Introduction

Wind erosion, as an environmental problem, has many adverse effects on the economics of societies and the health of terrestrial and marine ecosystems1,2,3. Therefore, predicting land susceptibility to wind erosion hazards such as dust emissions from land surfaces is essential for mitigating its effects. Literature review shows that different tools and techniques have been proposed for investigating different aspects of wind erosion and its consequences, uniquely identifying regions prone to generating sediments for wind erosion, including remote sensing, data mining, and sediment fingerprinting4,5,6,7. However, these techniques require intensive field sampling with expensive laboratory analyses8, and as a result, they are not efficient for large spatial domains.

Recently, together with developments of geospatial technology and computer sciences, machine learning (ML) has received considerable attention with many successful applications in the spatial mapping of different environmental hazards such as land subsidence, gully erosion, landslides, and dust provenance, as well as mapping of soil properties (microbial dynamics, moisture, shear strength, soil taxa, bulk density, total nitrogen, organic carbon). However, to the best of our knowledge, exploration of the utility of advanced ML techniques in predicting land susceptibility to wind erosion has not been undertaken.

Typical ML models applied to date in different areas of environmental research include decision tree and linear equation models, the particle swarm optimization-adaptive network-based fuzzy inference system (PANFIS), genetic algorithms, support vector regression (SVR), artificial neural networks (ANN), hybrid models, random forest (RF), Wang and Mendel's (WM), partial least square regression (PLSR), principal component regression (PCR), Cubist, Bayesian additive regression trees (BART), radial basis function (RBF), extreme gradient boosting (XGBoost) and regression tree analysis8,9,10,11,12,13,14,15. Since, to date, a comprehensive study applying regression-based ML models to mapping wind erosion hazard has not been investigated, there remains a need for such work since wind erosion hazards are a major socio-economic challenge for some parts of the world. Accordingly, this work aimed to address this gap in the existing literature by providing a comprehensive assessing of the prediction performance of 16 regression-based ML models (robust linear regression (RLR), Cforest, non-convex penalized quantile regression (NCPQR), neural network with feature extraction (NNFE), monotone multi-layer perception neural network (MMLPNN), ridge regression (RR), boosting generalized linear model (BGLM), negative binomial generalized linear model (NBGLM), boosting generalized additive model (BGAM), spline generalized additive model (SGAM), spike and slab regression (SSR), stochastic gradient boosting (SGB), support vector machine (SVM), relevance vector machine (RVM), Cubist and adaptive network-based fuzzy inference system (ANFIS)) for mapping land susceptibility to the wind erosion hazard in the Isfahan province, central Iran. Using this case study, we provide more generic recommendations.

Results

Multicollinearity test

Table1 shows the values of the tolerance coefficient (TC) and the variance inflation factor (VIF) for the controlling factors for wind erosion. VIF > 10 and TC < 0.1 indicate multicollinearity among the effective factors. Based on our results, the lowest TC value was obtained for electrical conductivity (EC), while the highest VIF value (5.93) value was calculated for bulk density. The results indicated the absence of any multicollinearity between the 13 factors controlling wind erosion in the study area.

Table 1 Values of the TC and VIF for examining multicollinearity among the effective factors for wind erosion using the training dataset.
Full size table

Relative importance of the factors affecting wind erosion

The model with the highest performance (MMLPNN) was applied to quantify the relative importance of the effective factors for wind erosion. Based on Fig. 1, three factors, DEM (with relative importance 0.95), precipitation (with relative importance 0.8), and NDVI (with relative importance 0.54), were recognized as the most important factors controlling wind erosion in the study area. Wind erosion has been shown to be affected by many factors such as wind, precipitation, temperature, soil properties (texture, composition, and aggregation), topography, aerodynamic roughness, vegetation, and land use practice16.

Figure 1
figure 1

The relative importance of the effective factors for wind erosion estimated by MMLPNN. DEM, PR, NDVI, AWC, CCP, ESP, OCC, EC, GE, BD, WS, ST, and LU indicate digital elevation model, precipitation, normalized difference vegetation index, available water content, calcium carbonate content, exchangeable sodium percentage, organic carbon content, electrical conductivity, geology, bulk density, wind speed, soil texture, and land use, respectively.

