Research on the method of improving the autonomous spatial processing ability of GIS Copilot

Geospatial agents with large language models as the core of decision-making have preliminarily demonstrated the ability of spatial analysis for autonomous planning and execution, which can effectively reduce the professional threshold and improve the efficiency of spatial data analysis and processing. However, geospatial agents represented by GIS Copilot still face problems such as inaccurate geospatial semantic understanding and weak geospatial analysis task planning capabilities, resulting in low accuracy when autonomously completing complex geospatial analysis tasks. To address the above issues, this paper proposes a performance improvement method integrating prompt optimization, Retrieval-Augmented Generation, and supervised fine-tuning, aiming to improve the accuracy of geospatial agents through explicit constraints, external knowledge support, and domain knowledge internalization. Specifically, it includes: 1) improving data understanding rules through prompts to provide more rigorous constraints for model reasoning; 2) constructing the GeoGuide Vault, and organizing geoprocessing workflow cases to provide a transferable paradigm reference; 3) applying LoRA fine-tuning to optimize easily confused tool pairs and semantic mapping deviation scenarios, thereby reducing the risk of tool mis-selection. Experimental results based on GIS Copilot show that the proposed method improves the hit rate of geospatial agent tool selection from 32.39% to 65.49%, and the F1 score from 0.3470 to 0.6150, significantly enhancing the autonomous geospatial analysis capability of the agent.