Journal of Hydrology 618 (2023) 129229

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Soil water erosion susceptibility assessment using deep learning algorithms

Khabat Khosravi a, Fatemeh Rezaie b, c, James R. Cooper d, *, Zahra Kalantari e,
Soroush Abolfathi f, Javad Hatamiafkoueieh g, *
a
Department of Earth and Environment, Florida International University, Miami, USA
b
Geoscience Data Center, Korea Institute of Geoscience and Mineral Resources (KIGAM), 124 Gwahak-ro, Yuseong-gu, Daejeon 34132, Republic of Korea
c
Department of Geophysical Exploration, Korea University of Science and Technology, 217 Gajeong-ro, Yuseong-gu, Daejeon 34113, Republic of Korea
d
Department of Geography & Planning, School of Environmental Sciences, University of Liverpool, Liverpool, UK
e
Department of Sustainable Development, Environmental Science and Engineering (SEED), KTH Royal Institute of Technology, Stockholm, Sweden
f
School of Engineering, University of Warwick, CV4 7AL Coventry, UK
g
Department of Mechanics and Control Processes, Academy of Engineering, Peoples’ Friendship University of Russia (RUDN University), Miklukho-Maklaya Str. 6,
Moscow 117198, Russian Federation

A R T I C L E I N F O A B S T R A C T

This manuscript was handled by Marco Borga, Accurate assessment of soil water erosion (SWE) susceptibility is critical for reducing land degradation and soil
Editor-in-Chief, with the assistance of Fran­ loss, and for mitigating the negative impacts of erosion on ecosystem services, water quality, flooding and
cesco Comiti, Associate Editor infrastructure. Deep learning algorithms have been gaining attention in geoscience due to their high performance
and flexibility. However, an understanding of the potential for these algorithms to provide fast, cheap, and
Keywords:
accurate predictions of soil erosion susceptibility is lacking. This study provides the first quantification of this
Soil erosion
potential. Spatial predictions of susceptibility are made using three deep learning algorithms – Convolutional
Deep learning
Land degradation Neural Network (CNN), Recurrent Neural Network (RNN) and Long-Short Term Memory (LSTM) – for an Iranian
CNN catchment that has historically experienced severe water erosion. Through a comparison of their predictive
RNN performance and an analysis of the driving geo-environmental factors, the results reveal: (1) elevation was the
LSTM most effective variable on SWE susceptibility; (2) all three developed models had good prediction performance,
with RNN being marginally the most superior; (3) maps of SWE susceptibility revealed that almost 40 % of the
catchment was highly or very highly susceptible to SWE and 20 % moderately susceptible, indicating the critical
need for soil erosion control in this catchment. Through these algorithms, the soil erosion susceptibility of
catchments can potentially be predicted accurately and with ease using readily available data. Thus, the results
reveal that these models have great potential for use in data poor catchments, such as the one studied here,
especially in developing nations where technical modeling skills and understanding of the erosion processes
occurring in the catchment may be lacking.

1. Introduction erosion, in which particles are detached by raindrop impact, uncon­

centrated flow or concentrated flow, and transported via rainsplash,
Soil water erosion (SWE) is a major cause of global land degradation interrill flow or concentrated flow (Cooper et al., 2012). These processes
and soil loss (Tang et al., 2015) through its negative impact on the cause direct and indirect damage to ecosystems services through soil
organic, physical and chemical characteristics of soils (Aslam et al., nutrient loss, decreasing soil productivity and reducing food production.
2021). Various parts of the world are under the hazard of SWE, and these Thus SWE is often the main driver for a number of global social and
areas are growing in extent as a result of climate and land use change, economic issues such as food insecurity, higher food prices and loss of
with areas extending beyond those of arid and semi-arid regions into biodiversity (Phinzi et al., 2021). Furthermore, SWE can have consid­
humid regions (Boudjemline and Semar, 2018; Wijitkosum, 2021). erable off-site impacts, including decreasing water quality through
Soil is eroded by water through processes which vary in temporal increasing water turbidity, leading to irreparable damage to the aquatic
and spatial scales. These processes are splash, interrill, rill and gully system, reducing dam reservoir capacity, changing river morphology

* Corresponding authors.
E-mail addresses: khabat.khosravi@gmail.com, kkhosrav@fiu.edu (K. Khosravi), rezaie@kigam.re.kr (F. Rezaie), james.cooper@liverpool.ac.uk (J.R. Cooper),
zahrak@kth.se (Z. Kalantari), Soroush.Abolfathi@warwick.ac.uk (S. Abolfathi), khatamiafkuiekh-d@rudn.ru (J. Hatamiafkoueieh).

https://doi.org/10.1016/j.jhydrol.2023.129229
Received 20 September 2022; Received in revised form 13 January 2023; Accepted 2 February 2023
Available online 6 February 2023
0022-1694/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

and increasing flood risk. the physics behind the processes of detachment, transportation and
One country in which these negative impacts of SWE are being felt deposition. These models usually require many variables and detailed
severely is Iran, the focus for this paper. Iran is a semi-arid country, spatial and temporal catchment data for the model build, calibration and
covering a land area of 165 million hectares. The country experiences a validation. Such information is rarely available in developing countries
mean annual soil erosion rate of 25 to 30 ton/ha/year, over 20 times because the monitoring required, especially for large catchments, is
higher than the mean global rate and four times higher than any other costly and time consuming (Conoscenti et al., 2008). For example, in
country (Afshar et al., 2010; Khalili Moghadam et al., 2015; Sadeghi, Iran, many of the hydrometric stations only measure sediment concen­
2017). On average, 1 to 5 million tons of erosion takes places annually tration for a few days during some severe storms, and the temporal
over the country’s land mass (Mohammadi et al., 2021). Thus, although pattern of sediment yield is rarely provided (Darvishan et al., 2010).
50 million hectates of land is available for agriculture, only 1.3 million Furthermore, physically-based simulations are complex, time-
hectares is classed as being suitable for growing crops (Laylin, 2018). consuming and require highly-skilled end users that do not always
SWE is causing 400 million m3 of sedimentation each year in reservoirs, reside in these countries. This issue has limited their use and success in
resulting in an annual volume reduction of 0.5 % in reservoir capacity Iran (Akhavan et al., 2010; Amiri, 2010). Moreover, concerns have been
(Emadodin et al., 2012; Sadeghi, 2017) at an annual cost of US $0.6 raised about their tendency to overestimate small runoff events and
billion (World Bank, 2005). This sedimentation reduces power genera­ underestimate large runoff amounts, even when calibrated (Kinnell,
tion and land irrigation capacity, further reducing the availability of 2017), and the physical basis of many commonly used modes has been
arable lands. As a consequence, these combined impacts of SWE are questioned (Wainwright et al., 2008).
costing the country US $56 to US $112 billion every year, higher than Recently, development of machine learning (ML) approaches have
the revenue generated by oil production (Sadeghi, 2017). opened up new and exciting ways to predict environmental behavior,
Enhanced rates of SWE in Iran are due to the combined effects of including soil erosion susceptibility modeling (e.g., Mosavi et al., 2020;
over-exploitation by communities and industry, and changes in climate Vu et al., 2020). These approaches have a non-linear structure and seek
patterns (Sadeghi, 2017). This expliotation includes improper and un­ to find a robust relationship between input and output readily available
necessary infrastructure development, land use changes and unlawful parameters. ML techniques do not seek to explain the physical processes
exploitation of resources. An intensification in agricultural activities and mathematical reasoning for changes in environmental behavior but
over the past five decades has took place in tandeem with this unsus­ to recognize patterns, both expected and unexpected, within data. These
tainable land use change. From the 1950 s until 2008, around 5 million patterns can highlight environmental relationships in space and time
hectares of forest was converted to farmland, pasture, and urban areas that may unveil critical details about behavior, reveal previously un­
(Emadodin and Bork, 2012). There are also a number of indirect causes suspected relationships, or mitigate uncertainty in estimates. Further­
for these high SWE rates, including an inattention to soil value, a small more, ML models are insensitive to missing data. Thus, these types of
number of hydrometry stations, limited short-term studies of erosion, techniques are at their most beneficial in situations when physically-
unreliable monitoring data, a misunderstanding of the SWE processes based models cannot be applied (e.g. lack of understanding of the un­
and apathy (Sadeghi, 2017). All of these indirect causes have combined derlying physics of the process) or that suffer from inadequacies due to
to mask the severity and intensity of the problem. The consequence is the limitation of data. Therefore, the key advantage of ML approaches is
that soil is lost due to erosion approximately 19 times faster than it forms that some parameters which are difficult or expensive to measure, such
(Emadodin et al., 2012). Nearly $75 million is thought to be needed as soil erosion, could be easily predicted using other readily available
annually in soil improvement projects to reduce SWE (Akbari, 2017). factors, such as rainfall and those gained from satellite data.
Thus, in Iran, like many other countries experiencing the impacts of Artificial Neural Network (ANN) is the oldest and most widely used
SWE, there is a critical need to target soil erosion control to those areas ML technique in geoscience due to its computational efficiency (Abra­
most susceptible, as well to identify areas most suitable for sustainable hart et al., 2012). For example, Ebtehaj and Bonakdari (2013) applied
agricultural development. ANN algorithms to predict sediment transport in sewers, revealing that
Modeling plays a key role in identifying these areas, understanding ANN had a higher prediction power than existing empirical transport
the factors that lead to high erosion susceptibility and in testing the formulas. Similar results have also been found within other areas of
efficacy of soil erosion control strategies. Thus far, three main types of hydrology, such as river suspended load prediction and evaporation
approaches have been used: empirical, physically-based and data- modeling (Kisi et al., 2016; Melesse et al., 2011). However, these algo­
driven/machine learning models (Raza et al., 2021). Empirical rithms have slow coverage speed during the training procedure, high
models, such as the Universal Soil Loss Equation (Wischmeier and errors in the modeling phase, and a low convergence and generalization
Smith, 1965), and its associated family of derivative models (e.g. power (Kisi et al., 2012). Thus, ANN algorithms have poor prediction
Revised USLE (RUSLE; Renard et al., 1997), Modified USLE (MUSLE; power when the range of the testing dataset is outside of the range of the
Williams, 1975), and the modified Mediterranean Desertification and training data (Kisi et al., 2016; Melesse et al., 2011), and they require a
Land Use (MEDALUS; Abuzaid et al., 2021) estimate SWE using a pre­ long-term dataset to achieve a reasonable result. Thus, to solve this
fixed set of physical parameters representing the main factors thought to weakness, ANN algorithms have been ensembled with fuzzy logic al­
affect erosion (Conoscenti et al., 2008). These models have a linear and gorithms to create Adaptive Neural Fuzzy Inference System (ANFIS)
simple structure and do not aim to simulate the physical processes of models. ANFIS models take the advantages from both ANN and fuzzy
SWE. Instead, they are based on producing a mathematical algorithm logic methods to generally produce a better predictive capability (Ebt­
that best describes the relationship between these parameters and ehaj and Bonakdari, 2014) but they are poor at finding the best weight
measured erosion rates. These empirical models can perform well for the parameters which heavily influence prediction accuracy (Tien Bui et al.,
conditions upon which they are calibrated, but typically perform poorly 2016). Furthermore, ANFIS algorithm suffer from the need for a large
for conditions outside those used in calibration (e.g. Cerdan et al., 2010; number of model operators, each of which need to be set accurately,
Kinnell, 2010; Rapp, 1963; Tan et al., 2018; Zhang et al., 1996). especially the weights of membership function. Also ANFIS algorithms
Furthermore, they do not simulate soil deposition (e.g., sedimenta­ lack a systematic approach in the design of fuzzy rules and in the choice
tion) and, in most cases, insufficient measured data exist to rigorously of membership functions variables (Khosravi et al., 2018; Tien Bui et al.,
determine the single factors for all needed situations and scenarios 2016).
(Auerswald et al., 2006; Wischmeier and Smith, 1978, 1965). Another powerful ML technique commonly used is support vector
Physically-based models, such as Watershed Erosion Prediction machine (SVM). This algorithm has been used in a range of applications
Project (WEPP; Nearing et al., 1989) and KINematic runoff and EROSion in hydrology, such as rainfall-runoff forecasting (e.g. Dibike et al., 2001)
(KINEROS; Woolhiser et al., 1990), attempt to mathematically represent and suspended sediment load prediction (Çimen, 2008). The developed

