤ − 0.114072 & CNBL > 1039.14

85%, 51%, 49% and 35%, respectively. In the linear models, PCA3,
Classification rules and corresponding models.

CNBL > 1039.14 & CNBL ≤ 1139.32

B10 > 10.7678 & CNBL > 1139.32


GDVI > 0.0476 & CNBL ≤ 1039.14

PCA2, TC1, TC2, and TC3 had higher contribution with values of 100%,
GDVI ≤ 0.0476 & CNBL ≤ 1039.14

96%, 96%, 96% and 96%, respectively. The other factors contributions
were < 90%. The results showed that factors with a higher contribution
in the rules were not always of high contribution in the linear models,
GDVI > 0.05259998 &

which is basically consistent with the result of Viscarra Rossel et al.


(2014). Under the specific rules, the factors selected in the linear
B10 ≤ 10.7678

models in different regions were different (Ma et al., 2017). Even if the
selected factors were the same, the regression coefficient of each factor
was different indicating a variable contribution to model EC.
Rules

The spatial distribution of the sub-regions was presented in Fig. 3.


Combing with the results from visual interpretations and ground sur-
Regions

veys, we conduct the statistics in each sub-regions on land types, ve-


Table 4

getation coverages, and so on, shown as Table 5. From the statistical


1

2
3

6
7

results, the Cubist model could distinguish the land surface

1314
J. Peng et al. Geoderma 337 (2019) 1309–1319

Fig. 3. Distribution map of seven region subsets derived from Cubist model.

Table 5 Table 6
The characteristic of different sub-regions. Accuracy comparison of PLSR and Cubist models.
Regions Land use types Vegetation coverage/% Tone Models Calibration Validation

1 Ecotone of shrubs and trees 30–60 Dark R2 RMSE MAE R2 RMSE MAE RPD
2 Farmland 70–100 Dark
3 Mountains, Gobi Desert ≤5 Bright PLSR 0.71 9.07 7.32 0.66 10.46 8.21 1.56
4 Sparse herbaceous covering area 10–30 Dark Cubist 0.92 4.55 3.12 0.91 5.18 3.76 3.15
5 Fallow farmland, desert 0 Bright
6, 7 Salt crusts 0 Bright
3.5. Estimation of EC with Cubist and PLSR models and validation

characteristics in the corresponding sub-regions, such as the informa- The R2, RMSE and MAE of PLSR estimation in the calibration set
tion of land types and vegetation coverages. As for the vegetation were 0.71, 9.07 dS m−1 and 7.32 dS m−1, respectively (Table 6). The
coverages, the vegetation types are the main reason resulting in dif- Cubist model showed a much better performance than PLSR, with a
ferent coverages, for example, trees, shrubs, and herbs, which were also higher R2 of 0.92, a lower RMSE of 4.55 dS m−1 and a lower MAE of
attach to different sub-regions. Additionally, the human activities are 3.12 dS m−1 (Table 6). For the validation set, the assessment indicators
also important reason leading to divide the study area into sub-regions. were changed slightly, yet the R2 of the Cubist model still reached to
For instance, croplands in growing season, free farm lands, new re- 0.91. The R2 and RPD of the Cubist model were 0.91 and 3.15, re-
claimed lands, and deserts were divided into different sub-regions. spectively, and of the PLSR model 0.66 and 1.56, respectively (Table 6).
Therefore, the sub-regions from Cubist model were a comprehensive Comparative better performance of the Cubist model over PLSR may be
result by considering various factors.

