GeoAutoModuler: a knowledge–enhanced large language model for heterogeneous geospatial code understanding and knowledge extraction

In the era of geospatial big data, massive geospatial modeling scripts have accumulated across the Internet and have become essential carriers of modeling knowledge. However, these scripts are often unstructured and heterogeneous in terms of semantics and logic, making their embedded knowledge difficult to interpret and reuse. To address this issue, this paper proposes GeoAutoModuler, a specialized large-language model (LLM) for functional semantic knowledge extraction from geospatial modeling scripts. GeoAutoModuler adopts a structure-aware continued pre-training strategy combined with the QLoRA approach to improve geospatial code understanding, and a skeleton-of-thought (SoT)-based domain instruction tuning framework to guide staged knowledge extraction tasks, including functional module summarization, input/output element identification, dependency extraction, and code snippet mapping. Three datasetsGeoModeling-PT, GeoModeling-SFT, and GeoModeling-Eval were constructed for training and evaluation. Experimental results show that GeoAutoModuler achieves a 15.7%–30.8% improvement in knowledge extraction metrics, such as F1, SMR-Graph, ROUGE-L, and Levenshtein, over models of similar scale, with several metrics approaching or surpassing GPT-4o. A prototype system further demonstrates the potential of GeoAutoModuler in supporting knowledge graph construction, workflow visualization, and geospatial code generation, which lays a foundation for intelligent and autonomous geospatial modeling. Supplementary examples, implementation details, and additional analyses are provided in the Appendix.