A semi-supervised classification algorithm named DE-self-training for hyperspectral remote sensing images was proposed. Firstly, taking a few labeled samples as initial training set, the initial classification model was constructed by using improved Self-training algorithm to classify unlabeled samples. Then, partial samples and corresponding labels were selected randomly as a proportion from classification results into training set, and the augmented training set was used to retrain the model to classify the unlabeled samples. Then, the algorithm continued the process of training-classifying-picking out samples to augment training set iteratively. During this process, in order to ensure the training set’s quality and the correct labeling of new increased samples, the algorithm edited and purified mislabeled samples by using data editing strategy based on the nearest neighbor rule. Finally, the proposed algorithm trained classification model iteratively to get a more accurate result until the unlabeled samples set was empty. In the experiments, AVIRIS Indian Pines and Hyperion EO—1 Botswana data were used to test the algorithm. According to the comparison with SVM classification results, the DE-self-training algorithm can get higher accuracy and Kappa coefficients by utilizing unlabeled samples information under limited labeled samples.