AN EXPLAINABLE PHYSICS-INFORMED XGBOOST FRAMEWORK FOR REAL-TIME LANDSLIDE PREDICTION USING WIRELESS SENSOR NETWORKS
Pawan Kumar, Anu Chaudhary, Sandeep Yadav, Vivek Agarwal, Deepika Yadav, Kirti Singh
Abstract: Globally, landslides pose a significant risk to infrastructure and human life, claiming over 4,500 lives and causing $12 billion in economic losses in 2025. The proposed research methodology introduces a unique physics-informed XGBoost approach for improving explanatory power and landslide prediction by combining WSN data with relevant geotechnical classes. The proposed research methodology introduces three main key innovations: physically aware composite features (saturation-slope interaction terms) that effectively capture the importance of hydro-mechanical coupling, a physics-constrained objective that incorporates FoS as a regularisation and provides a better way to express the physicochemical aspects of landslide occurrence, and improved feature engineering with domain-knowledge-based interactions. The excellent results are demonstrated by experimental validation using 35 environmental parameters from multi-sensor WSN deployments: 97.36% accuracy, 97.57% precision, 97.18% recall, and 0.9707 ROC-AUC. This new framework effectively connects established geotechnical principles with data-driven machine learning, offering improved interpretability and superior predictive performance, both of which are crucial for operational landslide early warning systems.
Keywords: Landslide prediction; Physics-informed machine learning; SHAP analysis; Wireless Sensor Networks; XGBoost
DOI: 10.24874/PES08.03.018
Recieved: 15.06.2026 Revised: 26.07.2026 Accepted: 27.08.2026
UDC:
Reads: 3 
