Abstract: Accurate segmentation is one of main researches in remote sensing image segmentation. To this end, a remote sensing image segmentation method based on the Polya urn model is proposed. In this method, the Polya urn model is used to simulate the relationships between the pixel and its neighbor in segmentation processes. First, the label field model is built by combining Markov random field and a new potential function which is defined by the Polya urn model. Then, the feature field is modeled on the assumption that spectral measures of pixels in each homogeneous region follow an identical and independent Gaussian distribution. Finally, the statistical image segmentation model is built based on Bayesian theorem, and the Metropolis-Hastings (M-H) algorithm is designed to simulate this model to obtain segmentation results. Simulated and real remote sensing images are both segmented by the proposed method. The segmentation results indicate that this method could remove the error pixels and improve the accuracy of segmentation. The qualitative and quantitative analyses demonstrate the feasibility and efficiency of the proposed method.