Landslide susceptibility assessment using machine learning with a novel SHAP-based sampling strategy

Abstract

The landslide and non-landslide samples are important inputs for machine learning-based landslide susceptibility assessment. Compared to landslide samples, non-landslide samples generally present higher uncertainty due to random sampling. However, most sampling strategies (e.g., the feature space-based) for non-landslides only consider the characteristics of a single factor or the overall characteristics of all factors, which subsequently leads to either excessive artificial concentration of non-landslide samples or sampling information redundancy. To address these issues, a SHapley Additive exPlanations (SHAP) based sampling strategy considering combined characteristics of landslide conditioning factors (LCFs) is proposed. This strategy sorts the importance of LCFs based on SHAP algorithm and generates multiple sampling spaces using different numbers of LCFs in the sense of importance order. The optimal sampling space is selected according to the Bayesian optimization algorithm. Then, random forest (RF) and extreme gradient boosting (XGBoost) models are utilized to assess the susceptibility of Chaling County, Yanling County, and Guidong County, China, based on the proposed strategy and traditional random sampling. The results indicate that, compared with the traditional RF and XGBoost models, the improved models show better performance with an 8.2% and 9.0% increase in the AUC, respectively. Furthermore, the SHAP-based sampling framework demonstrates good adaptability across the study areas with different geological and geomorphic conditions, suggesting its potential transferability to other regions, although local optimization of parameter settings may still be required.