Abstract Urban informatics is pivotal for smart city development, yet the integration of multi-source spatial data remains hindered by fundamental scale-related challenges. First, the heterogeneity of spatial resolution and boundary representation across datasets complicates cross-scale information transformation. Second, conventional aggregation methods often dilute critical local variations, obscuring essential spatial heterogeneity. Third, the lack of interpretability in existing models undermines their utility for transparent urban governance. Current approaches, typically reliant on rigid hierarchical scaling, fail to capture the dynamic nesting, overlapping, and cross-scale interactions inherent in real urban systems—limiting both performance and adaptability. To address these limitations, this study establishes a systematic framework for multi-scale urban information fusion. We first formalize the relationship between Urban Spatial Scale (USS) and Urban Computing Scale (UCS), mapping their connections to common urban tasks and data sources. Building on this foundation, we introduce the Urban Multi-scale Information Graph (UMIG), a structured representation that preserves cross-scale dependencies while maintaining spatial fidelity. Further, we develop the Urban Multi-scale Information Fusion Model (UMIFM), which leverages attention mechanisms to enhance interpretability and adaptability across diverse urban configurations. Rigorous experiments demonstrate UMIFM’s efficacy in multi-scale fusion tasks, outperforming traditional methods in preserving spatial granularity and computational flexibility. Comparative analyses also reveal its robustness under varying cross-scale block implementations. Moreover, we use the Moran's I analysis to investigate the spatial dependence of prediction residuals. By bridging theoretical rigor with practical applicability, this work advances scalable urban computing methods while providing actionable and interpretable tools for smart city governance.
Ma et al. (Mon,) studied this question.