New techniques for old fires: Using deep learning to augment fire maps from the early satellite era

Abstract

Disturbances such as fire play a critical role in forest ecosystems. However, anthropogenic fires can profoundly impact forests to the point of destabilizing ecosystems. In addition, fires have legacy effects on environments which may be observed in forests for decades after the fire is extinguished. Thus, understanding the extent of historic fires in a landscape is vital to understanding current forest structure and ecological processes (e.g., carbon sequestration capacity and provision of habitat) and, therefore, essential for informing land-management decisions. However, little work has been done to map forest fires pre 1980s due to the challenges of interpreting imagery from the 1970s-era Landsat Multispectral Scanner (MSS) platform. MSS imagery is distinguished from recent satellite missions through lower temporal, spatial, and spectral resolutions. Recent advances in image processing have brought the goal of high-quality MSS classifications within reach. In this study, we use deep learning, specifically UNet (a fully convolutional neural network (CNN)), to detect historic forest fires in MSS imagery for the forest-dominated regions of Quebec, Canada. While other studies have applied deep learning to present-day satellite data for land cover classification, hardly any work has specifically applied deep learning to MSS data for fire detection. We trained our UNet model on 206 MSS images that were labelled by applying thresholds to the Burned Area Index inside polygons drawn by the authors around burned areas. We then used the trained model to label burns in 5104 MSS images that were compiled to generate annual burned area maps. Our results identified (with a 95% confidence interval) 3503.95 ± 484.90 km2 of burns not previously reported in any database; this represents a 35.30 ± 3.94% increase in the total known burned area across the forest-dominated regions of Quebec between 1973 and 1982.