Results of the model ablation experiment.

Forest fires pose a significant threat to public safety and the environment, and harmful pollutants spread rapidly in areas covered by vegetation. Early detection is very important for preventing forest fires from evolving into catastrophic fires. The traditional prediction methods have relatively low accuracy. They can only identify fires clearly after they occur, making it difficult to meet the requirements of precise real-time detection. The YOLOv5-PSG model proposed in this paper improves the YOLOv5 model. After 300 rounds of training, the average recognition accuracy rate of mAP can reach 93.1%, and the accuracy rate can reach approximately 0.802. After 300 rounds of training and learning, the confidence level can reach about 0.965. This improvement makes fire early warning and prediction more comprehensive and effective, ultimately protecting human life and the environment by mitigating the impact of wildfires.