A GeoAI framework for landslide susceptibility mapping using Google Earth Engine, ensemble machine learning, and SHAP explainable AI

Landslides are among the most destructive natural hazards in mountainous regions, yet susceptibility mapping in data-scarce areas remains challenging. This study proposes a reproducible Geospatial Artificial Intelligence (GeoAI) framework that integrates Google Earth Engine (GEE), ensemble machine learning, and explainable AI (XAI) for landslide susceptibility mapping in central Nepal. Using a balanced dataset of 9,479 observations from 4,743 landslide and 4,736 non-landslide locations, extracted seven conditioning factors: topographic (elevation, slope, aspect), climatic (rainfall), ecological (NDVI, land cover), and anthropogenic (population density) within the GEE cloud computing environment. Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models using stratified train-test splitting and 10-fold cross-validation. Both models showed strong predictive performance, with XGBoost achieving superior results (Accuracy: 88.3%, ROC-AUC: 0.952) compared to RF (Accuracy: 87.9%, ROC-AUC: 0.945). SHapley Additive exPlanations (SHAP) identified slope as the most influential predictor associated with model predictions followed by rainfall (16%) and elevation (15%), while lower NDVI values consistently correlated with higher susceptibility. The susceptibility map identified high-risk zones concentrated in Sindhupalchok, Dolakha, Dhading, Rasuwa, Gorkha, and Nuwakot districts. This framework reveals that freely available satellite data, cloud computing, and explainable machine learning can be effectively integrated for susceptibility mapping and community resilience initiatives.