Self-Supervised Cross-View Graph Search Framework for Ground-to-Satellite Geo-Localization

Traditional geo-localization primarily focuses on same-view matching, whereas cross-view geo-localization (CVGL) addresses the challenges of extreme viewpoint disparities, offering greater practical value and demonstrating broad development prospects through tasks such as ground-to-satellite matching. However, current ground-to-satellite matching tasks still face two major challenges:1) significant semantic and viewpoint differences between ground and satellite images, making accurate matching difficult and 2) heavy reliance on large-scale labeled data, limiting scalability, and generalization. To address these challenges, this article introduces a novel self-supervised cross-view graph search (CVGS) framework. First, to bridge the semantic gap between drastically different views, an intermediate drone feature relay mechanism is incorporated to enhance cross-view feature consistency. Second, to tackle the absence of labeled data, a tri-view hard-mined contrastive learning (HMCL) strategy is proposed, where a semantic adapter is trained with pseudo-label constraints derived from dynamically clustered drone features, enabling precise modeling of fine-grained cross-view correspondences. Finally, a k-core-based drone relationship graph construction method is developed during inference to suppress noisy drone data and a graph-optimized path search (GOPS) algorithm achieves stable matching results in practical settings. Without relying on any labeled data, CVGS achieves state-of-the-art performance with minimal parameter overhead, surpassing several supervised baselines. Specifically, it attains 0.89% Recall@1 and 2.80% average precision (AP) on the University-1652 benchmark, and 13.90% Recall@1% and 28.57% AP on our newly introduced CVGL dataset (CVGD).