Explainable Machine Learning Model for Rockfall Susceptibility Evaluation

Common machine learning has limited application in rockfall susceptibility mapping by a lack of interpretability. Represented by SHAP, explainable machine learning has been developed recently. This study proposed a novel interpretable hybrid-optimized model based on SHAP and XGBoost to interpret rockfall susceptibility evaluation results at both global and local levels. After hybrid-optimized by grid searching hyperparameters and recursive feature elimination (RFE) screening factors, only nine main factors selected by RFE from the 23 initial conditioning factors influence the occurrence of rockfall in the study area. The developed rockfall susceptibility evaluation model provided the accuracy, precision, and AUC value improved by 0.0846, 0.0809, and 0.0616, respectively, for the test data sets. The hybrid-optimized model has a good performance for rockfall susceptibility. The SHAP-based global and local interpretation could further insight into the rockfall occurrence mechanics.