Large language models in geospatial science

Purpose Large language models are opening new ways to work with geospatial data, from asking questions about places to building analytical workflows. This study aims to examine how these applications fit together and what they suggest about the developing role of LLMs in geospatial science. Design/methodology/approach A systematic review of 55 studies was conducted, covering GIScience and related spatial applications. Dictionary-based coding, supported by manual review, was combined with term frequency–inverse document frequency analysis and non-negative matrix factorisation to examine research roles, tasks and themes. Comparisons between core GIScience and adjacent fields helped place the overall patterns in context. Findings Benchmarking was the most common research role (41.8%), followed by multimodal interpretation (20.0%). Across the literature, two complementary directions emerge: using language models to organise geospatial operations and using vision-language models to interpret spatial observations. The abstracts place particular emphasis on methods, while the thematic analysis connects general GIS evaluation with code generation, remote sensing and applied spatial reasoning. Practical implications The synthesis points to the value of combining flexible language interfaces with established GIS computation. Practical evaluation should consider how a system turns a request into a verified spatial result, including the time and expert involvement needed to complete that process. Originality/value By bringing GIScience and adjacent spatial research into one analytical framework, the review shows how model capabilities, data representations and external tools contribute to geospatial workflows. This perspective offers a basis for comparing systems and identifying productive connections between research communities.