2
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

models have a high prediction power (Ganguli and Reddy, 2014; Goyal 2. Study area
et al., 2014), but suffer from having a lot of hyper-parameters making
model implementation difficult (Ahmad et al., 2018). Noor-Rood catchment is one of the main sub-catchments of the Haraz
Group Method of Data Handling (GMDH) is another powerful and watershed, located southwest of Amol city, Mazandaran province,
flexible neuron-based algorithm. The GMDH algorithm relates to the Northern Iran (Fig. 1; Eastern longitudes 51◦ 26′′ 13′′ to 52◦ 18′′ 21′′ and
deterministic self-organizing method group, where the principle of a Northern latitudes 36◦ 00′′ 58′′ to 36◦ 16′′ 18′′ ). The catchment is fully
black box, connectionism and induction is used (Anastasakis and Mort, mountainous (elevation between 721 and 4333 m), has an area of 1298
2001). Applications include land subsidence susceptibility mapping km2 and experiences a mean annual precipitation of 504 mm (Solaimani
(Panahi et al., 2022) and flood modeling (Dodangeh et al., 2020). and Hadian Amr, 2008). The study area is mostly covered by the central
However the weakness of the GMDH algorithm lies in its fixed config­ Elborz and Shemshak (shale, marl and sandstone) and Karaj (tuff and
uration, using a deterministic approach to find the optimal partition of shale) formations. Historically the catchment has suffered intense SWE
datasets and parameters (Robinson, 1998). area due to flooding, and human-induced processes (e.g., land use
Although these ML models have been applied to model behaviors in a change, over-grazing, intensive agriculture and etc.) causing high
wide range of environmental settings, they all suffer from the need to turbidity in Haraz River, high sedimentation and a decrease in water
determine accurate weights in the membership function and the optimal capacity in the Haraz dam, located downstream of the catchment. Since
values for hyper-parameters. Since trial and error is not possible to the building of large dams in the Haraz watershed are costly and of
determine the exact weights, bio-inspired algorithms have been applied critical importance to irrigation and energy generation, delineating SWE
to optimize these ML techniques. For example, Angileri et al. (2016) prone-areas in the study area is necessary for aiding the location of new
applied the stochastic gradient tree boost (SGT) for SWE modeling in dams and mitigating the impacts of SWE on existing dam capacity.
Italy, revealing excellent reliability accuracy. Sajedi-Hosseini et al.
(2018) applied the Fuzzy Decision-Making Trial and Evaluation Labo­ 3. Methodology
ratory (Fuzzy DEMATEL) approach for SWE modeling and mapping at
Noor-Rood catchment in Iran and found this approach had a reasonable The conceptual framework for modeling SWE susceptibility is shown
prediction power. In the same catchment, Mosavi et al. (2020) utilized in Fig. 2. Four main steps are used:
several machine learning models – weighted subspace random forest
(WSRF), Gaussian process with a radial basis function kernel (Gausspr­ 1. Data collection of SWE historical data;
radial) and Naive Bayes (NB) – for SWE susceptibility mapping. Their 2. Extraction of SWE geo-environmental variables;
results showed the WSRF model outperformed the other models, fol­ 3. Generating SWE susceptibility maps using deep learning CNN, LSTM
lowed by Gaussprradial, and NB. Yousefi et al. (2021) applied three ML and RNN algorithms;
models of Random Forest (RF), classification and regression tree 4. Model evaluation and comparing the results across algorithms using
(CART), and SVM to model land degradation in Alborz Mountains, Iran; the area under the receiver operating characteristic curve (AUC).
RF outperformed the CART and SVM models. Although tree-based
models have higher performance than ANN, ANFIS and SVM models,
trees are sensitive to noisy data (Tien Bui et al., 2012) making them less 3.1. SWE inventory map
suitable to catchments that lack continuous erosion monitoring or where
this monitoring is sparse in spatial extent. Spatial modeling through the machine learning approach has a bi­
Since these studies were conducted, a new type of ML model has been nary classification system (i.e., SWE occurrence/non-occurrence). The
developed, namely deep learning algorithms. These algorithms have a first and vital step in this binary modeling was determining the spatial
greater flexibility than traditional ML models and thus a higher pre­ relationship between SWE historical data and geo-environmental factors
dictive performance (Ghorbanzadeh et al., 2019). Convolutional Neural in order to determine the effectiveness of each on SWE susceptibility.
Network (CNN), Recurrent Neural Network (RNN) and Long-Short Term The historical data was derived from a survey of soil erosion in 2017 by
Memory (LSTM) are amongst the popular deep learning algorithms. Sajedi-Hosseini et al. (2018), in which recordings of sheet, rill, interrill,
Examples of application of the CNN algorithm include flood modeling gully erosion, and mass movements were made at 116 locations, (Fig. 1);
(Khosravi et al., 2020) and landslide susceptibility assessment (Thi Ngo and equally occurrences of no erosion were recorded at 116 locations.
et al., 2021). The RNN algorithm has been applied for landslide sus­
ceptibility assessment (Li et al., 2021), and LSTM for flood modeling 3.2. Construction of the training and testing datasets
(Fang et al., 2021). A significant gap exists in understanding the po­
tential of these deep learning algorithms, and in the identification of the In the next step, values of 0 and 1 were allocated to locations that
most flexible and accurate algorithm for SWE susceptibility prediction. have experienced no SWE and SWE, respectively. Then, all data was split
The present paper, therefore, aims to fill this gap in understanding by into two parts (Chung and Fabbri, 2003) in a ratio of 70:30 for training
achieving the following objectives, using the heavily eroded Noor-Rood and testing, respectively, based on the most commonly used ratio. The
catchment in Iran as a case study: (1) Delineate SWE susceptibility areas first 70 % was used for model development, and the remaining 30 % for
using three deep learning algorithms techniques, namely CNN, RNN and model evaluation. In the final step, the training dataset was overlaid
LSTM; (2) compare the predictive power of these data-driven models; with all the SWE geo-environment conditioning factors to extract their
and (3) perform an analysis of effectiveness of the geo-environmental attribute values for modeling.
driving variables through Information Gain Attribute Evaluator (IGAE)
feature selection. The performance of these particular algorithms is 3.3. SWE geo-environmental factor
tested for the following reasons: (1) a CNN model automatically detects
the important features without any human supervision. (2) RNN can Based on a body of previous work on SWE susceptibility modeling
process inputs of any length; and (3) LSTM that enables a model to (Sajedi-Hosseini et al. 2018; Aslam et al., 2021; Conoscenti et al., 2008;
“remember” past information. The research offers new insight into Mosavi et al., 2020; Rosskopf et al., 2020), the characteristics of the
which deep learning algorithms offer the potential to provide relatively study area and data availability, 14 SWE geo-environmental condi­
cheap and fast predictions of SWE susceptibility in situations when tioning factors were considered (Fig. 3a–n). After initial selection of
understanding of the physical processes at play may not be well un­ these input factors, their effectiveness was investigated. Parameters with
derstood, such as the catchment studied here. a low or null value can cause lower model performance. To identify if
this was the case within the training dataset, an Information Gain

