PE&RS January 2018 Full - page 44

at both field and landscape scales (Allred et al., 2008; Corwin,
2008). ECa is influenced by many soil properties, including ion
concentrations in soil solutions, clay and water content and
temperature (McNeill, 1980). ECa is also associated with bulk
density, soil structure, soil organic carbon, and pH (Doolittle
et al., 2014). In the salt-affected area, the concentration of
soluble salts is the dominant factor that influences ECa, and
can therefore be measured effectively using EMI sensors (Wil-
liams et al., 2006; Heilig et al., 2012). Therefore, integrating
various data sources (i.e., remote and proximal sensing data
and ancillary measurements) at different scales (i.e., field,
local and regional) using data assimilation/fusion techniques
and simulation models may offer possibilities for better
understanding soil salinity dynamics (Psjr et al., 1991; Met-
ternicht, 2003; Amezketa, 2006; Farifteh et al., 2006; Mulder
et al., 2011; Horta et al., 2015). Modeling approaches for
combining different data sources have been applied to assess
and monitor soil salinity and other soil properties (Table 1).
Most of the studies focused on the development of integrative
retrieval methods or models for soil salinity estimation using
both proximal and remote sensing data. For instance, Fernan-
dez-Buces et al. (2006) developed an exponential regression
model using measured ground radiance and Landsat images
to measure soil EC and sodium absorption ratios in the Mexico
Basin. Semih and Cankut (2008) measured soil reflectance
spectra in northeastern Konya and combined such data with
Landsat TM images to establish an empirical model for soil
salinity estimation.
In addition, hyperspectral remote sensing data (e.g., Hy-
map and Hyperion) have become useful dataset to assess soil
salinity due to the high spectral resolution with the capability
for detecting specific absorption bands of soil salt content (De-
haan and Taylor, 2003; Farifteh et al., 2007; Weng et al., 2008
and 2010). High spatial resolution remote sensing data (e.g.,
Ikonos, QuickBird, and WorldView) are also promising data
sources due to the sufficient resolution to characterize the
spatial variability of soil salinity, and used in many relevant
studies (Eldeiry and Garcia, 2008; De Benedetto et al., 2013;
Sidike et al., 2014). In recent studies, proximal sensing (e.g.,
EM38) used as a primary data source, and combined with
secondary remote sensing images (e.g., Landsat, WorldView)
for conducting large scale soil salinity estimation using vari-
ous interpolation approaches. For example, Wu et al. (2009)
characterized the spatial variability of soil salinity in Fengqiu
County, China using EMI measurements and Landsat TM data
jointly analyzed by Kriging, Exponential Regression and
Regression-Kriging. Ding and Yu (2014) measured ECa of sa-
line soils in Weigan-Kuqa River Delta Oasis, China, and built
regression models with spectral indices derived from Landsat
images for addressing spatial variability of soil salinity in
wet and dry seasons. Nevertheless, the relatively low spatial
Table 1. A short summary of retrieval methods for soil salinity and other properties.
Soil attributes Data type Approach Regression statistics Study area Reference
Spatial extent
and magnitude
of saline soils
IRS-1B data and
Landsat TM image
Visual Interpretation Accuracy = 86%,
Kappa Coefficient = 0.86
Nagarjunsagar Left Bank
canal command area
Dwivedi et
al., 1999
Soil salinity Measured reflectance spectra
and Hymap image
Mixture-tuned matched
filter approach
Accuracy = 86% Tragowel Plains
Irrigation Region
Dehaan and
Taylor, 2003
EC, sodium
absorption ratio
Measured ground radiance and
Landsat TM and ETM+ image
Exponential regression
model
Statistical
confidence = 80%
Basin of Mexico Fernandez-
Buces et al.,
2006
Soil salinity Measured reflectance spectra
and Hymap image
PLSR and ANN R2 = 0.860,
R2 = 0.940
Texel Island, NW
Netherlands
Farifteh et
al., 2007
Soil salinity Measured reflectance spectra
and Landsat TM image
Empirical model R2 = 0.950 Northeast of Konya Semih and
Cankut, 2008
Soil salinity Ikonos and Landsat TM image OLS, Spatial AR and
modified residual Kriging
R2 = 0.840 The southeastern
Colorado, USA
Eldeiry and
Garcia, 2008
Soil salinity Landsat TM image and
EM38 data
Kriging, exponential
regression and RK
0.161 < RMSE < 0.357 Fengqiu County, China Wu et al.,
2009
Soil salinity Measured reflectance
spectra and Hyperion data
PLSR R2 = 0.873,
RMSE = 0.986
Yellow River delta,
China
Weng et al.,
2010
Clay, silt, sand,
EC, TOC, TN , CN
and pH
EM38 data, measured
reflectance spectra
SMLR, PLSR and
PCA-SMLR
0.59 ≤ R 2 ≤ 0.94 west of Wageningen,
Netherlands
Mahmood et
al., 2012
Spatial
heterogeneity of
soil properties
Measured reflectance spectra,
EM38DD data, WorldView-2
and GPR data
PCA, LMC, LMK and non-
parametric clustering
Capitanata plain,
southern Italy
De Benedetto
et al., 2013
Soil salinity QuickBird data and soil
reflectance spectra
PLSR and SMR R2 = 0.992,
RMSE = 0.195
Pingluo County, China Sidike et al.,
2014
Soil salinity Landsat TM image and EM38 UK, SIR and RK R2 = 0.435,
R2 = 0.391
Weigan and Kuqa River
Delta, China
Ding and Yu,
2014
Note: BD is bulk density; EC is electrical conductivity; TOC is total organic content; TN is total nitrogen; CN is carbon to nitrogen ratio; OLS is
ordinary least squares; REML is residual maximum likelihood; OK is ordinary kriging; PLSR partial least squares regression; ANN is artificial
neural network; SWR is stepwise regression; AR is autoregressive model; RK is regression-kriging; SMLR is stepwise linear multivariate regres-
sion; PCA is principal component analysis; PCA-SMLR is the integrated method of PCA and SMLR ; LMC is linear model of coregionalization;
LMK is linear model kriging; SMR is stepwise multiple regression; UK is universal kriging; SIR is Spectral index regression; R 2 is the determi-
nation of coefficient; RMSE is root mean square error.
44 January 2018 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING