In order to enhance the capability of local feature representation of remote sensing images and to make full use of sparse decomposition of over-complete dictionary, this paper proposes a new method of remote sensing image retrieval based on sparse representation feature. In this method first the local invariant features are extracted from the training remote sensing image database and trains over-complete dictionary based on local features, and thus the sparse representation will be obtained under the dictionary update; the authors regard the sparse representation as the image’s final feature description. Secondly, the authors construct a visual dictionary using sparse representation features, and obtain the sparse histograms by spatial pyramid matching algorithm. Finally, the SVM classification model is trained based on the sparse features, by using the classification model, the images classified as one category with the query image to be output. The similarity matching is carried out in the output image set, and an image with the largest similarity is returned to achieve the image retrieval. Experimental result shows that the features extracted by the new method not only possess the robustness of local invariant features but also provide the necessary semantic information, which is of great practicality and applicability in image retrieval research field.