TPFormer: Robust Wildfire Segmentation via Thermal Prior Integration and Dual-Decoder Supervision

Wildfire identification aims to achieve robust monitoring under all weather conditions. Although significant progress has been made in segmentation methods based on dual-modal Optical (RGB) and Thermal Infrared (TIR) imagery, these methods often struggle to effectively suppress modal noise in complex scenarios (e.g., light reflections) and fail to achieve refined boundary reconstruction due to the lack of explicit thermodynamic physical guidance. To address these challenges, this paper proposes an end-to-end segmentation network named the Thermodynamic Prior-Guided Transformer (TPFormer). The proposed architecture integrates two key modules: (1) a Hybrid Attention Fusion Module, which effectively calibrates cross-modal features and suppresses non-thermal interference by combining the global modeling capability of linear transformers with the local detail perception of CNNs in parallel, while incorporating a confidence gating mechanism; and (2) a TIR-Prior Gated Decoder, which utilizes raw thermal radiation intensity as prior cue to guide the upsampling process, thereby effectively eliminating visual artifacts and sharpening edge details. Experimental results demonstrate that the proposed method achieves remarkable performance on public wildfire segmentation datasets, reaching 79.72% in Intersection over Union (IoU) and 88.72% in F1-Score. This represents IoU improvements of 3.32% and 1.56% compared to the classic RTFNet and the state-of-the-art RegionNet, respectively. Furthermore, the method achieves an effective balance between inference speed and accuracy, validating its effectiveness in complex scenarios.