Data Synthesis Agent for Urban Modeling

• Data Synthesis Agent is a computational framework that synthesizes micro-level urban population data using open data and agent-based modeling.
• It sequentially refines spatial units, applies probabilistic urban parcel classification, and utilizes census-based microdata synthesis tools.
• Empirical validation demonstrates strong fidelity with ground truth data, supporting robust urban microsimulation and policy analysis.

A Data Synthesis Agent refers to a class of computational frameworks, agent architectures, or automated pipelines that generate synthetic, structured, or annotated data for downstream modeling, analysis, or system training. These agents leverage open data, distributed computing, rule-based or statistical modeling, and agent-based approaches to synthesize micro-level datasets—often combining spatial, categorical, and attribute data—according to explicit modeling, consistency, and validation constraints. The approach described in “Population spatialization and synthesis with open data” provides a canonical instantiation: an end-to-end, automated framework that spatializes and synthesizes urban populations at fine spatial scales by integrating disparate open data sources, probabilistic spatial models, and individual-level attribute synthesis via agent-based tools (Long et al., 2014).

1. Methodological Foundations and Workflow

The core Data Synthesis Agent methodology hinges on sequential, modular processing: (i) spatial unit delineation; (ii) probabilistic urban parcel classification using points of interest (POIs) and cellular automata; (iii) inference of residential parcels using auxiliary housing activity signals; and (iv) attribute-level population synthesis using census data and agent-based microdata generation tools.

The workflow operates as follows:

1. Spatial Unit Delineation: Road networks from OpenStreetMap (OSM) are cleansed (trimming short segments, extending endpoints) and used to buffer and extract parcels, i.e., spatial polygons representing continuously developed areas bounded by roads.
2. Urban Parcel Inference: A vector-based cellular automata (CA) model is applied to parcels, characterizing their probability of being “urban” via logistic regression potentials and incorporating neighborhood effects, spatial constraints, and stochasticity:

$$S_{ij}^{(t+1)} = f(S_{ij}^{(t)}, Q_{ij}', \text{Con}, N)$$

and in probabilistic form:

$$P_{ij} = (P_D)_{ij} \cdot (P_{\Omega})_{ij} \cdot \text{con}(\cdot) \cdot P_\text{stoch}$$

1. Residential Parcel Classification: Urban parcels are further filtered using residential POI densities or check-in data, standardized as $\log(\text{raw})/\log(\text{max})$, to discriminate residential from all urban parcels.
2. Population Synthesis: Populations are proportionally allocated to residential parcels using relative residential POI densities, followed by synthetic attribute generation using a microdata synthesis tool (Agenter) that expands aggregated census distributions into synthetic, cross-tabulated individual profiles.

This highly modular approach distinguishes the Data Synthesis Agent paradigm by its capacity to sequentially refine data granularity and attribute detail, from coarse infrastructure maps to high-fidelity agent microdata.

2. Data Inputs and Analytical Toolchain

The Data Synthesis Agent is intrinsically data-driven, aggregating heterogeneous sources and applying specialized analytical transformations:

Data Source	Analytical Role	Tool/Process
OSM Road Network	Parcel geometry extraction	Network merging, buffering
POIs (crowd-sourced, check-ins)	Urban function and residential inference	Vector CA, density metrics
Census Aggregates	Attribute distribution priors	Constraint-based synthesis (Agenter)
Ground Truth (e.g. BICP)	Validation	Overlap, correlation metrics

The central technological lever is the integration of open, volunteered geographic data (VGI), which enables the derivation and refinement of base spatial units otherwise unavailable at city scale—particularly in data-scarce environments. Attribute data (census) is utilized downstream by specialized synthesis engines (e.g., Agenter) that enforce statistical consistency with empirical distributions and covariances at the agent level.

3. Probabilistic Spatialization and Synthesis Models

Data Synthesis Agents operationalize spatialization and synthesis via coupled probabilistic models. In parcel urbanization, the conversion likelihood is decomposed as a product of a logistic regression-derived potential (capturing parcel size, compactness, accessibility, and POI intensity), neighborhood urbanization effects, hard spatial constraints (e.g., masking steep or waterlogged parcels), and controlled stochastic variability. Residential classification applies a standardized POI density thresholding against known residential parcels. For population synthesis, a proportional allocation scheme draws on inferred densities, while Agenter’s microdata generator performs iterative attribute sampling constrained by both 1st-order marginals and empirically derived attribute cross-tabulations.

The result is a hybrid stochastic-deterministic pipeline that produces high-resolution spatially explicit agent datasets, each agent encoded as an individual microrecord adhering to real-world distributions and spatial logic.

4. Empirical Validation

Multi-level validation is integral to the Data Synthesis Agent framework:

• Spatial Unit Validation: The overlap between OSM-derived parcels and a Beijing Institute of City Planning (BICP) ground truth dataset reached 71.2% in area coverage, underscoring the method’s fidelity given input data quality.
• Residential Density Validation: Inferential population estimates per parcel, aggregated via the POI method, exhibited a Pearson’s $r \approx 0.858$ correlation with building floor area, reinforcing the assumption that POI density proxies built density at fine spatial scales.
• Attribute Synthesis: Agenter’s synthetic microdata achieved a 72.6% similarity indicator (SI) with survey microdata, versus 43.9% for a null model, confirming high-fidelity replication of census attribute relationships.

These quantitative results support the internal consistency and external validity of the agent-based synthesis pipeline, enabling systematic confidence in downstream modeling applications.

5. Applications in Urban Modeling and Policy

Fine-grained, attribute-rich population datasets synthesized by this agentic approach have become critical for:

• Urban microsimulation (e.g., spatial micro-simulation, agent-based models): Providing necessary “resident” agents for policy scenario evaluation.
• Data-sparse regions: Enabling planning and resource allocation in locales where survey microdata and parcel-level distributions are otherwise unavailable.
• Socioeconomic exposure studies: Facilitating analyses of air quality, services, mobility, or market exposure at the resolution of parcels or neighborhoods.
• Scalable, replicable synthesis: The reliance on global open data (OSM, VGI, open census) ensures portability to other urban regions subject to data coverage, with the agentic process amenable to parallelization, update, and extension as new sources arise.

6. Limitations and Generalization

Potential limitations derive from the input data resolutions and coverage. OSM-derived parcels may be overly coarse in regions with missing or outdated tertiary road coverage, inflating block sizes relative to true built form. Assumptions (e.g., POI density as a proxy for residential density) may be less accurate in mixed-use or non-standard urban morphologies. The method’s generalization relies on the availability and quality of core open datasets but is robust to be extended or updated as new volunteered geographic and activity data (e.g., mobile check-ins, social sensing signals) become commonplace. The underlying agentic strategy—automated, modular, and data-driven—remains broadly applicable.

7. Significance and Broader Impacts

The end-to-end, openly powered Data Synthesis Agent paradigm described in this work exemplifies a scalable architectural pattern for producing micro-level, spatially explicit agent data in the absence of comprehensive administrative records. This approach addresses key challenges in urban data science, spatial microsimulation, and agent-based modeling, directly supporting fine-scale scenario evaluation, real-time monitoring, and rapid assessment in contextually constrained domains. The confluence of open data, modular agentic modeling, and probabilistic synthesis provides a foundational template for future developments in large-scale, granular synthetic data generation pipelines for social, economic, and environmental applications.