Gaussian mixture model (GMM) clustering algorithm is widely used in image segmentation during recent years. The algorithm is however quite sensitive to speckle noise since spatial correlations between pixels are ignored. This paper presents a region-based GMM clustering algorithm for SAR image segmentation featured by incorporating spatial correlations. The watershed algorithm is first used to generate primitive homogeneous regions. Regional mean values are then calculated as input samples of the GMM clustering process. The impact of noise on the segmentation result can therefore be reduced in the space of regions instead of pixels. A feedback mechanism is further introduced into the expectation-maximization (EM) algorithm to improve the precision of parameter estimation. The efficiency of the proposed algorithm has been demonstrated on the segmentation of synthetic SAR images and real SAR images

where the segmentation accuracy has been substantially improved in contrast to pixel-based the GMM algorithm.