Forest Fire Detection Method Based on Multi-Scale Feature Fusion and Attention Residual in UAV Perspective

Abstract

With the rapid advancement of drone technology and deep learning, drone-based object detection has become increasingly important for forest fire monitoring. However, significant challenges remain, particularly in accurately detecting the initial fire stage (small-scale flames) in aerial images. To address these challenges, this paper proposes YOLOv8s-MS, a model based on multi-scale feature fusion and attention residual, from the perspective of drones, aiming to enhance the performance of forest fire detection. First, a multi-scale feature fusion (MSFF) strategy is adopted to efficiently fuse feature maps of different scales. On this basis, a shallow detection head is introduced and the deep detection head is removed, thereby enhancing the perception ability of small targets and reducing the computational cost. Second, a SKAttention-based residual module (SKRM) is proposed by integrating the residual design into the Res-RepViT block and combining the efficient SKAttention mechanism to achieve adaptive adjustment of the receptive field size, thus more accurately capturing multi-scale features in complex image space and improving the accuracy of feature extraction. Finally, a systematic experimental evaluation was conducted on the FLAME&FLAME2, and fire datasets. The results showed that the proposed method achieved a significant performance improvement in the flame object detection task, with mAP of 81.8% and 95.7%, respectively, which were 10.1% and 4.2% higher than the baseline model. In addition, in the generalization experiments conducted on the foggia and firedetector public datasets, the overall performance of this method ranked second, further verifying its good generalization ability and adaptability.