Background Low-altitude small object detection faces prominent challenges: unmanned aerial vehicles (UAVs) and birds share high similarity in appearance and motion characteristics, which frequently causes misclassification in detection models. Meanwhile, small targets under long-distance imaging have limited pixel occupancy and insufficient semantic information, and continuous downsampling in deep neural networks further leads to the loss of fine-grained feature information, raising the risks of missed detection and false detection in complex backgrounds.

Purpose This study aims to improve the discriminative ability between UAVs and birds, and boost the detection performance of ultra-small targets in complex low-altitude environments.

Methods A BD-UAV dataset containing real co-occurrence samples of UAVs and birds was established, and controlled variable experiments verified that real co-occurrence samples can facilitate the learning of discriminative features for similar flying targets and effectively reduce cross-misclassification. Based on YOLO11n, an improved GSA-YOLO model is proposed for low-altitude small object detection: a P2 high-resolution detection head is introduced to enhance shallow detail feature representation, a Ghost-SPD module is embedded in the backbone network to mitigate spatial information loss during downsampling, and an ADown structure is applied in the neck network to optimize multi-scale feature fusion.

Results Experimental results indicate that GSA-YOLO reaches 93.3% mAP@0.5 and 59.5% mAP@0.5:0.95, up 3.2% and 2.1% respectively compared with the YOLO11n baseline. For ultra-small targets smaller than 16×16 pixels in complex backgrounds, compared with YOLO11n-P2, its Precision, mAP@0.5 and mAP@0.5:0.95 are improved by 12.57%, 1.82% and 2.11% respectively.

Conclusions The BD-UAV dataset and GSA-YOLO model effectively enhance the UAV-bird discrimination capacity and tiny target detection performance in low-altitude complex scenarios, providing reliable technical support for low-altitude airspace surveillance applications.