Geo-environmental controls on gully erosion in loess-covered regions: a geomorphological analysis and machine learning approach

This study highlights the need for interpretable, high-accuracy models of gully erosion susceptibility in the Kalat-e-Naderi basin, northeast Iran, an area characterized by unstable loess deposits that are highly prone to both surface and subsurface erosion processes. A gap remains in regional literature due to the lack of prior Machine Learning (ML) assessments for this understudied basin and the absence of explainable ML applications in existing Iranian loess studies. We introduce a framework for gully susceptibility in the Kalat-e-Naderi basin, combining four advanced ML algorithms (Random Forest (RF), LightGBM, XGBoost, and CatBoost) with the SHapley Additive exPlanations (SHAP) as an explainable method to identify specific hydrological and topographical thresholds that lead to loess failure. A detailed gully inventory comprising 652 points was created through high-resolution Google Earth image analysis and field assessments. Fourteen geo-environmental factors were selected for susceptibility modeling. Model performance was tested using multiple metrics including Root Mean Square Error, Mean Absolute Error, Kappa coefficient, Overall Accuracy and the Area Under the Receiver Operating Characteristics Curve (AUROC), with RF achieving the highest accuracy and discriminative ability (AUROC = 0.90), confirming its superior performance for spatial classification. SHAP analysis validated the process-driven approach, revealing precipitation, Topographic Position Index (TPI), and elevation as the main drivers of gully initiation. These results show that precipitation acts as the key hydrological trigger, promoting piping and hydro-consolidation within the loess, while TPI and elevation shape spatial vulnerability by gathering moisture in low-lying convergence zones. The resulting maps quantify the risk, indicating that the High and Very High susceptibility classes jointly account for approximately 22% of the total basin area. The derived gully erosion susceptibility maps serve as essential and immediate tools for regional land use planning, supporting sustainable agriculture and boosting socio-economic stability in the region. This methodological approach provides a solid and adaptable framework for erosion risk assessments in similar under-researched, semi-arid, loess-covered regions worldwide.