GeoForge: Geospatial Analysis with Large Language Models (GeoLLMs)
I: From ‘Geo+ for ChatGPT’ to ‘GeoForge’
At Ageospatial, our vision since our ideation in October 2023 has been to make Geospatial accessible through Geo-AI. Our journey began with ‘Geo+ for ChatGPT’, an OpenAI GPT extension/plugin designed for spatial data generation on the OpenAI platform.
‘Geo+ for ChatGPT’ showcased significant potential, engaging over 1000 users and securing a steady weekly user base of 600, along with 1600 LinkedIn followers within its launch month of January 2024. We encountered limitations of the text format, which pushed us to move our development outside of the ChatGPT platform for an independent web-based Geospatial platform that we like to call ‘GeoForge’.
II: Development of ‘GeoForge’ — Geospatial Analysis with GeoLLMs
‘GeoForge’ makes mapping and geospatial analysis easy for everyone, whether you’re just starting out or you’re already an expert. Our approach is built on three key principles:
- LLM into GeoLLM: We fine-tuned the capabilities of OpenAI’s GPT-4 language model, by training it with additional geospatial vocabulary and knowledge. This training ensures that our GeoLLM can accurately interpret and respond to spatial queries while minimizing the inaccuracies often seen in standard language models. Additionally, we are developing new methods for data generation such as allowing the language model to generate a solution graph with each node representing data(input, intermediary, output) and data processing operations. Once the solution graph is generated, the model can execute each node in sequential order to answer the user’s query using our algorithms. This approach allows the model to understand and process complex geospatial tasks and queries.
- Data quality: Ensuring data integrity is at the core of our priorities for this project. For this reason, we ensure our data is of high quality and up to date by utilizing open-source databases that are actively updated. For our database, we have decided to integrate next-generation database management systems like DuckDB Spatial, to ensure high performance and speed without compromising data integrity. Internal protocols running through servers based in our home country of Switzerland are implemented to ensure the privacy of the data.
- User-friendly interface: A user interface that’s ‘simple but not simplistic’, with features like client-side tile rendering and customizable data layers. ‘GeoForge’ is designed to be both simple and sophisticated, catering to users’ needs without overwhelming them. Direct interactions with GeoLLMs further improve the user’s experience by enabling dynamic map adjustments, base map style customization, automatic camera adjustments, and interactive visualizations.
III: Demonstrating ‘GeoForge’ with a Bangladesh case study
Note: The following is an ongoing development example.
Prompt: “Please show me the administrative boundaries of Bangladesh and provide the names of the five most important cities within the country”
→ ‘GeoForge’ quickly returns a map of the requested administrative boundaries and a list of key city locations, accompanied by a downloadable .geojson file for further GIS exploration.
Prompt: “Geocode the cities from the previous prompt”
→ Our chatbot can remember questions and responses throughout your sessions. Here, we request the geocoding features of city names generated in a previous response, the program will plot markers corresponding to the geocoded cities and generate a CSV file containing all the relevant information
Prompt: “Generate me a Sentinel-2 Image of Chittagong of the last week”
→ ’GeoForge’ retrieves near real-time Sentinel images of the specified area from the Microsoft Planetary Computer, maintaining the original resolution of 10 meters.
→ Using the interactive features of our website, we draw an Area of Interest (AOI) that matches the boundary of the TIFF image. Then, we request the program to generate buildings within this AOI, which corresponds to approximately 100,000 polygons.
IV: What’s next for GeoForge ?
- GeoForge early release in Spring 2024: This test version will offer limited requests and functionalities, but it will provide a glimpse into the potential of this technology.
- Development of features: We will enhance the user experience with improved features like LLM nodes workflow, Complexe dataset, dynamic styling, and more.
- Extensions: Specialized extensions are in development with PhD students to address near-real-time geospatial needs, including damage assessments for emergencies and flood event analysis, integrating data on infrastructure, buildings, and populations to support critical decisions.
Learn More and Get Involved
- Our LinkedIn (where we post all our updates !)
- Our Medium
- Our Website
- contact@ageospatial.com
- GeoForge Dev Journal n°1
Attributions
- DuckDB: Raasveldt, M., & Muehleisen, H. (2024). DuckDB. Retrieved from https://github.com/duckdb/duckdb
- Microsoft Building footprints: https://github.com/microsoft/IdMyPhBuildingFootprints. This data is licensed by Microsoft under the Open Data Commons Open Database License (ODbL).
- And all our sources for Geo+: https://www.ageospatial.com/attributions
- Contains modified Copernicus Sentinel data 2024’ for Sentinel data
Made with passion in Switzerland