3
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 1. The location of Noor-Rood catchment and the SWE sites used to train and test the deep learning models.

Attribute Evaluator (IGAE) feature selection technique was applied to records can be found in Mosavi et al. (2020). HSG (Table A in Supple­
determine input variable importance. The IGAE technique evaluates the mentary material) and soil texture data was acquired in a shape-file
worth of an attribute by measuring the information gain with respect to format from Mazandaran Regional Water Authority. NDVI and land
the class, as follows: use factors were measured using Landsat 8 OLI/TIRS (Operational Land
Imager/Thermal Infrared Sensor) imagery from June 2016 (Sajedi-
IGAE(Class, Attribute) = H(Class) − H(Class|Attribute) (1)
Hosseini et al., 2018). The geology dataset was obtained from the na­
where H is the information entropy. More information about this tech­ tional geology map of Iran at a scale of 1:100,000 (Table B in Supple­
nique can be found in Novakovic (2009) and Trabelsi et al. (2017). mentary material).
Through the use of this technique, all 14 geo-environmental factors
were considered effective and all considered in the next step of model
3.4. Model description
development. These factors were elevation, ground slope, slope aspect,
plan curvature, topographic wetness index (TWI), stream power index
3.4.1. Convolutional Neural Network (CNN)
(SPI), distance from river, slope length and steepness factor (LS factor),
The CNN method has the structure of ordinary feed-forward neural
rainfall erosivity (RE), hydrologic soil group (HSG), Normalized Dif­
networks, with each hidden layer made of neurons fully connected to all
ference Vegetation Index (NDVI), land use, soil texture and geology.
neurons in the previous and next layer. Neurons in a single layer operate
Ground slope, elevation, slope aspect, LS, plan curvature, TWI, and SPI
completely independently and do not share any connections. Weight
were gained directly from ASTER Global Digital Elevation Model (DEM)
sharing and local connection are two special characteristics of the CNN
of the study area with a 30 m × 30 m resolution (https://earthexplorer.
model, enhancing the network training efficiency and improving its
usgs.gov/). The distance from river factor was determined in ArcGIS
capability to effectively decrease the network’s free parameters (Du
10.2 using a digital river network. The RE factor was calculated from
et al., 2022).
recorded rainfall in the study area using a rainfall gradient approach,
The generalized architecture of the CNN model consists of three main
due to the lack of rainfall intensity data, according to the following
types of layers: convolutional, pooling, and fully-connected layers. In
equation (Yu and Rosewell, 1996):
the convolution layer, a linear operator, such as convolution, or a
RE = 0.0483(P)1.61 (2) nonlinear activation function, such as rectified linear unit (ReLU;
defined as f(x) = max(0, x); Yamashita et al., 2018) separates and
where P is the mean annual precipitation (mm). The mean annual pre­ identifies the various features of the dataset for analysis. These functions
cipitation was estimated based on data recorded from 1976 to 2016 by are able to capture the spatial and temporal dependencies between data.
six rain gauges in the catchment. Further information about the rainfall The primary aim of the pooling layer is to decrease the size of the

4
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 2. Conceptual model development framework.

convolved feature map to reduce the computational costs, and to handle calculated and then sent back to the modeling process as an adjusted/
overfitting issues. This step is performed by decreasing the connections modified input. This approach continues until a fixed and steady output
between layers, independently operating on each feature map. Max, Min is obtained. The mathematical approaches are as follows:
and average pooling are the most common form of pooling layer used to ∑ ∑
reduce data size, whilst still preserving the most important features ai (t) = φji Yi (t) + σji hi (t − 1), j = 1, ..., nH (3)
within them (Pally and Samadi, 2022). For example, max pooling, a
nonlinear downsampling operation, computes the largest value in each hi (t) = F(ai (t)), j = 1, ..., nH (4)
patch of a feature map (Venkatappareddy et al., 2021). Thus, the pooling ∑
layer generalizes the features extracted by the convolution layer, and bi (t) = aji hi (t), j = 1, ..., nj (5)
helps the networks to recognize the features independently. The fully-
connected layer is the final part of CNN model which maps the extrac­ Xi (t) = G(bj (t)), j = 1, ..., nj (6)
ted features into final output using a softmax activation function to
convert the vector of numbers into the vector of probabilities. The where φji , σji and aji are weights, F is defined as the non-linear trans­
output of the function reveals the probability of each class (Li et al., formation parameter, and nH is the hidden number of neurons. Gener­
2021). ally, hidden layers through networks input [i.e., Y(t)] calculate the
adjusted network weights.
3.4.2. Recurrent Neural Network (RNN)
RNN is a class of ANN model, derived from feed-forward networks 3.4.3. Long-Short term memory (LSTM)
(FFN), where connections between nodes form a directed or undirected LSTM is another type of RNN, capable of learning order dependence
graph along a temporal sequence. RNN algorithms are powerful and in sequence prediction problems. This algorithm extends the internal
robust because they use their internal state (memory) to process variable memory and helps the model to recall events from a long period of time.
length sequences of inputs. RNNs are best applied to sequential or time- Standard RNN models, due to the issue of gradient vanishing, are unable
series data, but by extracting the contextual information from the data, to prepare long memories. In contrast, LSTM models are able to train and
RNN models can be efficaciously applied to data classification. The RNN learn from important events with lags of unknown duration between
structure is composed of several successive recurrent layers. A FFN al­ important events in a time series. LSTM uses three ways to allocate the
locates a weight to the input parameters, similar to other deep learning weights: forget the information, let in new information and hold the
models, but the key difference is that RNN algorithms, using the internal information that affects the output. The model also has three main layers
memory, assign this weight to both current and previous inputs. or gates: input (control the input information in a memory cell), output
RNN models benefit from having loops in hidden layers, and these (maintain control over the outgoing information throughout the
loops have a significant effect on the training capability of the model. remainder of the networks) and forget (control the input from the pre­
Feedback loops from the output layer (X1(t), …, Xnj) are evaluated using vious memory and determine if it should be deleted based on the pre­
the networks input (Y(t)). For each new input, the output values are ceding cell condition). This information is calculated through the
following equations:

5
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 3. Spatial variation in SWE geo-environmental factors used for model development, a) elevation, b) aspect, c) slope degree, d) plan curvature, e) NDVI, f) RE, g)
landuse, h) SPI, i) TWI, j) LS, k) distance from river, l) hydrologic soil group, m) soil texture, and n) geology.

Ft = σ (Wxf Xt + Whf Ht− 1 + Bf ) (7)

Ct = Ft *Ct− 1 + It *Ct (10)
It = σ (Wxi Xt + Whi Ht− 1 + Bi ) (8)
Ot = σ(Wxo Xt + Who Ht− 1 + Bo ) (11)

Ct = σ(Wxc Xt + Whc Ht− 1 + Bc ) (9)

6
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 3. (continued).

