Regional Water-Erosion Screening in Northeastern Algeria Using RUSLE and Machine-Learning Classification

Water erosion poses a persistent challenge for agricultural soils and catchment management in Mediterranean landscapes. We examined Skikda Province, northeastern Algeria, using a revised RUSLE screening framework that combines a CHIRPS-based rainfall-erosivity estimate for 2000–2024, the soil-erodibility dataset, terrain-derived LS, land-cover-based C and P = 1. The reference map covers 26,821,605 valid output cells (3798.21 km2); Model A has a median of 6.05 and a mean of 137.36 in nominal t ha−1 yr−1, with a strongly skewed upper tail. A matched, spatially buffered evaluation of Random Forest, LightGBM, XGBoost and CatBoost used 60,599 complete samples. LightGBM achieved macro-F1 = 0.942 with all 12 predictors but only 0.469 when the four factors defining the RUSLE target were excluded. This contrast underscores that the classifiers mainly reproduce model-derived erosion classes. An external field comparison was possible at 77 of 100 georeferenced sites; at an exploratory threshold of 20 t ha−1 yr−1, site-level accuracy was 87.0%, sensitivity 58.8% and specificity 95.0%. Field observations indicate visible erosion features rather than measured annual soil loss. Uncertainty in empirically estimated erosivity, crop-specific C values and terrain-derived LS limits the quantitative interpretation of absolute erosion rates.