You are on page 1of 20

remote sensing

Article
Integrating Active and Passive Remote Sensing Data for
Mapping Soil Salinity Using Machine Learning and Feature
Selection Approaches in Arid Regions
Sayed A. Mohamed 1 , Mohamed M. Metwaly 1 , Mohamed R. Metwalli 1 , Mohamed A. E. AbdelRahman 2
and Nasem Badreldin 3, *

1 Data Reception, Analysis, and Receiving Station Affairs Division, National Authority for Remote Sensing and
Space Sciences, Cairo 11769, Egypt
2 Land Use Department, Division of Environmental Studies and Land Use, National Authority for Remote
Sensing and Space Sciences (NARSS), Cairo 11769, Egypt
3 Department of Soil Science, Faculty of Agricultural and Food Science, University of Manitoba, 13 Freedman
Crescent, Winnipeg, MB R3T 2N2, Canada
* Correspondence: nasem.badreldin@umanitoba.ca

Abstract: The prevention of soil salinization and managing agricultural irrigation depend greatly
on accurately estimating soil salinity. Although the long-standing laboratory method of measuring
salinity composition is accurate for determining soil salinity parameters, its use is frequently con-
strained by the high expense and difficulty of long-term in situ measurement. Soil salinity in the
northern Nile Delta of Egypt severely affects agriculture sustainability and food security in Egypt.
Understanding the spatial distribution of soil salinity is a critical factor for agricultural development
and management in drylands. This research aims to improve soil salinity prediction by using a
combined data collection method consisting of Sentinel-1 C radar data and Sentinel-2 optical data
acquired simultaneously via integrated radar and optical sensor variables. The modelling approach
focuses on feature selection strategies and regression learning. Feature selection approaches that
include the filter, wrapper, and embedded methods were used with 47 selected variables depending
on a genetic algorithm to scrutinize whether regions of the spectrum from optical indices and SAR
Citation: Mohamed, S.A.; Metwaly, texture choose the optimum combinations of selected variables. The sub-setting variables resulting
M.M.; Metwalli, M.R.; AbdelRahman, from each feature selection method were used to train the regression learners’ random forest (RF),
M.A.E.; Badreldin, N. Integrating
linear regression (LR), backpropagation neural network (BPNN), and support vector regression (SVR).
Active and Passive Remote Sensing
Combining the BPNN feature selection method with the RF regression learner better predicted soil
Data for Mapping Soil Salinity Using
salinity (RME 0.000246; sub-setting variables = 18). Integrating different remote sensing data and
Machine Learning and Feature
machine learning provides an opportunity to develop a robust prediction approach to predict soil
Selection Approaches in Arid
Regions. Remote Sens. 2023, 15, 1751.
salinity in drylands. This research evaluated the performances of various machine learning models,
https://doi.org/10.3390/rs15071751 overcame the limitations of conventional techniques, and optimized the variable input combina-
tions. This research can assist farmers in soil-salinization-affected areas in better managing planting
Academic Editor: Xianjun Hao
procedures and enhancing the sustainability of their lands.
Received: 9 February 2023
Revised: 20 March 2023 Keywords: machine learning; feature selection; soil salinity; arid regions
Accepted: 22 March 2023
Published: 24 March 2023

1. Introduction
Copyright: © 2023 by the authors.
Soil salinization is a prevalent form of land degradation, which is a global phenomenon
Licensee MDPI, Basel, Switzerland. that impacts soil productivity and health in arid and semi-arid ecosystems [1–8]; it limits
This article is an open access article the land’s suitability for cultivation or reclamation while also increasing soil dispersion
distributed under the terms and and erosion [9]. Furthermore, due to the initial high water-soluble salt, sparse rainfall,
conditions of the Creative Commons high evaporation, high groundwater levels, and inadequate agricultural practices, this
Attribution (CC BY) license (https:// type of land degradation requires a dynamic soil salinity mapping approach and moni-
creativecommons.org/licenses/by/ toring technique to provide spatiotemporal information that potentially improves the soil
4.0/). conservation management and planning [10].

Remote Sens. 2023, 15, 1751. https://doi.org/10.3390/rs15071751 https://www.mdpi.com/journal/remotesensing


Remote Sens. 2023, 15, 1751 2 of 20

Traditional soil salinization assessment methods depend on field soil sampling, which
requires a huge sampling size due to the high spatial variation in soil salinity, followed by
extensive laboratory analysis, which is time-consuming and labor-intensive, as well as not
operational or applicable for monitoring multitemporal large-scale case studies [11–13].
Remote sensing technology has progressively supplanted the traditional method as a
more effective way of monitoring soil salinity from the same region across regional or large
scales [14–16]. The interaction between soil reflectance and salinity indicators is used to
quantify and measure the spatiotemporal dynamics [17]; this technology has the potential
to obtain accurate chemical and physical soil properties [12].
Various space-borne remote sensing data with different spatial and temporal resolu-
tions have been utilized in soil salinity mapping, thereby resulting in the development
of new techniques for qualitative research methods on soil characteristics [18], such as
soil texture [19], soil bulk density [20], anion and cation exchange [21], soil organic car-
bon [19,22], total available nitrogen [23,24], pH [25], available potassium [26], and total
phosphorus [27]. Recently, several studies succeeded in mapping and monitoring soil
salinity using accessible optical remote sensing data, such as Landsat [8,28,29], SPOT [30],
ASTER [31], RapidEye [30], IKONOS [32], QuickBird [33], and Sentinel 2 [34,35].
The limitation of using solely optical sensors to measure soil properties lies in its
sensitivity to meteorological conditions such as clouds and dust storms, which deteriorate
its radiometric quality [10]. Synthetic aperture radar (SAR) remote sensing has merits under
all weathering conditions and the ability to penetrate vegetation and the soil surface to a
depth of 150 cm, depending on the frequency and bandwidth (e.g., X, S, L, Ka, and P) [22].
Numerous researchers studied the possibility of utilizing SAR technology for mapping and
monitoring the spatial variation in soil attributes such as texture [36], moisture [37], and
salinity field [38–40]. SAR images’ diversity and complexity, in addition to the scarcity of
publicly accessible high-resolution radar and multispectral images, limit their use in digital
soil mapping [41].
Integrating SAR and optical remote sensing can enhance the satellite-based detection
approach for soil surface features, thereby compensating for optical remote sensing limi-
tations. Sentinel-1 and Sentinel-2 have reasonably high spatial and temporal resolutions
to map and assess soil salinization impact at a regional scale [10]. Sentinel-2 has been
used by numerous researchers in assessing soil salinity [35,42,43], and Sentinel-1 has also
been used by several researchers in the soil attribute mapping research field [38,40,44]. In
addition, several researchers combined sensors with varied capabilities and concluded that
combining multi-source sensors can significantly increase the accuracy of the digital soil
mapping field [24,45,46]. To create a spatial prediction model for mapping saline soils in
the study region, we need information on the characteristics impacting the accuracy of the
digital soil map output between the optical and radar data [10,47,48].
For predicting and mapping soil salinity, several statistical models from classical artifi-
cial neural networks (ANN) and deep learning (DL) were applied in the past few years [49];
now, artificial intelligence (AI) and regression models are becoming the new reliable tools in
digital soil mapping. Hosseini et al. [50] indicated that numerous researchers have recently
used new machine learning (ML) approaches, such as decision trees (DT), random forests
(RF), and ANN, to predict some chemical and physical soil properties [51,52]. Achieng [53]
employed a multilayer perceptron neural network (MLP-NN) to forecast soil salinity and
used an artificial and deep neural network to model soil moisture. Analyzing big geospatial
data using ML analyses has special workflows depending on data type and size; it starts
by utilizing semi-automated datasets to find linear and nonlinear correlations between
salinities and other input variables (e.g., salinity indices, vegetation indices, intensity in-
dices, and terrain derivatives), to build statistical relationships for a continuous spatial soil
salinity prediction [10,54]. RF [10,35,55] classification and regression trees (CART) [10], as
well as boosted regression trees (BRT), are among the most common ML models that have
been shown to have robust performance in soil salinity prediction [24,56].
Remote Sens. 2023, 15, 1751 3 of 20

