Forest fires: a CatBoost model detects fires from GOES imagery more effectively than the official product, even at night

Trained on more than 300,000 VIIRS detections, the model identified 26 out of 51 historical fires before existing systems did, compared with 6 for the operational product.

Researchers trained a gradient boosting (CatBoost) model to detect wildfires in images from the ABI imager of the geostationary GOES satellites. The training set combines thousands of images with more than 300,000 fire detections from the VIIRS instrument, which is carried on low-Earth orbit satellites that see better but pass through less often.

Evaluated on separate data from five regions, the model outperforms the operational product GOES FDC (Fire Detection and Characterization) in both accuracy and recall, both inside and outside the training area, with higher F1 scores of 0.16 to 0.38. Out of 51 historical fires, it reported 26 before VIIRS and GOES FDC, while the official product was the first only six times.

The largest discrepancy is at night, where the recall of GOES FDC drops to about 0.03, while the learned model continues to correctly detect fires. The authors see this as evidence that machine learning can make geostationary products more accurate, less prone to false alarms, and faster, a valuable gain for fire services.