14 May 2026 in IEEE Transactions on Geoscience and Remote Sensing
Recent advances in cross-view geo-localization (CVGL) serve as a critical foundation for UAVs to deliver reliable remote sensing applications in GNSS-denied complex environments. However, current CVGL methods primarily focus on RGB images, neglecting the potential advantages of incorporating additional modalities. In this paper, we propose DepthGeoLoc, a novel multimodal training CVGL framework that integrates depth-aware structural features to address the significant cross-view disparity challenges. Concretely, we propose two key innovations: i) a Depth-aware Feature Fusion (DAFF) module to enable joint geometry-semantics learning of target landmarks, effectively bridging geometric depth cues with high-level semantic representations. ii) Moreover, inspired by the depth distribution patterns of salient regions in real-world urban scenarios captured from UAV and satellite perspectives, we propose two novel components: Regional Depth Distribution Masking (RDDM) and Target Region Structural Alignment (TRSA). These mechanisms synergistically address the cross-view alignment challenges by leveraging depth correlation priors. Benefiting from our proposed coarse-to-fine knowledge distillation paradigm, we achieves a favorable balance between accuracy and computational efficiency. Extensive experiments demonstrate that our multimodal training framework significantly outperforms the state-of-the-art models. The depth-distilled version of our DepthGeoLoc achieves 1.5× higher throughput (vs. SDPL) and a 52.68% reduction in FLOPs (vs. Safe-Net), while maintaining competitive performance.
30 July 2025 in Applied Soft Computing