Alamdar et al. [57] evaluated the performance of three machine learning (ML) algo-
rithms, including random forest (RF), gradient boosting machine (GBM), and multi-layer
perceptron (MLP), in detecting and monitoring saline soils using an 18-indicator dataset,
including seven Landsat-8 OLI bands, vegetation, and salinity indices. The GBM method
exhibited the highest performance with an R2 of 0.89 and a RMSE of 0.63, followed by the
RF model with an R2 of 0.85 and a RMSE of 0.71, and the MLP model with an R2 of 0.75 and
a RMSE of 0.88. The study revealed that satellite data could potentially estimate soil salinity
with the appropriate method and acceptable in situ data. Another study [58] evaluated
three ML models’ performances in predicting soil salinity parameters using inputs such
as soil temperature, potential hydrogen, soil water content, and electrical conductivity.
The study used 467 soil samples from northwest China for model training, testing, and
validation. The RF and support vector machine (SVM) models performed well with the
EC, soil temperature, and pH, while the extreme gradient boosting (XGB) model was better
with the EC, soil water content, and soil temperature. Mzid et al. [59] used Sentinel-2
multispectral imagery to predict soil EC in salt-affected soils in central Tunisia, where they
achieved the best results using several ML algorithms. Other studies [60,61] developed
various ML-based models that accurately predicted the spatial distributions of soil salinity.
Wang et al. [62] and Rafik et al. [63] found that integrating different optical satellite data
such as Landsat 4–5, Landsat 7, Landsat 8, and Sentinel 2 had a substantial impact on the
accuracy of soil salinity prediction, with an R2 and Nash–Sutcliffe model efficiency coeffi-
cient (NSE) of 0.93 and 0.86, respectively. Other studies found improvements in soil salinity
prediction when integrating optical data (e.g., Landsat) with a digital elevation model
(DEM) using a deep extreme learning machine (DELM), a sine cosine algorithm—Elman
(SCA-Elman) [64], a DT, and a RF [65].
Recent studies [66–69] have stated that the efficiency of ML models depends not
only on the accuracy of prediction, but also on the combination of input parameters and
the model adopted. Currently, the prediction of soil characteristics is still rare. More
importantly, soil alkalinity and salinity are often determined by the relative content of soil
content [70]. However, predicting soil salinity parameters with machine learning models is
crucial for guiding planting and irrigation. However, there are still few studies in this area,
and there is a lack of specific evaluation and management. Interestingly, using physical
parameters such as conductance, temperature, soil water content, and pH measured by
sensor technology as predictive indicators could improve machine learning efficiency and
reduce the cost of laboratory tests [71,72].
Several ML algorithms have been used for various applications, with optical or radar
sensors as the basis. In [73], The study objective was to compare five machine learning
algorithms for predicting soil salinity: the least absolute shrinkage and selection operator
(LASSO), the multiple adaptive regression splines (MARS), the classification and regression
trees (CART), the RF, and the stochastic gradient tree boost (SGT); This presented us with
a challenge to identify the most effective factor from optical and radar data that was the
impact factor for predicting soil salinity in the northern Nile Delta of Egypt area. However,
there has been no advice about which ML algorithms and satellite sensor settings may be
ideal for predicting soil salinity in various regions [74]. A subset of remote sensing variables
were used to test the soil salinity prediction models using feature selection techniques.
Therefore, the main goal of this study is to construct a robust and flexible spatial
prediction model using optical and SAR satellite data to assess and map soil salinity in the
northern Nile Delta of Egypt as an example of an arid ecosystem. This goal will be achieved
in two steps: (i) using feature selection techniques (e.g., filter, wrapper, and embedded)
depending on statistical and ML algorithms to scrutinize whether regions of the spectrum
from optical indices and SAR texture contain essential information for soil estimation to
choose the optimum combinations of selected variables, and (ii) using the selected variables
to train and optimize the parameters of linear regression (LR), support vector regression
(SVR), RF, and the backpropagation neural network (BPNN). The performance of regression
learners will be measured using descriptive statistics, such as RMSE and R2 (Adj-R2 ).
Remote Sens. 2023, 15, 1751 4 of 20

2. Materials and Method


2.1. Study Area Description and Field Survey
The study area is located on the northwestern side of the Nile delta. It is surrounded
from the north by the Mediterranean Sea, from the east by the Rashid branch, from
the west by the governorates of Alexandria and Matrouh, and from the south by the
southern part of El-Beheira Governorate, see Figure 1. The study area covers an area of
3941 km2 (394,096.5 hectares (ha)). Agricultural fields represent an area of 3311.16 km2
(331,116.12 ha), and elevation ranges from 1 to 69 m above sea level; it is located 5under
Remote Sens. 2023, 15, x FOR PEER REVIEW of 22 arid
conditions, which are characterized by mean minimum and maximum annual tempera-
tures, which are 12.6 ◦ C and 26.2 ◦ C, respectively, which indicates a thermic temperature
regime. The annual precipitation is 83.7 mm, showing torric moisture regime [75], and the
mean annual relative humidity is 57% [76].

Egypt

4 1
5
3 2

Five-point sampling 0 – 30 cm
Two depth
30 – 60 cm
30 m x 30 m
Soil profile

Figure 1. The case study area is located on the northwestern side of the Nile Delta of Egypt; black
Figure 1. The case study area is located on the northwestern side of the Nile Delta of Egypt; black
pointsare
points arethe
thelocation
locationofofsoil
soil samples
samples that
that were
were used
used in the
in the MLML training
training and and testing.
testing.

Table The
1. SoilRosetta
salinity Branch provides
classification a constant
and crop water
growth based onsource for the city’s vast farmed lands.
EC [78,79].
It has a stable climate and abundant fishing opportunities along the coast on the Rosetta
Salinity
Branch Class
and at Edko SoilLake. Salinity (dS/m) Effect
It is well-known ondiverse
for its Crop Plants
agricultural commodity pro-
Non-saline <2 Negligible
ductions, including cotton, rice, wheat, maize, and potatoes. salinity effects.
It leads the governorates in
Slightly saline
food production 2–4
(e.g., fruits and vegetables) Yields of sensitive
and the export of crops maybe
citrus, affected.
potatoes, tomatoes,
Moderately saline 4–8
artichokes, melons, string beans, and pepper. Yields of many crops are affected.
Strongly saline to Afifi
According 8–16
and Darwish [77], theOnly tolerantmaterial
soil parent crops will survive.
of the study area is mainly
Very strongly
clayey. salinesoil
The main >16orders are Entisols and OnlyAridisols,
a few tolerant
these crops
twowill survive.
different soil orders
illustrate the distribution of different subgroups of soil: Vertic Torrifluvents, Typic Quartizp-
2.2. Remote Sensing
samments, Data AcquisitionTypic
Typic Torripsamments, and Pre‐Processing
Torrifluvents, Typic Torriorthents, Typic Haplocalsids,
TypicInHaplosalids,
this research,Typic
SAR Aquisalids,
(Sentinel-1) Typic
andand Calcigypsids.
optical (Sentinel-2) variables were chosen be-
causeAoftotal
theirofeffectiveness
100 soil sampling
in ML locations
modelingwere collectedtofrom
performance 68 agricultural
predict soil salinity fields
in dry-(flood-
lands suchfarms)
irrigated as the Northern Nile Delta
in the research areaofduring
Egypt. a field survey conducted from February to
March To 2018.
incapacitate
At eachweather
samplinglimitations
location,that affect the radiometric
a five-point quality from
sampling method was Sentinel-
conducted to
2collect
(optical remote
five sensing data),
soil samples usingSAR an data
auger were integrated
at each into the
sampling sitemodeling
(30 m × process. The
30 m) (collected
Catband sensor
the four on theand
corners Sentinel-1 (A, B)ofsatellite
the center platform
each plot), which operates
were at 5.405
then GHz them
mixed and has an to
on-site
incidence angle of 20 to 45 degrees. Sentinel-2 (A, B) was launched in April 2016 and con-
tinues to collect data regularly [80,81] and at an altitude of 693 kilometers (km). This plat-
form travels in a sun-synchronous orbit toward the pole. The Copernicus Open Access
Hub was used to obtain SAR single look complex (SLC) products, and pre-processing was
done with the SNAP (Sentinel Application Platform) software. Calibration, multi-looking,
Remote Sens. 2023, 15, 1751 5 of 20

create a representative composite sample (topsoil at 0–30 cm and subsoil at 30–60 cm)
while recording their geographic locations with a portable GPS (UniSrong G120); with this
sampling approach, we collected 1000 soil samples, which were air-dried, pulverized, and
sieved through a 2 mm sieve to collect soil particles for further lab analysis. The pipette
method determined the fractions of soil particles (clay, sand, and silt). In a 1:5 soil water
diluted extract procedure, the multi-parameter method determined the soil electrical con-
ductivity (ECe) (WTW multi-3430). According to the international standards of soil salinity
measurements, the soil sample solutions were extracted using a multiparameter measuring
device (Multi 3420 Set B, WTW GmbH, Weilheim, Germany), which was equipped with a
composite electrode (TetraCon 925) at a normal laboratory temperature of 25 ◦ C to measure
electrical conductivity (EC) (1:5 soil–water extraction solution). The measured ECe was
used to calculate the soil salinity [78]. The soil salinity in this research was categorized as
shown in Table 1. Among the collected soil samples, the general descriptive statistics of
soil salinity (dS/m) in the case study were a minimum of 0.4, a maximum of 16, a mean of
6.2, a standard deviation of 4.3, and a coefficient of variation (CV) of 69.3%.

Table 1. Soil salinity classification and crop growth based on EC [78,79].

Salinity Class Soil Salinity (dS/m) Effect on Crop Plants


Non-saline <2 Negligible salinity effects.
Slightly saline 2–4 Yields of sensitive crops maybe affected.
Moderately saline 4–8 Yields of many crops are affected.
Strongly saline 8–16 Only tolerant crops will survive.
Very strongly saline >16 Only a few tolerant crops will survive.

2.2. Remote Sensing Data Acquisition and Pre-Processing


In this research, SAR (Sentinel-1) and optical (Sentinel-2) variables were chosen be-
cause of their effectiveness in ML modeling performance to predict soil salinity in drylands
such as the Northern Nile Delta of Egypt.
To incapacitate weather limitations that affect the radiometric quality from Sentinel-2
(optical remote sensing data), SAR data were integrated into the modeling process. The
C band sensor on the Sentinel-1 (A, B) satellite platform operates at 5.405 GHz and has
an incidence angle of 20 to 45 degrees. Sentinel-2 (A, B) was launched in April 2016 and
continues to collect data regularly [80,81] and at an altitude of 693 kilometers (km). This
platform travels in a sun-synchronous orbit toward the pole. The Copernicus Open Access
Hub was used to obtain SAR single look complex (SLC) products, and pre-processing was
done with the SNAP (Sentinel Application Platform) software. Calibration, multi-looking,
filtering, and geometry correction were among the pre-processing procedures, along with
layover and shadowing analysis. As a result, a series of geocoded intensity images were
created, which were then transformed into backscattering coefficients ranging from 0 to 1.
(Sigma Nought δ0). The software converts radar reflectivity into a radar cross-section with
an area normalization aligned with the ground range plane. The final products are encoded
as 16-bit integers (signed for SLC, unsigned for GRD) [82]. As a result, the Sentinel-1 images
were transformed into Sigma Nought (δ0). Also, gray-level co-occurrence matrix (GLCM)
factors included contrast, energy, dissimilarity, homogeneity, angular second moment, and
correlation [83].
Sentinel-2 multi-spectral instrument (MSI) level-1C (L1C) images were acquired on
14 June 2019, which had reflectance data from the top of the atmosphere (TOA). The Sen2Cor
technique was used to transform L1C data into level-1A (L1A) data; TOA reflectance was
transformed to the bottom of the atmosphere or Earth’s surface reflectance following
atmospheric adjustment. In this investigation, four (4) bands (b) with a resolution of
10 m (b2, b3, b4 and b8) and six (6) bands with 20 m spatial resolutions (b5, b6, b7, b8A,
b11, and b12) were used from Sentinel-2. The selected bands were stacked and trimmed
using SNAP software to create a subset of the research region. An ML approach was
used to better comprehend the sensitivity of SAR and optical bands to the soil salinity
Remote Sens. 2023, 15, 1751 6 of 20

