Novel approaches are needed to support content-based retrieval on geographic image databases. Motivated by the success in the text-based information retrieval, where the indexing and retrieval is based on keywords, we present the framework of geoblock-based image retrieval, where the retrieval is conducted based on the concept of visual keyword: "geoblock". In this paper, we will introduce a Learning Vector Quantization (LVQ) based approach to construct the dictionary of semantic-oriented geoblocks. Based on the frequency and correlation of these geoblocks in the images, comprehensive geographic image features are extracted to facilitate image retrieval. The experimental results demonstrate the effectiveness of this approach.