Ht = ot tanh(Ct ) (12)
Table 1
Interpreting AUC values.
Yt = σ(Why Ht + By ) (13)
AUC quantitative values Quality interpretation
1 0.5–0.6 Unsatisfactory
σ (x) = (14)
0.6–0.7 Satisfactory
1 + exp− x

0.7–0.8 Good
where Xt, Yt, It, Ft, Ot, Ct, Ct and σ are input vector, output vector, input 0.8–0.9 Very good
0.9–1 Excellent
gate, forget gate, output gate, finishing state in memory block, tempo­
rary and sigmoid function respectively. Wxf ,Wxi ,Wxc , and Wxo are input
weight matrices, Whf , Whi , Whc , and Who are recurrent weights matrices, values can be interpreted for classifying model quality.
Why is the output weight and Bf, Bi, Bc, Bo and By are related bias vectors.
4. Result and analysis
3.5. SWE susceptibility map generation
4.1. Effectiveness of input variables on SWE susceptibility
After developing the three models using the training data, they were
used to calculate SWE indices (SWEI) for all pixels in the study area. The effectiveness of the 14 input variables on SWE susceptibility,
These SWEI values were classified based on the quantile classification assessed using the IGAE feature selection technique, is presented in a
scheme to very low, low, moderate, high, and very high to produce a Radar-chart (Fig. 4). This chart shows that elevation was the parameter
SWE susceptibility map. with the highest effectiveness on SWE (0.30), followed by RE (0.16),
NDVI (0.143), TWI (0.142), plan curvature (0.141), ground slope (0.12),
geology (0.099), aspect (0.091), SPI (0.073), distance from river
3.6. Model performance evaluation
(0.070), land use (0.067), soil texture (0.030), LS (0.020), and HSG
(0.016).
The prediction power of each of the developed models was evaluated
quantitatively using the powerful and reliable receiver operating char­
acteristic (ROC) curve method, using Statistical Package for the Social 4.2. SWE susceptibility maps
Sciences (SPSS) software. Model evaluation was assessed during both
model development (success-rate ROC curve) using the training dataset, The delineated SWE susceptibility maps showed the middle and
and during model evaluation using the testing dataset (prediction-rate eastern part of the catchment had a high and very high susceptibility to
ROC curve). Success-rate ROC curves only reveal the efficiency of the SWE, while the rest of the study area, which surrounded the susceptible
built model, whereas prediction-rate ROC curves, using data not used in areas, had a lower susceptibility (Fig. 5a-c).
model development, reveals how good the model is at prediction and According to the ROC model evaluation technique, all models, in
thus shows the model generalization power. Area Under the ROC curve terms of success rate and prediction rate, had a very good prediction
(AUC) measures the entire two-dimensional area underneath the ROC power (0.8 < AUC < 0.9) (Fig. 6). All models had the same performance
curve, providing an aggregate measure of performance across all in the training phase (AUC = 0.85), but RNN was marginally superior in
possible classification thresholds. AUC varies between 0 (the lowest the testing phase (AUC = 0.83). Thus, the RNN model had the highest
model prediction power) and 1 (ideal model with the highest perfor­ generalization power, albeit only slightly.
mance). Table 1, based on Yesilnacar (2005), shows how ranges of AUC Differences in the three maps are subtle. The maps produced by the

7
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

experience less erosion than the lower and mid-elevations. Thus RE had
a poor correlation with SWE (Fig. 3f). Furthermore livestock farming is
denser in low elevations, resulting in overgrazing, a reduction in vege­
tation cover (especially during early spring when immature is
damaged), soil compaction, lower infiltration rates, lower soil retention
and higher runoff, combining to make soil more susceptible to rain­
splash, sheet, rill and gully erosion (Sternberg et al., 2000; Orgill et al.,
2018; Narantsetseg et al., 2018; Donovan and Monaghan; 2021).
Areas with low NDVI, representing soil with little or no vegetation,
were strongly positively correlated with SWE corresponding with higher
susceptibility (Fig. 8). Whereas areas with high NVDI were negatively
correlated because vegetation cover reduced the volume and intensity of
rainfall reaching the soil surface (e.g. Vaezi et al., 2016) and the
detachment of sediment in overland flow. In the long term, vegetation
also increases soil organic matter, and improves soil physical properties
(e.g. Puigdefabregas et al., 1999), further reducing soil erosivity.
High values of TWI were correlated strongly with SWE, since areas
with high TWI can indicate areas where there is an accumulation of
water, possibly due to the routing of overland flow, resulting in a higher
likelihood of particle detachment through rilling and gullying. Flat
Fig. 4. Radar-chart showing variable effectiveness on SWE.
slopes were also associated with areas with high SWE susceptibility,
since, with all things constant (such as soil thickness and vegetation
LSTM and RNN models were most alike, producing similar predictions cover), lower angle slopes are most often associated with lower infil­
for the percentage of catchment occupied by very low, high, very high tration capacities, and thus higher rates of runoff (e.g. Chen and Young,
susceptibility. The CNN model predicted a higher percentage of very 2006). Two lithological units, Pldv (Rhyolitic to rhyodacitic volcanic
high susceptibility areas than the other two models, but a slightly lower rocks) and K2l2 (Thick-bedded to massive limestone (Maastrichtian))
proportion of very low susceptibility (Fig. 7). The CNN model over­ were highly correlated with areas of high SWE, with all other units
estimated the very high susceptibility classes, while underestimated the displaying a weak association (Fig. 8). Slopes with different aspect
moderate classes. The LSTM model overestimated the very low suscep­ receive differing amounts of solar radiation; southerly aspects corre­
tibility classes while underestimated low and moderate susceptibility spond strongly with areas of high SWE, since these areas receive more
classes (Figs. 5 and 7). According to the predictions made by the RNN solar radiation, reducing soil wetness, aggregate stability and vegetation
model, 19.8 % of the study area had a very high SWE susceptibility, 19.7 cover, and more crusting of the soil surface, reducing infiltration ca­
%, 20.8 %, 20.8 %, and 18.9 % had a high, moderate, low and very low pacities and increasing runoff (e.g. Marques and Mora, 1992; Fang and
susceptibility (Fig. 7). Overall, 39.5 % of the catchment had a high and Guo, 2015). All other factors had a poor correlation with SWE
very high susceptibility to SWE. susceptibility.

5. Discussion 5.2. Comparison in model performance

5.1. Effectiveness of geo-environmental factors Since inputs were the same for each of the models, the difference in
performance between algorithms is attributable to their different
To examine the effectiveness of each of the geo-environmental fac­ computational structures. RNN and LSTM models outperformed the
tors, a Frequency Ratio (FR) approach was used. FR is a bivariate sta­ CNN model because they both have an internal memory that allows
tistical approach that calculates the probabilistic relationship between them to ‘remember’ important information about the input they receive,
SWE susceptibility and each of the factors. The FR is the ratio of the area enhancing their precision in predicting what’s coming next. Although
where SWE occurs to the total study area for a given attribute. For each the generalization ability of CNN is enhanced due to the use of a pooling
range or type of factor, FR was calculated as follows: layer, their fitting ability is reduced when the inputs are low-
∑
Di / Ni=1 Di dimensional vectors, leading to information loss (Liu et al., 2019).
FR = ∑ CNN and RNN algorithms operate similarly by introducing sparsity
Ai / Ni=1 Ai
and reusing the same neurons and weights. However in the CNN model
where Di is the area of SWE of the i-th category (very low, low, medium, latent patterns in data are detected using convolution operations,
high, very high), Ai is the area of the i-th category for a certain factor and whereas in the RNN model the specific sequences in input data are found
N is the category number of the factor. A derived FR value of more than 1 by considering the relationship between the current and the previous
indicates a strong and positive relationship between the concerned class state, allowing RNNs to feed results back into the network. Also, in CNN
of the selected factor data layer and high SWE susceptibility; on the models, the input size and resulting output are fixed but in RNN models
other hand, a frequency ratio value of<1 suggests a poor and negative they vary. Further, the RNN algorithm benefits from a parameter sharing
correlation between SWE susceptibility and the concerned class of the technique, allowing RNN models to share parameters across different
factor data layer, and low susceptibility. time steps. In addition, the incorporation of convolutional layers can
The FR values and the elevation map (Fig. 3a) show that areas with stretch its efficiency to nearby pixels.
lower and mid-elevations (727–2404 m) had a strong correlation with The CNN and LSTM algorithms are feed forward neural networks,
the highest SWE susceptibility, particularly due to mass movement. but a backpropagation training algorithm is used in the RNN model
Areas with higher elevations may usually be thought to more suscepti­ providing the ability to more effectively interpret temporal information
ble, due to higher rainfall and steeper slope, thinner soils and bedrock and capture long-term dependencies in data, as well as obtain a mini­
outcrops. However precipitation falls as snow in these higher elevations. mum error function value (Ajitha et al., 2022). When these differences
Since snow has less erosivity than rainfall and can acts as an armor to are taken together, this greater complexity in the RNN computational
protect soil from erosion, these higher elevation areas therefore structure accounts for the enhanced predictive power of RNN models
over CNN models (Mutlu et al., 2019).

