(BigData 2026) Multi-Agent Collaborative Data Governance System for Geographic Big Data

Urban geographic big data integrated from OpenStreetMap, commercial POI platforms, satellite imagery, and IoT streams routinely contains coordinate drift, missing attributes, spatial duplicates, and inconsistent schemas, which propagate into downstream analytics. Conventional GIS governance depends on hand-written ETL rules that break silently when source schemas evolve and offer no end-to-end quality accountability. To address these limitations, we present GeoGov, a reference implementation of a multi-agent data governance system in which four domain agents (Collection, Cleaning, Integration, Application) cover the full geographic data lifecycle, while a Global Coordination Agent iteratively re-triggers them through a closed-loop quality feedback mechanism until configurable quality targets are met or Kmax iterations are reached. On real-world OSM data from the Wudaokou sub-district of Haidian, Beijing (19,629 records across three thematic layers: POIs, buildings, roads), GeoGov reduces the POI-layer near-duplicate rate from 63.3% to 25.6% (a 37.7 pp reduction), outperforming a rule-based baseline by 24.9 pp on this metric in under 10 seconds and without manual intervention. An agent ablation shows that Cleaning is non-substitutable for the spatial-compliance gain (w/o Cleaning, Spatial Compliance drops 3.9 pp from the full configuration), and Integration is non-substitutable for the near-duplicate reduction (w/o Integration leaves 64.4%, vs. 25.6% with full); the Coordinator and feedback loop are designed to provide a quality-gated termination guarantee under non-deterministic cleaning, a regime targeted by ongoing companion work (§5). The five agents are deployed behind a single governance endpoint as composable services, illustrating an Agentic AI as a Service (AAIaaS) layout for geographic data governance. The current implementation uses deterministic algorithmic operators; an LLM-based Coordinator extension is currently under active development, with the agent-contract layer designed for this swap (§5).