Supervised image classification refers to the task of extracting information classes from a multi-band remote sensing image. The selection of training samples is critical and directly influences supervised classification accuracy. However, some impure training samples are possible selected because of human mistakes or limited labeling conditions, which leads to a reduction in the classification accuracy. To solve this issue, median absolute deviation (MAD) is adopted to refine training samples. A comparison of the full and refined training samples is conducted for the same classifier, i.e., maximum likelihood classification (MLC) or support vector machine (SVM), through experimental evaluation with two sets of experiments. The results of experiments show that the overall accuracy and the kappa coefficient of the refined training samples significantly outperform those of the full training samples for the same classifier (MLC or SVM). It shows that refining training samples using the MAD can effectively eliminate the influence of impure training samples so that the more reliable and accurate results can be obtained.