Content-based remote sensing image retrieval has been receiving more attention. Current research on content-based remote sensing image retrieval is mainly about feature extraction and feature fusion. But all these methods have ignored such a fact that retrieval objectives of different types ought to adopt different features. Besides

small size of the training set is also a challenging issue. Considering the abovementional two issues

a new remote sensing image retrieval method

which is based on feature selection and semi-supervised learning

is proposed in this paper. The new method includes four steps: 1) Determine the number of clusters automatically by using MDL criterion; 2) According to a clustering validity index

we select the optimal features which can describe the retrieval objectives best

meanwhile make some improvements in the original DB index; 3) Dynamically determine the weights of the best color feature and the best texture feature; 4) Automatically select an appropriate semi-supervised learning method and conduct image retrieval. Experiment results show that

in contrast to the retrieval performance of relevance feedback method

the method proposed in this paper obtains similar retrieval performances both in the applications of soil erosion area retrieval and general land cover retrieval in remote sensing image

as well as solving the small sample size problem.