Enhanced wildfire detection and semantic segmentation via fine-tuned deep learning models

Wildfires pose significant environmental and societal impacts worldwide. Deep learning models in computer vision have emerged as effective tools for improving early wildfire detection and semantic segmentation of flames and smoke. However, the scarcity of wildfire training datasets poses a challenge. To address this, we propose using human-induced tiny wildfire models to generate a comprehensive training dataset comprising 3,185 images. We further evaluate our models on real wildfire events in Gangneung City, South Korea, using 1,415 images. By fine-tuning popular deep learning backbones, including YOLOv8, U-Net(ResNet50), DeepLab(ResNet50), and DeepLab(EfficientNet), we achieve accurate wildfire segmentation with an average accuracy of approximately 90%. Notably, the YOLOv8 model exhibits superior performance in terms of accuracy, speed, and computational efficiency. Through extensive experiments, we demonstrate the robustness of the YOLOv8 model in detecting and segmenting wildfires, even in real-time scenarios using CCTV surveillance footage. Our findings highlight the potential of deploying such models for early wildfire detection and management.