Deep Learning-Based Fire Segmentation Using U-Net

Wildfires are dangerous threats to ecological systems, infrastructure, livestock, and human life. Therefore, reliable and scalable fire detection methods are required to mitigate their impact. This study proposes a deep learning-based framework for fire segmentation using RGB images. A U-Net convolutional neural network architecture was implemented and trained on a dataset of approximately 11,100 images paired with corresponding pixel-wise binary segmentation masks to accurately segment fire-affected regions under diverse visual conditions. To evaluate generalization, the model’s performance was analyzed across varying dataset sizes (100%, 50%, and 25%) and training durations (10 and 20 epochs). Evaluation was conducted using four segmentation metrics: Dice coefficient, Intersection over Union (IoU), precision and recall. The results demonstrate that the proposed model effectively identifies fire-affected regions, even in complex visual scenes. However, a noticeable decline in performance was observed as the dataset size decreased, emphasizing the importance of large training datasets in deep learning-based applications. Overall, these findings highlight the effectiveness of convolutional neural networks for fire segmentation. This research highlights the potential of models such as the one described above for future applications in wildfire monitoring and related computer vision tasks.