DynamicSwin-Fire: A Dynamic Swin Transformer with Fire-Specific Attention and Edge-Guided Gradient Fusion for Robust Fire Segmentation

Accurate and precise identification of forest fires is essential to minimizing environmental and economic loss. Existing fire segmentation methods are prone to struggling with uncertain fire edges, scale-varying fires, and complex background clutter. The proposed model overcomes these challenges with the integration of edge-aware fusion, adaptive attention, and dynamic window scaling mechanisms for better localization and segmentation quality. In this paper, we propose a DynamicSwin-Fire, a transformer-based segmentation model for accurate fire area segmentation of challenging forest environments. Based on the Swin Transformer backbone, the proposed model consists of four domain-aware modules: Gradient-Guided Feature Fusion (GGFM), Fire-Aware Attention (FAA), Dynamic Window Scaling (DWS), and a Fire Region Confidence Estimator (FRCE) to sharpen fire boundary detection, contextual correctness, and adaptive spatial modeling. We also enhance Near-Infrared (NIR) image channels wherever available, which assist in fire vs vegetation discrimination, further improving model robustness in ambiguous areas. The model is trained using a composite loss function of BCE and Dice Loss, with objectives of pixel-level accuracy and region-level overlap. DynamicSwin-Fire Evaluated on a benchmark fire segmentation dataset, the proposed model achieves precision of 86.87%, recall of 89.22%, F1-score of 91.02%, and IoU of 82.31%, outperforming state-of-the-art baselines by margins of 1.56% in precision, 9.52% in F1-score, and 0.87% in IoU. These gains reflect the model’s capacity for segmenting fire areas with higher accuracy and reliability, making it a valuable module of real-time wildfire monitoring systems.