Full size image

Discussion

Maps of wind erosion hazard

The wind erosion hazard maps generated by 16 individual ML models are presented in Figs.2, 3, and 4. Table 2 indicates the area (percentage and km2) of the four land susceptibility classes (low, moderate, high, and very high) for wind erosion hazard estimated by the 16 ML models. Based on the results of all 16 models, areas of land susceptibility to the low susceptibility class ranged between 15.5% (RVM and BGLM models) and 32.8% (MMLPNN model). The minimum and maximum areas of moderate land susceptibility to wind erosion were estimated by the SGB (0.6%) and SSR (15.7%) models, respectively. The area of land categorized into the high susceptibility class ranged from 1.2% (MMLPNN model) to 20.2% (NCPQR model). Corresponding areas assigned to the very high class of land susceptibility to wind erosion hazard ranged from 41% (NBGLM model) to 65.2% (SGB).

Figure 2
figure 2

Maps of wind erosion hazard generated by: (a) RLR, (b) Cforest, (c) NCPQR, (d) NNFE, (e) MMLPNN, and (f) RR. The values for pixels was estimated by R software (https://CRAN.R-project.org/doc/FAQ/R-FAQ.html) and then, values of pixels were mapped by ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image
Figure 3
figure 3

Maps of wind erosion hazard generated by: (a) BGLM, (b) NBGLM, (c) BGAM, (d) SGAM, (e) SSR, and (f) SGB. The values for pixels was estimated by R software (https://CRAN.R-project.org/doc/FAQ/R-FAQ.html) and then, values of pixels were mapped by ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image
Figure 4
figure 4

Maps of wind erosion hazard generated by: (a) SVM, (b) RVM, (c) Cubist and (d) ANFIS. The values for pixels was estimated by R software (https://CRAN.R-project.org/doc/FAQ/R-FAQ.html) and then, values of pixels were mapped by ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image
Table 2 Land susceptibility classes to wind erosion hazard calculated by 16 individual ML models.
Full size table

Model performance assessment

Model performance for mapping wind erosion hazard was assessed by three indices (MAE, MBE, and RMSE; (Fig. 5)). Additionally, a Taylor diagram for both the training and evaluation datasets were constructed (Fig. 6). MMLPNN was selected as the most accurate model for mapping wind erosion hazard, while according to the RMSE and MAE, NBGLM was the weakest predictive model, and NCPQR was recognized as the overall worst model.

Figure 5
figure 5

The values of the statistical indicators were used to evaluate model performance; (a) training dataset and (b) evaluation dataset.

Full size image
Figure 6
figure 6

Taylor diagrams for assessing the performance of the models in this research; (a) training dataset, and (b) evaluation dataset.

Full size image

Based on all three statistical indicators of model performance and the Taylor diagram for the evaluation dataset, five models (MMLPNN, SGAM, Cforest, BGAM, and SGB) returned low errors. SSR and NBGLM had the lowest accuracies among the 16 models. Based on the Taylor diagram drawn for the training dataset, five models (MMLPNN, Cforest, SGAM, SGB and NNFE) were identified as the most accurate predictive ML models for mapping wind erosion hazard in the study area, whereas NBGLM and RVM were the weakest predictive models.

Overall, MMLPNN, SGAM, Cforest, BGAM, and SGB were identified as the most accurate models for predicting land susceptibility to wind erosion. Based on MMLPNN (Fig. 2e), the four susceptibility classes covered 32.8%, 1.1%, 1.2% and 64.9% of the total area of Isfahan province, respectively. The land susceptibility map to wind erosion hazard generated using SGAM shows the high and very high susceptibility classes covered 5.4% and 61.5% of the total area, respectively, whereas the low and moderate susceptibility classes occupied 27.4% and 5.6%, respectively (Fig. 3d). According to Cforest (Fig. 2b), 26%, 6.4%, 6.6%, and 61% of the total area belonged to the low, moderate, high and very high susceptibility classes, respectively. Using BGAM (Fig. 3c), the very high susceptibility class covered 62% of the study area, whereas the low, moderate, and high classes occupied 23.2%, 7.8% and 7% of the total area, respectively. Finally, in the case of the SGB model (Fig. 3f), the results classified 32%, 0.6%, 2.2% and 65.2% of the study area as low, moderate, high, and very high susceptibility, respectively.

