Feature selection for landslide forecasting models in Southern Andes
Abstract. Rainfall-induced landslides (RIL) are a major hazard in the Southern Andes, threatening lives, infrastructure, and ecosystems. Early warning systems require accurate predictive models; however, their effectiveness is constrained by heterogeneous data availability and the lack of universal design standards. This study develops a systematic framework to identify the most influential features controlling landslide generation by integrating local soil, climatic, and topographic datasets. A national landslide inventory was expanded using Buffer Control Sampling and PUBagging to improve the representation of non-landslide cases, yielding a robust database of 3148 instances with 136 variables. Feature selection was performed using Classification and Regression Trees (CART) and, in parallel, Genetic Algorithms (GA), with both approaches evaluated using Support Vector Machines, Random Forest, and XGBoost classifiers. Results highlight precipitation, slope, and soil hydraulic properties – particularly bulk density and saturated water content – as recurrent critical predictors. GA-based models significantly outperformed CART, with GA-RF and GA-XGB achieving the lowest error rates (10.95 %) while using compact feature sets. These findings underscore the potential of evolutionary feature selection to enhance predictive accuracy while reducing data complexity, and they provide actionable insights into which variables should be prioritised in monitoring networks. This work contributes to the design of more reliable and region-specific early warning systems for rainfall-induced landslides, emphasising the role of shallow and deep water storage features in mountainous environments.