Urban Image Geolocation via Knowledge-Driven Cue Mining and Reasoning

Traditional geo-localization techniques such as GPS and inertial navigation systems rely either on external signals or on specialized onboard sensors, which limits their applicability in signal-denied or privacy-sensitive environments. Vision-based geo-localization offers an alternative by inferring location directly from images, but most existing methods formulate the task as large-scale visual matching, requiring extensive annotated data and exhibiting degraded performance in complex urban scenes with high visual ambiguity. Moreover, these approaches lack explicit reasoning mechanisms, resulting in limited robustness and poor interpretability. To address these limitations, we propose URBAN-REASON, a knowledge-driven framework for image location reasoning based on visual cue mining and multimodal reasoning. The framework is designed to systematically overcome the weaknesses of conventional image matching approaches through three key components. First, a modular multimodal geo-knowledge base is constructed to associate fine-grained urban visual cues with structured geographic semantics. Second, a dynamic hybrid retrieval mechanism retrieves relevant external geographic knowledge guided by visual cues during inference, reducing reliance on exhaustive visual exemplars. Third, a chain-of-thought-driven hierarchical reasoning strategy decomposes localization into progressive stages of visual parsing, cue association, and location inference, enabling gradual narrowing of the candidate location space. Experiments on a newly constructed dataset of large-scale non-landmark urban street images demonstrate that URBAN-REASON consistently outperforms a representative state-of-the-art image-matching baseline in both fine-grained and coarse-grained localization tasks. In addition, the proposed framework produces explicit and interpretable reasoning chains, supporting more reliable decision-making in complex urban environments.