The precision of training samples has important influence on the accuracy of Supervised Classification of Remote Sensing Images. It is easy to omit the ground feature which covers a small amount by visual selection directly on the original image. This leads to a large number of wrong points and leakage points which influences the Classification Accuracy of Remote Sensing Images. Based on what was said above, we first use K-means clustering algorithm to initialize the samples in this paper. According to the clustering results, the training samples selected by visual labeling and density-based k-means algorithm participated in the SVM classification model. The results show that selecting training samples by the k-means algorithm based on density can reduce the influence of visual labeling on the subjective factors of training sample selection. Meanwhile,through the optimization of the radius weighting factor, we have found optimal combination of weighting factors and then the classification accuracy of remote sensing images can be increased from 81.70% to 88.89%.