Landslide Susceptibility Assessment Based on Machine Learning Models Inbailong River Basin, China

AbstractAccuratesusceptibility assessment of geological disaster is crucial for disaster prevention and urban spatial planning. The Bailong River basin, characterized by steep slopes, high relative relief, deep incised valleys, and weak lithologic features, has a complex regional dynamic environment and extreme climatic conditions. In this study, nine evaluation factors were selected, including elevation, lithology, relief, river buffer distance, fault buffer distance, slope, aspect, and slope type. The Weight of Evidence (WOE) model, Support Vector Machine (SVM), WOE-SVM coupled model, Random Forest (RF) model, and WOE-RF coupled model were employed to develop landslidesusceptibility assessment models. The prediction effects of these models were compared and analyzed. The results indicate that the prediction accuracy of the machine learning-based susceptibility models is significantly higher than that of the WOE model. The Area Under the Curve (AUC) of the coupled models reaches a maximum of 0.9999, with 98% of geological disasters occurring in very high and high susceptibility zones. The assessment results are consistent with the spatial distribution of the disasters. The WOE-RF coupled model demonstrates the highest prediction accuracy, exhibiting superior predictive performance and generalization capability. The assessment results are more suitable as references for disaster prevention and long-term major project planning and construction.

Keywords: Susceptibility assessment, Machine learning, The weight of evidence model, Support vector machine, Random forest.