Coupling-based optimization method of gradient boosting machine for landslide susceptibility mapping

The complex interaction of multiple geological and environmental factors significantly influences the reliability and precision of landslide susceptibility mapping, posing challenges to both the generalization capability and interpretability of Gradient Boosting Machines (GBM). To address these challenges, this study focuses on Yongjia County in southeastern China, where frequent landslides are driven by rugged topography and intensive anthropogenic activities. Using Spearman's rank correlation coefficient and Shapley values, ten key conditioning factors were identified. Four hybrid models were subsequently developed by integrating GBM with Frequency Ratio (FR), Information Value (IV), Certainty Factor (CF), and Weight of Evidence (WOE). The results reveal that the dominant contributors to landslide occurrence in the study area are, in descending order of importance: elevation, normalized difference vegetation index, distance to roads, slope, distance to rivers, land use, stream power index, vegetation, lithology, and rainfall. Model validation indicates that the CF-Logit Boost model outperforms the others, achieving the highest Area Under the Curve (AUC) and Kappa coefficient. Compared to the standalone GBM model, it improved the Kappa coefficient by 4% to 14% and AUC by 1% to 7%. These results underscore the effectiveness of CF-based variable weighting in enhancing GBM performance and provide a robust framework for improving landslide susceptibility assessments.