Deep-learning-based burned area mapping using the synergy of Sentinel-1&2 data.

Around 350 million hectares of land are affected by wildfires every year influencing the health of ecosystems and leaving a trail of destruction. Accurate information over burned areas (BA) is essential for governments and communities to prioritize recovery actions. Prior research over the past decades has established the potentials and limitations of space-borne earth observation for mapping BA over large geographic areas at various scales. The operational deployment of Sentinel-1 and Sentinel-2 constellations significantly improved the quality and quantity of the imagery from the microwave (C-band) and optical regions on the spectrum. Based on that, this study set to investigate whether the existing coarse BA products can be further improved by the synergy of optical surface reflectance (SR), radar backscatter coefficient (BS), and/or radar interferometric coherence (COR) data with higher spatial resolutions. A Siamese Self-Attention (SSA) classification strategy is proposed for the multi-sensor BA mapping and a multi-source dataset is constructed at the object level for the training and testing. Results are analyzed by test sites, feature sources, and classification strategies to appraise the improvements achieved by the proposed method. • Sentinel-1 and Sentinel-2 data are synergistically used for burned area mapping. • A small object dataset is constructed over 12 California fires in 2018–2019. • A deep-learning-based method is applied to discriminate burns and unburns. • Performance is assessed over different data sources, test sites, and classifiers. • The flexibility of the trained classifier is evaluated over independent sites. [ABSTRACT FROM AUTHOR]