Impact of seasonality on rain-triggered landslide hazard modeling

Abstract

Rain-triggered landslides represent a critical geohazard in regions with intense and seasonally variable precipitation.This thesis investigates the influence of rainfall seasonality on landslide hazard modeling, combining empirical rainfall thresholds and machine learning (ML) approaches in the Paraba Valley and North Coast (PVNC) of So Paulo, Brazil.The study includes a systematic review of 21 scientific publications on ML models applied to landslide hazard assessment, revealing a prevalence of Random Forest, Support Vector Machine, and Gradient Boosting models.Validation of satellite rainfall estimates (IMERG and CHIRPS) against gauge data showed strong correlations (r > 0.75) and acceptable error margins (MAE/day < 2 mm) for use in predictive models.Empirical rainfall thresholds were calculated for each season and source, and five probability classes (R1 to R5) were established based on 1-day, 3-day, and 7-day accumulations.Gradient Boosting classifiers trained with seasonally differentiated thresholds outperformed undifferentiated models, achieving higher AUC values (up to 0.91) and improved precision-recall balance.These findings demonstrate that integrating rainfall seasonality can enhance landslide prediction and supports more effective Early Warning Systems.