Landslide Reference Data (Sentinel-1 & Sentinel-2)

This dataset contains a curated landslide inventory compiled to support supervised machine learning applications, particularly deep learning models such as U-Net. It includes labeled reference data for large landslides that occurred worldwide from July 2015 onwards, enabling the availability of both Sentinel-1 and Sentinel-2 satellite data for each event.

The majority of landslides were identified through reports on The Landslide Blog and subsequently verified via visual interpretation of pre- and post-event Sentinel imagery. An exception is the dataset subset from southern Kyrgyzstan, based on previous studies and manual corrections.

The inventory features landslides of varying sizes and types across diverse environmental and climatic conditions, with rainfall-triggered events being the most common. It encompasses both single-event landslides and clusters of multiple landslides within a single event.

Update: The reference dataset (h5 files) was updated to a newer version that contain “IDs_order” attribute. The readme with more details will follow.