wtf is impactmesh?

ImpactMesh is a dataset created by IBM, the German aerospace center DLR, and the European Space Agency's Φ-lab to help predict and map floods and wildfires. It covers over 400 disaster events around the world, with multiple satellite snapshots taken over time for each event. The goal is to give researchers and builders a rich, ready-to-use source of data to train AI models that can identify where fires or floods are spreading. The dataset pulls together several types of satellite and terrain data: radar images from Sentinel-1, regular optical imagery from Sentinel-2, and elevation data from the Copernicus DEM. Each event comes with four observations captured at different times, so a model can learn how a situation changes over the course of a disaster. The flood and wildfire portions are available separately on Hugging Face, and the repository itself provides Python code to load the data or fine-tune an existing model using a framework called TerraTorch. People who would use this include climate scientists, disaster response teams, or startups building geospatial AI tools. For example, an organization could take the wildfire dataset, fine-tune a model to recognize burn scars from satellite imagery, and then deploy that model to assess damage after new fires. Similarly, flood mapping models trained on this data could help agencies understand flood extents and prioritize rescue efforts. The data is released under a permissive license (CC-BY 4.0), making it broadly accessible for research and commercial use. The project is built around PyTorch and integrates with TerraTorch, which handles the training and prediction pipeline. The README includes ready-to-run configuration files for fine-tuning and instructions for running predictions on new data. Notably, prediction can be done with or without elevation data, and the repo provides an alternative inference script to avoid memory issues when processing large images. The dataset itself uses a format called Zarr, which is designed for efficiently storing and reading large arrays of scientific data.