The map of wind erosion hazard produced by MMLPNN is the most accurate. Overall, multi-layer perception networks (MLPS) as universal estimators are well-known techniques for system identification. The monotonicity of MMLPNN does not depend on the quality of the training data because it is guaranteed by its structure17. GAM with spline function (SGAM) was one of the 5 most accurate models for wind erosion hazard mapping. The spline functions allow the flexible representation of non-linear marginal relationships of the explanatory and response variables without the necessity to define a specific function18. Cforest, as a random forest (RF) model, uses conditional inference trees for prediction19. Several studies confirm the performance of RF as a suitable model for spatial predictions of environmental hazards. For example20, reported that the RF model is the best model for digital mapping of soil carbon fractions.

Some studies21 have also argued that RF has the highest predictive capability for modelling landslide susceptibility in comparison with other ML models. Some previous studies22 have also reported that in comparison with other methods, RF has better performance in estimating PM2.5 monthly concentration. In this study, we applied the boosting with generalized additive model (BGAM), and based on the indicators for examining model performance, this model exhibited satisfactory performance and was selected as one of the five most accurate models for mapping wind erosion hazard. Boosting is a technique for improving prediction rules, and it can be applied to classification and regression methods to increase the accuracy of the predictions23. SGB is related to both boosting and bagging24,25. Previous research26 has reported that SGB provides stable predictions for tree species presence.

Conclusions

This research assessed the performance of 16 individual regression-based ML algorithms for mapping land susceptibility to wind erosion hazard in an arid region in central Iran. In all, 13 effective factors for wind erosion were considered and regions with active wind erosion were mapped using a "wind erosion inventory map". Based on three statistical indicators and a Taylor diagram, the MMLPNN model was the most accurate model. We conclude that MMLPNN is powerful tool for mapping wind erosion hazard in arid and semi-arid region ecosystems worldwide. We recommend that future work should focus on testing and comparing the performance of regression-based and classification-based ML models for the mapping and spatial modelling of wind erosion and dust sources to ensure that robust evidence is provided to support management decisions.

Material and methods

Study area

Isfahan province (Fig. 7), an arid region, is located in central Iran, between the latitudes 30°45′59.51" to 34°27′13.27" N, and between the longitudes 49°41′53.86" to 55°30′13.67" E. It is experiencing intensive wind erosion on the southeastern side (Segzi plain) and its northern parts. Based on a digital elevation model (DEM), there is high variability in altitude with maximum and minimum elevations ranging between 686 m (in the northern part of the study area and southern parts of Dasht-e-Kavir) to 4398 m (in the vicinity of the Dena Mountain in the southwestern part of the study area). The average annual precipitation ranges between 72 mm (in the eastern part with a corresponding annual mean temperature of 18 °C) and 320 mm (in the western part with an average annual temperature of 13 °C).

Figure 7
figure 7

Location of the study area in Iran and sampling sites used for this study. Soil sampling sites were extracted from the world soil map30 and then, these sites were mapped in ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image

Factors controlling wind erosion

Different environmental and climatic factors are controlling wind erosion phenomena in drylands. Environmental variables affecting wind erosion include soil properties, lithology, land use, vegetation cover, topography, and elevation1,8,27. Previous research28 introduced a local wind erosion climatic index based on the wind speed and effective precipitation index developed by29 for applying in the Chepil wind erosion equation (WEQ).

Soil characteristics

Seven soil characteristics (e.g., available water content (AWC) (Fig. 8a), bulk density (Fig. 8b), calcium carbonate percentage (Fig. 8c), electrical conductivity (EC) (Fig. 8d), exchangeable sodium percentage (ESP)(Fig. 8e), organic carbon content (OCC)(Fig. 8f) and soil texture (Fig. 9a)) were extracted from the world soil map30 and mapped by interpolation in ArcGIS 10.4.1. It should be noted that a total of 803 points (Fig. 7) were used for generating spatial maps.

