Learning to Reason over Multi-Granularity Knowledge Graph for Zero-Shot Urban Land-Use Mapping
Zero-Shot Land-Use Mapping in 1080P: mKGR conducts zero-shot metropolitan land-use mapping across complex urban sectors. The task transforms high-resolution remote sensing imagery into detailed semantic functional zones via multi-granularity spatial-semantic knowledge graph reasoning, transferring seamlessly without manual target labels.
Abstract
Accurate urban land-use mapping is an essential undertaking for various urban issues, such as urban planning, disease transmission, and climate change. Recently, learning-based method has emerged as a prevalent approach for urban land-use mapping, although it relies heavily on abundant labeled data. However, since land-use categories are jointly determined by physical and social attributes, obtaining such labels is challenging. This scarcity of labeled data often leads existing learning-based methods to overfit, resulting in models that struggle to generalize to unseen areas. To address this, we formulate zero-shot urban land-use mapping (ZSLUM) as a new task and develop an innovative method, namely mKGR, which aims to conduct reasoning on a Multi-granularity Knowledge Graph (MKG) for ZSLUM. Specifically, we first construct an MKG that integrates multi-source geospatial data across different levels of spatial units (Grid, Block, Parcel, POI, ROI, and AOI) to comprehensively model urban spatial–semantic structures. Then, we formulate multi-hop reasoning over the MKG as a fault-tolerant relational embedding optimization to infer the land-use categories of target areas without manual training labels. Extensive experiments on representative metropolitan regions and nationwide scale demonstrate that mKGR achieves remarkable zero-shot transfer performance.
Method Overview
The Overall Architecture of mKGR: (a) The end-to-end workflow consisting of Multi-Granularity Knowledge Graph (MKG) construction and multi-hop reasoning; (b) Multi-granularity spatial units hierarchically structured from Grid, Block, and Parcel to POI, ROI, and AOI; (c) MKG schema connecting spatial and semantic entities across heterogeneous relational triples; (d) Fault-tolerant relational embedding and spatial-semantic reasoning for zero-shot urban land-use mapping.
Quantitative Results
| Region | PA (%) | UA (%) | OA (%) | MRR | Hit@1 | Data | Model |
|---|---|---|---|---|---|---|---|
| Wuhan | 47.8 | 47.4 | 71.6 | 85.5 | 78.4 | Zenodo | Zenodo |
| Guangzhou | 56.5 | 49.5 | 79.2 | 82.9 | 74.7 | Zenodo | Zenodo |
| Shanghai | 53.0 | 52.3 | 81.7 | 87.0 | 80.2 | Zenodo | Zenodo |
| Lanzhou | 21.6 | 33.1 | 59.9 | 83.4 | 77.8 | Zenodo | Zenodo |
| Yulin | 30.2 | 35.4 | 63.5 | 77.0 | 68.3 | Zenodo | Zenodo |
| Whole China | 35.6 | 37.9 | 71.1 | — | — | Zenodo | Zenodo |
Knowledge Graph & Mapping Visualizations
Multi-Granularity Knowledge Graph (Wuhan City): Constructed graph topology containing over 120,000 multimodal geospatial entities and 320,000 spatial-semantic relational links, capturing hierarchical interactions from building footprints to functional urban zones.
Zero-Shot Urban Land-Use Mapping: Predicted full-scale land-use classification across metropolitan areas under zero-shot transfer without manual label annotations.
Citation
@article{li2025learning,
title={Learning to reason over multi-granularity knowledge graph for zero-shot urban land-use mapping},
author={Li, Yansheng and Wang, Yu and Yu, Lei and Dang, Bo and Xu, Gang and Zhong, Zhenyu and Wu, Yuning and Guo, Xin and Wu, Kang and Li, Zheng and others},
journal={Remote Sensing of Environment},
volume={330},
pages={114961},
year={2025},
publisher={Elsevier}
}