estimation [84]. Sentinel-1 efficiency of Sigma Nought (δ0) feature extraction of level 1
images and Sentinel-2 indices were employed to predict and determine the areas under soil
salinity using a feature selection and prediction method. These raster factors (variables)
are shown in Table 2. Fifteen spectral indices—normalized difference vegetation index
(NDVI), green normalized difference vegetation index (GNDVI), difference vegetation index
(DVI), weighted difference vegetation index (WDVI), transformed normalized difference
vegetation index (TNDVI), soil-adjusted vegetation index (SAVI), modified soil-adjusted
vegetation index (MSAVI), second modified soil-adjusted vegetation index (MSAVI2),
modified chlorophyll absorption ratio index (MCARI), infrared percentage vegetation
index (IPVI), normalized difference water index (NDWI), modified normalized difference
water index (MNDWI), color index (CI), brightness index (BI)—were calculated using the
pre-processing Sentinel-2 bands and SNAP software 9.0.0.

Table 2. The stacked pre-processed raster variables of Sentinel-1 and Sentinel-2 that were used in ML
training for mapping soil salinity in the study area.

Descriptions Selected Variable


Sentinel-2, Band (b) selection b2, b3, b4, b5, b6, b7, b8, b8A, b11 and b12.
WDVI, TNDVI, SAVI, NDWI, NDVI, MSAVI, MSAVI2,
Sentinel-2, Indices
MNDWI, MCARI, IPVI, GNDVI, DVI, CI, BI and BI2.
Sentinal-1, sigma nought (δ0) db VV and VH.
Contrast_VV, Dissimilarity_VV, Homogeneity_VV,
AngularSecondMoment_VV, Energy_VV, Entropy_VV,
V V_ GLCM
MaximumProbability_VV, Correlation_VV, Mean_VV
and StandardDviation_VV.
Contrast_VH, Dissimilarity_VH, Homogeneity_VH,
AngularSecondMoment_VH, Energy_VH, Entropy_VH,
VH_ GLCM
MaximumProbability_VH, Correlation_VH, Mean_VH
and StandardDeviation_VH.

The feature selection (FS) and ML algorithms were analyzed using MATLAB R2021b
software. We used MATLAB Bayesian optimizer with 50 iterations and 3K-fold cross-
validation to fine-tune the different ML models’ hyperparameters (SVR, RF, BPNN) and
avoid overfitting, see Figure 2.

2.3. Feature Selection Techniques


The feature selection (FS) approach was used as a data reduction technique in the pre-
processing data stage to locate precise data models by deleting irrelevant and redundant
characteristics from a dataset to improve ML algorithms’ performance. It is divided
into several methods depending on several criteria, which include (i) the assessment
mechanism of the combined features and (ii) the technique that evaluates features at each
learning algorithm.
The pros and cons of feature selection strategies will be examined in this analysis stage.
The prime task of these various approaches is to select a subset (s) of the best-performing
inputs (features/variables).

2.4. Filter Methods


This approach is based solely on the data’s intrinsic qualities and computes a signifi-
cant value between one independent variable and the dependent variable. The independent
and dependent variables are used to explain the statistical coefficients. Both variables are
linearly dependent; their correlation coefficient should be close to one, and the correlation
coefficient is 0 if the variables are uncorrelated [84]. Only positive values are included
when the Pearson coefficient is employed as a filter mechanism, as shown in Equation (1):

∑ni=1 cov (xi , yi )


r = (1)
∑ni=1 σxi σyi
Remote Sens. 2023, 15, 1751 7 of 20

where xi and yi are the ith observations of independent and dependent variables respec-
tively; cov corresponds to covariance, and σ indicates the standard deviation of x and y.

Figure 2. A workflow illustration describes the analysis steps using various sources of remote sensing
data inputs and ground truthing (soil sampling).

Using a monotonic function, the Spearman coefficient quantifies the association be-
tween two variables [85]. A monotonic function is either wholly growing or entirely
declining. It is comparable to the Pearson coefficient; only it works with data ranks rather
than raw data. Equation (2) defines the Spearman correlation rank:

6 ∑ni=1 d2
ρ = 1− (2)
n(n2 − 1)

where di is the difference between ranks for each xi , yi data pair, and n is the number of
data pairs.
Remote Sens. 2023, 15, 1751 8 of 20

Entropy-based information gain: discretizes the independent variable, then computes


the entropy between x and the continuous y variable using Equation (3) [86]:

InfoGain = H(y) + H(x) − H(y, x) (3)

where H(x) and H(y) correspond to Shannon’s Entropy for x and y variables, and H(y, x)
is a joint Shannon’s Entropy for a variable y with a condition to x.
F-tests univariate feature ranking algorithm independently examines each predictor’s
value alone using an f-test to determine its importance. Each f-test compares the choice
hypothesis and states that the response values are grouped by estimated values (variable)
Remote Sens. 2023, 15, x FOR PEER REVIEW
taken from populations with the same mean, see Figure 3. The negative logs of9 the
of 22
p-values
are used to calculate the scores [87].

Theselected
Figure 3.3.The
Figure selected variables
variables by filter
by filter methods
methods of entropy,
of entropy, F-testsF-tests univariate
univariate feature feature
ranking,ranking,
Pearson, and Spearman.
Pearson, and Spearman.

2.5. Wrapper Methods


The main function of this stage is to evaluate the relative utility of each distinct subset
of features using prediction performance assessments of a given ML methodology (e.g.,
forward selection or backward elimination). The recursive feature elimination (RFE) is one
Remote Sens. 2023, 15, 1751 9 of 20

2.5. Wrapper Methods


The main function of this stage is to evaluate the relative utility of each distinct subset
of features using prediction performance assessments of a given ML methodology (e.g.,
forward selection or backward elimination). The recursive feature elimination (RFE) is
one of the wrapping techniques, which uses an ML algorithm and sequential search to
choose a subset of characteristics to best to predict response values. It constructs models
repeatedly, removing one unnecessary feature at a time until the predictor performance
decreases the least, which builds best-performing models using left features until all
variables have been studied. During the analysis, the selected features were frequently
ranked using cross-validation, which limits overfitting by defining a dataset to “test” the
model during the training phase. To compute the criterion for each candidate feature subset,
a performance metric such as root mean squared error (RMSE) was utilized to evaluate
the feature subset selection performance [88]. This study employed four ML regression
algorithms, namely: LR, SVR, RF, and BPNN. LR fits a linear equation to the observed
data to establish the relationship between dependent and independent variables using
the least squares approach. SVR utilizes kernel functions to describe complex nonlinear
interactions. RF employs an ensemble method to integrate multiple DTs and improve
prediction performance, while BPNN mimics signal transmission in human brains to
achieve high prediction accuracy for complex problems with both large and small samples.
Selected variables used different ML algorithms, including RFE in this study and
RMSE for regression performance evaluation, which are shown in Table 3.

Table 3. The recursive feature elimination base learners of LR, SVR, RF, LASSO, and BPNN.

No. of Subsetted
Learner Subsetted Variables RMSE
Variables

• band3, band6, band8, band8A, band11.


• TNDVI, SAVI, NDVI, MSAVI, MSAVI2, MCARI, DVI, BI2,
• VV, VH.
LR • Contrast_VV, Dissimilarity_VV, Homogeneity_VV, Correlation_VV, 29 0.48174263
Mean_VV, Standard_Deviation_VV.
• Contrast_VH, Dissimilarity_VH, Angular_Second_Moment_VH,
Energy_VH, Entropy_VH, Maximum_Probability_VH,
Correlation_VH, Standard_Deviation_VH.
• band3, band4, band5, band7, band8, band12.
• MNDWI, MCARI, SAVI, WDVI, BI2, NDVI.
• Mean_VV, Homogeneity_VV, Maximum_Probability_VV,
RF Angular_Second_Moment_VV, Energy_VV, Correlation_VV. 24 0.52825744
• Mean_VH, Homogeneity_VH, Angular_Second_Moment_VH,
Entropy_VH, Standard_Deviation_VH, Dissimilarity_VH.
• band6, band7, band11.
• BI2, WDVI, MNDWI, NDVI, DVI, CI, GNDVI.
SVR • VV. 14 0.55513133
• Mean_VV.
• Mean_VH, Dissimilarity_VH.
• band3, band4, band5, band6, band7, band12.
• SAVI, MSAVI, MNDWI, MCARI, GNDVI, DVI.
BPNN • VV. 18 0.302719033
• Mean_VV.
• Contrast_VH, Dissimilarity_VH, Mean_VH, Standard_Deviation_VH.
Remote Sens. 2023, 15, 1751 10 of 20

2.6. Embedded Methods


This method incorporates variable selection as part of the learning process [89], which
has the benefit of being reasonably quick, because the selection process is integrated into
the model fitting process, and no external feature selection tool is required. In addition,
this stage provides an immediate link between picking characteristics and performance
measures that the model tries to optimize, such as the root mean squared error (RMSE) [90].
RF [91] and least absolute shrinkage and selection operator (LASSO) were employed as
embedded approaches [92], see Table 4.

Table 4. The Feature selection and variables statistical assessment using embedded methods for soil
salinity predictions.

