MultiFireSeg: A Multitask Network With Interactive Prompts and Boundary Supervision for Wildfire Segmentation

Under the combined impacts of global climate change and human activities, extreme disasters such as wildfires have become increasingly frequent, resulting in severe casualties and property losses. Wildfire remote sensing monitoring typically involves multiple targets—flames, smoke, and burned areas—that exhibit substantial differences in spatial morphology and require costly fine-grained annotations, thereby limiting the performance of conventional semantic segmentation networks in multi-task dis aster information extraction. To address these challenges, we pro pose MultiFireSeg, a multi-task wildfire semantic segmentation network that integrates a multi-class expert prompt encoder–de coder framework with boundary-guided mechanism. Specifically, MultiFireSeg adopts GLMamba as the backbone, it employs CNN–Mamba collaborative architecture to efficiently model local details and global contextual features while maintaining controlled computational complexity. To achieve effective multi-task feature fusion and adaptive target modeling, we design a semantic–spatial prompt interaction module to efficiently fuse multi class expert prompt features with visual features. In addition, a dynamic weighting module is introduced to adaptively regulate the importance of different target features under multi-task wild fire settings. Finally, a boundary extraction module is incorporated into the decoder branch to enhance fine-grained boundary delineation in multi-task wildfire semantic segmentation. Extensive experiments on multi-source wildfire remote sensing datasets demonstrate that MultiFireSeg consistently outperforms baseline methods across flames, smoke, and burned-area tasks, achieving average gains of at least 1.6% on key evaluation metrics. Further-more, evaluations on wildfire generative and cross-domain datasets confirm its strong generalization capability and robustness under complex and unseen conditions. Overall, MultiFireSeg enables accurate multi-class wildfire damage mapping from remote sensing data with reduced annotation cost, offering a practical reference for timely fire extraction across diverse scenarios.