Learning to reason over multi-granularity knowledge graph for zero-shot urban land-use mapping

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 recognize diverse land-use categories. To bypass these limitations, this paper for the first time advocates knowledge graph to leverage indirect supervision from related tasks for zero-shot land-use mapping. Toward this goal, this paper introduces a multi-granularity knowledge graph reasoning (mKGR) framework.R Only with indirect supervision from other tasks, mKGR can automatically integrate multimodal geospatial data as varying granularity entities and rich spatial–semantic interaction relationships. Subsequently, mKGR incorporates a fault-tolerant knowledge graph embedding method to establish relationships between geographic units and land-use categories, thereby reasoning land-use mapping outcomes. Extensive experiments demonstrate that mKGR not only outperforms existing zero-shot approaches, but also exceeds those with direct supervision, achieving improvements of 0.16, 0.08, and 0.20 on PA, UA, and OA. Furthermore, this paper reveals the superiority of mKGR in large-scale holistic reasoning, an essential aspect of land-use mapping. Benefiting from mKGR’s zero-shot classification and large-scale holistic reasoning capabilities, a comprehensive urban land-use map of China is generated with low-cost. In addition, a nationwide assessment of 15-minute city walkability over the land-use map provides insights for urban planning and sustainable development. The code and data are available at https://github.com/vvangfaye/mKGR.