A Semantic–Spatial Collaborative-Driven UAV View Geolocalization Method

Abstract

Uncrewed aerial vehicle (UAV) view geolocalization plays a crucial role in applications, such as autonomous navigation and precise positioning. However, the significant viewpoint discrepancies among heterogeneous images pose substantial challenges to accurate UAV geolocalization. Existing methods predominantly focus on global semantic feature modeling, while often neglecting the intrinsic coupling between semantic information and spatial structural relationships. Moreover, environmental domain shifts caused by weather variations and temporal changes during flight further undermine the generalization capability of existing methods in complex scenarios. To address these challenges, we propose a dynamic cross-region semantic interaction and fusion network (DCRS) from a novel perspective of semantic–spatial collaborative perception. In addition, we propose a DINOv2-based geographic visual perception encoder (GVPE) and a multi-region collaborative modeling and joint feedback module (MRCM) as the core components of DCRS. In particular, GVPE learns high-level semantic representations that are robust to viewpoint variations and environmental disturbances through multiscale visual modulation and channel reconstruction mechanisms. The MRCM module enhances the model’s cross-region semantic interaction capability and spatial consistency modeling ability by learning the structural relationships among semantic features from different spatial regions. Extensive experimental results on the University-1652 and DenseUAV datasets demonstrate that the proposed method achieves competitive state-of-the-art performance in UAV view geolocalization tasks. Moreover, it preserves stable and robust localization accuracy under dynamic environmental disturbances, offering an effective and reliable solution for the advancement of UAV view geolocalization technologies.