1315
J. Peng et al. Geoderma 337 (2019) 1309–1319

due to the fact that the Cubist model partitioned the study area ac- was relatively weak.
cording to the environmental conditions and optimized the parameters
of the maximum contribution for each sub region before modeling and 3.7. Comparison of precision of Cubist and PLSR models under different soil
predicting the EC content (Viscarra Rossel et al., 2014). salinity, vegetation coverage, soil moisture levels
The contribution of the surface parameters on EC modeling changed
with the environmental conditions. In areas with high vegetation cov- The EC prediction errors of the Cubist and PLSR models in different
erage, vegetation spectral index showed higher contribution as an in- EC, NDVI and SMC segments were presented in Table 7. From the
dicator of soil salinization. While in areas with sparse vegetation or segmentation statistics of EC, the error of the PLSR model in the low
bare soil, the salinity spectral index showed stronger contribution. The and high values was large. The RMSE were 11.6 and 15.43 dS m−1, and
PLSR model assumed that all modeling factors had the constant con- MAE were 8.80 and 12.83 dS m−1, respectively for the soil with salinity
tribution to model EC no matter what environmental conditions pre- of 0–10 and > 40 dS m−1. For soil with a salinity of 10–20, 20–30 and
vailed in the study area. Due to lack of factor optimization based on 30–40 dS m−1, the RMSE was < 10 dS m−1 and the MAE were < 8.00
specific environmental conditions in PLSR, the modeling performance dS m−1. The RMSE and MAE were the lowest (5.90 and 4.89 dS m−1,
was lower than Cubist. Compared with the results of other similar respectively) when the EC of soil was 30–40 dS m−1. The results of the
studies, both PLSR and Cubist models showed satisfied estimation of Cubist model showed that the RMSE (from 2.66 to 7.1 dS m−1) and
precision in our study. For example, Ding and Yu (2014) constructed a MAE (from 1.67 to 5.86 dS m−1) increased as the EC increased from 0
stepwise regression model using vegetation spectral index and salinity to 10 to > 40 dS m−1 (Table 7). For the PLSR model, the RMSE and
spectral index, and reported an R2 of 0.39 to 0.44, respectively. Allbed MAE of all the intervals were higher than those of the Cubist model
et al. (2014) developed linear models to assess soil salinity using soil except the 30–40 dS m−1 interval, where the PLSR model showed ap-
salinity and vegetation indices derived from IKONOS high spatial re- proximately the same accuracy as that of the Cubist model.
solution imageries, and reported a R between −0.78 and 0.78. Harti NDVI is a key factor for vegetation coverage calculation with higher
et al. (2016) used soil salinity index (OLI-SI) for soil EC estimation and values indicating larger vegetation coverage (Zhang et al., 2016).
reported an R2 between 0.55 and 0.77. The main reasons for the sa- Therefore, this paper used NDVI to illustrate the effect of vegetation
tisfactory results of our study over others are; (1) the matching degree coverage on EC prediction accuracy. The RMSE and MAE of the PLSR
of satellite pass and ground collected salinity data, (2) the time of data model increased with the increase of NDVI (Table 6). When the
collection at the end of the vegetation growth period with maximum NDVI > 0.2, the RMSE reached 10.95 dS m−1 and the MAE reached
biomass resulting in minimum influence of vegetation cover difference 8.73 dS m−1. While the Cubist model did not show the same pattern,
caused by drought factors, (3) severe soil salinity in the study area with RMSE < 4.8 dS m−1 and MAE < 3.6 dS m−1 at all intervals. At
made surface salt accumulation obvious leading to easy distinction of the same time, the RMSE and MAE of the Cubist model in three inter-
soil salinization, and (4) variable land surface parameter indices in- vals did not change much, indicating that the partition of the region had
cluding topographic attributes, satellite spectra, vegetation spectral significantly weakened the interference of vegetation coverage on the
index and salinity spectral index, providing complementary effects. EC prediction. The NDVI interval between 0.1 and 0.2 exhibited the
largest error as the soil of this interval was located in the sparse ve-
3.6. Soil salinity maps derived from Cubist and PLSR models getation area, where the satellite image showed a mixture of vegetation
and soil spectral information. However, NDVI between 0 and 0.1
There was no significant difference between the trend of EC dis- and > 0.2 were observed in the area with bare soil and farmland, with
tribution maps from PLSR and Cubist (Fig. 4). The soils with EC of relatively pure soil and vegetation spectral information, respectively.
50–70 dS m−1 were concentrated in the northeastern part of the study For soil moisture content (SMC), both PLSR and Cubist showed an
area in both maps. Located at the foot of the mountain, the micro-ter- increasing trend of RMSE and MAE with an increase of soil moisture
rain in this area was low. The salt ran off and accumulated in this area and decrease thereafter. It showed better prediction accuracy when soil
was not further discharged and led to the high EC. The soils with EC of moisture was in low (0–100 g kg−1) or high (> 200 g kg−1) intervals.
0–10 dS m−1 were mainly distributed in the farmland where irrigation Nevertheless, the RMSE and MAE of the Cubist model were lower than
and drainage management were prominent, or at the Gobi Desert and those of the PLSR model in each interval. From the distribution of
nearby desert area with weak soil adsorption ability to salt. The soils samples in each interval, the SMC of 0–100 and > 200 g kg−1 were
with EC of 20–30 dS m−1 concentrated in the southwestern and mainly distributed in the low EC area with the mean value of 10.2 and
southeastern area that mainly covered the newly reclaimed farmland. 14.32 dS m−1, respectively. The samples with SMC of 100–150 and
These places, after years of irrigation, the salt content still did not 150–200 g kg−1 were mainly distributed in the area with high EC and
achieve the criterion for planting crops. The soils with EC of the mean value was 24.84 and 26.32 dS m−1, respectively and might
30–50 dS m−1 were mainly distributed in the middle of the study area have attributed to this pattern.
which was a desert with minimal human disturbance with very sparse Soil moisture and vegetation are crucial factors affecting the mon-
salt-tolerant vegetation. itoring accuracy of soil salinization. Different degrees of salinization
Comparing the fine features of the EC distribution, the map derived can produce a variable error due to the difference in vegetation cov-
from the Cubist model showed subtle changes and clearer textures. erage. Soil moisture can lead to the decrease of spectral reflectance in
Obvious differences in EC content was observed between the roads, the visible and near infrared bands (Stoner and Baumgardner, 1981;
river beds and the surrounding backgrounds in the Cubist based maps Muller and Decamps, 2001), while soil salinization usually leads to an
while these features in the PLSR based map were relatively vague. The increase in reflectivity (Bouaziz et al., 2011). Due to the influence of
percentages of the area calculated in the PLSR based map for 0–10 and soil moisture on the reflectance of near-infrared and red-light bands,
10–20 dS m−1 were lower than that of the Cubist based map. The area there will be a notable change in the spectral index commonly used in
coverages in the PLSR based map were 23.93% and 13.97% respec- remote sensing monitoring of soil salinization under different water
tively, while in the Cubist based map were 30.38% and 14.69%. For the contents (Fang et al., 2015). The study results of Ding and Yu (2014)
EC range of 20–30, 30–50 and 50–70 dS m−1, the PLSR-based map showed that the monitoring accuracy of soil salinization was better
exhibited larger area than that of the Cubist based map. The corre- during the dry seasons than that in the wet seasons. The most direct
sponding values for PLSR were 26.69, 24.65 and 10.76%, respectively, effect of vegetation on salinization monitoring is that it obscures the
and for the Cubist were 22.63, 23.47 and 8.83%. This led to the greater soil information. Second, the vegetation spectral index and salt spectral
error from PLSR model, over Cubist model, especially in the low and index are the most commonly used factors in soil salinization mon-
high values, since its ability to characterize local subtle feature changes itoring, but the vegetation index is usually more sensitive to changes in

