Water Erosion Susceptibility Assessment Using RUSLE, AHP and ANN: A Comparative Study in the Cheliff Basin—Algeria

Abstract

Water erosion poses a significant threat to the global environment, reducing both the quality and quantity of water and adversely impacting food production by affecting soil fertility. This study aims to assess water erosion susceptibility within the Cheliff basin, located in the northern region of Algeria. To achieve this objective, three different approaches were employed and compared: the Revised Universal Soil Loss Equation (RUSLE) as an empirical model, the Analytical Hierarchy Process (AHP) as a multi-criteria decision-making model, and the artificial neural network (ANN) as a machine learning model. The AHP and ANN were applied to six erosion-conditioning factors, namely: rainfall, slope, hillshade, soil type, Normalized Difference Vegetation Index (NDVI), and distance to river. Each model was integrated into a Geographical Information System (GIS) to generate water erosion susceptibility maps that were classified into five levels of susceptibility. Results revealed that high and very high erosion susceptibility levels cover 8.60, 12.31, and 16.02% of the Cheliff basin, according to the RUSLE, AHP, and ANN respectively. These were mainly found in mountain areas showing steep slopes in the north of the study area. The Receiver Operating Characteristic (ROC) and its Area Under the Curve (AUC) were employed to validate the models against observed data. Validation results demonstrated the superiority of RUSLE over ANN and AHP models, achieving AUC values of 94.3, 92.9, and 92.5%, respectively, demonstrating excellent predictive accuracy across all three models.