450M On-Board VLM Wildfire Detection Pipeline with Sentinel-2 and LFM2.5-VL
Frontier models can't run on satellites. Here's an end-to-end wildfire detection pipeline using a 450M on-board Vision-Language Model (Sentinel-2 + LFM2.5-VL)
PauLabartaBajo shared a wildfire detection PoC using 450M LFM2.5-VL on Sentinel-2 imagery. It pairs RGB and SWIR tiles, simulates orbit with SimSat, and covers 22 fire-prone sites. The key constraint is bandwidth: on-board inference downlinks only a JSON risk profile.
Why it matters: The author deployed a 450M vision-language model on the satellite itself for wildfire inference and downlinks only JSON risk assessments, sidestepping bandwidth limits on multispectral imagery.