MODELING AND ASSESSMENT OF WATER EROSION IN THE AGGAY WATERSHED: A MACHINE LEARNING-BASED APPROACH

Abstract

In this article, we investigate the susceptibility to water erosion in the Aggay watershed, located in the southeastern Sebou basin region, with a spatial area of approximately 405 km.The goal herein is to identify the soil erosion risk in our area according to two ML classification techniques: Extreme Gradient Boosting (XGBoost) and Random Forest (RF).The models were built using 12 predictor variables that represent environmental, climatic, and topographic aspects.The sites classified as the fourth class with very high soil loss rates (>32.18t/ha/year), belonging to the RUSLE model, have been inventoried, which are used as reference data in erosion susceptibility modelling.Using the sample over 70-30%, we performed 10-fold cross-validation with three repetitions to test the model's robustness and reliability.The average improvement in model accuracy computed during decision tree splits was used to identify the relative importance of each variable to the prediction of the erosive zones.Using the ROC (Receiver Operating Characteristic) curve for model analysis with data on Kappa index, sensitivity, specificity, and overall accuracy statistical indices, AUC values reached the average of 0.86 for XGBoost and 0.92 for RF, indicating better predictive performance of the Random Forest model, and the models provide the best compromise in precision, stability, and ability to predict.Both results underline the applicability of machine learning techniques for spatial modelling of water erosion susceptibility and their relevance for sustainable land usage and environmental planning.