A new AI model can detect fires caused by armed conflict in Sudan using commercial satellite data, flagging affected areas within roughly a day of the event.
Researchers adapted a Variational Auto-Encoder — an unsupervised deep learning architecture that learns what normal conditions look like rather than being trained on labeled examples — to analyze 4-band imagery from Planet Labs satellites at 3-meter resolution. Instead of requiring annotated fire data, the model identifies burn signatures by comparing pairs of images taken at different times and quantifying how much a patch of ground has changed from its learned baseline. In tests across five case studies in Sudan, the approach outperformed three conventional change-detection methods on recall and F1-score — the metrics that matter most when missing a fire is worse than a false alarm. Detection ran between 24 and 30 hours under favorable observational conditions.
The unsupervised design is the quietly important part. Conflict zones are among the hardest places to collect labeled training data — you cannot send field teams to verify burn sites while fighting is ongoing. A model that generalizes from "what land normally looks like" sidesteps that bottleneck entirely. And because it runs on commercially available 4-band imagery rather than specialized sensors, it could be deployed anywhere Planet Labs has coverage without waiting on government satellite access.
The 24-to-30-hour window is promising, but the paper leaves one question unanswered: how often do "favorable observational conditions" — read, cloud-free skies — actually occur over active conflict areas in Sudan, and what the detection timeline looks like when they do not.