Event-based rainfall windows and topographic controls on landslide susceptibility in West Sumatra: A machine-learning analysis

Landslides caused by rainfall are a chronic cause of land degradation in tropical mountainous areas, where steep terrain and heavy precipitation work together to destabilize slopes and undermine land usability. This study elucidated the combined effects of event-based rainfall windows and topography on the landslide susceptibility in West Sumatra, with the view of supporting degraded land management and mitigation planning. Daily rainfall data from CHIRPS were used with 137 landslide events (2014 to 2024) and an equal number of non-landslide points to create a balanced dataset based on the 7 x 7 m DEMNAS. Geomorphological predictors included slope, aspect, profile curvature, and plan curvature, whereas rainfall was measured at 0, 1, 3, 7, 14, and 30 days. The 3-day and 7-day rainfall windows, slope, and the profile curvature were identified as the most discriminative variables using the Mann-Whitney U test. Afterwards, logistic regression, random forest, and XGBoost models were built, and each achieved high predictive accuracy (AUC>0.93; AP>0.95). The feature importance and SHAP analyses consistently showed that slope was the most influential control factor, with short-term rain windows making a meaningful marginal contribution. Subsequent susceptibility maps have consistently identified the Bukit Barisan range as a high-risk area. This research also shows how event-based rainfall-terrain models can be operationalised to inform degraded land management by focusing mitigation efforts, zoning land use, and supporting rainfall-based early warning strategies in data-sparse tropical areas.