Democratizing Multi-Granularity Spatio-Temporal Intelligence with Multi-Agent Systems

Abstract

We propose a system that democratizes multi-granularity spatio-temporal analysis by integrating a Discrete Global Grid System (DGGS) data pipeline with a Multi-Agent System (MAS). Unlike existing single-agent spatial AI solutions that primarily target experts and lack support for heterogeneous data, persistent memory, and validation, our platform converts diverse datasets into standardized H3-indexed cells, enabling consistent analysis across scales. To enhance usability for non-experts, the system interactively guides users to refine queries, which are decomposed into sub-tasks managed by specialized agents for data retrieval, transformation, analysis, and visualization. Agents communicate through a decentralized framework with shared memory, supporting persistent reasoning and multi-turn dialogue. Reflection modules and human-in-the-loop validation further strengthen robustness. Demonstrated through real-world scenarios, such as analyzing the relationship between aging rate patterns and average income to inform social welfare policy in Taiwan, the system illustrates how natural language queries, combined with intuitive map- and chart-based visualizations, can support evidence-based decision-making.