Integrating AI with GIS: A Novel System for Efficient Task Execution

Geographic Information Systems (GIS) serve as a vital tool in human exploration of the natural world. To lower the threshold for using GIS technology, the integration of artificial intelligence (AI) and GIS is crucial for individuals who lack the professional GIS knowledge required to efficiently solve GIS-related tasks. Existing models, such as general large language models (LLMs) and fine-tuned small language models (SLMs), lack the capability to handle complex GIS tasks effectively. Consequently, we designed an AI-GIS system that organizes multiple agents with fine-tuned SLMs in a tree structure and links external APIs of GIS tools to intelligently provide a GIS workflow for consulting tasks. The AI-GIS system processes tasks by decomposing them into subtasks and then matching appropriate GIS tools for each subtask. Each step is executed by a distinct multi-agent structure. To ensure that the SLMs within these multi-agent structures possess adequate knowledge reserves, we employed a generation method combining prompt engineering and expert supervision. We generated and augmented two relevant GIS corpora and used them to fine-tune the SLMs. Comparative results demonstrate that our fine-tuned SLMs outperform seven other SLMs/LLMs. Especially, the fine-tuned SLMs improved the ability to decompose GIS task by about 18% compared to GPT-4 and about 14% compared to Deepseek-R1; the matching ability of GIS tools is about three times higher than GPT-4 and about four times higher than Deepseek-R1. Furthermore, we tested the AI-GIS system using a practical case study. The workflow generated by the system aligns with that by domain experts, and the execution results are consistent.