A Geo-consistent Weather Augmentation Method for Improving UAV Localization Robustness under Adverse Meteorological Conditions

Reliable global visual localization of unmanned aerial vehicles (UAVs) in GNSS-denied environments remains vulnerable to fog, snow, and nighttime domain shifts, while generative augmentation can invalidate pixel-level labels by altering object geometry. In this work, we propose Geo-consistent Weather Augmentation (GCWA), a fail-closed Generate–Verify–Filter framework that accepts image-to-image weather transformations only when mask intersection over union (IoU) and the structural similarity index measure (SSIM) applied to edge maps (Edge-SSIM) preserve building geometry and a perceptual-shift criterion confirms a nontrivial weather change. The accepted samples are used to fine-tune the YOLOv11 model and construct a compact structural descriptor for aerial visual place recognition (VPR). In cross-domain segmentation, baseline training degraded from 0.826 mean average precision at an IoU threshold of 0.50 (mAP50) on original images to 0.397 in fog, whereas GCWA mixed-domain training achieved 0.824, 0.815, 0.724, and 0.753 mAP50 on original, fog, night, and snow tests, respectively. On the VPAIR urban subset, YOLOv11+GCWA obtained top-1 recall (Recall@1) values of 0.121, 0.139, and 0.125 in fog, night, and snow and retained 65.8% of clean-condition performance, compared with 46.6% for the same YOLOv11 descriptor without GCWA and 39.4% for MixVPR; MixVPR remained best in clean conditions at 0.210 versus 0.195. Verification required 33.07 ms per candidate, and online segmentation averaged 6.54 ms on the tested desktop graphics processing unit. The results demonstrate that verified geometric consistency is an effective training-data constraint for adverse-weather UAV localization, while real-weather and embedded-platform validation remain necessary.