Number of Subsetted
Learner Subsetted Variables RMSE
Variables
WDVI, TNDVI, SAVI, NDWI, NDVI, MSAVI,
MSAVI2, IPVI, GNDVI, DVI, CI.
RF VV, VH 15 0.52825744
Mean_VV.
Mean_VH.
band3, band6, band8, band8A, band11, band12,
TNDVI, MNDWI, CI, BI2,
VV.
Contrast_VV, Dissimilarity_VV,
Homogeneity_VV,
Angular_Second_Moment_VV, Correlation_VV,
LASSO Mean_VV. 23 0.5093330
Contrast_VH, Dissimilarity_VH,
Homogeneity_VH,
Angular_Second_Moment_VH,
Maximum_Probability_VH, Correlation_VH,
Angular_Second_Moment_VH,
Maximum_Probability_VH, Correlation_VH.

3. Results
3.1. Modeling Assessment
The feature selection methods that chose a subset of variables were assessed, which
facilitated the soil salinity prediction models evaluation that was based on the selected
variables. The regression learners’ results after training with selected (subsets) variables
were chosen using different feature selection methods. Descriptive statistics were used
to estimate the performance of the regression learners using parameters such as the root
mean square error (RMSE), the coefficient of determination (R2 ), and the coefficient of
determination (adj-R2 ). The underlined values corresponded to the best two learners and
the tested feature selection method. As is shown in Figure 4, these statistical metrics were
commissioned to measure the prediction accuracy of the regression learners. A total of
47 variables were used to train the regression learners, and 456 examples were used in
the modeling assessment. The best two learners obtained the highlighted values and the
feature selection procedure. The Pearson, Spearman, and f-test filter techniques chose
21 variables, while the entropy coefficients selected 25 variables.
Remote Sens.
Remote Sens. 2023,
2023, 15,
15, 1751
x FOR PEER REVIEW 12
11 of 22
of 20

A B

Figure
Figure 4.4. Regression learners training results with a subset of variables selected by feature selection
methods; feature selection approaches were (i) recursive feature elimination (RFE), which included
LR, RF,
LR, RF,SVR,
SVR,and
andBPNN;
BPNN;(ii)
(ii)filter
filterincluded
includedentropy,
entropy, Pearson,
Pearson, Spearman,
Spearman, and
and f-Test;
f-Test; andand
(iii)(iii) embed-
embedded
ded included
included LassoLasso and
and RF; RF;
(A) (A) coefficient
coefficient of determination
of determination (R2 ), (B)(RAdjusted
2), (B) Adjusted coefficient of deter-
coefficient of determination
mination
(adj-R2 ), and (C) ),root
(adj-R 2 and (C) root mean square error (RMSE).
mean square error (RMSE).

3.2. Feature Selection Approaches for Digital Soil Salinity Mapping Mapping
In the filter
filter method,
method,February’s
February’sstudy studydates
dateswere
werechosen
chosen toto avoid
avoid rice
rice cultivation,
cultivation, as
as
ricerice is grown
is grown in in
thethe
area area
by by immersion.
immersion. Four
Four associated
associated methods
methods werewere included
included for for
ex-
extracting
tracting variousvarious soilsalinity
soil salinityranges
ranges(classes)
(classes)[93,94].
[93,94].The
The first
first method
method (entropy method
with
with aa RFRFlearner)
learner)waswasthe theleast
leastmixing
mixingbetween
betweensaline
salinelands,
lands, urban
urban areas, and
areas, andwater bodies.
water bod-
In
ies.the
In second
the second method
method(Spearman
(Spearman method
methodwithwith
a RFa learner),
RF learner),there waswas
there some confusion
some confu-
with somesome
sion with spotsspots
in urban and water
in urban bodies,
and water whichwhich
bodies, makesmakes
this method the most
this method theintegrating
most inte-
for water bodies with saline land. Similarly, the third method
grating for water bodies with saline land. Similarly, the third method (Pearson (Pearson method with a RF
method
learner) was similar to the first method in avoiding urban areas. In
with a RF learner) was similar to the first method in avoiding urban areas. In contrast, thecontrast, the fourth
method (f-test method
fourth method with a RF
(f-test method withlearner) was comparable
a RF learner) to the second
was comparable to themethod
secondin merging
method in
some
merging urbansomeareas withareas
urban the lands
with theaffected
landsby salt (see
affected byFigure 5).Figure 5).
salt (see
Remote Sens. 2023, 15, x FOR PEER REVIEW 13 of 22
Remote Sens. 2023, 15, 1751 12 of 20

A 4.62% 12.36% Non-saline B 4.08%


13.96%

Slightly saline
Moderately saline
49.01% 34.02% Strongly saline 47.51% 34.45%

C 1.37% 9.76% D 7.18% 10.81%


Non-saline
Slightly saline
Moderately saline

40.93%
Strongly saline
47.94% 43.97% 38.04%

Figure
Figure The
5. 5. predicted
The soilsoil
predicted salinity spatial
salinity distributions
spatial and the
distributions andtotal
theoftotal
classified salinitysalinity
of classified share areas
share
(%) at four (4) filter methods, (A) entropy method with RF learner; (B) Spearman
areas (%) at four (4) filter methods, (A) entropy method with RF learner; (B) Spearman methodmethod with RF
with
RF learner;
learner; (C) Pearson
(C) Pearson method
method with
with RF RF learner;
learner; andf-test
and (D) (D) f-test
methodmethod
with RFwith RF learner.
learner.

InInthethewrapper
wrapper method,
method, four methods
four methods successfully
successfully extracted
extractedthe soil
the salinity
soil salinity ranges
ranges
inin
the case study. The first method (BPNN with a RF learner) minimized
the case study. The first method (BPNN with a RF learner) minimized mixing between mixing between
saline
salinelands,
lands, urban
urbanareas, and
areas, and water bodies.
water bodies. InIn
the
thesecond
second method
method (LM
(LMmethod
method with
with a RF
a RF
learner),
learner), there
therewas some
was some ambiguity
ambiguity with
withsome
some urban
urban areas
areasand
andwater
waterbodies.
bodies.It It
is is
noted
noted
that
thatthis
thismethod
method was
wasthe
themost
mostintegrating
integratingforforwater
waterbodies
bodieswith saline
with salinesoils, which
soils, which was
was
identical to the third method (RF method with a RF learner) in the northern
identical to the third method (RF method with a RF learner) in the northern part of the part of the
area.
area.The
The fourth
fourthmethod
method (SVR
(SVR method
method with
witha RF
a RFlearner)
learner) performed
performed a prediction
a prediction that was
that was
similar to the third method (the south part of the study area) in merging some
similar to the third method (the south part of the study area) in merging some urban areas urban areas
with
with the
thesalinity-affected
salinity-affected soil (see
soil Figure
(see 6).6).
Figure
Remote Sens. 2023, 15, x FOR PEER REVIEW 14 of 22
Remote Sens. 2023, 15, 1751 13 of 20

A 3.27% 9.45% Non-saline B 4.58% 13.99%


Slightly saline
Moderately saline
Strongly saline
51.63% 35.64%
50.14% 31.29%

C 8.31% 9.87% D 7.09% 11.47%


Non-saline
Slightly saline
Moderately saline
43.07% 38.75% Strongly saline 45.57% 35.86%

Figure
Figure6. 6.The
Thepredicted soil
predicted soilsalinity
salinityspatial distribution
spatial and
distribution andthe total
the ofofclassified
total salinity
classified share
salinity areas
share areas
(%) at at
(%) four
four(4) wrapper methods,
(4) wrapper methods,(A) BPNN method
(A) BPNN with RF
method learner;
with (B) LM method
RF learner; (B) LM with RF learner;
method with RF
(C) RF method
learner; (C) RFwith RF learner;
method with RF and (D) SVR
learner; andmethod
(D) SVR with RF learner.
method with RF learner.

The
Theembedded
embeddedmethod
methodwas wasunblemished
unblemishedcompared
comparedwith withthe
thepreviously
previouslymentioned
mentioned
two
two methods, which were analogous to the extent of congruence in the outputs,but
methods, which were analogous to the extent of congruence in the outputs, butwhat
what
was very clear was the superiority of the first method in the small percentage
was very clear was the superiority of the first method in the small percentage of overlap of overlap
between
betweenurban
urbanareas
areasand
andwater
waterbodies
bodieswith
withsaline
salinelands.
lands.As
Asisisshown
shownininFigure
Figure7,7,the
the
results
results show the extent of the spread of saline lands in the study area. They, thus,indicate
show the extent of the spread of saline lands in the study area. They, thus, indicate
the
theimportance
importanceofof
finding
findingananoptimal
optimalsolution
solution toto
track
tracksaline lands
saline lands andandcheck
checkthe
theextent
extentofof
the increase and decrease in areas to build an optimal strategy to solve the salinity
the increase and decrease in areas to build an optimal strategy to solve the salinity prob- problem
by avoiding the processes of its formation and, thus, limiting it and working on choosing
lem by avoiding the processes of its formation and, thus, limiting it and working on choos-
appropriate management methods to work on soil conservation.
ing appropriate management methods to work on soil conservation.
Remote Sens. 2023, 15, x FOR PEER REVIEW 15 of 22
Remote Sens. 2023, 15, 1751 14 of 20

A 6.27% 10.8% Non-saline B 6.43% 13.22%


Slightly saline
Moderately saline
Strongly saline
48.03% 34.89% 46.01% 34.35%

Figure7.7. The
Figure The predicted
predictedsoil
soilsalinity
salinityspatial
spatialdistribution
distributionand
andthe
thetotal
totalof
ofclassified
classifiedsalinity
salinityshare
shareareas
areas
(%)
(%)atat two
two (2) embedded methods,
methods, (A)(A)LASSO
LASSOmethod
methodwith
withRF
RFlearner,
learner,and
and(B)(B)
RFRF method
method with
with RF
learner.
RF learner.

