Abstract
Numerous existing studies have employed slope stability analyses to investigate rainfall-induced landslides and estimate their timing. However, related studies have shown that calculated slope instability, indicated by a safety factor below unity, does not precisely align with actual landslide occurrences, often exhibiting significant temporal discrepancies. To address these limitations, this study not only applies machine learning models but also introduces a novel hybrid workflow that structurally integrates physically based hydrological and slope stability simulations with CNN, DBSCAN, and LSTM architectures. Specifically, CNNs are applied to factor-of-safety (FS) maps derived from physical models, enabling the extraction of geotechnically meaningful spatial features. DBSCAN then organizes these CNN outputs into discrete instability classes, moving beyond binary FS thresholding. Finally, these evolving classes, combined with rainfall and hydrological inputs, are processed by LSTM to capture the temporal delay between rainfall triggers and actual landslide occurrences. This architectural innovation bridges the gap between static susceptibility mapping and dynamic forecasting, offering both interpretability and improved predictive accuracy. The results highlight that this integrated approach significantly enhances the accuracy and reliability of landslide predictions. Furthermore, it facilitates detailed risk prioritization and provides insights into the spatial-temporal variability of landslides influenced by watershed geomorphological characteristics.
Showing the abstract — retrieve the full paper via the Exa API.