Automated Landslide Mapping with Open-Source Satellite Data in GEE: A Sri Lankan Case Study.

Abstract

Rainfall-induced landslides are the most frequent natural hazards in Sri Lanka. This study developed a scalable, cloudbased landslide detection method that integrates satellite data, topographic parameters, and a machine learning framework implemented on Google Earth Engine (GEE) to support rapid, reliable landslide mapping in disaster scenarios and datascarce regions. The methodology was applied to the central highlands and the south-western parts of Sri Lanka. Multitemporal Sentinel-1 and Sentinel-2 images, along with a 12.5 m digital elevation model, were used to generate four composite datasets (M1–M4) for pixel-based classification. M1 includes spectral index differences between pre-event and post-event optical images, M2 incorporates spectral distance and spectral angle data, M3 contains topographic attributes, and M4 combines all datasets with synthetic aperture radar (SAR) amplitude ratio bands. Inter-area model training and validation were conducted using RF, CART, SVM, and GTB with independent datasets from study area-1, followed by transferability testing in a second area. Model performance was evaluated using overall accuracy, Precision, Recall, F1- Score, Kappa, and AUC. SVM-M4 performed best in interarea analysis, while GBT–M4 and CART–M3 showed supe Cloud contamination and false detections remained key challenges.