Can Large Language Models Automatically Generate GIS Reports?

Abstract

Geographical Information Systems (GIS) are essential tools used for storing and performing analyses on spatio-temporal data. GIS reporting plays a crucial role in transforming the result of these analyses into actionable insights, enabling informed decision-making, by uncovering patterns and relationships and presenting them in a human-readable format suitable for the intended target audience. Nonetheless, the traditional process of creating reports involves manual analysis and interpretation of spatio-temporal data, a time-intensive task prone to human error. This paper aims to investigate the potential of Large Language Models (LLMs), particularly their natural language processing and generation abilities, to streamline the report generation process. To this end, three case studies are conducted, using the GPT-3.5 LLM to analyze real-world GIS data, extract key spatio-temporal insights and generate actionable, human-readable reports. The generated reports are then analyzed, to assess the model’s capacity for understanding complex spatio-temporal relationships and patterns and generating coherent reports. Results show that general-purpose LLMs can be remarkably effective in detecting spatio-temporal patterns and anomalies and in generating concise, effective human-readable reports. Despite this great potential, we also identify several key challenges of LLMs for GIS report generation, including a significant variability among different re-executions, a tendency to report incorrect data in some scenarios, and difficulty in understanding more complex spatial data such as polygons.

Data and code

The data employed in the case studies, as well as the prompts and the generated reports, are available in a dedicated replication package at .