Physical guided machine learning landslide susceptibility assessment method based on QGIS secondary development

Abstract Landslide susceptibility assessment is challenging due to the complex interplay of environmental factors and the limitations of traditional machine learning models, which often lack interpretability and rely on uncertain negative sample selection. This study introduces a physics-informed data-driven approach that integrates physical slope stability models (SINMAP and TRIGRS) with machine learning via a QGIS-based framework to address these challenges. Results demonstrate that the physics-informed approach significantly enhances model performance, with AUC improvements of 5.54–12.01% over traditional random sampling. The TRIGRS-constrained Random Forest model achieved the best overall performance, with notable gains in accuracy (11.33%), precision (4.18%), recall (27.99%), F1-score (16.70%), and AUC (12.01%). SHAP analysis further elucidated the influence of key factors, showing that low NDVI values, lower elevations, and proximity to roads heighten landslide risk—insights crucial for targeted mitigation. This study’s physics-informed framework not only improves prediction accuracy but also enhances model interpretability, offering a scalable solution for landslide susceptibility assessment in typhoon-prone areas. The method provides a robust technical reference for geohazard risk management, supporting disaster preparedness and infrastructure planning in vulnerable regions.