An automated and explainable machine learning model for landslide susceptibility mapping

Complexity of model training and difficulty in explaining prediction results greatly restrict development of machine learning in landslide susceptibility assessment. In this study a comprehensive explainable landslide susceptibility assessment model is constructed based on SHAP-XGBoost algorithm, introducing “Explainable Artificial Intelligence (XAI)” and “Automated Machine Learning (AutoML)” into landslide susceptibility assessment． This model achieves automated operation of complex model training, hyperparameter optimization, landslide susceptibility assessment mapping, and model explanation． Testing in Fengjie county in Three Gorges Reservoir Area on two scales (grid units, and slope units) demonstrates explainable automated landslide susceptibility assessment with high predictive accuracy． The model based on grid and slope units achieves AUC values of 0.875 and 0.873 respectively, with accuracy, precision, recall, and F1 scores all significantly higher than 0.5． SHAP algorithm provides explanations for the model from both global and local perspectives, facilitating understanding of distribution characteristics of causative factors in model building and occurrence patterns of landslide disasters． SHAP algorithm can explain prediction results of individual evaluation units with high credibility． This work provides some reference for automated machine learning and explainable models．