YOLO-SODE: enhanced YOLOv11 model for small object detection in UAV imagery

Small object detection in drone images faces challenges such as complex backgrounds, limited feature extraction, and occlusion interference, which significantly impact the performance of existing object detection algorithms. To address these challenges, we propose a small object detection enhanced algorithm based on you only look once (YOLO), called YOLO-SODE, which aims to improve the detection accuracy of small objects. First, an adaptive kernel convolution fused with spatial context awareness is proposed to enhance feature extraction capabilities. Next, a small object enhancement pyramid structure is designed, combining space-to-depth convolution and global-local joint feature extraction. By introducing the P2 layer feature map output, it improves global feature extraction and multiscale feature fusion. Finally, an occlusion-aware attention mechanism is incorporated to improve the accuracy of feature detection in target regions. Experimental results show that YOLO-SODE outperforms the baseline YOLOv11n by 7.6% and 6.1% in terms of mAP50 and mAP50-95, with the parameter count remaining at a relatively low level, on the VisDrone-DET2019 dataset. In addition, generalization experiments on multiple representative datasets, including tiny object detection in aerial images (AI-TOD), indicate consistent gains of about 5–6 mAP50 points over the baseline, further validating the effectiveness of the proposed method. The proposed algorithm provides an effective and adaptable solution for small object detection in unmanned aerial vehicle aerial images.