Unsupervised Burn-Attentive Method for Global Burned Areaschange Detection from Remote Sensing Imagery

Abstract

Accurate information over burned areas (BAs) is critical for government post-fire recovery, forest regulation and climate change research. The emergence of moderate resolution imagery provides rich spatial details for global BAs mapping. Traditional rule-based method is less robust while introducing uncertainty and noise in various wildfire. Deep learning-based methods have made huge progress, but these methods rely on plenty of pre-labeled references, while global fires have different geographic environments, burn severity, and regional scales, leading to weak migratory of supervised learning models. In contrast, unsupervised approaches are more applicable in timely post-fire recovery actions. However, BAs exhibit irregular structures and unstable spectral features, unsupervised methods lack the utilization of these prior knowledge. In addition, fine-grained assessment of BAs is less considered. In this research, we develop an unsupervised burn-attentive BAs change detection method.  1) A pseudo label generation algorithm based on uncertainty-aware mechanism is proposed to provide reliable samples in various fire scenarios and improve the generalization performance of the model. 2) To eliminate irrelevant landcover features and efficiently extract the spectral features and multi-scale spatial information of BAs, we design burn-attentive change detection network for extracting BAs. 3) We propose a boundary-constrained land-use/land-cover BAs estimation strategy to provide a basis for post-disaster damage assessment. We analyzed and validated the algorithm on 17 recent fires with different scales, burn severity and backgrounds on a global scale, ours can more precisely delineate BAs and better balance false and missed alarms. Furthermore, compared with publicly coarse resolution BAs products, the Dice coefficient of our products was enhanced. The above demonstrates the robustness of the proposed method for timely mapping and post-fire damage assessment of BAs globally, thus laying the foundation for global climate change studies.

Keywords: Burned areas mapping, remote sensing images, Sentinel-2, Unsupervised deep learning, Burn-attentive change detection, Pseudo label