UPUS-18: A large-scale geospatial vector dataset of urban public underground spaces across 18 major Chinese cities
Description
UPUS-18 is a large-scale geospatial vector dataset of urban public underground spaces covering 18 major Chinese cities: Beijing, Shanghai, Guangzhou, Shenzhen, Chengdu, Chongqing, Hangzhou, Nanjing, Wuhan, Suzhou, Xi'an, Shenyang, Qingdao, Zhengzhou, Tianjin, Hefei, Ningbo, and Changsha. The dataset contains 41,840 underground spatial objects across three categories: underground parking facilities, metro stations, and underground commercial spaces. Urban public underground spaces are defined as underground spaces that are publicly accessible and provide public services. The dataset includes spatial boundaries, areas, functional categories, and associated attributes. Metro stations are further divided into station halls and access corridors; underground commercial spaces include floor-level information; and underground parking facilities include estimated underground levels and parking capacities. All data are provided in Shapefile format.
Files
Steps to reproduce
The dataset was developed using a human-in-the-loop (HITL) iterative mapping framework that integrates multiple data sources, including online maps, points of interest (POIs), publicly available web information, building footprints, road networks, metro lines, and built-up area boundaries. GIS spatial analysis, machine learning, and human–AI collaborative rule refinement were employed to transform fragmented digital information into underground spatial vector data with geometric boundaries and associated attributes. Separate identification models were developed for the three categories of underground spaces. For underground parking facilities, an AI agent was used to retrieve parking capacity information. Combined with POI locations, aboveground buildings, and urban block boundaries, underground geometries and capacities were estimated through building matching, capacity constraints, and horizontal and vertical expansion. For metro stations, initial geometries were generated based on the spatial relationships among metro lines, stations, and entrances. LabelMe-annotated samples and a random forest algorithm were then used to predict the spatial extents of station halls and access corridors separately, followed by geometric and topological optimization. For underground commercial spaces, POI semantic filtering and floor-level grouping were combined with DBSCAN clustering and the AlphaShape algorithm to reconstruct spatial boundaries. Building footprints, road networks, and address information were further incorporated to refine the identification results. Subsequently, researchers collaborated with large language models to diagnose identification errors and translate proposed improvements into executable rules, progressively refining the identification models through multiple iterations. The results were evaluated using manually annotated reference samples, bootstrap statistical tests, and geometric and topological validation. This process ultimately produced a geospatial vector dataset of urban public underground spaces covering 18 Chinese cities.
Institutions
- Southwest Jiaotong UniversitySichuan, Chengdu