Assessment of gully erosion susceptibility using four data-driven models AHP, FR, RF and XGBoosting machine learning algorithms

Abstract

Gully erosion is a significant global threat to socioeconomic and environmental sustainability, making it a widespread natural hazard. Developing spatial models for gully erosion is crucial for local governance to effectively implement mitigation measures and promote regional development. This study applied two machine learning (ML) models, RF and XGB, alongside an AHP-based multi-criteria decision method and FR bivariate statistics, to assess gully erosion susceptibility (GES) in the Kangsabati River basin in eastern India's Chotonagpur plateau fringe. A GIS database was created, incorporating recorded gully erosion incidents and 20 conditioning variables, which were evaluated for multicollinearity. These variables served as predictive factors for assessing gully erosion presence in the study area. The models' performance was evaluated using metrics such as RMSE, MAE, specificity, sensitivity, and accuracy. The XGB model outperformed the others, achieving a predictive accuracy of 90.22%. The study found that approximately 6.56% of the Kangsabati catchment is highly susceptible to gully erosion, with 12.39% moderately susceptible and 81.05% not susceptible. The XGB model had the highest ROC value of 85.5 during testing, indicating its superiority over the FR (ROC ​= ​81.7), AHP (ROC ​= ​79.8), and RF (ROC ​= ​83.8) models. These findings highlight the XGB model's efficacy and potential for large-scale GES mapping.