Figure 8
figure 8

Spatial maps of soil characteristics: (a) AWC; (b) bulk density; (c) calcium carbonate percentage; (d) EC; (e) ESP, and; (f) OCC. All these factors were mapped spatially in ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image
Figure 9
figure 9

Spatial maps of: (a) soil texture; (b) geology; (c) land use; (d) NDVI; (e) DEM, and; (f) wind speed. All these factors were mapped spatially in ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image

Lithology and land use

Lithology (Fig. 9b) and land use (Fig. 9c) were mapped spatially based on the maps produced by the Forests, Rangelands, and Watershed Management Organization of Iran (FRWMOI).

Vegetation cover

The normalized difference vegetation index (NDVI) (Fig. 9d)31 as the most common index used for the spatial mapping of vegetation cover was applied in our study. NDVI is the difference between the red (RED) and near-infrared (NIR) band combination divided by the sum of the red and near-infrared band combination (Eq. 1).

$$ {\text{NDVI }} = \, \left( {{\text{NIR}}_{{{\text{b4}}}} {-}{\text{ RED}}_{{{\text{b3}}}} } \right) \, / \, \left( {{\text{NIR}}_{{{\text{b4}}}} + {\text{ RED}}_{{{\text{b3}}}} } \right) $$
(1)

Elevation

A digital elevation model (DEM) (Fig. 9e) for the study area was generated using shuttle radar topography mission (SRTM) images with a 30*30 m resolution8.

Climatic variables

Wind speed (Fig. 9f) and precipitation (Fig. 10a) were used as climatic factors affecting wind erosion. The spatial maps of these variables were generated based on the daily average wind speed and total annual precipitation data from 23 meteorological stations located in the Isfahan province. All spatial maps of factors controlling wind erosion were generated in ArcGIS 10.4.1.

Figure 10
figure 10

Spatial maps of: (a) total annual precipitation; (b) locations of the pixels with active wind erosion, and; (c) locations of the training and validation data points. All these characteristics were mapped spatially in ArcGIS 10.4.1 (https://www.esri.com/en-us/about/about-esri/overview).

Full size image

Inventory map of wind erosion

An inventory map shows regions with active three-stage processes, comprising detachment, transportation, and sedimentation due to wind erosion. An inventory map is needed for predicting land susceptibility to wind erosion hazard. We used a map of regions with active wind erosion produced by the Forest, Rangeland and Watershed Management Organization of Iran (FRWMOI) (Fig. 10b). Based on the inventory map, wind erosion active regions covered ~ 10,961 km2 (440 pixels) in the study area. Pixels with active wind erosion were randomly selected for the training (70% or 308 pixels) and validation (30% or 112 pixels) datasets for the ML models (Fig. 10c). Based on field work and FRWMOI, inventory map of wind erosion was generated in ArcGIS 10.4.1.

Multicollinearity among the factors controlling wind erosion

The tolerance coefficient (TC) (Eq. 2) and variance inflation factor (VIF) (Eq. 3) 8,15,32 were applied to examine multicollinearity among the factors for wind erosion in the Isfahan province.

$$ {\text{TC }} = { 1 }{-}{\text{ R}}^{{2}} $$
(2)
$$ {\text{VIF }} = \left[ {\frac{1}{{{\text{TC}}}}} \right] $$
(3)

where R2 is the regression coefficient. If the TC is < 0.1 and the VIF is > 10, both coefficients signify a multicollinearity problem.

Background of the ML algorithms used

This section briefly describes the 16 individual regression-based ML algorithms, which were adopted for mapping wind erosion hazard. These algorithms are available in the caret package, in R software.

Robust linear regression (RLR)

Robust regression is designed to overcome some limitations of traditional parametric and non-parametric methods. Available robust regression methods include M-estimates33, R-estimates34, least median of squares (LMS) estimates35, least trimmed squares (LTS) estimates and S-estimates36, generalized S-estimates (GS-estimates)37 and MM-estimates38. We used a robust linear regression model with M-estimates for predicting land susceptibility to wind erosion.

