A remote sensing image change detection approach based on greedy Expectation Maximization (EM) algorithm for Hidden Markov Random Field (HMRF) is proposed. The difference image is constructed by Principal Component Analysis (PCA) and subtraction operation. Firstly, the HMRF model is applied to characterize the contexture-dependent information, and the energy function of system is defined. Secondly, the greedy EM algorithm is used to overcome the disadvantage of the standard EM algorithm that assumed the number of the mixture components is a known priori, the performance of the overall parameter estimation process depends on the given good initial settings excessively, and the estimated parameter can be resulted from some local optimum points. The distribution model structure and parameters are learned accurately to find the best fit of the given data. Finally, the changed area is obtained by using Iterated Conditional Modes (ICM) to optimize the energy function. Experiments show that the proposed method has virtues of preserving structural change and filtering noises.