1316
J. Peng et al. Geoderma 337 (2019) 1309–1319

Fig. 4. EC map derived from Cubist and PLSR models.

Table 7 5.64 dS m−1 for non-, slightly, moderately, strongly, and extremely
Performance of PLSR and Cubist models under different soil salinity, vegetation saline soils, respectively. Similarly, Sidike et al. (2014) reported RMSE
coverage and soil moisture levels. of 0.308, 0.334 and 0.473% for soils with salinity ≤0.6%,
Factors Intervals n PLSR model Cubist model 0.6% < SSC < 4% and ≥4%, respectively clearly showing an in-
crease of prediction error with an increase of soil salinity.
RMSE MAE RMSE MAE

−1
EC(dS m ) 0–10 20 11.60 8.80 2.66 1.67 4. Conclusions
10–20 24 9.89 7.92 3.68 2.91
20–30 12 6.28 4.85 5.32 4.83 With the rapid population increase in southern Xinjiang, the ex-
30–40 9 5.90 4.89 6.13 5.38
> 40 10 15.43 12.83 7.10 5.86
isting land has been difficult to meet the needs of local residents.
NDVI 0–0.1 47 8.70 7.06 4.61 3.46 Reclamation of saline soil has become the main way to solve the con-
0.1–0.2 15 10.83 7.65 4.79 3.58 tradiction between man and land. Not only, how to improve the utili-
> 0.2 13 10.95 8.73 4.49 3.25 zation rate of limited water resources is the key to eliminate the con-
SMC(g kg−1) 0–100 13 9.20 7.75 2.25 1.37
tradiction between agricultural water and ecological water, it is also
100–150 23 11.40 8.77 3.80 2.99
150–200 24 11.79 9.49 7.09 6.06 related to the sustainable development of agriculture and ecological
> 200 15 7.19 5.71 3.29 2.48 environment in the southern Xinjiang. It is one of the most effective
ways to solve the contradiction between man and land by using remote
sensing technology to obtain the information and digital maps of soil
soil salinity under high vegetation coverage (Alhammadi and Glenn, salinization for decision-making about land reclamation, utilization,
2008; Zhang et al., 2011). However, in soil salinization remote sensing, and prioritization for protection and management. Based on remote
often the regionalization of the study area and the corresponding se- sensing images and ground-based survey data, the Cubist and PLSR
lection of the optimal parameters were not carried out according to the models of soil EC were constructed using principal component, tasseled
difference of environmental conditions, thus leading to uncertainty of cap, terrain attribute, vegetation spectral index and salt spectral index
the monitoring results. calculated from satellite images. The R2, RMSE, MAE and RPD were
Different degrees of salinization also produce different monitoring 0.91, 5.18 dS m−1, 3.76 dS m−1 and 3.15 for the Cubist model, and
errors. Scudiero et al. (2015) considered many indices including canopy 0.66, 10.46 dS m−1, 8.21 dS m−1 and 1.56 for the PLSR model. Both
response salinity index, field management type, annual total rainfall, models showed good predictive performance, however, the prediction
soil silt percentage and annual average minimum temperature to re- accuracy of the Cubist model was better than that of the PLSR model.
trieve the soil salinity from a farmland from an arid area and reported Under the environmental conditions with a large spatial variation of
an increase in error with the increase of soil salinity. For example, the vegetation, soil moisture and salinization, the Cubist model could ef-
validation Mean absolute errors were 2.94, 2.12, 2.35, 3.23, and fectively divide data into subsets according to the similarity of the

