Boussalim et al. (2026) Integrated RUSLE-machine learning modeling for water erosion risk assessment under climate change in a Mediterranean semi-arid region: a comparison of LR, SVM, and RF models

Identification

• Journal: Arabian Journal of Geosciences
• Year: 2026
• Date: 2026-03-24
• Authors: Youssef Boussalim, Youssef Dallahi
• DOI: 10.1007/s12517-026-12478-4

Research Groups

• Laboratory of Plant Biotechnology and Physiology, Center of Plant and Microbial Biotechnology, Biodiversity and Environment, Faculty of Science, Mohammed V University in Rabat, Rabat, 10000, Morocco

Short Summary

This study integrates the RUSLE model with machine learning (LR, SVM, RF) to predict future water erosion risk in the Ksob watershed, Morocco, under climate change scenarios (SSP2-4.5, SSP5-8.5), demonstrating that Random Forest best models rainfall erosivity (R) and vegetation cover (C) factors, leading to a projected dominant downward trend in erosion risk by the 2030s and 2050s due to decreased rainfall erosivity and improved vegetation.

Objective

• To analyze existing and future water erosion risk in the Ksob watershed using the RUSLE model, accounting for the dynamic nature of rainfall erosivity (R) and vegetation cover (C) through machine learning algorithms.
• To compare Linear Regression (LR), Support Vector Machine (SVM), and Random Forest (RF) models for predicting R and C factors in the 2030s and 2050s, testing the hypothesis that non-linear correlations govern the relationship between vegetation cover and environmental variables in semi-arid environments.

Study Configuration

• Spatial Scale: Ksob watershed, Moroccan Western High Atlas (1746 km²). Data resolutions: 30 meters (topography, vegetation), 1 kilometer (climate), 250 meters (soil).
• Temporal Scale:
  • Reference period: 2001–2020 (referred to as 2010s).
  • Future projection periods: 2021–2040 (2030s) and 2041–2060 (2050s).
  • Climate change scenarios: Shared Socioeconomic Pathways (SSP) SSP2-4.5 (medium) and SSP5-8.5 (pessimistic).
  • Global Climate Models (GCMs): FIO-ESM-2-0 and INM-CM5-0 (from CMIP6).

Methodology and Data

• Models used:
  • Revised Universal Soil Loss Equation (RUSLE) for soil loss estimation.
  • Machine Learning algorithms: Linear Regression (LR), Support Vector Machine (SVM) with linear kernel, and Random Forest (RF) for predicting R and C factors.
• Data sources:
  • Rainfall Erosivity (R): TerraClimate database (2001–2020 precipitation), WorldClim database v2.1 (19 bioclimatic variables for future projections).
  • Soil Erodibility (K): USDA surface soil texture data (Hengl 2018).
  • Topographical factor (LS): NASADEM (30 m resolution Digital Elevation Model).
  • Vegetation Cover (C): Landsat 7 and 8 data (30 m resolution) for Normalized Difference Vegetation Index (NDVI) calculation.
  • Conservation Practice (P): Derived from Land Use Land Cover (LULC) and slope.
  • Predictor variables for R and C modeling: WorldClim v2.1 (bioclimatic variables), SoilGrids v2.0 (11 edaphic variables, 250 m resolution), NASADEM (topographic variables, 30 m resolution).
  • Land Use Land Cover (LULC): Moroccan Forest National Inventory (HCEFLCD 2014) and ESRI LULC (Karra et al. 2021).

Main Results

• The Random Forest (RF) model demonstrated superior performance in predicting both the R factor (R² = 0.989, MAE = 0.930 MJ·mm/(ha·h·year), RMSE = 1.328 MJ·mm/(ha·h·year)) and the C factor (R² = 0.515, MAE = 0.085, RMSE = 0.109) compared to LR and SVM.
• For the 2001–2020 baseline, 72.05% of the Ksob watershed had a low water erosion risk, with bare ground lands, Tetraclinis articulata, Argania spinosa stands, and rangelands identified as the most vulnerable LULC classes.
• Sensitivity analysis revealed the LS factor as the most decisive variable controlling soil loss spatial variability (Pearson r = 0.86), followed by the C factor (r = 0.61) and R factor (r = 0.39). Non-linear analysis confirmed LS and C as dominant.
• Future projections indicate a decrease in rainfall erosivity (R) across 70% to 80% of the watershed, with more significant reductions under the SSP5-8.5 scenario, particularly in upstream and transition zones.
• Vegetation cover (C) is projected to improve (decrease in C factor) in 59.5% to 67.5% of areas, with the annual temperature range (bio7) being the most important predictor. Most LULC types show a significant downward trend in C, except agricultural lands which show a significant upward trend (vegetation decline).
• Overall, 15% to 15.5% of the watershed areas in the 2030s and 17.5% to 19% in the 2050s are projected to experience a shift in water erosion risk, predominantly towards lower levels (e.g., from high to moderate-high, moderate-high to moderate-low, or moderate-low to low).
• Bare lands, Tetraclinis articulata, Argania spinosa stands, and rangelands are the main occupations expected to experience shifts in erosion risk, while Juniperus phoenicea stands and agricultural areas are predicted to remain stable.

Contributions

• Developed an integrated RUSLE-machine learning modeling framework that explicitly projects both rainfall erosivity (R) and vegetation cover (C) factors under climate change, addressing a common limitation in previous studies.
• Demonstrated the superior predictive capability of Random Forest over Linear Regression and Support Vector Machine for modeling dynamic RUSLE factors in a Mediterranean semi-arid environment, highlighting the importance of non-linear relationships.
• Provided detailed spatiotemporal projections of water erosion risk shifts for a vulnerable semi-arid watershed under CMIP6 Shared Socioeconomic Pathways (SSP2-4.5 and SSP5-8.5) and two GCMs.
• Identified the "double-sided effect" of future climate aridity, where reduced rainfall erosivity and decreased annual temperature range can lead to vegetation recovery in rugged terrains, while low-slope areas may experience vegetation decline due to reduced climatic humidity.
• Offers practical, differentiated management recommendations for various LULC types to mitigate future erosion risk, including cost-effective soil conservation, ecological restoration, and targeted erosion control structures.

Funding

• Centre Nationale pour la Recherche Scientifique et Technique (CNRST) as part of the “PhD-ASsociate Scholarship – PASS” program.

Citation

@article{Boussalim2026Integrated,
  author = {Boussalim, Youssef and Dallahi, Youssef},
  title = {Integrated RUSLE-machine learning modeling for water erosion risk assessment under climate change in a Mediterranean semi-arid region: a comparison of LR, SVM, and RF models},
  journal = {Arabian Journal of Geosciences},
  year = {2026},
  doi = {10.1007/s12517-026-12478-4},
  url = {https://doi.org/10.1007/s12517-026-12478-4}
}