A Spatio-Temporal Dataset for Satellite-Based Landslide Detection

Abstract

The capability to accurately detect and monitor landslides is essential for understanding their dynamics and reducing associated risks. However, existing deep learning models often struggle to effectively capture temporal dynamics from satellite imagery, limiting their reliability in analyzing landslide behavior over time. To address this limitation, Sen12Landslides is introduced, a large-scale, multi-modal, multi-temporal dataset designed for satellite-based landslide monitoring and spatio-temporal anomaly detection. Sen12Landslides contains 75,000 landslide annotations from 15 diverse regions globally and over 12,000 patches derived from Sentinel-1 SAR, Sentinel-2 optical imagery, and Copernicus DEM. Each patch includes pixel-level annotations and precise event dates with pre- and post-event timestamps. The dataset supports advanced deep learning approaches, capturing spatial features and temporal changes critical for landslide detection. Benchmark experiments using established models, including U-ConvLSTM, 3D-UNet, and U-TAE, demonstrate the dataset’s utility for landslide detection, with the best-performing model achieving an F1-score exceeding 83% on Sentinel-2 data. By providing this comprehensive resource, Sen12Landslides enables more robust model training and promotes generalization across regions, advancing research in Earth observation and geohazard monitoring.