Mapping water erosion susceptibility in Northeastern Algeria using two machine learning models

Water erosion is a major driver of land degradation and environmental issues in many parts of the world. Efficient soil protection from erosion requires a comprehensive understanding of its causative factors, which vary spatially and temporally. This study employed 21 erosion conditioning factors in two machine learning models, Smile Random Forest (sRF) and Smile Gradient Tree Boost (sGTB) to predict water erosion susceptibility in the study area, and identify the main driving factors of water erosion susceptibility in Sétif province, Northeastern Algeria. To this end, erosion sites were inventoried by extensive field surveys. The most relevant driving factors were determined using the Boruta feature selection algorithm. The results show that areas with high water erosion susceptibility were predominantly located in mountainous regions, while the plateau areas exhibited relatively lower susceptibility. Slope length had a key role with a contribution of ~53.94% in the sGTB model, whereas in the sRF model, the importance of variables was more evenly distributed, with slope length (8.26%), slope (8.05%), and NDVI (7.63%) being the most influential factors. The AUC values indicated excellent performance, with scores of 0.993 and 0.918 for the sRF and sGTB models, respectively. These findings provide valuable insights for land use planners and decision–makers in controlling soil erosion, particularly in highly vulnerable areas.