A multimodal data fusion framework for urban functional zone mapping based on local-global information enhancement and adaptive spatial units: from standard datasets to global city validation

Accurate Urban Functional Zone (UFZ) mapping is a fundamental prerequisite for analyzing the spatial structure of cities and optimizing urban planning. Very High Resolution (VHR) remote sensing image and crowd sourcing geospatial data have emerged as critical data sources for UFZ mapping due to their capability to represent multidimensional spatial semantics. However, current research on UFZ mapping faces several limitations, including insufficient modeling of both global and local features in VHR image, unclear spatial units for functional zones, and a lack of cross-city validation. To address these issues, this study proposes a UFZ mapping framework that integrates VHR image and crowd sourcing geospatial data. First, a Global-Local Scale Attentive Network (GLSANet) is developed to extract deep functional semantic features from VHR image by modeling both local and global contextual information. Second, a Multi-scale Road-constrained Feature Segmentation (MRFS) algorithm is proposed, which incorporates OpenStreetMap (OSM) road constraints. By integrating geometric shape constraints with functional semantic features, this algorithm enables the generation of spatial units for functional zones. Finally, within the constructed spatial units, the framework integrates multiple data features, including low-light remote sensing data from SDGSAT-1, 3D building morphology features, POI kernel density features, and deep functional semantic features, and employs an ensemble learning model to classify urban functions. To validate the effectiveness of the proposed framework, we conducted UFZ mapping studies in 25 cities worldwide. The results demonstrate that the proposed framework exhibits strong adaptability across diverse urban regions, achieving an average overall accuracy of 0.906 in UFZ mapping. In addition, comparative analyses with other related methods further highlight the superior performance of the proposed framework.