SWUAV-DANet: A Severe-Weather UAV Dataset and Dynamic AlignAir Network for Robust Aerial Vehicle Detection

Unmanned aerial vehicle (UAV) aerial object detection is increasingly important for traffic monitoring, emergency rescue, and environmental perception. However, vehicle detection in heavy rain, dense fog, blizzards, and backlit night scenes suffers from target information loss, feature misalignment, and unstable performance. We, therefore, construct a new severe-weather UAV dataset, Severe-Weather UAV (SWUAV), and propose the real-time Dynamic AlignAir Network (DANet). SWUAV contains 18,195 red–green–blue (RGB) aerial images covering 12 adverse weather/illumination conditions with 236,392 vehicle instances. After the high-resolution backbone features, we insert a cross-scale adaptive alignment module that performs adaptive channel calibration, contrastive self-attention, and geometric/semantic remapping to reduce scale drift/mismatch, suppress noise, and strengthen degraded target cues; we then design a dynamic adaptive alignment head (DAAH) with a shared encoder and a deformable regression branch to mitigate classification–regression mismatch under adverse conditions while further reducing complexity. On SWUAV, DANet raises the YOLOv11-s baseline average precision (AP)/AP50 (AP at intersection over union, IoU = 0.50) from 43.9%/62.6% to 46.9%/64.8%, with only 8.65 M parameters, 22.7 giga floating-point operations (GFLOPs), and a 323.47 frames-per-second (FPS) end-to-end throughput (3.09 ms per image at batch size 16), outperforming EdgeYOLO-s and RT-DETR. The dataset and code are publicly available.