Direction-Robust UAV–Satellite Cross-View Geo-Localization via Spatial Adaptive Alignment and Circular Adjacent-Direction Gated Interaction

Cross-view geo-localization (CVGL) matches unmanned aerial vehicle (UAV) images with geotagged satellite imagery to support visual localization in Global Navigation Satellite System (GNSS)-challenged environments. However, substantial cross-view discrepancies between UAV and satellite images, together with variations in UAV viewing direction, induce pronounced shifts in feature distributions, thereby undermining the stability of cross-view matching. To address these issues, we propose a progressive spatial-to-directional framework. The Spatial Adaptive Alignment Module first adaptively balances height–channel and width–channel responses and applies a fine-grained consistency constraint to reduce cross-view feature discrepancies. On the spatially adapted features, the Direction-Aware Module performs radial directional decomposition and organizes the resulting directional features into a circularly ordered sequence. A channel-wise gated interaction mechanism then selectively transfers information between each direction and its circularly adjacent neighbors after semantic-space projection. This local interaction captures gradual structural transitions between neighboring directions while preserving direction-specific discriminability and avoiding indiscriminate global mixing. Across three independent training runs on University-1652, our method achieves mean R@1 and AP values of 94.63% and 95.46% for UAV-to-satellite retrieval. The corresponding means for satellite-to-UAV retrieval are 96.86% and 93.97%. These results demonstrate that our method effectively improves UAV–satellite cross-view retrieval performance and maintains stable retrieval performance under different UAV viewing directions.