# UrbanGraphEmbeddings: Learning and Evaluating Spatially Grounded Multimodal Embeddings for Urban Science Source: [UrbanGraphEmbeddings: Learning and Evaluating Spatially Grounded Multimodal Embeddings for Urban Science]() Collection index: [llms.txt]() Data updated: 2026-02-11T05:00:13.386Z Authors: Jie Zhang, Xingtong Yu, Yuan Fang, Rudi Stouffs, Zdravko Trivic Published: 2026-02-09T07:28:49.000Z ## Abstract Learning transferable multimodal embeddings for urban environments is challenging because urban understanding is inherently spatial, yet existing datasets and benchmarks lack explicit alignment between street-view images and urban structure. We introduce UGData, a spatially grounded dataset that anchors street-view images to structured spatial graphs and provides graph-aligned supervision via spatial reasoning paths and spatial context captions, exposing distance, directionality, connectivity, and neighborhood context beyond image content. Building on UGData, we propose UGE, a two-stage training strategy that progressively and stably aligns images, text, and spatial structures by combining instruction-guided contrastive learning with graph-based spatial encoding. We finally introduce UGBench, a comprehensive benchmark to evaluate how spatially grounded embeddings support diverse urban understanding tasks -- including geolocation ranking, image retrieval, urban perception, and spatial grounding. We develop UGE on multiple state-of-the-art VLM backbones, including Qwen2-VL, Qwen2.5-VL, Phi-3-Vision, and LLaVA1.6-Mistral, and train fixed-dimensional spatial embeddings with LoRA tuning. UGE built upon Qwen2.5-VL-7B backbone achieves up to 44% improvement in image retrieval and 30% in geolocation ranking on training cities, and over 30% and 22% gains respectively on held-out cities, demonstrating the effectiveness of explicit spatial grounding for spatially intensive urban tasks. ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [Geolocation Ranking on New York](): Geolocation Ranking; New York; Hit@5: 58.85. - [Nearest POI Spatial Grounding on SG](): Nearest POI Spatial Grounding; SG; Hit@5: 70.77. - [Nearest POI Spatial Grounding on PA](): Nearest POI Spatial Grounding; PA; Hit@5: 48.85. - [Distance-Direction Spatial Grounding on NY](): Distance-Direction Spatial Grounding; NY; Hit@5: 44.35. - [Distance-Direction Spatial Grounding on SG](): Distance-Direction Spatial Grounding; SG; Hit@5: 34.23. - [Distance-Direction Spatial Grounding on BJ](): Distance-Direction Spatial Grounding; BJ; Hit@5: 58.95. - [Distance-Direction Spatial Grounding on PA](): Distance-Direction Spatial Grounding; PA; Hit@5: 50.38. - [Geolocation Ranking on Singapore SG](): Geolocation Ranking; Singapore SG; Hit@5: 64.23. - [Geolocation Ranking on Beijing](): Geolocation Ranking; Beijing (BJ); Hit@5: 71.92. - [Geolocation Ranking on Paris PA](): Geolocation Ranking; Paris PA; Hit@5: 65.77. - [Image Retrieval on New York (NY)](): Image Retrieval; New York (NY); Hit@5: 65.43. - [Image Retrieval on Singapore (SG)](): Image Retrieval; Singapore SG; Hit@5: 0.711. - [Image Retrieval on Beijing (BJ)](): Image Retrieval; Beijing (BJ); Hit@5: 62.57. - [Image Retrieval on Paris (PA)](): Image Retrieval; Paris PA; Hit@5: 57.22. - [Urban Perception on New York (NY)](): Urban Perception; New York (NY); Hit@5: 41.69. - [Urban Perception on Singapore (SG)](): Urban Perception; Singapore SG; Hit@5: 35.35. - [Urban Perception on Beijing (BJ)](): Urban Perception; Beijing (BJ); Hit@5: 34.17. - [Urban Perception on Paris (PA)](): Urban Perception; Paris PA; Hit@5: 40.59. - [Distance Spatial Grounding on NY](): Distance Spatial Grounding; NY; Hit@5: 29.62. - [Nearest POI Spatial Grounding on BJ](): Nearest POI Spatial Grounding; BJ; Hit@5: 75.38. ## Original sources and implementation [Original paper]() [PDF]()