The
The optimal
optimal method
method showed
showed aa superior
superior ability
ability to
to reach
reach the
the saline
saline lands
lands instantly,
instantly,
thereby
therebyovercoming
overcomingthe theproblem
problemof ofvegetation
vegetationcover,
cover,asaswell
well as
as scattered
scattered and
and dense
dense build-
build-
ings.
ings. Also, the bare lands in the northern and southern regions helped reach the lands
Also, the bare lands in the northern and southern regions helped reach the saline saline
with
landshigh accuracy
with to identify
high accuracy the spectrum
to identify characteristics
the spectrum with the fallow
characteristics areas
with the easily.
fallow areas
easily.
4. Discussion
In this research, the analysis showed that the best-performing algorithms that predict
4. Discussion
soil salinity using the feature selection approach were the RF (regression learner) with
In this research, the analysis showed that the best-performing algorithms that predict
BPNN, entropy, f-Test, lasso, LR, Pearson, RF, Spearman, and SVR. The lowest performance
soil salinity using the feature selection approach were the RF (regression learner) with
was the LR and BPNN; these comparisons were based on the statistical assessment of the
BPNN, entropy, f-Test, lasso, LR, Pearson, RF, Spearman, and SVR. The lowest perfor-
RMSE, R2 , and R2 -adj (see Figure 4). The ML algorithms’ robustness resulted from the
mance was the LR and BPNN; these comparisons were based on the statistical assessment
importance of the variables from Sentinel 1 and Sentinel 2. The feature selection approach
of the RMSE, R2, and R2-adj (see Figure 4). The ML algorithms’ robustness resulted from
discovered the variables that significantly impacted detecting soil salinity in the case study.
the importance of the variables from Sentinel 1 and Sentinel 2. The feature selection ap-
They were six (6) bands from Sentinel-2 (band3, band4, band5, band6, band7, and band12),
proach
six discovered
(6) spectral the (SAVI,
indices variables that significantly
MSAVI, MNDWI, MCARI, impacted detecting
GNDVI, andsoil salinity
DVI), and sixin the
(6)
case bands
SAR study. and
Theyderivatives
were six (6)(VV,
bands from Sentinel-2
Mean_VV, (band3,Dissimilarity_VH,
Contrast_VH, band4, band5, band6, band7,
Mean_VH,
andStandard_Deviation_VH).
and band12), six (6) spectral indices (SAVI,
Therefore, MSAVI,
as was MNDWI,
consistent withMCARI,
what has GNDVI, and DVI),
been listed from
and six (6) SAR bands and derivatives (VV, Mean_VV, Contrast_VH,
the results, the areas of the best method (BPNN method with a RF learner) used according Dissimilarity_VH,
Mean_VH,
to andanalysis
the statistical Standard_Deviation_VH).
and land inventory in Therefore, as was
the field were consistent
as follows: with
3.27% forwhat
Stronglyhas
been listed from the results, the areas of the best method
saline > 8 dS/m and 51.63% of the total area for Moderately saline 4–8 dS/m. (BPNN method with a RF
learner)
The used
chosen according
model was,to theaccording
statisticalto analysis
the mostandaccurate
land inventory in the field were
spatial distribution of the as
follows: 3.27% for Strongly saline > 8 dS/m and 51.63% of the total
areas affected by salinity, capable of differentiating between the results and problems area for Moderately
saline 4–8from
resulting dS/m. all the models proposed in the previous studies in terms of the inability
The chosen
to differentiate betweenmodel was,salineaccording
lands and to urban
the most accurate
areas, as wellspatial distribution
as overlap of the ar-
with bodies of
eas affected
water and wetlands. This selection was built according to the statistical investigations re-
by salinity, capable of differentiating between the results and problems of
sulting
the from
results, and,allaccordingly,
the models the proposed in the previous
investigations were builtstudies
for thein terms
field of the inability
distribution of saline to
differentiate
lands. On thisbetween
basis, thesaline lands and
comparison urban areas,
between as wellis
the methods asclear,
overlap as iswith bodies
shown of water
in Figure 4.
and The
wetlands. This selection
salinity-affected soilswas
were built according
impeding thetogrowth
the statistical
of mostinvestigations
crops due to the of the
highre-
sults, and, of
percentage accordingly, the investigations
dissolved salts. These soils arewere built forproductive
considered the field distribution of saline
croplands; however,
the drought impacts and other abiotic stressors limit land productivity and soil health. In
Remote Sens. 2023, 15, 1751 15 of 20

this research, the salt-affected soils were widely spread in the northern regions of the study
area. The total areas of each method are shown in Figures 5–7, which depend entirely on
surface irrigation.
A further finding from the soil salinity classification mapping approach is that most
severely and excessively salinized soils were distributed in the northern region near the
lake. As a result of severe tidal influence, wind erosion, marine sedimentation, and
supratidal environment in the late Quaternary, it was suggested that the high saline soils
and sabkhas in this region were developed owing to the flattening of Pleistocene sand
dunes. Shamal winds (northerly winds) eroded inland dunes down to the groundwater
table level, while interdune voids simultaneously filled up during sea incursion to result
in carbonate sedimentation. Most of the non-saline soils were found in the southern
inland region.
Irrigation water management is the foremost factor in agricultural intensification and
horizontal expansion, which constitutes an essential pillar in agricultural development in
Egypt. The results showed that the distribution of salt-affected soils was highly dependent
on environmental factors such as climate, geology, geochemistry, and hydrological condi-
tions. The composition of the different types of salt-affected lands in the irrigated areas
is directly related to the concentration of chloride, sulfate, and bicarbonate that prevail.
The seawater has an additional impact on the nearby soils, and salty lakes accelerate the
salinity of the soil solution. Also, magnesium chloride and sulfate salts are the main salinity
sources in the Edko Lake.
Soil degradation processes in the study area are affected by several factors such as soil
depth, exposure time, regularity of impacts, and groundwater fluctuation, and, usually,
these soils are rich in gypsum veins that arise under conditions of deep groundwater, while
the alkaline type forms in areas that suffer from high groundwater levels. The main reason
for the emergence of salinity in that area is the predominance of sodium chloride salts,
and irrigation with saline water with a salinity of 2500 ppm significantly drains the water.
Alkaline soils in the study area are characterized by destructive natural properties such as
poor ventilation, permeability, and leaching, which are directly related to the dominance of
sodium cations on the exchange bench and the presence of magnesium silicate precipitated
during the formation of alkaline soils. The soils rich in gypsum veins are characterized
by low water permeability due to deaf layers under the layers rich in gypsum veins at
different depths. Additionally, when these soils are exposed to enough soil moisture, these
layers are dismantled, which turn into deaf layers by drought. The crop rotation system
in the study region where salinity-affected soils are includes rice/cotton in the summer
and wheat/alfalfa in the winter. All these crops have a degree of salinity tolerance to some
extent. The sugar beet crop is also cultivated in that area, and it is a salinity-resistant crop
that feeds the sugar factory in the northern Delta region.
The used dataset contained relatively higher information dimensions, including veg-
etation type and vitality, surface reflectance, surface texture (different scales), terrain
variables that indirectly represent hydrological changes, parent materials, and so on. Fur-
thermore, the results also show that the selected inputs were obviously adequate and useful
information from the complex variable datasets in this study. In areas covered by vegeta-
tion, several studies have shown that vegetation vitality (which can be indirectly reflected
by the vegetation index) can mitigate the extent to which soil is affected by salinization
in areas with high proportions of bare soil. Peng et al. [95] proved that soil reflectance
increased with increased electrical conductivity in the coastal to SWIR1 band. This was the
basis for constructing a soil salinity index on bare land. Therefore, the performances of
selected parameters were outstanding in the study area.

5. Conclusions
Soil salinity poses a major threat to land resources in the region, but some methods
work to mitigate the effects of this problem or reverse it completely. To this end, it is
necessary to scientifically diagnose the problem to develop economically viable measures
Remote Sens. 2023, 15, 1751 16 of 20

to mitigate salinity. In the absence of sufficient information, salinity mitigation measures


cannot be applied. Accordingly, this project is related to diagnosing salinity using tradi-
tional and modern methods (remote sensing) using spectral analysis procedures for saline
lands in the northwest Egypt Delta.
Soil salinity on the northern coast of the Nile Delta of Egypt is considered an essential
challenge facing decision makers in developing effective agriculture sustainability and
soil conservation policies. Hence, using advanced remote sensing technology to map soil
salinity with the highest accuracy and spatial resolution is paramount in Egyptian food
security and rural development.
ML and remote sensing data provide a unique opportunity to develop a robust predic-
tion approach. The selected subset of variables by RF and base learners, such as LR, SVR,
RF and BPNN, which were assessed using RMSE for modeling performance, were highly
efficient in soil salinity prediction. Clayey soils were the primary component in abundance
in the study area due to the Nile Delta formation; also, the study found that the amount
of silt had decreased in the north of the Edko Lake, which encouraged the soil salinity
increase due to the Mediterranean Sea and saltwater intrusion. Thus, all of this resulted in
forming the area of lands affected by salt adjacent to the coastline along the north of the
study area. Using the proposed methodology, these lands have become easy to track and
predict for their salinity hazard.
As a result of this research’s findings, seven recommendations were proposed 1—
Reclamation of the pristine lands affected by salts with high ground and saline water levels,
such as the lands of the bottom of Lake Edko; 2—Reclamation of saline lands with low
productivity or virgin land with a low groundwater level, such as lands parallel to the
coastline in the study area; 3—Reclamation of waterlogged lands in the north and east of the
area around Lake Edko; 4—Use chemical enhancers such as (a) soluble calcium salts such
as gypsum and calcium chloride, (b) acidic substances such as sulfuric acid, iron sulphate,
aluminum sulphate, lime, sulphur, pyrite, and (c) calcium salts with a low solubility
product (limestone); 5—Improving the lands that suffer from secondary salinity resulting
from irrigation with water of low validity; 6—Cultivating crops that improve sod lands,
such as rice, sugar beet, and pasture plants; 7—Maintaining and preserving productive
agricultural lands from the emergence of secondary problems of salinity, waterlogging, and
the maintenance of covered drainage networks.
Given the aggravation of the problem of alkalinity associated with the salinity of
the land, it is recommended to add soil conditioners such as agricultural gypsum (at a
rate of 50% of gypsum needs), lime, and municipal fertilizer at a rate of 20 m3 /feddan
(1 feddan = 4200 m2 ), in addition to mixing gypsum and lime separately with municipal
fertilizer. It is also recommended to irrigate with fresh water and mixed water. It is also
recommended to periodically analyze other soil properties such as cation exchange capacity
(CEC), pH, nitrogen, and phosphorus. It is recommended with these treatments to grow
rice and wheat crops under highly alkaline soil conditions.
The study succeeded in drawing a map of soil characteristics and salinity to use soil
resources sustainably, as well as extrapolating the results and publishing the salinity map,
which will contribute to linking land uses with the spread of soil salinity. Also, future
research avenues may entail expanding the proposed approach to encompass other regions
impacted by soil salinization and examining the effect of diverse crop types on soil salinity
levels. Furthermore, additional research could investigate the influence of climate change
on soil salinity and strategies to manage these changes in agricultural environments. There
is also a need to develop a cost-effective, user-friendly tool that incorporates this approach
to facilitate farmers in predicting soil salinity levels in their fields and implementing
suitable management techniques. As for policy formulation, governmental bodies could
allocate resources toward advancing and implementing remote sensing technologies and
ML models for soil salinity prediction in agricultural areas. Incentives could be offered
to farmers who adopt sustainable management practices to mitigate soil salinity levels.
Governments could also promote education and awareness programs to educate farmers
Remote Sens. 2023, 15, 1751 17 of 20

