Advancing Canadian wildfire technology through onboard processing and on the ground collaboration

Abstract. Wildfire regimes are changing at a time when new satellite platforms can run machine learning analysis to detect and refine wildfire data products onboard, allowing explorations of new concepts of operations. Combining wildfire operations and science expertise with indigenous community knowledge and a technology demonstration satellite designed to deploy and test machine learning algorithms on orbit we characterise Fire Band Analysis Networks (FireBAN) models for wildfire detection from optical, multispectral, and hyperspectral data. Testing on a custom training dataset from Sentinel-2 and AVIRIS data prior to launch of the onboard models shows a test accuracy of 95.1% and mean intersection over union (mIoU) of 64.6% on Sentinel-2 and accuracy of 94.6% and mIoU of 72.6% for the best performing model architecture, which shows band-dependency effects under an ablation study. On-orbit performance demonstrates the ability to detect series of wildfire smoke plumes and gather data necessary to retrain models on the ground to counter sensor-based domain shifts and provide continuous updates of model weights to the satellite platform.