Rainfall-Induced Landslide Detection in the Central–Southern Andes of Chile Using Integrated SAR, Optical, and Machine-Learning Approaches in Google Earth Engine

Abstract

This article presents a framework for postevent detection of rainfall-induced landslides in the south-central Andes of Chile by combining Sentinel-1 C-band synthetic aperture radar backscatter changes (∼10-m resolution), Sentinel-2 multispectral variables (10-m resolution), and advanced land observing satellite PALSAR-derived topographic attributes within a machine-learning workflow implemented in Google Earth Engine. The framework comprises three stages: 1) multisource feature extraction; 2) classification using random forest (RF), gradient tree boosting (GTB), support vector machine (SVM), classification and regression trees (CART), and Naïve Bayes (NB), evaluated through nested leave-one-area of interest-out spatial cross-validation; and 3) probability mapping. The approach was applied to landslides triggered by an extreme rainfall event between 20 and 25 June 2023, in the Andean sector of the Biobío Region. Under the reference configuration, RF achieved the best performance (F1 = 0.883, Precision = 0.947, Recall = 0.829), followed by GTB (F1 = 0.869); CART, SVM, and NB reached F1-scores of 0.825, 0.822, and 0.753, respectively. Sensitivity analyses showed strong effects of temporal sampling and optical data quality. The 7-day postevent scenario yielded spatially unreliable results (F1 = 0.463), whereas a 50-day period improved performance (F1 = 0.650). Restricting Sentinel-2 composites to the dry season improved results relative to longer periods affected by cloud and snow contamination (F1 = 0.883 versus 0.818). Overall, normalized difference vegetation index percentage change combined with RF provided the most robust performance for operational landslide probability mapping.