可变类空间约束高斯混合模型遥感图像分割

摘要

针对传统高斯混合模型(GMM,Gaussian mixture model)难以自动获取类属数和对噪声敏感问题,提出了一种基于可变类空间约束GMM的遥感图像分割方法.首先在构建的GMM中,将像素类属性建模为马尔可夫随机场(MRF,Markov random field),并在此基础上定义其先验概率;结合邻域像素类属性的后验概率和先验概率,定义噪声平滑因子,以提高算法的抗噪性;在参数求解过程中,分别采用可逆跳变马尔可夫链蒙特卡罗(RJMCMC,reversible jump Markov chain Monte Carlo)方法和最大似然(ML,maximum likelihood)方法估计类属数和模型参数;最后以最小化噪声平滑因子为准则获取最终分割结果.为了验证提出的分割方法,分别对模拟图像和全色遥感图像进行了可变类分割实验.实验结果表明提出方法的可行性和有效性.%In view of the traditional Gaussian mixture model (GMM), it was difficult to obtain the number of classes and sensitive to the noise. A remote sensing image segmentation method based on spatially constrained GMM with unknown number of classes was proposed. First, in the built GMM, prior probability that represented the membership between a pixel and one class was modeled as a Markov random field (MRF). In order to improve the sensitivity of noise, the smoothing factor was defined by combining the a posterior probability and the prior probability of neighboring pixels. For estimating the number of classes and the parameters of model, the reversible jump Markov chain Monte Carlo (RJMCMC) and maximum likelihood (ML) estimation were employed, respectively. Finally, by minimizing the smooth-ing factor the final segmentation was obtained. In order to verify the proposed segmentation method, the synthetic and real panchromatic images were tested. The experimental results show that the proposed method is feasible and effective.