Study on Improving Detection Performance of Wildfire and Non-Fire Events Early Using Swin Transformer

The increasing prevalence of wildfires globally has resulted in significant loss of life and extensive property damage, underscoring the urgent need for improved detection systems. Recent research has prioritized minimizing false fire detections and reducing non-fire alarms using deep learning technologies. However, existing studies predominantly focus on fire detection after significant spread, with limited emphasis on early detection and the differentiation between fire and non-fire scenarios. This study introduces a wildfire detection model employing the Swin Transformer as the backbone for Mask-RCNN, enabling simultaneous detection of bounding boxes and segmentation masks. The proposed model enhances the management of non-fire alarms, reduces false fire detections, and significantly improves early fire detection capabilities. Evaluations conducted across various fire and non-fire scenarios—including clouds, fog/haze, chimney smoke, flames, and wildfire smoke—demonstrated the model’s superior performance. The Swin Transformer-based model achieved a Segm mAP50 of 0.842, surpassing the baseline ResNet-50 model (0.792). Furthermore, it exhibited robust detection across object sizes, with mAP_S, mAP_M, and mAP_L values of 0.166, 0.494, and 0.699, respectively. Confusion Matrix analysis highlighted the model’s reduced false positives for non-fire classes and improved detection of small-scale fire indicators, such as early wildfire smoke trails. These findings underscore the Swin Transformer’s effectiveness in enhancing wildfire detection, particularly during the critical early stages when timely intervention is crucial. Future research will focus on integrating the proposed model into one-stage architectures like YOLO, expanding datasets, and employing data augmentation techniques to further optimize its real-world applicability in wildfire management systems.