An enhanced feature fusion method for urban functional zone mapping with SDGSAT-1 day-night imagery and multi-dimensional geospatial data

Abstract

Urban functional zones (UFZs) are well-planned spatial units characterized by distinct socioeconomic activities and composite land uses, such as residential areas, industrial zones, and blue-green spaces. Fine-grained UFZ mapping has played an increasingly crucial role in supporting targeted urban renewal and transformation of development mode in megacities, facilitating spatial structure optimization to enhance urban livability and sustainability. Prior UFZ mapping methods that focus on two-dimensional (2D) features of point of interest and multi-spectral imagery, pay little attention to three-dimensional (3D) features of building height and digital surface model, mostly with the absence or underutilization of emerging nighttime light imagery. Given the availability of high-quality day-night spectral signatures provided by the Sustainable Development Science Satellite 1 (SDGSAT-1) in a single sensor observing mode, it has become possible to effectively perform UFZ mapping with day-night feature enhancement. In this study, we proposed a progressive and cross-scale deep fusion architecture for generating UFZ maps at the block scale, enhancing spectral and spatial information through sequential refinement—from feature representation and relationship extraction to context modeling. To verify the effectiveness and generalizability of the proposed method, experiments were conducted in two Chinese megacities with distinct UFZ landscapes. Results demonstrated that the medium-resolution SDGSAT-1 imagery could be used as a reliable data source for deriving day-night features, enabling the generation of fine-grained UFZ maps when combined with 2D—3D features from other geospatial big data. Cross-method comparisons also showed that this approach could significantly improve both semantic segmentation and topological interpretation across different UFZ types. Notably, our method could not only achieve acceptable levels of mapping performance (overall accuracy > 0.91 and average F1-score > 0.91), but also realize the accurate extraction of purer UFZ blocks with a small sample size (training-testing ratio = 1:4), further indicating considerable potential in large-scale UFZ mapping. The source codes are available at: https://github.com/Sustainable-City-Lab/UFZ-data-fusion.