An improved buffer-controlled sampling strategy for landslide susceptibility assessment considering the spatial heterogeneity of conditioning factors

The selection of landslide and non-landslide samples significantly influences the performance of machine learning (ML)-based landslide susceptibility assessment (LSA). The commonly used buffer-controlled sampling (BCS) strategy for selecting non-landslide samples overlooks the spatial heterogeneity of the geological environment and lacks a standardized method for determining buffer radius. As a result, the sampling process introduces significant uncertainty to ML models. This paper proposes an improved BCS strategy that incorporates the spatial heterogeneity of conditioning factors to address this issue. The proposed strategy generates a buffer zone for each landslide by merging all neighboring areas with the same attributes as the landslide and then calculates the average equivalent radius of those zones for comparative analysis. The random forest (RF) and the support vector machine (SVM) models are employed to predict the landslide susceptibility of Taojiang County, China, using both the improved and the traditional BCS strategy. Furthermore, the impact of different buffer radii on the model prediction is thoroughly investigated to provide guidance for the selection of buffer radius. The results demonstrate that a buffer radius of less than 3,000 m is optimal in Taojiang County. Compared with the traditional RF and SVM model, the corresponding improved models exhibit superior performance, with higher AUC values and increased peak frequency ratios in areas of very high susceptibility. These findings confirm the effectiveness of the proposed strategy, offering valuable guidance for buffer radius selection and improving the ML-based LSA.