TransFireNet: A novel dual-branch transformer network for burn severity estimation using bi-temporal Sentinel-2 imagery

Most burn severity mapping faces challenges in effectively capturing the complex spatiotemporal characteristics of pre- and post-fire imagery, leading to difficulties in accurately assessing the varying degrees of fire impact. In response, we aim to develop an efficient deep learning architecture that leverages bi-temporal Sentinel-2 images for accurate burn severity estimation. To this end, we construct a new burn severity dataset comprising 45 wildfire events across 8 European countries, using bi-temporal Sentinel-2 images and reference data from Copernicus Emergency Management Service products. Then, we propose a novel dual-branch transformer-based architecture, referred to as TransFireNet, which employs Pyramid Vision Transformers for feature extraction from pre- and post-fire images and a convolutional decoder for fine-grained spatial detail preservation. The proposed model incorporates multi-scale feature differences between image pairs through progressive feature fusion, enabling the model to focus on temporal changes while maintaining spatial context. Experimental results demonstrate that TransFireNet achieves the highest performance among state-of-the-art methods, evidenced by superior F1-score and IoU, particularly for severely damaged areas. In addition, the proposed network maintains computational efficiency with significantly fewer parameters than other transformer-based models. This study offers a robust solution for burn severity assessment that can effectively support post-fire management and ecological restoration efforts across diverse geographical regions.