The timely and accurate delineation of wildfire-affected regions is critical for effective post-disaster management, strategic planning, and operational response. This paper introduces CA-BAFormer, a connectivity-aware Transformer-based model designed for burned area mapping using uni-temporal Sentinel-2 satellite imagery at 20-meter spatial resolution. Unlike conventional bi-temporal change detection methods, the proposed uni-temporal framework reduces commission errors arising from spatiotemporal heterogeneity in landscape features, while effectively capturing intrinsic burn scar dynamics. To enhance the detection of contiguous burn scars, the model incorporates a semantic connectivity-aware learning (SCL) scheme. This scheme leverages connected components from semantic labels to guide feature extraction and improve segmentation completeness by explicitly modeling the geometric continuity of fire progression through connectivity-constrained learning. By integrating object-contextual representations with Transformer mechanisms, CA-BAFormer demonstrates superior precision in delineating compact burn scars. Quantitative evaluation across 58 major Canadian wildfires shows that the model achieves 82.04% IoU for burned area segmentation, surpassing the uni-temporal UNet baseline (75.83%) and approaching the 82.60% benchmark established by bi-temporal UNet. Notably, this single-image solution provides an automated, high-fidelity, and bias-mitigated capability for large-scale mapping, making it particularly suitable for integration into next-generation Earth observation platforms with constrained onboard processing resources. • Unstructured burned area segmentation from uni-temporal Sentinel-2 imagery. • Semantic connectivity-aware learning scheme to enhance compact segmentation. • 6% IoU increase over uni-temporal UNet and close to the bi-temporal approach. • Comprehensive evaluation against popular CNN and Transformer architectures.
Hu et al. (Sat,) studied this question.