Scholarlink

<p>Soil erosion represents a critical global environmental challenge, undermining agricultural productivity, degrading water resources, and threatening ecosystem sustainability. Semi-humid regions such as the Safsaf watershed in northeastern Algeria are particularly susceptible, where fragile soils, steep topography, and progressive land cover degradation interact to accelerate sediment loss. Addressing this issue, the present study aims to improve the spatial prediction of water-induced soil erosion by integrating the Revised Universal Soil Loss Equation (RUSLE) with Machine Learning (ML) algorithms, Random Forest (RF), and CatBoost, within a Geographic Information System (GIS) environment. The five RUSLE factors (rainfall erosivity (R), soil erodibility (K), slope length and steepness (LS), cover management (C), and support practices (P)) were derived from multisource datasets including Landsat imagery, a 30&#xa0;m resolution Digital Elevation Model (DEM), and WorldClim climate data. Model performance was evaluated across four benchmark years (1990, 2000, 2010, and 2020) using a 70% data for training and 30% data for testing split and statistical indicators such as the coefficient of determination (R²), Root Mean Square Error (RMSE), and the area under the Receiver Operating Characteristic (ROC) curve (AUC). The findings revealed that both RF and CatBoost delivered high predictive accuracy, with R² values ranging from 0.85 to 0.91 and RMSE as low as 5.81 t ha⁻¹ yr⁻¹. RF consistently outperformed CatBoost in regression tasks, whereas CatBoost demonstrated superior classification ability, attaining the highest Area Under the Curve (AUC) values (0.73–0.78) in 2010 and 2020. Shapley Additive Explanations (SHAP) analysis further highlighted the dominant influence of topographic factors, particularly LS, followed by vegetation cover (C) and rainfall erosivity (R). The resulting erosion susceptibility maps identified persistent high-risk zones on steep, sparsely vegetated slopes, offering crucial guidance for soil conservation, land management, and climate adaptation strategies. Overall, this research underscores the added value of coupling process-based and ML models to enhance erosion prediction in data-scarce semi-humid environments.</p> Graphical Abstract <p>The graphical abstract illustrates the methodological workflow and main findings of the study, integrating the Revised Universal Soil Loss Equation (RUSLE) with advanced machine learning algorithms (Random Forest and CatBoost) within a GIS framework to assess water-induced soil erosion in the Safsaf watershed, northeastern Algeria. The upper section shows the study area location, key physiographic features, and input datasets, including DEM, Landsat imagery, soil data, and precipitation records. The left workflow highlights the derivation of the five RUSLE factors (R, K, LS, C, and P) from multisource data and their integration into machine learning models. Comparative performance metrics (R², RMSE, AUC) are displayed to emphasize the predictive improvements achieved by RF and CatBoost relative to the baseline RUSLE model. The central section presents erosion risk maps for 1990, 2000, 2010, and 2020, showing spatial and temporal dynamics of soil loss. Color-coded maps classify erosion severity from low to very high, clearly delineating high-risk zones on steep, sparsely vegetated slopes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Data-Driven Prediction of Water Erosion Risk Zones in the Safsaf Watershed: a RUSLE-Machine Learning Integration

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Asma Alliouche,
  • Chaouki Benabbas,
  • Amer Zeghmar,
  • Abdeldjalil Belkendil,
  • Haythem Dinar,
  • Nouh Rebouh,
  • Yacine Benzid,
  • Saddam Hussain

摘要

Soil erosion represents a critical global environmental challenge, undermining agricultural productivity, degrading water resources, and threatening ecosystem sustainability. Semi-humid regions such as the Safsaf watershed in northeastern Algeria are particularly susceptible, where fragile soils, steep topography, and progressive land cover degradation interact to accelerate sediment loss. Addressing this issue, the present study aims to improve the spatial prediction of water-induced soil erosion by integrating the Revised Universal Soil Loss Equation (RUSLE) with Machine Learning (ML) algorithms, Random Forest (RF), and CatBoost, within a Geographic Information System (GIS) environment. The five RUSLE factors (rainfall erosivity (R), soil erodibility (K), slope length and steepness (LS), cover management (C), and support practices (P)) were derived from multisource datasets including Landsat imagery, a 30 m resolution Digital Elevation Model (DEM), and WorldClim climate data. Model performance was evaluated across four benchmark years (1990, 2000, 2010, and 2020) using a 70% data for training and 30% data for testing split and statistical indicators such as the coefficient of determination (R²), Root Mean Square Error (RMSE), and the area under the Receiver Operating Characteristic (ROC) curve (AUC). The findings revealed that both RF and CatBoost delivered high predictive accuracy, with R² values ranging from 0.85 to 0.91 and RMSE as low as 5.81 t ha⁻¹ yr⁻¹. RF consistently outperformed CatBoost in regression tasks, whereas CatBoost demonstrated superior classification ability, attaining the highest Area Under the Curve (AUC) values (0.73–0.78) in 2010 and 2020. Shapley Additive Explanations (SHAP) analysis further highlighted the dominant influence of topographic factors, particularly LS, followed by vegetation cover (C) and rainfall erosivity (R). The resulting erosion susceptibility maps identified persistent high-risk zones on steep, sparsely vegetated slopes, offering crucial guidance for soil conservation, land management, and climate adaptation strategies. Overall, this research underscores the added value of coupling process-based and ML models to enhance erosion prediction in data-scarce semi-humid environments.

Graphical Abstract

The graphical abstract illustrates the methodological workflow and main findings of the study, integrating the Revised Universal Soil Loss Equation (RUSLE) with advanced machine learning algorithms (Random Forest and CatBoost) within a GIS framework to assess water-induced soil erosion in the Safsaf watershed, northeastern Algeria. The upper section shows the study area location, key physiographic features, and input datasets, including DEM, Landsat imagery, soil data, and precipitation records. The left workflow highlights the derivation of the five RUSLE factors (R, K, LS, C, and P) from multisource data and their integration into machine learning models. Comparative performance metrics (R², RMSE, AUC) are displayed to emphasize the predictive improvements achieved by RF and CatBoost relative to the baseline RUSLE model. The central section presents erosion risk maps for 1990, 2000, 2010, and 2020, showing spatial and temporal dynamics of soil loss. Color-coded maps classify erosion severity from low to very high, clearly delineating high-risk zones on steep, sparsely vegetated slopes.