Traditional natural language to structured query language (NL2SQL) models face significant challenges in spatial databases, especially in entity matching and query type recognition. The narrow coverage of existing corpora and the limited diversity of query types restrict the ability to effectively model spatial relations and geographic features, leading to insufficient training performance and weak generalization. To address these limitations, a large language model (LLM)-driven corpus construction strategy was proposed, focusing on corpus verification and corpus generation. In the verification stage, an error detection and repair tool was developed, in which LSTM networks were combined with customized entity and relation extraction algorithms built on natural language processing techniques. Preprocessed queries were modeled and corrected by the tool, thereby enhancing semantic recognition and grammatical correction. In the generation stage, a query-type-oriented, rule-driven mechanism was introduced, integrating LLM-based generation, manual collection, and template synthesis to develop a corpus generation tool supporting multiple databases and diverse query categories. Experiments were conducted on datasets derived from real spatial databases, and the results demonstrated that the query conversion rates and accuracy of NL2SQL models in spatial contexts were improved by the proposed strategy. Furthermore, higher generation efficiency and greater corpus diversity were achieved by this approach compared with GPT 4o, DeepSeek, and Gemini 2.0.