SAR sea ice image segmentation is essential for climate variation research and navigation safety. However, current region-level MRF-based algorithms have been widely used in SAR sea ice image segmentation. Speckle noise is a phenomenon of the SAR imaging system produced by inherent defects of the system. This phenomenon seriously affects the accuracy of polarimetric SAR image interpretation and subsequent segmentation, classification, target detection, and treatment. Thus, polarimetric SAR image speckle should be inhibited. To effectively suppress the interference of speckle noise and preserve edge information, We proposed a new segmentation algorithm in this paper.A novel segmentation algorithm called polarimetric SAR sea ice image with noise suppression (NS-RMRF) is introduced. Speckle Reduction Anisotropic Diffusion (SRAD) filtering is used in polarization total power span before splitting the original image. The simulated overflow watershed segmentation algorithm is applied to generate initialized areas. After the initialized segmentation, a region adjacency graph is constructed. With the difference between the area and the adjacent area, the difference degree is introduced to the region MRF model based on the Wishart distribution. Simulated annealing is the optimization algorithm used to minimize the objective function.Two polarimetric SAR of sea ice image data obtained by RADARSAT-2 and SIR-C were used to verify the effectiveness of the proposed method. The classical polarimetric segmentation ML method of Lee, the region-WRMF segmentation method of Wu, and the PolarIRGS segmentation method of Yu were compared with the NS-RMRF method. The result indicates that the proposed algorithm exhibit advantages over other image segmentation methods.In a subjective perspective, the results of the NS-RMRF segmentation algorithm can indicate the true distribution of surface features. In an objective perspective, the proposed algorithm achieves better segmentation results than the three other methods based on the overall accuracy and Kappa coefficient.The segmentation algorithm of NS-RMRF is proposed in this study and a new calculation method that measures the regional difference is also presented. Noise-reduction filtering algorithm is used to establish a valid initial segmentation and maintain edge information by considering the difference between regions in a spatial context model. The proposed algorithm can effectively capture and maintain details and smoothen homogeneous areas more effectively than ML, region-based WMRF, and PolarIRGS algorithms. However, the proposed method are limited by deficiencies; thus, the proposed algorithm should be improved.