Cforest

Random forest (RF), introduced by39, is the most popular method for regression and classification in decision tree learning40. RF makes a large number of decision trees in the training phase, and then by averaging the output values of the trees, the output of the model is finalized. Cforest is a type of RF commonly applied for prediction purposes19.

Non-convex penalized quantile regression (NCPQR)

Quantile regression (QR) has gained considerable attention in different fields of modelling since the work of41. In comparison with mean regression (MR), QR provides an alternative that is more efficient when the error term follows a non-normal heavy-tailed distribution42. We used a penalized QR with a non-convex function42 for mapping wind erosion hazard.

Neural networks (NN)

NN can accurately approximate complicated non-linear input/output relationships43. The NN structure includes a set of interconnected units or neurons that estimates the non-linear correlations between each variable. The input neurons or predictor variables are connected to a single or multiple layers of hidden neurons, which are then linked to the output neurons44. We used a NN with the feature extraction algorithm (NNFE)45 and a monotone multi-layer perception neural network (MMLPNN)46 for mapping wind erosion hazard. The feature extractors used textural features based on the spatial relationships between pixels45.

Ridge regression with variable selection

Ridge regression (RR), which was proposed by47, is expressed as follows (Eq. 4):

$$L\left(w\right)=\sum_{i=1}^{n}{\left({y}_{i}- {\tilde{y}}_{i}\right)}^{2}=\sum_{i=1}^{n}{({y}_{i}-w. {x}_{i})}^{2}$$
(4)

Given a set of n vectors, x1, … , xn in Rm, where m is the number of properties, and the dependent variable yi ∈ R, i = 1, …, n, the objective is to minimize the loss function, i.e., the discrepancy between the real values yi and the predicted values ỹi = w.x.

We applied a RR model with a kernel function48 as follows:

$$ {\tilde{y}}=f\left(x\right)=\sum_{i=1}^{n}{\ss }_{i} K \left(x, {x}_{i}\right)$$
(5)

where \(K \left(x, {x}_{i}\right)\) is the kernel function and βi is the weighting.

Generalized linear models (GLMs)

GLMs have been applied to a wide range of research49. GLMs have three components, comprising an observation model, a linear predictor, and an invertible link function50. Using boosting with GLMs can improve prediction accuracy23. We applied two GLMs; boosting GLM (BGLM) and negative binomial GLM (NBGLM)51.

Generalized additive models (GAMs)

GAMs52 can be expressed as follows:

$$e\left({\mu }_{i}\right)= {Z}_{i}^{*}.\ss + \sum_{j}{f}_{j} ({x}_{ij})$$
(6)

with

$${\mu }_{i}=E \left({Y}_{i}\right), and {Y}_{i} \sim EF \left({\mu }_{i}, \varnothing \right),$$
(7)

where \({Y}_{i}\) is the ith value of the response variable from an exponential distribution family (EF) with a location parameter (\({\mu }_{i})\) and a scale parameter (\(\varnothing \)),\({Z}_{i}^{*}\) indicates the ith row of a parametric model matrix with the vector β, fj shows unknown functions and \({x}_{ij}\) indicates the ith value of the jth variable. \(g\left({\mu }_{i}\right)\) is the link function. We applied two GAMs, comprising boosting (BGAM) and spline (SGAM)18.

Spike and slab regression (SSR)

SSR is one of the typical variable selection approaches in regression settings, and this model has been applied widely in challenging problems53. SSR was proposed by54 and can be expressed as follows53:

$${y}_{i}={\beta }_{\mathrm{1,0}}{x}_{i,1}+\dots + {\beta }_{p,0}{x}_{i,p}+ {\varepsilon }_{i}, i=1, \dots , n,$$
(8)

where (εi)1≤ i ≤n are independent random variables such as E(εi) = 0 and E (\({\varepsilon }_{i}^{2}\)) = \({\sigma }_{0}^{2}>0.\) Write X for the n × p design matrix corresponding to (1) and \({\beta }_{0}={({\beta }_{\mathrm{0,1}},\dots , {\beta }_{0,P})}^{T}\) for the true regression parameter. The variables \({x}_{i}={({x}_{i,1},\dots , {x}_{i,p})}^{T}\) and the response-vector \(y={({y}_{1},\dots , {y}_{n})}^{T}\) are assumed to the standardized such that:\(\sum_{i=1}^{n}{x}_{i,k}=0, \sum_{i=1}^{n}{x}_{i,k}^{2}=n, \sum_{i=1}^{n}{y}_{i}=0.\)

Stochastic gradient boosting (SGB)

SGB or gradient boosting machine, proposed by24 is a hybrid algorithm that combines both the advantages of bagging and boosting. This model makes additive regression models by the least-squares at each iteration.

