Machine learning-based global landslide susceptibility analysis: spatiotemporal variability and dominant environmental associations

Global landslide susceptibility assessments exist but exhibit three critical limitations: current products lack seasonal temporal resolution, systematic uncertainty quantification remains insufficient, and continental-scale heterogeneity in environmental associations is inadequately recognized, leading to oversimplified global models. We integrate 14,720 recorded landslides with seventeen topographic, geological, climatic, ecological, and anthropogenic variables to produce a seasonally resolved, 10 km landslide susceptibility atlas. Seven ML models were cross-validated; the top two (CNN, GBDT) captured > 75% of events while classifying < 10% of land as very high risk. Areas of greatest model disagreement coincide with clay-rich soils, moderate rainfall belts, and low hills, guiding uncertainty communication. Seasonal retraining reveals a global hazard “see-saw”: monsoonal Asia and Africa peak in summer, whereas the southern Andes, southern Africa, and south-east Australia peak in winter. Continental factor rankings invert common assumptions, with rainfall leading in Asia, clay soils in North America, and road proximity in South America. By incorporating seasonal hydro-ecological data and leveraging multiple advanced machine learning approaches, this work provides a nuanced, region-specific understanding of global landslide hazard patterns. The resulting susceptibility and uncertainty maps serve as an open, reproducible baseline to support disaster risk reduction, climate adaptation planning, and infrastructure development worldwide.