Geo-CLASS: Knowledge from Earth Engine Scripts

publications

Thousands of Google Earth Engine scripts shared online encode practical knowledge about how geospatial analyses are built, from data selection and cloud masking to index calculation and export. This knowledge is scattered across code and hard to reuse.

From crowdsourced scripts to a geoprocessing knowledge graph
Extracting themes, functions, data sources and spatial and temporal scopes from a script into a knowledge graph.

Geo-CLASS coordinates a large language model with small language models. Guided by a domain schema and a schema-aware, chain-of-thought prompt, the LLM recognises entities and relations in each script. The SLMs then map the free-text descriptions onto a geoscience knowledge base (GCMD, GeoNames, the GEE data catalogue and a functional dictionary) through semantic vectorisation, cosine similarity and best-match selection, so that synonymous descriptions from different scripts end up as the same standardised entity.

On 237 scripts sampled from 295,943, the framework improved entity and relation recognition accuracy by up to 31.9% and 12.0% over LLM baselines including Llama-3, GPT-3.5 and GPT-4o.

The overview figure at the top shows the full Geo-CLASS framework. Figures from Zhao et al. (2026), International Journal of Geographical Information Science, 40(3), 808–841, doi:10.1080/13658816.2025.2577252.

First Author LLM Knowledge Graphs
Anqi Zhao
Authors
Anqi Zhao
PhD Student in Geospatial AI
I am a PhD student at the Chair of Cartography and Visual Analytics at the Technical University of Munich, supervised by Prof. Liqiu Meng and mentored by Prof. Yu Feng. I came to AI from maps. I studied geodesy, cartography and GIS at Wuhan University, where I first used language models to make sense of geospatial code, and I now spend most of my time on the AI side of GeoAI. Besides research, I supervise Master’s students and maintain open learning resources on LLM agents and web scraping.