1317
J. Peng et al. Geoderma 337 (2019) 1309–1319

predicted variables and attain the independent prediction model, which and ANN). Remote Sens. Environ. 110 (1), 59–78.
would improve the prediction accuracy. The contribution of each pre- Gallant, J.C., Dowling, T.I., 2003. A multiresolution index of valley bottom flatness for
mapping depositional areas. Water Resour. Res. 39 (12), 291–297.
dictor to EC was not adjusted according to the difference of the en- Gitelson, A.A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote
vironmental conditions in the PLSR model. Therefore, it was difficult to sensing of global vegetation from EOS-MODIS. Remote Sens. Environ. 58 (3),
obtain satisfactory results using the PLSR model when the EC was at 289–298.
Goel, N.S., Qin, W., 1994. Influences of canopy architecture on relationships between
high or low values, NDVI ≥ 0.2, and the soil moisture was various vegetation indices and LAI and FPAR. Remote Sens. Rev. 10 (4), 309–347.
100–200 g kg−1. Thus, the Cubist model was a suitable method for Goossens, R., Van Ranst, E., 1998. The use of remote sensing to map gypsiferous soils in
obtaining high-precision soil salinization information in this study, and the Ismailia Province (Egypt). Geoderma 87 (1–2), 47–56.
Gorji, T., Tanik, A., Sertel, E., 2015. Soil salinity prediction, monitoring and mapping
the satisfactory results are expected to be achieved in salinization using modern technologies. Procedia Earth and Planetary Science 15, 507–512.
monitoring applications in similar arid or semi-arid areas outside this Gorji, T., Sertel, E., Tanik, A., 2017. Monitoring soil salinity via remote sensing tech-
study area. nology under data scarce conditions: a case study from Turkey. Ecol. Indic. 74,
384–391.
Gutierrez, M., Johnson, E., 2010. Temporal variations of natural soil salinity in an arid
Acknowledgements environment using satellite images (J. South Am. Earth Sci.). 30, 46–57.
Harti, A.E., Lhissou, R., Chokmani, K., Ouzemou, J., Hassouna, M., Bachaoui, E.M.,
This work was supported by the National Natural Science Ghmari, A.E., 2016. Spatiotemporal monitoring of soil salinization in irrigated Tadla
Plain (Morocco) using satellite spectral indices. Int. J. Appl. Earth Obs. Geoinf. 50,
Foundation of China (No. 41361048), (No. 41261083), (No. 64–73.
41561049), and National Key Research and Development Program Henderson, B.L., Bui, E.N., Moran, C.J., Simon, D.A.P., 2005. Australia-wide predictions
(2016FYC0501400). of soil properties using, decision trees. Geoderma 124 (3–4), 383–398.
Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25 (3),
295–309.
References Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview of
the radiometric and biophysical performance of the MODIS vegetation indices.
Remote Sens. Environ. 83 (1–2), 195–213.
Abbas, A., Khan, S., 2007. Using remote sensing techniques for appraisal of irrigated soil
Khan, N.M., Rastoskuev, V.V., Sato, Y., Shiozawa, S., 2005. Assessment of hydrosaline
salinity. In: MODSIM 2007: International Congress on Modelling and Simulation:
land degradation by using a simple approach of remote sensing indicators. Agric.
Land, Water and Environmental Management: Integrated Systems for Sustainability,
Water Manag. 77 (1–3), 96–109.
pp. 2632–2638.
Li, H.Y., Webster, R., Shi, Z., 2015. Mapping soil salinity in the Yangtze delta: REML and
Abbas, A., Khan, S., Hussain, N., Hanjra, M.A., Akbar, S., 2013. Characterizing soil sali-
universal kriging (E-BLUP) revisited. Geoderma 237-238, 71–77.
nity in irrigated agriculture using a remote sensing approach. Phys. Chem. Earth
Ma, Z.Q., Shi, Z., Zhou, Y., Xu, J.F., Yu, W., Yang, Y.Y., 2017. A spatial data mining
