Research on YOLOv8-based forest fire detection with adaptive feature fusion

Frequent forest fires represent a significant danger to both ecosystems and economic stability. Helicopter bucket firefighting is widely used because of its flexibility and efficiency, but its effectiveness largely depends on accurate fire detection under complex conditions. Therefore, a multi-scale forest fire detection model ADT-YOLOv8 based on YOLOv8 is proposed. The model combines Diversified Branch Block (DBB) for improved feature learning, DBB-C2F structure for efficient cross-dimensional feature enhancement, lightweight triple attention mechanism for adaptivemulti-scale feature fusion, and ASFF-Head for adaptive multi-scale feature fusion. Compared with YOLOv8s, ADT-YOLOv8 improves the precision by +3.0%, the recall by +2.7%, the performance of [email protected] by +2.5%,and the performance of mAP@(0.5:0.95) by +2.6%. The proposed approach notably enhances the accuracy of forest fire detection while minimizing the rate of false alarms, and it also shows strong effectiveness when applied to real-world forest fire situations.