LLM-Geo aims to automate Geographic Information System (GIS) tasks by leveraging Large Language Models (LLMs) for spatial data collection, analysis, and visualization. It targets GIS professionals and researchers seeking to reduce manual operations and make spatial analysis more accessible. The system demonstrates potential for next-generation AI-powered GIS by enabling autonomous goal achievement in spatial problem-solving.

How It Works

LLM-Geo utilizes an LLM, specifically GPT-4, as its reasoning core. Users define spatial problems through natural language tasks. The LLM then generates a solution graph and Python code to execute the analysis, including data retrieval, processing, and visualization. The system includes a review and debugging loop where the LLM attempts to correct errors in the generated code, aiming for an 80% success rate in executing the analysis.

Quick Start & Requirements

• Install by cloning the repository and renaming your_config.ini to config.ini, then adding your OpenAI API key.
• Requires Python and the geopandas package.
• GPT-4 API access is mandatory for effective reasoning and code generation.
• Setup involves configuring the API key and running a Jupyter notebook (LLM-Geo4.ipynb).

Highlighted Details

• Demonstrates autonomous GIS capabilities through three case studies involving population analysis, human mobility trends, and COVID-19 death rate analysis.
• Generates solution graphs, Python code, and visualizations (maps, charts) from natural language task descriptions.
• Includes a direct request notebook (Direct_request_LLM.ipynb) for simpler tasks.
• Under active development with ongoing improvements to prompt generation and data understanding modules.

Maintenance & Community

The project is under active development, with recent updates including the use of o3-mini as a default model and the addition of a data overview module. Related projects include agents for geospatial data retrieval and QGIS plugins. Community engagement can be found via GitHub.

Licensing & Compatibility

The repository does not explicitly state a license in the provided README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The system is a prototype and lacks modules for logging and code testing. GPT-4's debugging capabilities are noted as weak, and the LLM does not receive the full conversation history, potentially impacting robustness. Case study 2's API is shut down. The success rate of generated programs is around 80%, and results may vary due to LLM output variability.