Real-Time Wildfire Detection and Segmentation using an Optimized Attention U-Net

Forest fires, often caused by extreme weather conditions or electrical failures, pose serious threats to property and human life. Rapid and accurate fire detection is essential for effective disaster response. In this study, we propose a real-time fire segmentation method for drone-captured video streams, balancing accuracy and computational efficiency. Our approach integrates YOLOv4 for fire detection and an Attention U-Net for precise segmentation. YOLOv4 first identifies fire regions using bounding boxes, which are then padded and refined through the Attention U-Net to achieve pixel-level segmentation. Various hyperparameters were evaluated for optimizing performance, including learning rate, batch size, and number of training epochs, with the objective of enhancing training diversity and improving generalization. Our results demonstrate the importance of hyperparameter tuning and model architecture in real-time fire detection. The proposed model effectively identifies large forest fires, improving early detection and response capabilities. This system has the potential to mitigate wildfire damage and save lives by enabling faster and more precise results.