Simulation of erosion susceptibility in the Medjerda watershed under different climate change scenarios

Assessing future soil erosion risks, sediment budget, and delivery ratios can aid in developing effective mitigation strategies. However, limited research has explored the comparative performance of machine learning algorithms for predicting future soil erosion susceptibility under climate scenarios. To address this gap, the present paper employs four algorithms: deep learning neural network (DLNN), memetic programming (MP), random forest (RF), and gradient-boosted model (GBM) for assessing the spatial and temporal patterns of soil erosion susceptibility under current and future climate scenarios. The analysis utilises data from the Coupled Model Intercomparison Project Phase 6 (CMIP6) global circulation model across three shared socioeconomic pathway (SSP) scenarios, SSP1-2.6 and SSP3-7.0 while considering eight variables affecting soil erosion. The results show that the Rceiver Operating Characteristic (ROC) curve indicates strong performance across all models, with GBM model achieving the highest accuracy (0.975), followed by RF (0.969), DLNN (0.965), and MP (0.921). Findings suggest that GBM is the most reliable for predicting soil erosion within basins. The analysis indicates that areas with high susceptibility will likely increase significantly in the near to medium-term future under the SSP1-2.6 and SSP3-7.0. Results provide a foundation for policymakers to develop strategies, underscore adaptive land management, and support targeted conservation.