about soil salinity management strategies and the significance of avoiding soil salinization.
By adopting these measures, policymakers can help enhance food security and agricultural
sustainability in regions impacted by soil salinity, such as Egypt.

Author Contributions: Conceptualization, S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.; method-
ology, S.A.M., M.M.M., M.R.M. and M.A.E.A.; software, S.A.M., M.M.M. and M.R.M.; validation,
S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.; formal analysis, S.A.M., M.M.M., M.R.M., M.A.E.A.
and N.B.; investigation, S.A.M., M.M.M., M.R.M. and M.A.E.A.; resources, S.A.M., M.M.M., M.R.M.
and M.A.E.A.; data curation, S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.; writing—original draft
preparation, S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.; writing—review and editing, S.A.M.,
M.M.M., M.R.M., M.A.E.A. and N.B.; visualization, S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.;
supervision, S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.; project administration, S.A.M., M.M.M.,
M.R.M., M.A.E.A. and N.B.; funding acquisition, S.A.M., M.M.M., M.R.M., M.A.E.A. and N.B.; All
authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: We thank the academic editor and reviewers for their constructive comments to
improve the manuscript’s quality.
Conflicts of Interest: The authors declare conflict of interest.

References
1. AbdelRahman, M.A.E.; Metwaly, M.M.; Afifi, A.A.; D’Antonio, P.; Scopa, A. Assessment of Soil Fertility Status under Soil
Degradation Rate Using Geomatics in West Nile Delta. Land 2022, 11, 1256. [CrossRef]
2. AbdelRahman, M.A.E.; Engel, B.; Eid, M.S.M.; Aboelsoud, H.M. A new index to assess soil sustainability based on temporal
changes of soil measurements using geomatics—An example from El-Sharkia, Egypt. All Earth 2022, 34, 147–166. [CrossRef]
3. AbdelRahman, M.A.E.; Farg, E.; Saleh, A.M.; Sayed, M.; Abutaleb, K.; Arafat, S.M.; Elsharkawy, M.M. Mapping of soils and
land-related environmental attributes in modern agriculture systems using geomatics. Sustain. Water Resour. Manag. 2022, 8, 116.
[CrossRef]
4. Aboelsoud, H.M.; AbdelRahman, M.A.E.; Kheir, A.M.S.; Eid, M.S.M.; Ammar, K.A.; Khalifa, T.H.; Scopa, A. Quantitative
Estimation of Saline-Soil Amelioration Using Remote-Sensing Indices in Arid Land for Better Management. Land 2022, 11, 1041.
[CrossRef]
5. AbdelRahman, M.A.E.; Afifi, A.A.; D’Antonio, P.; Gabr, S.S.; Scopa, A. Detecting and Mapping Salt-Affected Soil with Arid
Integrated Indices in Feature Space Using Multi-Temporal Landsat Imagery. Remote Sens. 2022, 14, 2599. [CrossRef]
6. Ibrahim, M. Modeling Soil Salinity and Mapping Using Spectral Remote Sensing Data in the Arid and Semi-arid Region. Int. J.
Remote Sens. Appl. 2016, 6, 76. [CrossRef]
7. Mohamed, N.N. Management of salt-affected soils in the Nile Delta. In The Nile Delta; Negm, A., Ed.; The Handbook of
Environmental Chemistry; Springer: Cham, Switzerland, 2016; Volume 55, pp. 265–295. [CrossRef]
8. Hammam, A.; Mohamed, E. Mapping soil salinity in the East Nile Delta using several methodological approaches of salinity
assessment. Egypt J. Remote Sens. Space Sci. 2020, 23, 125–131. [CrossRef]
9. Alqasemi, A.S.; Ibrahim, M.; Al-Quraishi, A.M.F.; Saibi, H.; Al-Fugara, A.; Kaplan, G. Detection and modeling of soil salinity
variations in arid lands using remote sensing data. Open Geosci. 2021, 13, 443–453. [CrossRef]
10. Ma, G.; Ding, J.; Han, L.; Zhang, Z.; Ran, S. Digital mapping of soil salinization based on Sentinel-1 and Sentinel-2 data combined
with machine learning algorithms. Reg. Sustain. 2021, 2, 177–188. [CrossRef]
11. Mulder, V.; de Bruin, S.; Schaepman, M.; Mayr, T. The use of remote sensing in soil and terrain mapping—A review. Geoderma
2011, 162, 1–19. [CrossRef]
12. Allbed, A.; Kumar, L.; Aldakheel, Y.Y. Assessing soil salinity using soil salinity and vegetation indices derived from IKONOS
high-spatial resolution imageries: Applications in a date palm dominated region. Geoderma 2014, 230–231, 1–8. [CrossRef]
13. Nguyen, T.G.; Tran, N.A.; Vu, P.L.; Nguyen, Q.-H.; Nguyen, H.D.; Bui, Q.-T. Salinity intrusion prediction using remote sensing and
machine learning in data-limited regions: A case study in Vietnam’s Mekong Delta. Geoderma Reg. 2021, 27, e00424. [CrossRef]
14. Elnaggar, A.A.; Noller, J.S. Application of Remote-sensing Data and Decision-Tree Analysis to Mapping Salt-Affected Soils over
Large Areas. Remote Sens. 2009, 2, 151–165. [CrossRef]
15. Davis, E.; Wang, C.; Dow, K. Comparing Sentinel-2 MSI and Landsat 8 OLI in soil salinity detection: A case study of agricultural
lands in coastal North Carolina. Int. J. Remote Sens. 2019, 40, 6134–6153. [CrossRef]
16. Delavar, M.A.; Naderi, A.; Ghorbani, Y.; Mehrpouyan, A.; Bakhshi, A. Soil salinity mapping by remote sensing south of Urmia
Lake, Iran. Geoderma Reg. 2020, 22, e00317. [CrossRef]
17. Erkin, N.; Zhu, L.; Gu, H.; Tusiyiti, A. Method for predicting soil salinity concentrations in croplands based on machine learning
and remote sensing techniques. J. Appl. Remote Sens. 2019, 13, 034520. [CrossRef]
Remote Sens. 2023, 15, 1751 18 of 20

