Spatiotemporal clustering for landslide hazard prediction revealed by the path-dependent perspective

Traditional landslide hazard predictions often overlook the spatiotemporal clustering of events, resulting in static predictions that fail to capture the dynamic nature of landslide occurrences. To address this limitation, this study proposes a novel spatiotemporal framework that incorporates path-dependency into landslide hazard prediction. Using a 23-year landslide dataset (2002–2024) from the Renewable Energy Transmission Corridor (RETC) in southwestern China, Ripley’s K-function was used to determine the optimal temporal clustering window for landslide events. Based on this, two ensemble frameworks were introduced to assess the spatial probability of landslide occurrence across varying spatiotemporal clusters. Finally, the spatiotemporal evolution characteristics of landslide hazard were investigated by considering the interactions among spatial, temporal, and magnitude probabilities. The results indicate that a 10-year temporal clustering window best captures the time-sensitive nature of landslide path-dependency in the study area. Dynamic susceptibility modeling demonstrates the superior performance of the Blending ensemble framework (AUC > 0.9 for all periods) and illustrates the spatial evolution of landslide susceptibility across different levels under path-dependency. The integrated hazard assessment results effectively reveal the spatiotemporal distribution of landslide hazard of varying temporal occurrences and magnitudes, providing insights informed by path-dependency for realistic landslide hazard management. The proposed framework strengthens hazard zoning, enhances infrastructure resilience, and offers a scalable solution for dynamic risk management from the perspective of climate change.