B-Mamba: Burn severity mapping from large-scale Landsat imagery using a Mamba-based Siamese network

Operational burn severity mapping remains largely dependent on spectral index-based approaches that require manual threshold selection and expert interpretation, limiting scalability and consistency across events. Deep learning methods have improved change detection accuracy, but most have been developed for very high-resolution datasets, reducing their transferability to medium-resolution, globally available sensors such as Landsat. To address this gap of automating the operational burn severity mapping with deep learning using Landsat images, we propose burn severity Mamba (B-Mamba), a Siamese Mamba-based network that exploits the linear-complexity, long-range dependency modeling of state-space architectures while explicitly capturing pre- to post-fire changes through dual-branch feature extraction, hierarchical change fusion, and a progressive reconstruction decoder. On the large-scale Landsat-BSA benchmark dataset, B-Mamba achieves state-of-the-art performance, reaching a mean Intersection-over-Union (mIoU) of 79.31% and an overall accuracy of 95.70%, outperforming leading convolutional, transformer-based, and existing Mamba-based change detection models by 3.71–11.66% in mIoU. Ablation experiments confirm the importance of the learnable patch-expanding decoder, change fusion and the proposed composite loss function for accurate severity delineation. Additional evaluations on independent wildfire events within and outside the contiguous United States (Alaska and Hawaii), including evaluations on Sentinel-2 imagery, demonstrate the robustness and generalization of the approach. These results indicate that Mamba-based state-space architectures provide a scalable, effective pathway toward operational, large-area burn severity mapping from medium-resolution satellite imagery.