Comparative evaluation of machine learning model and PAP/CAR approach for water erosion prediction in the Beht watershed, Morocco
Więcej
Ukryj
1
Functional Ecology and Environmental Engineering Laboratory (LEFGE), FST-Fes, Sidi Mohammed Ben Abdellah University
2
Laboratory of Intelligent Systems, Energy, and Sustainable Development (SIEDD), Private University of Fez, Lotissement Quaraouiyine Route Ain Chkef, Fès 30000, Morocco.
3
Laboratory of Innovative Materials and Mechanical Manufacturing Processes (IMMM), ENSAM-Meknes, Moulay Ismail University, Marjane 2, BP: 15290 Meknes 50500, Morocco.
Autor do korespondencji
Fahed El amarty
Functional Ecology and Environmental Engineering Laboratory (LEFGE), FST-Fes, Sidi Mohammed Ben Abdellah University
SŁOWA KLUCZOWE
DZIEDZINY
STRESZCZENIE
Soil erosion is a major environmental concern, particularly in hydrologically unstable regions. Reliable prediction methods are essential for effective erosion risk assessment and mitigation. This study evaluates and compares two erosion prediction methodologies in the Beht watershed, Morocco: the traditional PAP/CAR model and an advanced machine learning technique, Extreme Gradient Boosting (XGBoost). Using Geographic Information Systems (GIS) and remote sensing data, we integrate various conditioning factors such as slope, elevation, rainfall intensity, and land cover. While the PAP/CAR model provides a qualitative assessment of erosion susceptibility, its limitations in spatial precision necessitate a comparison with data-driven approaches. The results indicate that the PAP/CAR model identifies high-risk erosion zones covering approximately 42.37% of the watershed, but it tends to overestimate spatial distributions. In contrast, the XGBoost model, trained on 70% of inventory data and validated on the remaining 30%, achieves an Accuracy of 90.02%, a Kappa coefficient of 0.6, and an AUC-ROC score of 0.96, demonstrating its superior predictive power. By leveraging optimized hyperparameters, XGBoost enhances classification stability, reducing bias and variance, thereby improving model reliability. These findings emphasize the necessity of integrating advanced computational techniques into geospatial analyses for erosion risk management, offering more precise tools for soil conservation strategies and watershed management.