8
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 5. SWE susceptibility maps based on the different deep learning models: a) CNN, b) LSTM and c) RNN.

Fig. 6. ROC curves for (a) success rate and (b) prediction rate.

9
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 7. A histogram showing the percentage of SWE classes that fall into the susceptibility class for each deep learning model.

A direct comparison between our findings with those of previous potential for use in SWE susceptibility assessment in data poor catch­
studies is not possible as deep learning algorithms have not yet to be ments, especially in developing nations where technical modeling skills
applied for SWE modeling and mapping. Several studies have applied and understanding of the hydrologic and erosion processes occurring in
different ML models for SWE mapping however. Yousefi et al. (2021) the catchment may be lacking.
applied three models of RF, CART and SVM model for SWE susceptibility The major disadvantages of these types of models however are
mapping in a different catchment in Iran and reported these models have twofold. First, like all statistical methods, the developed models only
accuracies of between 0.89 and 0.96 based on the ROC-AUC method. relate directly to the catchments being considered, and thus their
These values are higher than reported in this study, despite the use of application to other catchments may prove inappropriate. Future studies
more powerful algorithms, resulting from the greater complexity in the should apply the developed models to catchments with differing rainfall,
relationships between the geo-environmental factors and erosion pro­ soils, land use, land cover, geology and morphology to discover whether
cesses in the catchment studied in the present paper. In particular, snow this is the case. Second, due to their ‘black-box’ structure, they provide
in mountainous areas, high livestock density and deforestation in low­ poor explanatory power, and thus are unable to extract understanding of
land areas, and little variation in soil grain size, resulted in some factors, the processes that cause changes in SWE susceptibility. With these
such as RE, land use and soil texture, that would normally be well considerations in mind, the use of deep learning models may not simply
correlated with SWE to be not well correlated, and there was an unex­ lie in predicting erosion, but integrating these techniques into process-
pected relation with slope elevation. In the same catchment, Sajedi- based models to help identify and optimize model parameters and
Hosseini et al. (2018) applied a Fuzzy DEMATEL approach to SWE mitigate uncertainty in model estimates (e.g., Vojinovic et al., 2013),
mapping with an accuracy of 0.83. Although their method has higher help recognize patterns within satellite data to unveil critical details
performance, the generated map had higher uncertainty, because the about behavior, and possibly reveal new environmental relationships.
weights were calculated based on the expert opinion. In contrast, Mosavi Future studies should seek to explore this potential.
et al. (2020) applied WSRF, Gaussprradial, and NB methods for SWE in This study has considered geo-environmental factors that affect
the same study area and found predictive accuracy was lower; WSRF erosion. Where data is available, future studies should consider human-
with Kappa index of 0.81 had a higher performance than Gaussprradial related factors in deep learning models, such as livestock density (e.g.,
(0.76) and NB with Kappa (0.71). Also Abolhasani et al. (2022) imple­ Evans, 1997), arable yields (e.g., Gliessman, 2004) the location of road-
mented RF, boosted regression tree (BRT), SVM, and classification and networks (e.g., Deng et al., 2011; Keshkamat et al., 2013) and distance
regression tree (CART) algorithms, with AUC values of 0.81, 0.76, 0.71 from water resources (e.g., Mirzabaev et al., 2016). For example, Yousefi
and 0.63 for SWE mapping at Qazvin Plain, Iran. The three models et al. (2016) reported that the rate of soil degradation around water
developed in the current study performed more strongly. resources was higher than in other parts more distant from water sup­
plies. Deng et al. (2011) demonstrated that when areas are composed of
5.3. Applying deep learning models to map SWE susceptibility relatively high quality grassland, roads lead to soil erosion and degra­
dation, whereas when grassland resources are sparse, access to a road
The choice of the ‘best’ predictive model is most often a compromise results in soil restoration. In addition, due to a lack of rainfall intensity
between model prediction accuracy and model complexity, with the data, the current study had to take a simplified approach to account for
later, in deep learning models, most closely related to the data input rainfall as a factor, using mean annual rainfall data to generate the RE
requirements. The major advantages of the deep learning models factor. This approach does not account for the role of individual storm
developed in this paper are their simplicity, and their ease and in events, and their magnitude and duration, on soil erosion. Nor does it
expense to build and run, unlike physically-based models, whilst consider how preceding events affect susceptibility in proceeding
providing little compromise on model performance. In other words, the events. Thus a focus for future studies should be on applying deep-
deep learning models provided good prediction performance based on learning models to more data-rich catchments, using observations of
inputs that are readily available from satellite imagery and national rainfall intensity to generate the RE factor with higher accuracy
rainfall monitoring without the need for catchment or hillslope-scale (Capolongo et al. 2008; Panagos et al. 2015; Petroselli et al., 2021).
monitoring. Thus, the results reveal that these models have great Since deep learning models are data-driven, one can postulate that with

10
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Fig. 8. Effectiveness of each of the classes of the most important factors geo-environmental factors on SWE susceptibility, based on FR method.

hourly rainfall data incorporated with the kind of satellite data used in Convolutional Neural Network (CNN), Recurrent Neural Network (RNN)
this current study, it should be possible to build more powerful, more and Long-Short Term Memory (LSTM) – were applied to assess suscep­
accurate models of soil erosion susceptibility over large spatial extents, tibility in an Iranian catchment that has historically experienced severe
providing an important tool for targeting soil erosion control to those erosion. The main findings were as follows:
areas most susceptible, and for identifying areas most suitable for sus­
tainable agricultural development. 1. Elevation was the most effective geo-environmental variable on SWE
susceptibility, followed by rainfall erosivity, normalized difference
6. Conclusion vegetation index, topographic wetness index, plan curvature, ground
slope, geology, aspect, stream power index, distance from river, land
Soil erosion by water is a major cause of global land degradation and use, soil texture, slope length and steepness factor, and hydrologic
soil loss. Accurate predictions of erosion susceptibility are critical for soil groups
protecting soils and targeting efforts to mitigate the impacts of erosion 2. Model evaluation revealed that all three developed models had good
on ecosystem services, water quality, flooding and infrastructure. Using prediction performance, with RNN being marginally the most
satellite data, rainfall, soil and other several readily available data, this superior.
paper has quantified for the first time, the potential of deep learning 3. Maps of SWE susceptibility revealed that almost 40 % of the catch­
models to provide accurate predictions of soil water erosion (SWE) ment was considered to be highly or very highly susceptible to SWE
susceptibility. Three state-of-the-art deep learning algorithms –

