Detection of extreme rainfall-induced landslides in mountainous areas using ensemble machine learning and Sentinel-2 satellite image analysis
TL;DR
This study applied ensemble machine learning approaches using satellite imagery acquired from the Sentinel-2 satellite to detect landslide locations using the XGBoost model, demonstrating excellent predictive capability for all models.
Abstract
Landslide detection is a fundamental prerequisite for subsequent investigations, including landslide susceptibility mapping and landslide risk assessment. In mountainous regions with complex terrain conditions, this task is particularly challenging due to the limited accessibility of landslide sites for direct field investigation and data acquisition. Recent advances in remote sensing image analysis have enabled efficient detection and extraction of landslide information through the analysis of changes in land cover and topographic characteristics derived from pre- and post-event imagery. Nevertheless, the application of remote sensing techniques still faces several limitations, particularly in discriminating actual landslide areas from vegetation loss caused by anthropogenic activities such as deforestation and infrastructure construction. To overcome these limitations, this study applied ensemble machine learning approaches using satellite imagery acquired from the Sentinel-2 satellite to detect landslide locations. Pre- and post-event satellite images associated with the extreme rainfall-induced landslide events that occurred in late October 2020 in the mountainous region of Phuoc Son were collected and analyzed. In addition to satellite imagery, several conditioning factors, including slope and rainfall, were incorporated into the models to minimize the misclassification of non-landslide areas as landslides. The training and validation results obtained from three models, namely Logistic Regression, Random Forest, and XGBoost (eXtreme Gradient Boosting), demonstrated excellent predictive capability for all models. Among them, the XGBoost model achieved the highest predictive performance, with an AUC value of 0.999 on the validation dataset. Subsequently, the XGBoost model was employed to identify landslide locations across the entire study area. Validation based on field survey data and Google Earth imagery further confirmed the high effectiveness and reliability of the proposed approach.