Localization by Fine-Grained Map Retrieval: Geo-Localization of Remote Sensing Images Using Vector Map Tiles

Accurate geo-localization of remote sensing images without explicit geographic metadata remains a persistent challenge in the remote sensing field. Most existing approaches rely on matching query images against pre-referenced imagery; however, these often struggle with dynamic field changes, viewpoint variations, domain gaps, and scale inconsistencies. In contrast, vector maps, which have already undergone feature representation, selection, and generalization, can reduce redundant information and thus offer a valuable alternative. Inspired by map-based registration methods, we propose a novel framework that geo-localizes remote sensing images by retrieving map tiles containing only linear symbols at fine granularity. First, vector data are cartographically converted into multi-scale linear symbol map tiles embedding precise location information. This approach helps minimize overlap and reduce map size. Next, a Dual-Stage Attention Multi-Level Feature Model (DSA-MLM) uses the query remote sensing image to retrieve the most relevant map tiles. Finally, a cluster-based geo-localization strategy estimates the potential geographic location of the query image. Experimental results across various geographic regions, sensor types, resolutions, and map conditions demonstrate that the proposed method achieves high retrieval and geo-localization accuracy, as well as efficiency and robustness in real-world scenarios, highlighting its generalizability and reliability.