Using deep learning and multi-source satellite data to refine wildland fire progression estimates with boreal fire drivers

Abstract:

Each year, fire seasons in forested Canada are becoming increasingly variable due to the changing climate. Earth observation data are used to map fires and ignition dates to support monitoring fire season changes over time. While these satellite-derived maps are useful for reconstructing past fire seasons, variability in spatial resolution, temporal scale, and collection objectives can impact the accuracy of each satellite’s estimates. For these reasons, there is value in contextualizing single-source satellite-derived fire estimates with ecological-based drivers known to influence fire conditions, spread rates, and occurrences. To explore the impact that ecological drivers have on space-based fire estimates, we fuse MODIS burned area data (MCD64A1) with forest fuel, topographical, and climatic data using a deep-learning approach. Our first deep learning model predicts weekly regional burned area amount to examine how cumulative burned area varies across ecozones and years across forested Canada. Our second model predicts per-pixel fire ignition dates (day-of-burn) to investigate within-year dates for when regions burn across ecozones. From these deep-learning refined datasets, we estimate spatiotemporal fire season metrics from 2001 to 2020 for forested Canada. By synthesizing fire drivers and burned-area data, we can investigate the relationship between fire drivers and changing fire seasons over the last two decades in Canada. This research supports future advances in fire modelling to use deep learning to reconstruct spatiotemporal fire progressions. The ultimate target of this research is to synthesize multi-source space-based earth observation data to support existing efforts for wildland fire mapping and predictive modelling using ecologically-based fire drivers and burned-area datasets.