11
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

and 20 % moderately susceptible, indicating the critical need for soil complex hillslope in western Iran. J. Environ. Radioact. 101, 606–614. https://doi.
org/10.1016/j.jenvrad.2010.03.008.
erosion control in this catchment to reduce susceptibility.
Ahmad, M.W., Reynolds, J., Rezgui, Y., 2018. Predictive modelling for solar thermal
4. SWE mainly occurs in areas with elevations lower than 2700 m, bare energy systems: A comparison of support vector regression, random forest, extra
soils, high TWI, flat curvature, a southerly aspect, and rhyolitic to trees and regression trees. J. Clean. Prod. 203, 810–821. https://doi.org/10.1016/j.
rhyodacitic volcanic rocks jclepro.2018.08.207.
Ajitha, A., Goel, M., Assudani, M., Radhika, S., Goel, S., 2022. Design and development of
Residential Sector Load Prediction model during COVID-19 Pandemic using LSTM
The strength of these algorithms lies in their ease to implement, use based RNN. Electr. Power Syst. Res. 212, 108635 https://doi.org/10.1016/j.
of readily available satellite and rainfall data, and being inexpensive to epsr.2022.108635.
Akbari, A.-M., 2017. Soil erosion in Iran 2.5 times the world average [WWW Document].
build and run in comparison to physically-based models, whilst Tehran Times. URL https://www.tehrantimes.com/news/418381/Soil-erosion-in-
providing little compromise on model performance. Together, these Iran-2-5-times-the-world-average#:~:text=TEHRAN — Iran has a mean, deputy
findings reveal that deep learning models have great potential for use in agriculture minister has said.
Akhavan, S., Abedi-Koupai, J., Mousavi, S.-F., Afyuni, M., Eslamian, S.-S., Abbaspour, K.
SWE susceptibility assessment, especially in situations when under­ C., 2010. Application of SWAT model to investigate nitrate leaching in Hamadan-
standing of the physical processes at play may not be well understood or Bahar Watershed. Iran. Agric. Ecosyst. Environ. 139, 675–688. https://doi.org/
field monitoring data is lacking. Thus, understanding more about this 10.1016/j.agee.2010.10.015.
Amiri, F., 2010. Estimate of Erosion and Sedimentation in Semi-arid Basin using
potential for different catchments and input variables represents a vital Empirical Models of Erosion Potential within a Geographic Information System. Air.
research avenue for hydrologists. Soil Water Res. 3, ASWR.S3427. https://doi.org/10.4137/ASWR.S3427.
Anastasakis, L., Mort, N., 2001. The Development of Self-Organization Techniques in
Modelling: A Review of the Group Method of Data Handling (GMDH). United
Funding Kingdom.
Angileri, S.E., Conoscenti, C., Hochschild, V., Märker, M., Rotigliano, E., Agnesi, V.,
This publication has been supported by the RUDN University Sci­ 2016. Water erosion susceptibility mapping by applying Stochastic Gradient
Treeboost to the Imera Meridionale River Basin (Sicily, Italy). Geomorphology 262,
entific Projects Grants System, project No 202235-2-000. In addition, 61–76. https://doi.org/10.1016/j.geomorph.2016.03.018.
James Cooper was partially supported by two UK Natural Environment Aslam, B., Maqsoom, A., Salah Alaloul, W., Ali Musarat, M., Jabbar, T., Zafar, A., 2021.
Research Council grants (NE/S01697X/1 and NE/V008404/1). Soil erosion susceptibility mapping using a GIS-based multi-criteria decision
approach: Case of district Chitral, Pakistan. Ain Shams Eng. J. 12, 1637–1649.
https://doi.org/10.1016/j.asej.2020.09.015.
CRediT authorship contribution statement Auerswald, K., Kainz, M., Fiener, P., 2006. Soil erosion potential of organic versus
conventional farming evaluated by USLE modelling of cropping statistics for
agricultural districts in Bavaria. Soil Use Manag. 19, 305–311. https://doi.org/
Khabat Khosravi: Conceptualization, Methods, Software, Write the 10.1111/j.1475-2743.2003.tb00320.x.
manuscript, Review and editing. Fatemeh Rezaie: Methods, Software. Boudjemline, F., Semar, A., 2018. Assessment and mapping of desertification sensitivity
James R. Cooper: Write the manuscript, Review and editing. Zahra with MEDALUS model and GIS – Case study: basin of Hodna, Algeria. J. Water L.
Dev. 36, 17–26. https://doi.org/10.2478/jwld-2018-0002.
Kalantari: Write the manuscript. Soroush Abolfathi: Write the Capolongo, D., Diodato, N., Mannaerts, C.M., Piccarreta, M., Strobl, R.O., 2008.
manuscript. Javad Hatamiafkoueieh: Write the manuscript. Analyzing temporal changes in climate erosivity using a simplified rainfall erosivity
model in Basilicata (southern Italy). J. Hydrol. 356 (1–2), 119–130.
Cerdan, O., Govers, G., Le Bissonnais, Y., Van Oost, K., Poesen, J., Saby, N., Gobin, A.,
Declaration of Competing Interest Vacca, A., Quinton, J., Auerswald, K., Klik, A., Kwaad, F.J.P.M., Raclot, D., Ionita, I.,
Rejman, J., Rousseva, S., Muxart, T., Roxo, M.J., Dostal, T., 2010. Rates and spatial
variations of soil erosion in Europe: A study based on erosion plot data.
The authors declare that they have no known competing financial Geomorphology 122, 167–177. https://doi.org/10.1016/j.geomorph.2010.06.011.
interests or personal relationships that could have appeared to influence Chen, L., Young, M., 2006. Green-Ampt infiltration model for sloping surfaces. Water
Resour. Res. 42 (7) https://doi.org/10.1029/2005WR004468.
the work reported in this paper. Chung, C.-J.-F., Fabbri, A.G., 2003. Validation of Spatial Prediction Models for Landslide
Hazard Mapping. Nat. Hazards 30, 451–472. https://doi.org/10.1023/B:
Data availability NHAZ.0000007172.62651.2b.
Çimen, M., 2008. Estimation of daily suspended sediments using support vector
machines. Hydrol. Sci. J. 53, 656–666. https://doi.org/10.1623/hysj.53.3.656.
Data will be made available on request. Conoscenti, C., Di Maggio, C., Rotigliano, E., 2008. Soil erosion susceptibility assessment
and validation using a geostatistical multivariate approach: a test in Southern Sicily.
Nat. Hazards 46, 287–305. https://doi.org/10.1007/s11069-007-9188-0.
Acknowledgment Cooper, J.R., Wainwright, J., Parsons, A.J., Onda, Y., Fukuwara, T., Obana, E.,
Kitchener, B., Long, E.J., Hargrave, G.H., 2012. A new approach for simulating the
The authors wish to thank Dr. Bahram Choubin for sharing the SWE redistribution of soil particles by water erosion: A marker-in-cell model. J. Geophys.
Res. 117 (F4), n/a–n/a.
dataset. Darvishan, A., Sadeghi, S., Gholami, L., 2010. Efficacy of Time-Area Method in
simulating temporal variation of sediment yield in Chehelgazi watershed, Iran. Ann.
Warsaw Univ. Life Sci. - SGGW. L. Reclam. 42, 51–60. https://doi.org/10.2478/
Appendix A. Supplementary data
v10060-008-0064-8.
Deng, X., Huang, J., Huang, Q., Rozelle, S., Gibson, J., 2011. Do roads lead to grassland
Supplementary data to this article can be found online at https://doi. degradation or restoration? A case study in Inner Mongolia, China. Environ. Dev.
org/10.1016/j.jhydrol.2023.129229. Econ. 16, 751–773. https://doi.org/10.1017/S1355770X11000180.
Dibike, Y.B., Velickov, S., Solomatine, D., Abbott, M.B., 2001. Model Induction with
Support Vector Machines: Introduction and Applications. J. Comput. Civ. Eng. 15,
References 208–216. https://doi.org/10.1061/(ASCE)0887-3801(2001)15:3(208).
Dodangeh, E., Panahi, M., Rezaie, F., Lee, S., Tien Bui, D., Lee, C.-W., Pradhan, B., 2020.
Novel hybrid intelligence models for flood-susceptibility prediction: Meta
Abolhasani, A., Zehtabian, G., Khosravi, H., Rahmati, O., Alamdarloo, E.H.,
optimization of the GMDH and SVR models with the genetic algorithm and harmony
D’Odorico, P., 2022. A new conceptual framework for spatial predictive modelling of
search. J. Hydrol. 590, 125423 https://doi.org/10.1016/j.jhydrol.2020.125423.
land degradation in a semiarid area. L. Degrad. Dev. 33 (17), 3358–3374.
Donovan, M., Monaghan, R., 2021. Impacts of grazing on ground cover, soil physical
Abrahart, R.J., Anctil, F., Coulibaly, P., Dawson, C.W., Mount, N.J., See, L.M.,
properties and soil loss via surface erosion: A novel geospatial modelling approach.
Shamseldin, A.Y., Solomatine, D.P., Toth, E., Wilby, R.L., 2012. Two decades of
J. Environ. Manage. 287, 112206 https://doi.org/10.1016/j.jenvman.2021.112206.
anarchy? Emerging themes and outstanding challenges for neural network river
Du, R., Liu, W., Fu, X., Meng, L., Liu, Z., 2022. Random noise attenuation via
forecasting. Prog. Phys. Geogr. Earth Environ. 36, 480–513. https://doi.org/
convolutional neural network in seismic datasets. Alexandria Eng. J. 61, 9901–9909.
10.1177/0309133312444943.
https://doi.org/10.1016/j.aej.2022.03.008.
Abuzaid, A.S., AbdelRahman, M.A.E., Fadl, M.E., Scopa, A., 2021. Land Degradation
Ebtehaj, I., Bonakdari, H., 2013. Evaluation of Sediment Transport in Sewer using
Vulnerability Mapping in a Newly-Reclaimed Desert Oasis in a Hyper-Arid Agro-
Artificial Neural Network. Eng. Appl. Comput. Fluid Mech. 7, 382–392. https://doi.
Ecosystem Using AHP and Geospatial Techniques. Agronomy 11, 1426. https://doi.
org/10.1080/19942060.2013.11015479.
org/10.3390/agronomy11071426.
Afshar, F.A., Ayoubi, S., Jalalian, A., 2010. Soil redistribution rate and its relationship
with soil organic carbon and total nitrogen using 137Cs technique in a cultivated

