A question-answering framework for geospatial data retrieval enhanced by a knowledge graph and large language models

The rapid advancement of Earth observation technologies has resulted in an explosive growth of geoscientific data. Efficient retrieval of such data is essential but often challenging, especially for non-experts, due to decentralized data sources and limited semantic support in existing portals. To address these challenges, we propose the Geospatial Data service-oriented Question-Answering (GDQA) system, which combines a geospatial data knowledge graph (GDKG) with large language models (LLMs) to enable natural language-based data discovery. Our framework includes two key components. First, we construct the GDKG by aggregating data from five major portals and employing an information extraction model to automatically identify and structure fine-grained knowledge from unstructured texts. Second, we develop a method called Spatio-Temporal Reasoning on Knowledge Graphs (STRKG), which enables LLMs to perform reasoning and exploration over GDKG, enhancing their ability to understand and respond to complex geoscientific queries. While LLMs contribute to flexible natural language interaction, the GDKG complements them with domain-specific knowledge. To evaluate our system, we design a set of competency questions spanning four dimensions: constraint quantity, reasoning depth, spatiotemporal logic, and inference complexity. Compared to baseline approaches including KGQArr, Text2Cypher, KAPING, and ToG, our method achieves superior performance in both accuracy and efficiency, significantly reducing time and token costs.