Image segmentation model based on Markov Random Field (MRF), which combines prior knowledge and likelihood information, have the certain ability to preserve image boundary and reduce speckle noise. However, the method in SAR image with high noise background is difficult to balance reduction of noise and preservation of boundary because of the abilities based on neighborhood structure. If the neighborhood structure is simple, the algorithm can keep the image boundary, but cannot reduce noise well. Besides, the pixel processing unit in the iterative optimization stage of the method led to massive inefficiencies, and can not meet the requirements of practical application. Thus, a new segmentation strategy is proposed. Firstly, the original image is partitioned into many small segments which have similar grayscale and good boundary. Secondly, in order to enhance the ability of reducing noise of the method, each image pixel gray value is set as the average of the segment to which the pixel belongs. Finally, in iterative optimization stage of the algorithm, each segment in place of pixels was used as the processing units, which reduce the amount of processing unit and improve the efficiency of the algorithm. In order to demonstrate our algorithm’s performance, simulated and real SAR image’s segmentation results are presented.