Urban functional zone mapping by coupling domain knowledge graphs and high‐resolution satellite images

Abstract Timely and accurate mapping of urban functional zones (UFZs) is crucial to urban planning and management. Although existing methods for identifying urban functions have made remarkable progress, they still suffer from limitations, such as high sample dependency, insufficient semantic relationship representation, and poor interpretability due to being data‐driven models. To bridge this gap between these methods and the way humans identify functional areas, a new framework that couples domain knowledge and remote sensing images was proposed for the mapping of UFZs. First, to model the concepts, attributes, and spatial and semantic relationships of urban functional objects, a UFZ knowledge graph (UFZ‐KG) was constructed to assist in the mapping of UFZs. Then, the contrastive language‐image pretraining model was adopted to encode jointly the semantic features of UFZ‐KG and the visual features of UFZ images. In this model, a nonlinear embedding module was designed to achieve semantic alignment of these two different modal features in shared space. The effectiveness of the proposed method was verified in three test areas in Shenzhen, China. Results demonstrate that the proposed method of coupling UFZ‐KG with satellite images significantly enhances the UFZ classification accuracy compared to the method that solely relies on image features. Furthermore, it exhibits good generalization performance.