Modeling soil water erosion and sediment transport in Bouregreg watershed (Morocco) using machine learning and climate projections

Soil erosion presents a critical environmental challenge, particularly in regions exposed to climatic variability and anthropogenic pressures. In Morocco, where complex physiographic and climatic conditions prevail, the assessment of soil erosion and sediment transport is hindered by persistent data scarcity, limiting model accuracy and the effectiveness of watershed management. This thesis focuses on the development and application of robust soil erosion and sediment modeling approaches, combining both statistical and process-based methods.The study begins with a national-scale review of soil erosion assessments in Morocco. Results indicate that research is predominantly concentrated in the Rif and Atlas Mountains, with models based on the Revised Universal Soil Loss Equation (RUSLE) or its original version, the Universal Soil Loss Equation (USLE), representing over 51% of applications. However, many studies omit key parameters such as the support practice factor and have limited spatial coverage. Erosion rates were found to be strongly influenced by geomorphological, climatic, and land use factors, although methodological inconsistencies and data limitations contribute to significant variability.To address these challenges, an advanced modeling framework was applied in the Bouregreg watershed, a semi-arid basin in northwestern Morocco, with a focus on modeling suspended sediment concentration (SSC) and sediment yield (SY) over the period 09/01/2016–08/31/2021.The initial approach employed four machine learning (ML) algorithms, Extra Trees, Random Forest, CatBoost, and XGBoost, were integrated with Genetic Programming to enhance predictive accuracy and robustness. Model evaluation using Root Mean Square Error (RMSE), correlation coefficient (r), and Nash–Sutcliffe Efficiency (NSE) demonstrated strong performance (NSE: 0.53–0.86; RMSE: 1.20–2.55 g/L; r: 0.83–0.91) in predicting SSC. To improve interpretability, SHapley Additive exPlanations (SHAP) analysis was employed, revealing streamflow and seasonality as the most influential predictors.Subsequently, the RUSLE and the Modified Universal Soil Loss Equation (MUSLE) models were used to estimate soil loss and sediment yield using high-resolution soil data from INRA (1:50,000) and global FAO data (1:1,000,000). RUSLEINRA yielded more accurate results (15.56 t/ha/yr) than RUSLEFAO (10.24 t/ha/yr), while MUSLEINRA estimated 11.40 t/ha/yr. Sediment yield was validated using observed data from the Sidi Mohamed Ben Abdellah (SMBA) Dam. Projections under the SSP126 and SSP585 climate scenarios for the period 09/01/2021–12/31/2040 predict increased soil erosion (31.54–37.04 t/ha/yr), highlighting the urgent need for proactive soil conservation strategies.This thesis contributes a scalable and interpretable modeling framework that integrates machine learning and geospatial data for improved erosion prediction and watershed management under current and future climate conditions.