Integrating Physical and Machine Learning Models for Enhanced Landslide Prediction in Data-Scarce Environments

This study addresses the challenges associated with landslide prediction due to limited geotechnical data and high costs. A comprehensive approach is proposed to address these challenges. It integrates soil mechanics computations, rainfall infiltration analysis and synthetic aperture radar (SAR) time-series data to develop a transparent and unified model. The framework incorporates a physically based model, specifically the transient rainfall infiltration and grid-based regional slope stability model (TRIGRS), alongside a machine learning (ML) model, particularly leveraging the random forest (RF) algorithm. An explainable artificial intelligence technique, namely, the Shapley additive explanation (SHAP) algorithm, is employed to address the interpretability challenge associated with the model. Testing of this approach in the Himalayan terrain of Bhutan, leveraging historical landslide data and 36 features, shows promising results. The integrated model successfully obtains an area under the receiver operating characteristic curve (AUC-ROC) value equal to 0.936. Further performance improvement is achieved, with the AUC-ROC value increasing to 0.957, through feature optimisation using Harris’s hawk optimisation (HHO). Altitude, normalised difference vegetation index (NDVI)-2014, ALOS-topographic diversity (i.e. ALOS-TopographydiVE) and aspect, alongside NDVI-2018, are identified as the predominant features within the studied region. This study highlights the value of integrating physically based and ML models, particularly in data-scarce settings. Furthermore, the reliance on freely available satellite data, such as SAR, enhances the replicability and broad applicability of the proposed model.