Explainable Machine Learning for Wireless Sensor Network-Based Landslide Early Warning Systems

Abstract

Landslide events pose critical risks to human lives and infrastructure in mountainous regions. Wireless Sensor Networks (WSN) offer real-time multi-parameter monitoring capabilities that complement conventional geospatial data for early warning systems. However, most existing machine learning (ML) approaches for WSN-based landslide detection lack interpretability, limiting their practical adoption by field operators. This study presents an explainable ML framework for WSN-integrated landslide early warning. A balanced dataset of 9,864 samples—comprising seven in-situ WSN sensors and 27 geospatial and meteorological features—was evaluated across nine classification algorithms using 5-fold stratified cross-validation. Gradient Boosting achieved the highest area under the receiver operating characteristic curve (AUC) of 0.9731 ± 0.0014 with 98% test set accuracy. SHAP (SHapley Additive exPlanations) analysis revealed that WSN in-situ sensors collectively account for 6.4% of total model attribution, with Microseismic Activity and Soil Temperature identified as the most informative real-time sensors. The proposed framework demonstrates that explainable ML can provide both high accuracy and transparent decision support for landslide early warning.