55–57 (2), 43–52.
algorithm for downscaling TMPA 3B43 V7 data over the Qinghai-Tibet Plateau with
Abdel-Kader, F.H., 2011. Digital soil mapping at pilot sites in the northwest coast of
the effect of systematic anomalies removed. Remote Sens. Environ. 200, 378–395.
Egypt: a multinomial logistic regression approach. Egypt. J. Remote Sens. Space. Sci.
Ma, Z.Q., Xu, Y.P., Peng, J., Chen, Q.X., Wan, D., He, K., Shi, Z., Li, H.Y., 2018. Spatial
14 (1), 29–40.
and temporal precipitation patterns characterized by TRMM TMPA over the Qinghai-
Akramkhanov, A., Martius, C., Park, S.J., Hendrickx, J.M.H., 2011. Environmental factors
Tibetan plateau and surroundings. Int. J. Remote Sens. 39, 3891–3907.
of spatial distribution of soil salinity on flat irrigated terrain. Geoderma 163 (1–2),
Masoud, A.A., 2014. Predicting salt abundance in slightly saline soils from Landsat ETM
55–62.
plus imagery using spectral mixture analysis and soil spectrometry. Geoderma 217,
Alhammadi, M.S., Glenn, E.P., 2008. Detecting date palm trees health and vegetation
45–56.
greenness change on the eastern coast of the United Arab Emirates using SAVI. Int. J.
Mehrjardi, R.T., Minasny, B., Sarmadian, F., Malone, B.P., 2014. Digital mapping of soil
Remote Sens. 29 (6), 1745–1765.
salinity in Ardakan region, central Iran. Geoderma 213, 15–28.
Allbed, A., Kumar, L., Aldakheel, Y.Y., 2014. Assessing soil salinity using soil salinity and
Metternicht, G.I., Zinck, J.A., 2003. Remote sensing of soil salinity: potentials and con-
vegetation indices derived from IKONOS high-spatial resolution imageries: applica-
straints. Remote Sens. Environ. 85 (1), 1–20.
tions in a date palm dominated region. Geoderma 230, 1–8.
Moore, I.D., Grayson, R.B., Ladson, A.R., 1991. Digital terrain modeling - a review of
Bannari, A., Guedon, A.M., El-Harti, A., Cherkaoui, F.Z., El-Ghmari, A., 2008.
hydrological, geomorpholgical, and biological applications. Hydrol. Process. 5, 3–30.
Characterization of slightly and moderately saline and sodic soils in irrigated agri-
Muller, E., Decamps, H., 2001. Modeling soil moisture-reflectance. Remote Sens. Environ.
cultural land using simulated data of advanced land imaging (EO-1) sensor. Commun.
76 (2), 173–180.
Soil Sci. Plant Anal. 39 (19–20), 2795–2811.
Nawar, S., Buddenbaum, H., Hill, J., Kozak, J., 2014. Modeling and mapping of soil
Barbouchi, M., Abdelfattah, R., Chokmani, K., Ben Aissa, N., Lhissou, R., Harti, A., 2015.
salinity with reflectance spectroscopy and Landsat data using two quantitative
Soil salinity characterization using polarimetric InSAR coherence: case studies in
methods (PLSR and MARS). Remote Sens. 6 (11), 10813–10834.
Tunisia and Morocco. IEEE Journal of Selected Topics in Applied Earth Observations
Park, S.J., Ruecker, G.R., Agyare, W.A., Akramkhanov, A., Kim, D., Vlek, P.L.G., 2009.
and Remote Sensing 8 (8), 3823–3832.
Influence of grid cell size and flow routing algorithm on soil-landform modeling.
Bouaziz, M., Matschullat, J., Gloaguen, R., 2011. Improved remote sensing detection of
Journal of the Korean Geographical Society 44 (2), 122–145.
soil salinity from a semi-arid climate in Northeast Brazil. Compt. Rendus Geosci. 343
Peng, J., Ji, W.J., Ma, Z.Q., Li, S., Chen, S.C., Zhou, L.Q., Shi, Z., 2016. Predicting total
(11−12), 795–803.
dissolved salts and soluble ion concentrations in agricultural soils using portable
Bui, E., Henderson, B., Viergever, K., 2009. Using knowledge discovery with data mining
visible near-infrared and mid-infrared spectrometers. Biosyst. Eng. 152, 94–103.
from the Australian Soil Resource Information System database to inform soil carbon
Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems
