The frequency of wildfires has recently increased worldwide. Accordingly, studies are underway to reducing the false detection rate of wildfires or to improve the problem of non-fire reports using deep learning technology to prevent the significant loss of life and property damage caused by them. However, previous studies have mainly evaluated the performance of deep learning models in detecting fires based on flame and smoke image data after fire spread and fire detection studies based on various fire events in actual wildfires, such as early- and non-fire reports. In this study, a Swin transformer was applied to Mask-RCNN’s backbone, which can detect the bounding box and segmentation simultaneously to improve the non-fire reports, reduce the false fire detection, and increase the accuracy of early wildfire detection. In addition, different fire events (Cloud, Fog/haze, Chimney smoke, flames, and fire smoke) were configured to improve wildfire detection performance. The fire events were configured around the flames and smoke of the early fire and included non-fire events such as Clouds, Fog/haze, and Chimney smoke that could be Fog/hazeaken for a fire. Test results showed that the existing Mask-RCNN model’s Bbox mAP 75 performance was 0.642, and the Bbox mAP 75 performance of the model applying the Swin transformer was 0.703, indicating an improvement in detection performance of 5% or more. Moreover, the Segm mAP 75 of the existing model was 0.567, while the performance of the model applying the Swin transformer was about 0.640, indicating an improvement in detection performance of about 8%.
Flames and smoke vary in appearance across fire environments, limiting the ability of aggregate detection metrics to characterize model applicability. This study compared lightweight models from ten YOLO versions using an integrated dataset comprising five fire scenarios. Basic image statistics and DINOv2 feature embeddings characterized the scenario groups, while class- and scenario-specific evaluation examined performance differences obscured by aggregate metrics. All models used the same fixed data splits. Aggregate test results were summarized as means and sample standard deviations across three training runs using seeds 0, 21, and 42. The evaluation covered detection accuracy, object-size-specific average precision, inference speed, and fixed-threshold error characteristics. No single model achieved the highest class-specific AP50-95 across all scenarios. Wildland and outdoor fires yielded AP50-95 values of 0.285–0.358 for flame and 0.293–0.392 for smoke, whereas smoke AP50-95 exceeded 0.90 for all evaluated models in electrical equipment and industrial facility fires and in road and vehicle fires. Small objects also remained challenging across models. These findings demonstrate how scenario-stratified evaluation reveals detection weaknesses beyond aggregate rankings and supports baseline selection according to the target environment, object class, and computational requirements.