12
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Ebtehaj, I., Bonakdari, H., 2014. Performance Evaluation of Adaptive Neural Fuzzy Narantsetseg, A., Kang, S., Ko, D., 2018. Livestock grazing and trampling effects on plant
Inference System for Sediment Transport in Sewers. Water Resour. Manag. 28, functional composition at three wells in the desert steppe of Mongolia. J. Ecol.
4765–4779. https://doi.org/10.1007/s11269-014-0774-0. Environ. 42, 13. https://doi.org/10.1186/s41610-018-0075-2.
Emadodin, I., Bork, H.R., 2012. Degradation of soils as a result of long-term human- Nearing, M.A., Foster, G.R., Lane, L.J., Finkner, S.C., 1989. A process-based soil erosion
induced transformation of the environment in Iran: an overview. J. Land Use Sci. 7, model for USDA-water erosion prediction project technology. Trans. ASAE 32,
203–219. https://doi.org/10.1080/1747423X.2011.560292. 1587–1593.
Emadodin, I., Narita, D., Bork, H.R., 2012. Soil degradation and agricultural Novakovic, J., 2009. Using Information Gain Attribute Evaluation to Classify Sonar
sustainability: an overview from Iran. Environ. Dev. Sustain. 14, 611–625. https:// Targets. In: 17th Telecommunications Forum TELFOR 2009. Serbia, Belgrade,
doi.org/10.1007/s10668-012-9351-y. pp. 1351–1354.
Evans, R., 1997. Soil erosion in the UK initiated by grazing animals. Appl. Geogr. 17, Orgill, S.E., Condon, J.R., Conyers, M.K., Morris, S.G., Alcock, D.J., Murphy, B.W.,
127–141. https://doi.org/10.1016/S0143-6228(97)00002-7. Greene, R.S.B., 2018. Removing Grazing Pressure from a Native Pasture Decreases
Fang, H., Guo, M., 2015. Aspect-induced differences in soil erosion intensity in a gullied Soil Organic Carbon in Southern New South Wales, Australia. L. Degrad. Dev. 29,
hilly region on the Chinese Loess Plateau. Environ. Earth Sci. 74, 5677–5685. 274–283. https://doi.org/10.1002/ldr.2560.
Fang, Z., Wang, Y., Peng, L., Hong, H., 2021. Predicting flood susceptibility using LSTM Pally, R.J., Samadi, S., 2022. Application of image processing and convolutional neural
neural networks. J. Hydrol. 594, 125734 https://doi.org/10.1016/j. networks for flood image classification and semantic segmentation. Environ. Model.
jhydrol.2020.125734. Softw. 148, 105285 https://doi.org/10.1016/j.envsoft.2021.105285.
Ganguli, P., Reddy, M.J., 2014. Ensemble prediction of regional droughts using climate Panagos, P., Ballabio, C., Borrelli, P., Meusburger, K., Klik, A., Rousseva, S., Tadić, M.P.,
inputs and the SVM-copula approach. Hydrol. Process. 28, 4989–5009. https://doi. Michaelides, S., Hrabalíková, M., Olsen, P., Aalto, J., Lakatos, M., Rymszewicz, A.,
org/10.1002/hyp.9966. Dumitrescu, A., Beguería, S., Alewell, C., 2015. Rainfall erosivity in Europe. Sci.
Ghorbanzadeh, O., Meena, S.R., Blaschke, T., Aryal, J., 2019. UAV-based slope failure Total Environ. 511, 801–814.
detection using deep-learning convolutional neural networks. Remote Sens. 11, Panahi, M., Khosravi, K., Golkarian, A., Roostaei, M., Barzegar, R., Omidvar, E.,
2046. https://doi.org/10.3390/rs11172046. Rezaie, F., Saco, P.M., Sharifi, A., Jun, C., Bateni, S.M., Lee, C.-W., Lee, S., 2022.
Gliessman, S.R., 2004. Integrating Agroecological Processes into Cropping Systems A country-wide assessment of Iran’s land subsidence susceptibility using satellite-
Research. J. Crop Improv. 11, 61–80. https://doi.org/10.1300/J411v11n01_04. based InSAR and machine learning. Geocarto Int. 1–23 https://doi.org/10.1080/
Goyal, M.K., Bharti, B., Quilty, J., Adamowski, J., Pandey, A., 2014. Modeling of daily 10106049.2022.2086631.
pan evaporation in sub tropical climates using ANN, LS-SVR, Fuzzy Logic, and Petroselli, A., Apollonio, C., Luca, D.L., Salvaneschi, P., Pecci, M., Marras, T.,
ANFIS. Expert Syst. Appl. 41, 5267–5276. https://doi.org/10.1016/j. Schirone, B., 2021. Comparative Evaluation of the Rainfall Erosivity in the Rieti
eswa.2014.02.047. Province, Central Italy, Using Empirical Formulas and a Stochastic Rainfall
Keshkamat, S.S., Tsendbazar, N.E., Zuidgeest, M.H.P., Shiirev-Adiya, S., van der Veen, A., Generator. Hydrology 8 (4), 171. https://doi.org/10.3390/hydrology8040171.
van Maarseveen, M.F.A.M., 2013. Understanding transportation-caused rangeland Phinzi, K., Ngetar, N.S., Ebhuoma, O., 2021. Soil erosion risk assessment in the
damage in Mongolia. J. Environ. Manage. 114, 433–444. https://doi.org/10.1016/j. Umzintlava catchment (T32E), Eastern Cape, South Africa, using RUSLE and random
jenvman.2012.10.043. forest algorithm. South Afr. Geogr. J. 103, 139–162. https://doi.org/10.1080/
Khalili Moghadam, B., Jabarifar, M., Bagheri, M., Shahbazi, E., 2015. Effects of land use 03736245.2020.1716838.
change on soil splash erosion in the semi-arid region of Iran. Geoderma 241–242, Puigdefabregas, J., Sole, A., Gutierrez, L., del Barrio, G., Boer, M., 1999. Scales and
210–220. https://doi.org/10.1016/j.geoderma.2014.11.025. processes of water and sediment redistribution in drylands: results from the Rambla
Khosravi, K., Panahi, M., Tien Bui, D., 2018. Spatial prediction of groundwater spring Honda field site in Southeast Spain. Earth-Sci. Rev. 48, 39–70. https://doi.org/
potential mapping based on an adaptive neuro-fuzzy inference system and 10.1016/S0012-8252(99)00046-X.
metaheuristic optimization. Hydrol. Earth Syst. Sci. 22, 4771–4792. https://doi.org/ Rapp, J., 1963. Error assessment of the revised universal soil loss equation using natural
10.5194/hess-22-4771-2018. runoff plot data. University of Arizona.
Khosravi, K., Panahi, M., Golkarian, A., Keesstra, S.D., Saco, P.M., Bui, D.T., Lee, S., Raza, A., Ahrends, H., Habib-Ur-Rahman, M., Gaiser, T., 2021. Modeling Approaches to
2020. Convolutional neural network approach for spatial prediction of flood hazard Assess Soil Erosion by Water at the Field Scale with Special Emphasis on
at national scale of Iran. J. Hydrol. 591, 125552 https://doi.org/10.1016/j. Heterogeneity of Soils and Crops. Land 10, 422. https://doi.org/10.3390/
jhydrol.2020.125552. land10040422.
Kinnell, P.I.A., 2010. Event soil loss, runoff and the Universal Soil Loss Equation family of Renard, K.G., Foster, G.R., Weesies, G.A., Mccool, D.K., Yoder, D.C., 1997. Predicting soil
models: A review. J. Hydrol. 385, 384–397. https://doi.org/10.1016/j. erosion by water: a guide to conservation planning with the revised universal soil
jhydrol.2010.01.024. loss equation (RUSLE).
Kinnell, P.I.A., 2017. A comparison of the abilities of the USLE-M, RUSLE2 and WEPP to Robinson, C., 1998. Multi-objective optimisation of polynomial models for time series
model event erosion from bare fallow areas. Sci. Total Environ. 596–597, 32–42. prediction using genetic algorithms and neural networks. University of Sheffield.
https://doi.org/10.1016/j.scitotenv.2017.04.046. Rosskopf, C.M., Di Iorio, E., Circelli, L., Colombo, C., Aucelli, P.P.C., 2020. Assessing
Kisi, O., Dailr, A.H., Cimen, M., Shiri, J., 2012. Suspended sediment modeling using spatial variability and erosion susceptibility of soils in hilly agricultural areas in
genetic programming and soft computing techniques. J. Hydrol. 450–451, 48–58. Southern Italy. Int. Soil Water Conserv. Res. 8, 354–362. https://doi.org/10.1016/j.
https://doi.org/10.1016/j.jhydrol.2012.05.031. iswcr.2020.09.005.
Kisi, O., Genc, O., Dinc, S., Zounemat-Kermani, M., 2016. Daily pan evaporation Sadeghi, S.H.R., 2017. Soil erosion in Iran: State of the art, tendency and solutions.
modeling using chi-squared automatic interaction detector, neural networks, J. Agric. For. 63 https://doi.org/10.17707/AgricultForest.63.3.04.
classification and regression tree. Comput. Electron. Agric. 122, 112–117. https:// Sajedi-Hosseini, F., Choubin, B., Solaimani, K., Cerdà, A., Kavian, A., 2018. Spatial
doi.org/10.1016/j.compag.2016.01.026. prediction of soil erosion susceptibility using a fuzzy analytical network process:
Laylin, D., 2018. Environmental and wildlife degradation in Iran [WWW Document]. Atl. Application of the fuzzy decision making trial and evaluation laboratory approach.
Counc. URL https://www.atlanticcouncil.org/in-depth-research-reports/issue-brief/ L. Degrad. Dev. 29, 3092–3103. https://doi.org/10.1002/ldr.3058.
environmental-and-wildlife-degradation-in-iran-2/. Solaimani, K., Hadian Amr, M.A., 2008. Application of IRS-1D Data in Water Erosion
Li, H., Xu, Q., He, Y., Fan, X., Yang, H., Li, S., 2021. Temporal detection of sharp Features Detection (Case Study: Nour Roud Catchment, Iran). Pakistan J. Biol. Sci.
landslide deformation with ensemble-based LSTM-RNNs and Hurst exponent. 11, 1893–1900. https://doi.org/10.3923/pjbs.2008.1893.1900.
Geomatics. Nat. Hazards Risk 12, 3089–3113. https://doi.org/10.1080/ Sternberg, M., Gutman, M., Perevolotsky, A., Ungar, E.D., Kigel, J., 2000. Vegetation
19475705.2021.1994474. response to grazing management in a Mediterranean herbaceous community: a
Liu, H., Lang, B., Liu, M., Yan, H., 2019. CNN and RNN based payload classification functional group approach. J. Appl. Ecol. 37, 224–237. https://doi.org/10.1046/
methods for attack detection. Knowl.-Based Syst. 163, 332–341. https://doi.org/ j.1365-2664.2000.00491.x.
10.1016/j.knosys.2018.08.036. Tan, Z., Leung, L.R., Li, H., Tesfa, T., 2018. Modeling Sediment Yield in Land Surface and
Marques, M.A., Mora, E., 1992. The influence of aspect on runoff and soil loss in a Earth System Models: Model Comparison, Development, and Evaluation. J. Adv.
Mediterranean burnt forest (Spain). CATENA 19 (3–4), 333–344. Model. Earth Syst. 10, 2192–2213. https://doi.org/10.1029/2017MS001270.
Melesse, A.M., Ahmad, S., McClain, M.E., Wang, X., Lim, Y.H., 2011. Suspended Tang, Q., Xu, Y., Bennett, S.J., Li, Y., 2015. Assessment of soil erosion using RUSLE and
sediment load prediction of river systems: An artificial neural network approach. GIS: a case study of the Yangou watershed in the Loess Plateau, China. Environ.
Agric. Water Manag. 98, 855–866. https://doi.org/10.1016/j.agwat.2010.12.012. Earth Sci. 73, 1715–1724. https://doi.org/10.1007/s12665-014-3523-z.
Mirzabaev, A., Ahmed, M., Werner, J., Pender, J., Louhaichi, M., 2016. Rangelands of Thi Ngo, P.T., Panahi, M., Khosravi, K., Ghorbanzadeh, O., Kariminejad, N., Cerda, A.,
Central Asia: challenges and opportunities. J. Arid Land 8, 93–108. https://doi.org/ Lee, S., 2021. Evaluation of deep learning algorithms for national scale landslide
10.1007/s40333-015-0057-5. susceptibility mapping of Iran. Geosci. Front. 12, 505–519. https://doi.org/
Mohammadi, S., Balouei, F., Haji, K., Khaledi Darvishan, A., Karydas, C.G., 2021. 10.1016/j.gsf.2020.06.013.
Country-scale spatio-temporal monitoring of soil erosion in Iran using the G2 model. Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., 2012. Landslide Susceptibility
Int. J. Digit. Earth 14, 1019–1039. https://doi.org/10.1080/ Assessment in Vietnam Using Support Vector Machines, Decision Tree, and Naïve
17538947.2021.1919230. Bayes Models. Math. Probl. Eng. 2012, 1–26. https://doi.org/10.1155/2012/
Mosavi, A., Sajedi-Hosseini, F., Choubin, B., Taromideh, F., Rahi, G., Dineva, A., 2020. 974638.
Susceptibility Mapping of Soil Water Erosion Using Machine Learning Models. Water Tien Bui, D., Pradhan, B., Nampak, H., Bui, Q.T., Tran, Q.A., Nguyen, Q.P., 2016. Hybrid
12, 1995. https://doi.org/10.3390/w12071995. artificial intelligence approach based on neural fuzzy inference model and
Mutlu, B., Nefeslioglu, H.A., Sezer, E.A., Akcayol, M.A., Gokceoglu, C., 2019. An metaheuristic optimization for flood susceptibilitgy modeling in a high-frequency
experimental research on the use of recurrent neural networks in landslide tropical cyclone area using GIS. J. Hydrol. 540, 317–330. https://doi.org/10.1016/j.
susceptibility mapping ISPRS. Int. J. Geo-Inf. 8 (12), 578. jhydrol.2016.06.027.

