In Pursuit of High-Quality Urban Data

Abstract

Urban data are central to transportation planning, mobility modeling, accessibility analysis, infrastructure assessment, and many other forms of urban decision-making. However, the usefulness of urban data depends not only on their availability, but also on their quality. In practice, urban datasets are often too coarse to support fine-scale analysis, incomplete with respect to important infrastructure, inconsistent across sources, or insufficiently faithful to the real-world systems they are meant to represent. These limitations constrain what urban phenomena can be observed, modeled, integrated, and acted upon. This dissertation argues that high-quality urban data are a prerequisite for reliable and equitable urban analytics, and that improving urban data quality requires targeted methods designed around the specific deficiencies, structures, and intended uses of the data. This dissertation studies urban data quality through four complementary dimensions: granularity, coverage, coherence, and fidelity. Each dimension corresponds to a different kind of data deficiency and motivates a different methodological study. Granularity concerns whether urban phenomena are represented at sufficient spatial or temporal resolution. Coverage concerns whether important urban features are represented in the data at all. Coherence concerns whether heterogeneous data sources can be reconciled into consistent and unified representations. Fidelity concerns whether generated or extracted data preserve the real-world structure and function needed for downstream applications. To refine granularity, this dissertation introduces the Structurally-Aware Recurrent Network (SARN), a model that recovers fine-resolution urban signals from coarse, privacy-preserving aggregates by jointly modeling temporal dynamics, global spatial interactions, and hierarchical containment relationships. To increase coverage, it develops a zero-shot annotation pipeline that combines class-agnostic segmentation, candidate filtering, and Set-of-Mark prompting to identify underrepresented built-environment features from aerial imagery using vision-language models. To improve coherence, it investigates large-language-model-assisted spatial data integration for spatial join and spatial union tasks, showing that LLMs can support integration decisions more effectively when paired with human-interpretable geometric features. To enhance fidelity, it introduces TraversRL, a vision-conditioned iterative generator for pedestrian pathway graphs that is improved through reinforcement learning objectives aligned with graph-level geometry and traversability. Overall, these studies show that urban data quality is multidimensional and that different deficiencies require different computational strategies. More broadly, the dissertation demonstrates that urban data should not be treated only as fixed input to downstream models. Urban data themselves can be the object of methodological innovation. By improving how urban data are refined, expanded, reconciled, and structured, this dissertation advances a broader agenda for building high-quality urban data in support of more reliable and actionable urban analytics.