mapping in Australia. Glob. Biogeochem. Cycles 23, 1–15.
in the great plains with ERTS. In: NASA Special Publication. 351. pp. 309.
Cheng, J.L., Shi, Z., Zhu, Y.W., 2007. Assessment and mapping of environmental quality
Scudiero, E., Skaggs, T.H., Corwin, D.L., 2014. Regional scale soil salinity evaluation
in agricultural soils of Zhejiang Province, China. J. Environ. Sci. (China) 19 (1),
using Landsat 7, western San Joaquin Valley, California, USA. Geoderma Reg. 2-3,
50–54.
82–90.
Dehaan, R., Taylor, G.R., 2003. Image-derived spectral endmembers as indicators of
Scudiero, E., Skaggs, T.H., Corwin, D.L., 2015. Regional-scale soil salinity assessment
salinisation. Int. J. Remote Sens. 24 (4), 775–794.
using Landsat ETM plus canopy reflectance. Remote Sens. Environ. 169, 335–343.
Deng, X.Y., Yang, Z.F., Long, A.H., 2013. Ecological operation in the Tarim River basin
Shi, Y.L., Wang, R.S., Zhou, H.B., Chen, L., 2006. Land use change and its ecological
based on rational allocation of water resource. J. Glaciol. Geocryol. 35, 1600–1609
effects in the ecotone of Southern Xinjiang Uyghur Autonomous Region: a case study
(In Chinese).
of Akesu City area. Chinese Journal of Ecology 25 (7), 753–758 (In Chinese).
Ding, J., Yu, D., 2014. Monitoring and evaluating spatial variability of soil salinity in dry
Sidike, A., Zhao, S., Wen, Y., 2014. Estimating soil salinity in Pingluo County of China
and wet seasons in the Werigan–Kuqa Oasis, China, using remote sensing and elec-
using QuickBird data and soil reflectance spectra. Int. J. Appl. Earth Obs. Geoinf. 26,
tromagnetic induction instruments. Geoderma 235–236 (4), 316–322.
156–175.
Douaoui, A.E.K., Nicolas, H., Walter, C., 2006. Detecting salinity hazards within a
Stoner, E.R., Baumgardner, M.F., 1981. Characteristic variations in reflectance of surface
semiarid context by means of combining soil and remote-sensing data. Geoderma 134
soils. Soil Sci. Soc. Am. J. 45, 1161–1165.
(1–2), 217–230.
Turhong, M., Abudukeremu, M., Nishizaki, Y., Abuliz, T., Kurbanjan, 2008. Distribution
Fan, X., Liu, Y., Tao, J., Weng, Y., 2015. Soil salinity retrieval from advanced multi-
and characteristics of salinized soil in the south region of Xinjiang. Environ. Sci.
spectral sensor with partial least square regression. Remote Sens. 7 (1), 488–511.
Technol. 31, 22–26 (In Chinese).
Fan, X., Weng, Y., Tao, J., 2016. Towards decadal soil salinity mapping using Landsat
Viscarra Rossel, R.A., Webster, R., Bui, E.N., Baldock, J.A., 2014. Baseline map of organic
time series data. Int. J. Appl. Earth Obs. Geoinf. 52, 32–41.
carbon in Australian soil to support national carbon accounting and monitoring under
Fang, S., Yu, W., Qi, Y., 2015. Spectra and vegetation index variations in moss soil crust in
climate change. Glob. Chang. Biol. 20 (9), 2953–2970.
different seasons, and in wet and dry conditions. Int. J. Appl. Earth Obs. Geoinf. 38,
Wang, J.X., Pang, X.A., Zheng, D.M., Hu, Y.X., Liu, B., 2006. Present situation, existing
261–266.
problem and control countermeasures of Tarim river basin ecological environment.
Farifteh, J., Farshad, A., George, R.J., 2006. Assessing salt-affected soils using remote
In: System Science and Comprehensive Studies in Agriculture. 22. pp. 193–196 (In
sensing, solute modelling, and geophysics. Geoderma 130 (3–4), 191–206.
Chinese).
Farifteh, J., Van der Meer, F., Atzberger, C., Carranza, E.J.M., 2007. Quantitative analysis
Weng, Y.L., Gong, P., Zhu, Z.L., 2008. Soil salt content estimation in the Yellow River
of salt-affected soil reflectance spectra: a comparison of two adaptive methods (PLSR
delta with satellite hyperspectral data. Can. J. Remote. Sens. 34 (3), 259–270.