13
K. Khosravi et al. Journal of Hydrology 618 (2023) 129229

Trabelsi, M., Meddouri, N., Maddouri, M., 2017. A New Feature Selection Method for conservation. Agricultural Research Service, U. S. Dept of Agriculture in cooperation
Nominal Classifier based on Formal Concept Analysis. Proc. Comput. Sci. 112, with Purdue Agricultural Experiment Station.
186–194. https://doi.org/10.1016/j.procs.2017.08.227. Wischmeier, W.H., Smith, D.D., 1978. Predicting rainfall erosion losses: a guide to
Vaezi, A.R., Hasanzadeh, H., Cerdà, A., 2016. Developing an erodibility triangle for soil conservation planning. Science and Education Administration, U.S. Dept. of
textures in semi-arid regions, NW Iran. CATENA 142, 221–232. https://doi.org/ Agriculture.
10.1016/j.catena.2016.03.015. Woolhiser, D.A., Smith, R.E., Goodrich, D.C., 1990. KINEROS:a kinematic runoff and
Venkatappareddy, P., Culli, J., Srivastava, S., Lall, B., 2021. A Legendre polynomial erosion model: documentation and user manual.
based activation function: An aid for modeling of max pooling. Digit. Signal Process. World Bank, 2005. Islamic Republic of Iran, Cost Assessment of Environmental
115, 103093 https://doi.org/10.1016/j.dsp.2021.103093. Degradation.
Vojinovic, Z., Abebe, Y., Ranasinghe, R., Vacher, A., Martens, P., Mandl, D., et al., 2013. Yamashita, R., Nishio, M., Do, R.K.G., Togashi, K., 2018. Convolutional neural networks:
A machine learning approach for estimation of shallow water depths from optical an overview and application in radiology. Insights Imaging 9, 611–629. https://doi.
satellite images and sonar measurements. Journal of Hydroinformatics 15 (4), org/10.1007/s13244-018-0639-9.
1408–1424. Yesilnacar, E.K., 2005. The Application of Computational Intelligence to Landslide
Vu, D.T., Tran, X.-L., Cao, M.-T., Tran, T.C., Hoang, N.-D., 2020. Machine learning based Susceptibility Mapping in Turkey. University of Melbourne.
soil erosion susceptibility prediction using social spider algorithm optimized Yousefi, S., Moradi, H., Boll, J., Schönbrodt-Stitt, S., 2016. Effects of road construction on
multivariate adaptive regression spline. Measurement 164, 108066. https://doi.org/ soil degradation and nutrient transport in Caspian Hyrcanian mixed forests.
10.1016/j.measurement.2020.108066. Geoderma 284, 103–112. https://doi.org/10.1016/j.geoderma.2016.09.002.
Wainwright, J., Parsons, A.J., Müller, E.N., Brazier, R.E., Powell, D.M., Fenti, B., 2008. Yousefi, S., Pourghasemi, H.R., Avand, M., Janizadeh, S., Tavangar, S., Santosh, M.,
A transport-distance approach to scaling erosion rates: 1. Background and model 2021. Assessment of land degradation using machine-learning techniques: A case of
development. Earth Surf. Process. Landforms 33, 813–826. https://doi.org/10.1002/ declining rangelands. L. Degrad. Dev. 32, 1452–1466. https://doi.org/10.1002/
esp.1624. ldr.3794.
Wijitkosum, S., 2021. Factor influencing land degradation sensitivity and desertification Yu, B., Rosewell, C.J., 1996. Technical Notes: A Robust Estimator of the R-factor for the
in a drought prone watershed in Thailand. Int. Soil Water Conserv. Res. 9, 217–228. Universal Soil Loss Equation. Trans. ASAE 39, 559–561. https://doi.org/10.13031/
https://doi.org/10.1016/j.iswcr.2020.10.005. 2013.27535.
Williams, J.R., 1975. Sediment-yield prediction with Universal Equation using runoff Zhang, X.C., Nearing, M.A., Risse, L.M., McGregor, K.C., 1996. Evaluation of WEPP
energy factor. Runoff And Soil Loss Predictions Using Natural Runoff Plot Data. Trans. ASAE 39,
Wischmeier, W.H., Smith, D.D., 1965. Predicting rainfall-erosion losses from cropland 855–863. https://doi.org/10.13031/2013.27570.
east of the Rocky Mountains:guide for selection of practices for soil and water

14