Evaluating soil erosion zones in the Kangsabati River basin using a stacking framework and SHAP model: a comparative study of machine learning approaches

Abstract

Soil erosion is a major concern in the Kangsabati River basin, necessitating a comprehensive scientific approach for effective soil erosion management. This study aimed to predict soil erosion susceptibility zones in the basin using integrated soil erosion and deep learning (DL) based stacking framework. Additionally, the SHAP (SHapley Additive exPlanations) model was utilized to augment the interpretability of the DL model. The study employed the RUSLE model to estimate the soil loss. Through ArcGIS, 2000 erosion sites and 2000 non-erosion sites were randomly selected to generate an inventory map. The study considered 17 factors in four primary categories: topographic, climatic, soil, and land use/land cover (LULC). The Boruta algorithm assessed the importance of these variables. Random Forest (RF), (Deep Neural Networks) DNN, Convolution Neural Network (CNN), and stacking (Meta model) models were used to map soil erosion susceptibility based on the inventory map and controlling features. The RUSLE model revealed five erosion zones with soil loss rates ranging from very low (less than 9 t/ha/year) to very high (above 43 t/ha/year). The results demonstrated that 24.93% of the study area fell within the very high erosion susceptibility zone as predicted by DNN, while RF identified 34.32%, Meta model identified 24.84%, and CNN indicated 10.47% of the study area falling into the very high erosion susceptibility category. In terms of RMSE (value) and MSE, the Meta model demonstrates superior performance, whereas the DNN model excels in terms of MAE. The SHAP values output highlights the substantial impact of Land Use and Land Cover (LULC), the K factor, soil moisture, and elevation on the DNN model. These findings provide a scientific basis for developing strategies and policies to combat soil erosion in the Kangsabati River basin, aiding targeted interventions and sustainable land management decisions.