1318
J. Peng et al. Geoderma 337 (2019) 1309–1319

Weng, Y.L., Gong, P., Zhu, Z.L., 2010. A spectral index for estimating soil salinity in the hyperspectral vegetation indices as a proxy to monitor soil salinity. Ecol. Indic. 11
Yellow River Delta Region of China using EO-1 Hyperion data. Pedosphere 20 (3), (6), 1552–1562.
378–388. Zhang, T.T., Qi, J.G., Gao, Y., Ouyang, Z.T., Zeng, S.L., Zhao, B., 2015. Detecting soil
Wu, W., Al-Shafie, W.M., Mhaimeed, A.S., Ziadat, F., Nangia, V., Payne, W.B., 2014. Soil salinity with MODIS time series VI data. Ecol. Indic. 52, 480–489.
Salinity Mapping by Multiscale Remote Sensing in Mesopotamia, Iraq. IEEE J. Sel. Zhang, C., Lu, D., Chen, X., Zhang, Y., Maisupova, B., Tao, Y., 2016. The spatiotemporal
Top. Appl. Earth Obs. Remote Sens. 7 (11), 4442–4452. patterns of vegetation coverage and biomass of the temperate deserts in Central Asia
Zhang, T.T., Zeng, S.L., Gao, Y., Ouyang, Z.T., Li, B., Fang, C.M., Zhao, B., 2011. Using and their relationships with climate controls. Remote Sens. Environ. 175, 271–281.

1319

You might also like