Landslide Susceptibility Assessment and Interpretability Analysis Based on Causal Inference: A Case Study of Foshan, China

The growth of urbanization has increased the frequency of landslide hazards. Exploring the landslide susceptibility mapping (LSM) provides valuable scientific guidance for landslide prevention and mitigation efforts. However, existing studies only reveal the correlation between driving factors and landslides, while neglecting the complex interactions among landslide factors. To tackle these problems, this study introduced causal inference into machine learning methods to uncover causal relationships of driving factors. Firstly, the fast causal inference algorithm is employed to analyze the causal structure of driving factors and validate crucial factors using machine learning. Then, counterfactual inference is used to analyze the causal effects of driving factors. Foshan City, China, was selected as the study area. The experimental results indicate that causal methods are more effective in accurately identifying and explaining the key factors driving landslides. XGBoost can achieve an accuracy of 86.42% using the six crucial factors identified by the causal method. Correlation-based methods requiring more driving factors only attained lower performance. The counterfactual inference experiment indicates that different interventions exhibit spatial heterogeneity, and combined treatment strategies are the primary options for mitigating landslides. Therefore, causal inference offers a novel perspective to analyze and treat landslide susceptibility.