ChatGeoAI: Democratising Geospatial Analysis with Large Language Models

Topic: ChatGeoAI: Democratising Geospatial Analysis with Large Language Models

When: 29 October at 12.00-13.00

Where: Online

Speaker: Rachid Oucheikh, Physical Geography and Ecosystem Science, Lund university

Moderator: Pengxiang Zhao, Centre for Geographical Information Systems (GIS Centre)

Spoken language: English

Abstract

Large Language Models (LLMs) are transforming how we interact with technology by enabling natural language to drive complex computational tasks. In the geospatial domain, these advances open new possibilities for making Geographic Information Systems (GIS) more accessible, flexible, and intelligent. 

This seminar presents recent research on integrating LLMs with GIS environments while focusing on two complementary directions: 

1. frameworks that translate natural language queries into executable geospatial code in PyQGIS, and
2. multi-agent architectures that enhance task execution through reasoning, collaboration, and tool integration. Together, these approaches showcase how LLMs can empower non-expert users to perform spatial analyses, automate workflows, and interact with GIS software more intuitively. 

We will discuss methods such as fine-tuning, Retrieval-Augmented Generation (RAG), and Chain of Thought (CoT) reasoning and highlight their role in bridging user intent with geoprocessing operations. Practical demonstrations reveal both the strengths (such as high success rates in simple tasks, semantic alignment, and visualization support) and current challenges, including ambiguity handling, failure in complex multi-step workflows, and reasoning depth. 

These findings raise broader questions about the limits of current LLM reasoning in highly structured, domain-specific contexts like geospatial analysis. How can such systems balance automation with the need for expert oversight? And what strategies are necessary to move from prototype demonstrations toward robust, trustworthy GIS automation? 

The seminar aims to provide a critical overview of these opportunities and challenges and situate LLM-powered geospatial analysis within the wider vision of intelligent, user-friendly Digital Earth infrastructures.