Fusion of AI, Machine Learning, and Remote Sensing for Proactive Landslide Detection

Abstract

Landslides are one of the most destructive natural hazards in the Upper Mahaweli Catchment (UMC) of Sri Lanka, causing loss of life and infrastructure damage, particularly during the Northeast Monsoon season. Traditional detection methods are predominantly reactive and lack spatial precision for proactive disaster management. This study integrates Machine Learning (ML) algorithms with remote sensing-derived geospatial data to develop a proactive landslide susceptibility framework. Three ML classifiers – Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Logistic Regression (LR) – were trained on a National Building Research Organization (NBRO) static dataset and a temporal event-based dataset, balanced using Synthetic Minority Oversampling Technique (SMOTE). Twenty-two conditioning factors encompassing topography, hydrology, soil properties, land use, and proximity to infrastructure were derived from multi-source satellite imagery and Geographic Information System (GIS) processing. The best-performing model, Random Forest, achieved a Receiver Operating Characteristic–Area Under Curve (AUC-ROC) of 0.8763 according to the NBRO static data, while XGBoost attained 0.824 for the Event – Based data. The five most influential factors of the landslide occurring were slope, Distance to Railway, Distance to road, rainfall, and LST. Results confirm ML – remote sensing fusion can deliver scalable proactive landslide risk maps for tropical highland catchments.