Self-evolving Spatial Intelligence - Integrating Large Language Models with GIS for Automated Geospatial Knowledge Discovery and Decision Support

Abstract

Recent breakthroughs in natural language understanding, reasoning, and code generation have opened a new path for the intelligent transformation of geographic information systems (GIS). Existing large language models (LLM) -driven GIS methods. However, largely remain task-specific and static, lacking the ability to accumulate knowledge and evolve autonomously. This paper proposes a self-evolving spatial intelligence framework that deeply integrates LLM with GIS for automated geospatial knowledge discovery and decision support. The framework begins with the construction of a geospatial knowledge graph, extracting, organizing, and continuously updating spatial knowledge from multi-source heterogeneous data through a reflexion-enhanced self-learning loop. Building on this, an autonomous spatial analysis agent converts natural-language intents into executable geospatial workflows through tool calling and multi-agent collaboration. Finally, a knowledge-augmented decision-support layer produces interpretable, evidence-based recommendations for spatial planning and decision-making. Together, these three components form a closed loop of knowledge discovery, decision support, and self-evolution, moving GIS from a passive analysis tool toward a continuously self-improving spatial intelligence system. Case studies show that the framework is effective in automating geospatial analysis, improving the accuracy and completeness of knowledge extraction, and enhancing decision-making efficiency, while the hallucination and reliability of LLM remain open challenges.