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

Abstract:

This work presents a simplified implementation of a TPFormer-inspired framework for wildfire segmentation and detection using MATLAB. The proposed method processes RGB images by first generating a simulated thermal infrared (TIR) map to enhance fire-sensitive regions. A confidence map is then derived using a sigmoid activation function to highlight potential fire pixels. The algorithm applies a set of strict rule-based conditions, including color, brightness, saturation, and thermal constraints, to accurately identify fire regions while minimizing false detections from non-fire objects such as vegetation or sky. Morphological operations are employed to refine the segmentation mask and remove noise. The system further determines the presence of fire based on a threshold percentage of detected fire pixels. A proxy ground truth is generated for performance evaluation, and metrics such as IoU, F1-score, precision, recall, specificity, and accuracy are computed. The results are visually presented through segmentation masks, overlays, and performance graphs. Overall, the method offers an efficient and interpretable approach for wildfire detection using image processing techniques.

Keywords: Wildfire Detection, Image Segmentation, Thermal Infrared (TIR) Simulation, TPFormer, Fire Pixel Classification, RGB Image Processing, Morphological Operations, Confidence Map, Computer Vision, Performance Metrics (IoU, F1-Score, Accuracy).