Regional landslide susceptibility assessment via coupled modeling and sample optimization: A case study of Dujiangyan, China

Study region Dujiangyan City in Sichuan Province, China, represents a typical hydrologically complex mountainous region. It is characterized by dense river networks and abundant seasonal rainfall, and rainfall-induced landslides are frequent in this area. Study focus This study proposes a landslide susceptibility assessment framework based on optimized sample selection and coupled modeling. Gaofen-2 (GF-2) satellite imagery was used to extract landslide boundaries. Based on these boundaries, a Landslide Boundary Sampling Method (LBSM) was developed to restrict the selection of negative samples. The assessment was screened from 19 candidate factors to a final set of 15, including precipitation and distance to rivers. A coupled model integrating the Information Value Model (IVM) with a parameter-optimized Random Forest (RF) algorithm was constructed. Four sampling strategies and four positive-to-negative sample ratios from 1:1 to 1:3 were evaluated. New hydrological insights for the region The results show that traditional random sampling may not fully reflect the true spatial distribution of landslides in hydrologically complex areas. The proposed LBSM reduces sampling bias by using mapped landslide boundaries to constrain negative-sample selection. The coupled model based on LBSM achieved an AUC of 0.98. The 1:2 ratio yielded the highest AUC (0.99) and Balanced Accuracy (0.936). These results indicate that boundary-based sampling can improve high-risk zone identification in hydrologically complex mountainous regions.