Aimed at tackling the difficulty of achieving the prior probabilities of class samples in using the traditional Maximum-Likelihood (ML) and solving the problem of remote sensing image classification, the authors put forward a novel and improved ML method. This method can automatically obtain the optimal prior probability of class samples, thus overcoming the main defect in the traditional ML method. In the experiments the authors used six scenes of Radarsat-1 Fine mode SAR images, and the results demonstrate that good classification can be achieved by using the method proposed in this paper. In the case of single-band and single-polarization SAR data, the classification precision can be expected to reach 80%.