Satellite-based near-real time Burnt Area Segmentation of Wildfires: An Evaluation of the Algorithm's Efficiency with a focus on Portugal

Abstract

Wildfires continue to exert a profound impact on our planet causing damage to infrastructure, human lives, and the environment highlighting the need for fast fire assessment. The German Aerospace Center has developed an algorithm that derives satellite-based near-real time burnt areas (BA). This study compared the algorithm's output with reference layers to assess its geometric accuracy. For Portugal, 22 Sentinel 3 scenes each were used for August 2018, August 2019 and August 2020, totaling 66 scenes. Reference layers were created per scene based on a Portugal ground truth and an EFFIS ground truth combined with the Geotiff of each scene. The algorithm's output and the reference layers were then used to calculate the general Intersect over Union (IoU) indices, IoU regarding ESA landcover (trees, shrubland, grassland, cropland) and IoU regarding BA size (small, medium, large). The Portugal and EFFIS based references were very similar and for the remainder of the study only the former reference was used to reduce redundancy. The process showed false detections occurring mainly around water bodies (water-sun glint) and darker features (likely valley bottoms).