Maximum Likelihood (ML) classification method is based on the assumption that the data are normally distributed, which is not always true for the realistic remote sensing data, and may result in decrease of classification accuracy.The classification results are impacted directly by the prior probability.The selection of training samples is somewhat stochastic and subjective.The ML method uses the same prior probability for the whole image, which will also reduce the classification accuracy.Theoretically, every smooth density function can be approximated to within any accuracy by such a mixture of normal densities.Thus the first problem of ML can be solved by using a combination of several normal functions instead of one.In this way, a very general capability can be provided, while still maintaining the convenient properties of the normal assumption.For the second problem, ISODATA is used to make a clustering image of the original data, after that, one can select the training areas of the image by comparing with the reference image.At last, the result of experiment shows that the proposed methods can not only realize the classification of remote sensing image but also achieve very high accuracy visually and mathematically in overall accuracy and Kappa coefficient.