Global image geolocalisation via inference with large vision–language models

Images, the primary carrier of information in cyberspace, are widely present across the Internet, capturing environments, scenes, and events from diverse geographical locations, and containing substantial high-value information. However, most images lack explicit geolocation metadata, which notably limits their application potential in critical domains, such as cyberspace mapping, intelligence gathering, and public opinion monitoring. Traditional image geolocation methods are limited by database coverage and lack interpretability. To address these challenges, we propose an image reasoning and localisation dataset that leverages large vision–language models (LVLMs) to perform reasoning-driven image localisation. First, we constructed the GeoLocReason dataset in which each image was annotated with visual reasoning cues at multiple levels of granularity. Second, we designed a two-stage geolocation reasoning pipeline that jointly optimises the localisation accuracy and inference interpretability by explicitly modelling the perception and spatial reasoning stages, thereby enhancing both the model’s understanding of the visual context and its capacity for geographic inference. We evaluated GeoLocInfer on three benchmark datasets—Im2GPS3K, YFCC4K, OSV5M and GAEA-Bench—and demonstrated that our approach accurately identified geoinformative visual features, exhibited coherent and interpretable reasoning behaviour during clue integration, and achieved superior localisation performance across multiple distance thresholds compared with advanced methods and existing open-source LVLMs.