Support and relevance vector machine (SVM and RVM) algorithms

The relevance vector machine (RVM) is a probabilistic sparse kernel model identical in functional form to the support vector machine (SVM). SVM is a very successful approach to supervised learning, and it makes predictions based on the following function55:

$$y\left(x\right)=\sum_{n=1}^{m}{w}_{n} K \left(x, {x}_{n}\right)+ {w}_{0},$$
(9)

where \({w}_{n}\) indicates the model weights and K (. , .) is a kernel function. We applied two algorithms, SVM with linear kernel function and RVM with polynomial kernel function.

Cubist

Cubist, a rule-based regression tree algorithm, is based on the M5 theory56. This model involves four main steps as follows: (1) growing a tree by branching data, (2) developing the model, (3) pruning the tree, and (4) smoothing the tree57.

Adaptive network-based fuzzy inference system (ANFIS)

This model has been applied in different sciences. ANFIS works based on the fussy if/then rules58:

$$ {\text{Rule 1}}:{\text{ if }}\left( {{\text{x is A}}_{{1}} } \right){\text{ and }}\left( {{\text{y is B}}_{{1}} } \right){\text{ then }}\left( {{\text{f}}_{{1}} = {\text{ p}}_{{1}} {\text{x }} + {\text{ q}}_{{1}} {\text{y }} + {\text{ r}}_{{1}} } \right) $$
(10)
$$ {\text{Rule 2}}:{\text{ if }}\left( {{\text{x is A}}_{{2}} } \right){\text{ and }}\left( {{\text{y is B}}_{{2}} } \right){\text{ then }}\left( {{\text{f}}_{{2}} = {\text{ p}}_{{2}} {\text{x }} + {\text{ q}}_{{2}} {\text{y }} + {\text{ r}}_{{2}} } \right) $$
(11)

where x and y are as input parameters for FIS, f as FIS output, A and B are fuzzy sets, and p, q, and r are parameters.

In all 16 models, the predicted values for pixels ranged between 0–1. Therefore, we can divide susceptibility predictions into four classes (low (0–0.25), moderate (0.25–0.50), high (0.50–0.75) and very high (0.75–1)).

Assessment of model performance

In order to evaluate model performance in predicting land susceptibility to wind erosion hazard in the study area, three statistical methods comprising root mean square error (RMSE), mean absolute error (MAE)59,60 and mean bias error (MBE) were used:

$$RMSE=\sqrt{\frac{\sum_{i=1}^{m}({v}_{k}-{v}_{p}{)}^{2}}{m}}$$
(12)
$$MAE= \frac{\sum_{i=1}^{m}\left|{v}_{k}-{v}_{p}\right|}{m}$$
(13)
$$MBE= \frac{1}{m} \sum_{i=1}^{m}({v}_{k}- {v}_{p})$$
(14)

where m is number of the observations, \({v}_{k}\) and \({v}_{p}\) indicate the measured and predicted values, respectively. Also, a Taylor diagram was applied as a further test for assessing the performance of individual regression-based ML models14.

Prioritization of the factors controlling wind erosion

Among the 16 ML models tested, a model with the lowest error (RMSE, MAE, and MBE) was applied to quantify the relative importance of the factors controlling wind erosion. In this study, MMLPNN had the lowest error (with RMSE, MAE, and MBE < 0.002%) and was therefore applied for determining the relative importance of the factors for wind erosion.

A brief overview of the main steps used in our methods is presented in Fig. 11.

Figure 11
figure 11

Flowchart of the methodology for mapping of wind erosion hazard.

Full size image