This paper proposes a supervised classification method for remote sensing images based on locally automatically searching training samples and spectral matching technique. The best training samples are searched and selected on the whole image by the local space information and spectral matching, and then they are purified on spectral domains. Both spatial and spectral information are purified to enable the training sample to meet the requirement for classification at the best of times. An experiment for TM image classification based on BP neural networks has been conducted to validate the procedure. It can be seen from our experiment that the classifying results are improved from the observation of naked eye and the numerical analysis. So the proposed approach has practical application value to some extend for it's simple and high efficiency.