&Segloc: Dual-Decoder-Based Semantic-Enhanced Remote Sensing Localization Method

Cross-view remote sensing geolocation is among the most widely adopted solutions in GNSS-denied environments. The primary challenge lies in learning highly discriminative feature representations to bridge the domain gap between heterogeneous image sources. While existing single-encoder-single-decoder methods have made progress, they still have significant limitations: first, excessive reliance on deep network features leads to the loss of geometric information; second, the single decoding task limits the feature representation capacity. To address these issues and better leverage semantic information, this paper proposes a novel plug-and-play single-encoder dual-decoder collaborative learning framework. By introducing an auxiliary semantic decoder and a new semantic matching strategy, the proposed approach overcomes the performance bottleneck of traditional classification networks. Additionally, to address the scarcity of semantic annotations in cross-view remote sensing data, we have constructed a new semantically labeled dataset. To our knowledge, this is among the first studies in cross-view geolocation to unify decoupled feature learning with multi-task collaboration and to introduce multi-label semantic segmentation into the cross-view matching framework, providing a new research paradigm for cross-platform geolocation. Experimental results on public benchmarks demonstrate that the proposed method achieves substantial performance improvements and attains excellent performance in UAV-view target geolocation tasks, validating its effectiveness and robustness. Moreover, the method exhibits strong adaptability to challenging conditions, including variations in viewpoint, illumination, scale, and rotation.