18. Zheng, Z.; Zhang, F.; Ma, F.; Chai, X.; Zhu, Z.; Shi, J.; Zhang, S. Spatiotemporal changes in soil salinity in a drip-irrigated field.
Geoderma 2009, 149, 243–248. [CrossRef]
19. Gholizadeh, A.; Žižala, D.; Saberioon, M.; Borůvka, L. Soil organic carbon and texture retrieving and mapping using proximal,
airborne and Sentinel-2 spectral imaging. Remote Sens. Environ. 2018, 218, 89–103. [CrossRef]
20. Hengl, T.; De Jesus, J.M.; Heuvelink, G.B.M.; Gonzalez, M.R.; Kilibarda, M.; Blagotić, A.; Shangguan, W.; Wright, M.N.; Geng, X.;
Bauer-Marschallinger, B.; et al. SoilGrids250m: Global gridded soil information based on machine learning. PLoS ONE 2017,
12, e0169748. [CrossRef]
21. Ivushkin, K.; Bartholomeus, H.; Bregt, A.K.; Pulatov, A.; Kempen, B.; de Sousa, L. Global mapping of soil salinity change. Remote
Sens. Environ. 2019, 231, 111260. [CrossRef]
22. Zhang, Z.; Ding, J.; Zhu, C.; Wang, J.; Ma, G.; Ge, X.; Li, Z.; Han, L. Strategies for the efficient estimation of soil organic matter in
salt-affected soils through Vis-NIR spectroscopy: Optimal band combination algorithm and spectral degradation. Geoderma 2020,
382, 114729. [CrossRef]
23. Kalambukattu, J.G.; Kumar, S.; Raj, R.A. Digital soil mapping in a Himalayan watershed using remote sensing and terrain
parameters employing artificial neural network model. Environ. Earth Sci. 2018, 77, 203. [CrossRef]
24. Zhou, T.; Geng, Y.; Chen, J.; Pan, J.; Haase, D.; Lausch, A. High-resolution digital mapping of soil organic carbon and soil total
nitrogen using DEM derivatives, Sentinel-1 and Sentinel-2 data based on machine learning algorithms. Sci. Total Environ. 2020,
729, 138244. [CrossRef]
25. Roelofsen, H.D.; van Bodegom, P.M.; Kooistra, L.; van Amerongen, J.J.; Witte, J.-P.M. An evaluation of remote sensing derived
soil pH and average spring groundwater table for ecological assessments. Int. J. Appl. Earth Obs. Geoinform. 2015, 43, 149–159.
[CrossRef]
26. Dong, W.; Wu, T.; Luo, J.; Sun, Y.; Xia, L. Land parcel-based digital soil mapping of soil nutrient properties in an alluvial-diluvia
plain agricultural area in China. Geoderma 2019, 340, 234–248. [CrossRef]
27. Shen, Q.; Wang, Y.; Wang, X.; Liu, X.; Zhang, X.; Zhang, S. Comparing interpolation methods to predict soil total phosphorus in
the Mollisol area of Northeast China. Catena 2018, 174, 59–72. [CrossRef]
28. Elhag, M. Evaluation of Different Soil Salinity Mapping Using Remote Sensing Techniques in Arid Ecosystems, Saudi Arabia. J.
Sens. 2016, 2016, 7596175. [CrossRef]
29. Bian, L.; Wang, J.; Liu, J.; Han, B. Spatiotemporal Changes of Soil Salinization in the Yellow River Delta of China from 2015 to
2019. Sustainability 2021, 13, 822. [CrossRef]
30. Wu, W.; Mhaimeed, A.S.; Al-Shafie, W.M.; Ziadat, F.; Dhehibi, B.; Nangia, V.; De Pauw, E. Mapping soil salinity changes using
remote sensing in Central Iraq. Geoderma Reg. 2014, 2–3, 21–31. [CrossRef]
31. Arnous, M.O.; El-Rayes, A.; Green, D.R. Hydrosalinity and environmental land degradation assessment of the East Nile Delta
region, Egypt. J. Coast. Conserv. 2015, 19, 491–513. [CrossRef]
32. Elhaddad, A.; Garcia, L. Detecting soil salinity levels in agricultural lands using satellite imagery. In Proceedings of the Ameri-can
Society for Photogrammetry and Remote Sensing Annual Conference, Reno, NV, USA, 1–5 May 2006.
33. Setia, R.; Lewis, M.; Marschner, P.; Segaran, R.R.; Summers, D.; Chittleborough, D. Severity of salinity accurately detected and
classified on a paddock scale with high resolution multispectral satellite imagery. Land Degrad. Dev. 2013, 24, 375–384. [CrossRef]
34. Chen, L.; Xing, M.; He, B.; Wang, J.; Shang, J.; Huang, X.; Xu, M. Estimating Soil Moisture Over Winter Wheat Fields During
Growing Season Using Machine-Learning Methods. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 3706–3718. [CrossRef]
35. Wang, J.; Peng, J.; Li, H.; Yin, C.; Liu, W.; Wang, T.; Zhang, H. Soil Salinity Mapping Using Machine Learning Algorithms with the
Sentinel-2 MSI in Arid Areas, China. Remote Sens. 2021, 13, 305. [CrossRef]
36. Aubert, M.; Baghdadi, N.; Zribi, M.; Douaoui, A.; Loumagne, C.; Baup, F.; El Hajj, M.; Garrigues, S. Analysis of TerraSAR-X data
sensitivity to bare soil moisture, roughness, composition and soil crust. Remote Sens. Environ. 2011, 115, 1801–1810. [CrossRef]
37. Mohamed, E.S.; Ali, A.; El-Shirbeny, M.; Abutaleb, K.; Shaddad, S.M. Mapping soil moisture and their correlation with crop
pattern using remotely sensed data in arid region. Egypt J. Remote Sens. Space Sci. 2020, 23, 347–353. [CrossRef]
38. Zhang, Z.; Ding, J.; Wang, J.; Ge, X. Prediction of soil organic matter in northwestern China using fractional-order derivative
spectroscopy and modified normalized difference indices. Catena 2019, 185, 104257. [CrossRef]
39. Periasamy, S.; Ravi, K.P. A novel approach to quantify soil salinity by simulating the dielectric loss of SAR in three-dimensional
density space. Remote Sens. Environ. 2020, 251, 112059. [CrossRef]
40. Wang, N.; Xue, J.; Peng, J.; Biswas, A.; He, Y.; Shi, Z. Integrating Remote Sensing and Landscape Characteristics to Estimate Soil
Salinity Using Machine Learning Methods: A Case Study from Southern Xinjiang, China. Remote Sens. 2020, 12, 4118. [CrossRef]
41. De Bernardis, C.; Vicente-Guijalba, F.; Martinez-Marin, T.; Lopez-Sanchez, J.M. Contribution to Real-Time Estimation of Crop
Phenological States in a Dynamical Framework Based on NDVI Time Series: Data Fusion with SAR and Temperature. IEEE J. Sel.
Top. Appl. Earth Obs. Remote Sens. 2016, 9, 3512–3523. [CrossRef]
42. Farahmand, N.; Sadeghi, V. Estimating Soil Salinity in the Dried Lake Bed of Urmia Lake Using Optical Sentinel-2 Images and
Nonlinear Regression Models. J. Indian Soc. Remote Sens. 2020, 48, 675–687. [CrossRef]
43. Sahbeni, G. A PLSR model to predict soil salinity using Sentinel-2 MSI data. Open Geosci. 2021, 13, 977–987. [CrossRef]
44. Ravi, K.P.; Periasamy, S. Systematic discrimination of irrigation and upheaval associated salinity using multitemporal SAR data.
Sci. Total Environ. 2021, 790, 148148. [CrossRef]
Remote Sens. 2023, 15, 1751 19 of 20

45. Rudiyanto; Minasny, B.; Setiawan, B.I.; Saptomo, S.K.; McBratney, A.B. Open digital mapping as a cost-effective method for
mapping peat thickness and assessing the carbon stock of tropical peatlands. Geoderma 2018, 313, 25–40. [CrossRef]
46. Tripathi, A.; Tiwari, R.K. Synergetic utilization of sentinel-1 SAR and sentinel-2 optical remote sensing data for surface soil
moisture estimation for Rupnagar, Punjab, India. Geocarto Int. 2020, 37, 2215–2236. [CrossRef]
47. Hoa, P.V.; Giang, N.V.; Binh, N.A.; Hai, L.V.H.; Pham, T.-D.; Hasanlou, M.; Bui, D.T. Soil Salinity Mapping Using SAR Sentinel-1
Data and Advanced Machine Learning Algorithms: A Case Study at Ben Tre Province of the Mekong River Delta (Vietnam).
Remote Sens. 2019, 11, 128. [CrossRef]
48. Pittman, R.C. Improvement of Soil Property Mapping in Northern Ontario’s Great Clay Belt Using Multi-Source Remotely Sensed Data;
YorkSpace: Toronto, ON, Canada, 2020.
49. Song, T.; Wang, Z.; Xie, P.; Han, N.; Jiang, J.; Xu, D. A Novel Dual Path Gated Recurrent Unit Model for Sea Surface Salinity
Prediction. J. Atmos. Ocean. Technol. 2020, 37, 317–325. [CrossRef]
50. Hosseini, M.; Bahrami, H.; Khormali, F.; Khavazi, K.; Mokhtassi-Bidgoli, A. Artificial Intelligence Statistical Analysis of Soil
Respiration Improves Predictions Compared to Regression Methods. J. Soil Sci. Plant Nutr. 2021, 21, 2242–2251. [CrossRef]
51. Schulz, K.; Hänsch, R.; Sörgel, U. Machine learning methods for remote sensing applications: An overview. In Proceedings of the
Volume 10790 SPIE Remote Sensing, Berlin, Germany, 10–13 September 2018; Volume 10790, p. 1079002. [CrossRef]
52. Were, K.; Bui, D.T.; Dick, Ø.B.; Singh, B.R. A comparative assessment of support vector regression, artificial neural networks, and
random forests for predicting and mapping soil organic carbon stocks across an Afromontane landscape. Ecol. Indic. 2015, 52,
394–403. [CrossRef]
53. Achieng, K.O. Modelling of soil moisture retention curve using machine learning techniques: Artificial and deep neural net-works
vs support vector regression models. Comput. Geosci. 2019, 133, 104320. [CrossRef]
54. Sahour, H.; Gholami, V.; Vazifedan, M. A comparative analysis of statistical and machine learning techniques for mapping the
spatial distribution of groundwater salinity in a coastal aquifer. J. Hydrol. 2020, 591, 125321. [CrossRef]
55. Bokde, N.D.; Ali, Z.H.; Al-Hadidi, M.T.; Farooque, A.A.; Jamei, M.; Al Maliki, A.A.; Beyaztas, B.H.; Faris, H.; Yaseen, Z.M. Total
Dissolved Salt Prediction Using Neurocomputing Models: Case Study of Gypsum Soil Within Iraq Region. IEEE Access 2021, 9,
53617–53635. [CrossRef]
56. Muller, S.J.; Van Niekerk, A. An evaluation of supervised classifiers for indirectly detecting salt-affected areas at irrigation scheme
level. Int. J. Appl. Earth Obs. Geoinf. 2016, 49, 138–150. [CrossRef]
57. Alamdar, S.; Ghazban, F.; Zarei, A. Efficiency of Machine Learning Algorithms in Soil Salinity Detection Using LAND-SAT-8 Oli
Imagery. ISPRS Annals of the Photogrammetry. Remote Sens. Spat. Inf. Sci. 2023, 10, 49–55.
58. Xiao, C.; Ji, Q.; Chen, J.; Zhang, F.; Li, Y.; Fan, J.; Hou, X.; Yan, F.; Wang, H. Prediction of soil salinity parameters using machine
learning models in an arid region of northwest China. Comput. Electron. Agric. 2023, 204, 107512. [CrossRef]
59. Mzid, N.; Boussadia, O.; Albrizio, R.; Stellacci, A.M.; Braham, M.; Todorovic, M. Salinity Properties Retrieval from Senti-nel-2
Satellite Data and Machine Learning Algorithms. Agronomy 2023, 13, 716. [CrossRef]
60. Kaya, F.; Schillaci, C.; Keshavarzi, A.; Başayiğit, L. Predictive Mapping of Electrical Conductivity and Assessment of Soil Salinity
in a Western Türkiye Alluvial Plain. Land 2022, 11, 2148. [CrossRef]
61. Taghizadeh-Mehrjardi, R.; Schmidt, K.; Toomanian, N.; Heung, B.; Behrens, T.; Mosavi, A.; Band, S.S.; Amirian-Chakan, A.;
Fathabadi, A.; Scholten, T. Improving the spatial prediction of soil salinity in arid regions using wavelet transformation and
support vector regression models. Geoderma 2020, 383, 114793. [CrossRef]
62. Wang, Z.; Zhang, F.; Zhang, X.; Chan, N.W.; Kung, H.-T.; Ariken, M.; Zhou, X.; Wang, Y. Regional suitability prediction of soil
salinization based on remote-sensing derivatives and optimal spectral index. Sci. Total Environ. 2021, 775, 145807. [CrossRef]
63. Rafik, A.; Ibouh, H.; Fels, A.E.A.E.; Eddahby, L.; Mezzane, D.; Bousfoul, M.; Amazirh, A.; Ouhamdouch, S.; Bahir, M.;
Gourfi, A.; et al. Soil Salinity Detection and Mapping in an Environment under Water Stress between 1984 and 2018 (Case
of the Largest Oasis in Africa-Morocco). Remote Sens. 2022, 14, 1606. [CrossRef]
64. Jiang, X.; Duan, H.; Liao, J.; Guo, P.; Huang, C.; Xue, X. Estimation of Soil Salinization by Machine Learning Algo-rithms in
Different Arid Regions of Northwest China. Remote Sens. 2022, 14, 347. [CrossRef]
65. Merembayev, T.; Amirgaliyev, Y.; Saurov, S.; Wójcik, W. Soil Salinity Classification Using Machine Learning Algo-rithms and
Radar Data in the Case from the South of Kazakhstan. J. Ecol. Eng. 2022, 23, 61–67. [CrossRef]
66. Estévez, V.; Beucher, A.; Mattbäck, S.; Boman, A.; Auri, J.; Björk, K.-M.; Österholm, P. Machine learning techniques for acid sulfate
soil mapping in southeastern Finland. Geoderma 2021, 406, 115446. [CrossRef]
67. Taghizadeh-Mehrjardi, R.; Hamzehpour, N.; Hassanzadeh, M.; Heung, B.; Goydaragh, M.G.; Schmidt, K.; Scholten, T. Enhancing
the accuracy of machine learning models using the super learner technique in digital soil mapping. Geoderma 2021, 399, 115108.
[CrossRef]
68. Nabiollahi, K.; Taghizadeh-Mehrjardi, R.; Shahabi, A.; Heung, B.; Amirian-Chakan, A.; Davari, M.; Scholten, T. Assessing
agricultural salt-affected land using digital soil mapping and hybridized random forests. Geoderma 2020, 385, 114858. [CrossRef]
69. Gharaibeh, M.A.; Albalasmeh, A.A.; Pratt, C.; El Hanandeh, A. Estimation of exchangeable sodium percentage from sodium
adsorption ratio of salt-affected soils using traditional and dilution extracts, saturation percentage, electrical conductivity, and
generalized regression neural networks. Catena 2021, 205, 105466. [CrossRef]
70. Paz, A.M.; Castanheira, N.; Farzamian, M.; Paz, M.C.; Gonçalves, M.C.; Santos, F.A.M.; Triantafilis, J. Prediction of soil sa-linity
and sodicity using electromagnetic conductivity imaging. Geoderma 2020, 361, 114086. [CrossRef]
Remote Sens. 2023, 15, 1751 20 of 20

