Spatial co-location pattern mining, aiming at discovering the association of spatial features, can help users find the correlation knowledge from spatial datasets. However, with the growth of spatial datasets, the traditional framework for mining co-location patterns generates a mass of redundant co-location patterns, which makes it difficult for users to perform further analysis. This paper researches how to extract representative co-location patterns, i.e., a concise summarization of prevalent co-location patterns. Two similarity metrics - feature similarity and distribution similarity, are introduced to evaluate the redundancy between co-location patterns from the perspective of both features and instances. With the above metrics, a novel approach called Representative Co-location Patterns Extractor (RCPE) is further proposed to condense the prevalent co-location patterns. Two types of representative co-location patterns, called Representative Maximal Co-location Patterns (RMCPs) and Worthy Non-Maximal Co-location Patterns (WNMCPs), are mined from the prevalent co-location patterns set respectively. A series of experiments on both synthetic and real datasets confirm that RCPE has better performance in compression power and running time compared with similar state-of-the-art approaches.