Bag of visual words (BoVW) model has been widely used in content-based image retrieval for its good performance. The k-means clustering algorithm plays an important role in BoVW, but there are still some problems that need to be solved. Many methods are proposed to improve the clustering performance, such as fast search and find of density peaks (FSFDP), which has been successfully applied in image classification or image retrieval for its high efficiency but suffers from some limitations when it is applied to remote sensing image retrieval. A remote sensing image retrieval method based on an improved FSFDP-BoVW model is proposed. A feature selection method combining a chi-square test and a term frequency-inverse document frequency is proposed to select the features that can best represent image content and reduce the redundancy of remote sensing image features; a natural nearest neighbor algorithm is utilized to redefine the density of data points to reduce the interference of human subjectivity in FSFDP. Experimental results show that the proposed method can effectively reduce the time spent in constructing a codebook and improve the retrieval accuracy.