Wildfire Risk Prediction Fine-Tuning Tutorial

This introduces an end-to-end pipeline that takes Sentinel-2 satellite imagery as input to identify wildfire risk areas.

• The input consists of two satellite images: RGB (B4-B3-B2) and SWIR (B12-B8-B4).
• The model uses LFM2.5-VL-450M, emphasizing a setup where inference runs onboard the satellite and only a small JSON payload is transmitted to the ground.
• The simulation is built with a SimSat Docker service, a predict.py watch loop, and a local model run via llama-server.
• The system monitors 22 fixed locations, generating one prediction per satellite pass for each location.
• Data labeling uses claude-opus-4-6 to annotate pairs of satellite images, and training data is created by combining locations, spatial tiles, and temporal tiles.
• Example locations include Attica, Greece, Angeles National Forest, Napa Valley, and Mati.

The tutorial covers the full procedure—cloning, running Docker, uv sync, running the prediction watch, backfilling historical data, and running the Streamlit app—making it a reproducible, hands-on resource.