Multi-sensor, multi-scale, Bayesian data synthesis for mapping within-year wildfire progression

As freely available remotely sensed data sources proliferate, the ability to combine imagery with high spatial and temporal resolutions enables applications aimed at near-term disturbance detection. In this case study, we present methods for synthesizing burned-area information from multiple sources to map the active phase of the Elephant Hill fire from the 2017 fire season in British Columbia. We used the Bayesian Updating of Land Cover (BULC) algorithm to merge burned-area classifications from a range of remote-sensing sources such as Landsat-8, Sentinel-2, and MODIS. We created provisional classifications by comparing the post-fire Normalized Burn Ratio against pre-fire image composite within the fire boundary provided by the Province of British Columbia. BULC fused the classifications in Google Earth Engine, producing a cohesive time-series stack with updated burned areas for 19 distinct days. The fire burned unevenly throughout its lifespan: a rapid burn phase of 53,097 ha in two weeks by late July, a steady burn phase to 60,000 ha until late August, an accelerated burn phase of 95,766 ha until mid-September, and containment at 203,560 ha in October. The highly automated methods presented herein can synthesize multi-source fire classifications for active phase monitoring both retrospectively and in near-real-time.