Statistical learning theory-based support vector machine (SVM), which is a supervised learning mechanism, can get good class rate in remote sensing image classification. But manual obtaining of labeled training samples is a much time-consuming work because of the much greater class number of remote sensing image. In addition, there are some subjective factors in manual job by different operators. In order to overcome these shortcomings, a semi-supervised approach has been developed and implemented. The training samples are labeled automatically with fuzzy C-means clustering algorithm. Only the initial clustering centroid for each class is chosen manually. Using these automatically labeled samples, multi-class SVM classifier is trained for remote sensing images classification. The results of the experiment show that the method does upgrade the classification efficiency greatly with practicable class rate.