71. El Bilali, A.; Taleb, A.; Nafii, A.; Alabjah, B.; Mazigh, N. Prediction of sodium adsorption ratio and chloride concentration in a
coastal aquifer under seawater intrusion using machine learning models. Environ. Technol. Innov. 2021, 23, 101641. [CrossRef]
72. El Bilali, A.; Taleb, A.; Brouziyne, Y. Groundwater quality forecasting using machine learning algorithms for irrigation purposes.
Agric. Water Manag. 2020, 245, 106625. [CrossRef]
73. Wang, F.; Yang, S.; Yang, W.; Yang, X.; Jianli, D. Comparison of machine learning algorithms for soil salinity pre-dictions in
three dryland oases located in Xinjiang Uyghur Autonomous Region (XJUAR) of China. Eur. J. Remote Sens. 2019, 52, 256–276.
[CrossRef]
74. Wei, Q.; Nurmemet, I.; Gao, M.; Xie, B. Inversion of Soil Salinity Using Multisource Remote Sensing Data and Parti-cle Swarm
Machine Learning Models in Keriya Oasis, Northwestern China. Remote Sens. 2022, 14, 512. [CrossRef]
75. Abdellatif, A.D.; El Ghonamey, Y.K.; Shoman, M.M. The use of geographic information systems for monitoring some soil
properties: Case study Damanhur District, El-Beheira Governorate—Egypt. Ann. Agric. Sci. 2017, 55, 979–988. Available online:
http://aasj.bu.edu.eg/index.php (accessed on 12 August 2021).
76. Aboukila, E.; Abdelaty, E. Assessment of saturated soil paste salinity from 1:2.5 and 1:5 soil-water extracts for coarse tex-tured
soils. Alex. Sci. Exch. J. 2017, 38, 722–732.
77. Afifi, A.A.; Darwish, K.M. Detection and impact of land encroachment in El-Beheira governorate, Egypt. Model. Earth Syst.
Environ. 2018, 4, 517–526. [CrossRef]
78. Staff, U. Diagnosis and Improvement of Saline and Alkali Soils; Handbook 60; US Department of Agriculture, United States Salinity
Laboratory (USSL): Washington, DC, USA, 1954.
79. Abrol, I.; Yadav, J.S.P.; Massoud, F.I. Salt-Affected Soils and Their Management; Food and Agriculture Organization: Washington,
DC, USA, 1988.
80. Drusch, M.; Del Bello, U.; Carlier, S.; Colin, O.; Fernandez, V.; Gascon, F.; Hoersch, B.; Isola, C.; Laberinti, P.; Martimort, P.; et al.
Sentinel-2: ESA’s Optical High-Resolution Mission for GMES Operational Services. Remote Sens. Environ. 2012, 120, 25–36.
[CrossRef]
81. Spoto, F.; Sy, O.; Laberinti, P.; Martimort, P.; Fernandez, V.; Colin, O.; Hoersch, B.; Meygret, A. Overview of Sentinel-2.
In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Munich, Germany, 22–27 July 2012;
pp. 1707–1710. [CrossRef]
82. Rossi, M.; Chiarito, E.; Cigna, F.; Cuozzo, G.; Fontanelli, G.; Paloscia, S.; Santi, E.; Tapete, D.; Notarnicola, C. Multisensor SAR and
optical estimation of grassland above-ground biomass and LAI: A case study for the Mazia valley in South Tyrol. In Proceedings
of the EGU General Assembly 2021, online, 19–30 April 2021; Volume EGU21-11932. [CrossRef]
83. Haryanto, T.; Pratama, A.; Suhartanto, H.; Murni, A.; Kusmardi, K.; Pidanic, J. Multipatch-GLCM for Texture Feature Extraction
on Classification of the Colon Histopathology Images using Deep Neural Network with GPU Acceleration. J. Comput. Sci. 2020,
16, 280–294. [CrossRef]
84. Corrales, D.C.; Schoving, C.; Raynal, H.; Debaeke, P.; Journet, E.-P.; Constantin, J. A surrogate model based on feature selection
techniques and regression learners to improve soybean yield prediction in southern France. Comput. Electron. Agric. 2022,
192, 106578. [CrossRef]
85. Pearson, K. The fundamental problem of practical statistics. Biometrika 1920, 13, 1–16. [CrossRef]
86. Spearman, C. “General Intelligence” Objectively Determined and Measured. In Studies in Individual Differences: The Search
for Intelligence; Jenkins, J.J., Paterson, D.G., Eds.; Appleton-Century-Crofts: New York, NY, USA, 1961; Volume 15, pp. 59–73.
[CrossRef]
87. Largeron, C.; Christophe, M.; Mathias, G. Entropy based feature selection for text categorization. In Proceedings of the 2011 ACM
Symposium on Applied Computing, Taichung, Taiwan, 21–24 March 2011.
88. Kroehl, M.E. On the Use of Lasso Regression for Mediation Analysis with Application to Microbiota Data; University of Colorado Denver,
Anschutz Medical Campus: Denver, CO, USA, 2014.
89. Tang, J.; Alelyani, S.; Liu, H. Feature selection for classification: A review. In Data Classification: Algorithms and Applications; CRC
Press: Boca Raton, FL, USA, 2014; Volume 37, pp. 37–64. [CrossRef]
90. Guyon, I.; Elisseeff, A. An introduction to variable and feature selection. J. Mach. Learn. Res. 2003, 3, 1157–1182.
91. Kuhn, M.; Johnson, K. Feature Engineering and Selection: A Practical Approach for Predictive Models; CRC Press: Boca Raton, FL,
USA, 2019. [CrossRef]
92. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
93. Tibshirani, R. Regression Shrinkage and Selection Via the Lasso. J. R. Stat. Soc. Ser. B 1996, 58, 267–288. [CrossRef]
94. Liaw, A.; Wiener, M. Classification and regression by randomForest. R News 2002, 2, 18–22.
95. Peng, J.; Biswas, A.; Jiang, Q.; Zhao, R.; Hu, J.; Hu, B.; Shi, Z. Estimating soil salinity from remote sensing and terrain data in
southern Xinjiang Province, China. Geoderma 2018, 337, 1309–1319. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

You might also like