On the Use of LLMs for GIS-Based Spatial Analysis

Abstract

This paper presents an approach integrating Large Language Models (LLMs), specifically GPT-4 and the open-source DeepSeek-R1, into Geographic Information System (GIS) workflows to enhance the accessibility, flexibility, and efficiency of spatial analysis tasks. We designed and implemented a system capable of interpreting natural language instructions provided by users and translating them into automated GIS workflows through dynamically generated Python scripts. An interactive graphical user interface (GUI), built using CustomTkinter, was developed to enable intuitive user interaction with GIS data and processes, reducing the need for advanced programming or technical expertise. We conducted an empirical evaluation of this approach through a comparative case study involving typical GIS tasks such as spatial data validation, data merging, buffer analysis, and thematic mapping using urban datasets from Pesaro, Italy. The performance of our automated system was directly compared against traditional manual workflows executed by 10 experienced GIS analysts. The results from this evaluation indicate a substantial reduction in task completion time, decreasing from approximately 1 h and 45 min in the manual approach to roughly 27 min using our LLM-driven automation, without compromising analytical quality or accuracy. Furthermore, we systematically evaluated the system’s factual reliability using a diverse set of geospatial queries, confirming robust performance for practical GIS tasks. Additionally, qualitative feedback emphasized improved usability and accessibility, particularly for users without specialized GIS training. These findings highlight the significant potential of integrating LLMs into GISs, demonstrating clear advantages in workflow automation, user-friendliness, and broader adoption of advanced spatial analysis methodologies.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 13 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.