基于区域统计特征的PolSAR影像分割

本文提出一种利用区域统计特征的PolSAR影像分割方法，以解决目前研究中在降低斑点噪声和提高分割效率方面的不足。首先利用基于梯度分割影像的分水岭算法进行SAR影像初分割，针对差值梯度非恒虚警率(CFAR)，可能给出虚假边缘从而导致分割错误的问题，引入恒虚警率的均值比率梯度（ROA，Ratio of Average）；同时，考虑到梯度影像中存在大量局部极小值，直接用分水岭处理得到初分割结果，存在过度过分割，给出了一种利用形态学方法进行梯度重构以消除局部极小值、抑制过分割的方法。然后，基于初分割得到区域，计算区域相干矩阵的最大似然估计，结合假设检验和相干矩阵的Wishart分布，给出一种有效描述区域相似度的目标函数，通过建立区域邻接关系图(RAG，Region Adjacent Graph)，执行等级区域合并得到最终分割结果。利用模拟数据，德国奥伯法芬霍芬L波段实测数据和中国海南陵水黎族自治县X波段的高分辨率数据，验证文中方法，初分割结果证实梯度重构处理不会破坏原有梯度结构，并能有效抑制过分割；最终分割结果定性定量对比分析表明，文中给出的目标函数，在分割效率、信息保持和分割精度上都有较好表现。上述结果证实了文中方法能有效降低斑点噪声，提高分割效率，从而提供更加准确的分割结果。

PolSAR imagery segmentation based on region statistical characteristic

A method of PolSAR imagery segmentation based on region statistical characteristic is proposed in this paper to deal with the shortage of recent reasearch in speckle noise and sgementation efficiency. Firstly, watershed using gradient to segment imagery is used to get initial segmentation. The ROA gradient is introduced due to the differential gradient is not constant false alarm rate which probably provides false edges. Meanwhile, a method of gradient construction based on morphological operation is given to eliminate the local minimum of gradient which would induce massive over-segmentation after watershed transform and suppress over-segmentation. After that the maximum likelihood estimataion of region coherence matrix is estimated by the arithmetical mean of pixels in region. Then, an object function effectively weigh dissimilarity between regions is deduced in combination with Wishart distribution of coherence matrix and hypothesis test, and the RAG is build base on object function, the final segmentation is retrieved after hierarchical regions merging. The simulation data, L-band airborne provided by DLR of Oberpfaffenhofen Test Site and X-band PolSAR data aqucistion in Hainan Lingshui County are used to validate method of this paper. The initial segmentation proved that gradient reconstruction would not undermine the original gradient construction, and can effectively suppress over-segmentation. The qualitative and quantitative analysis of final segmentation show that object function can provide good segmentation result, and have good performance on segmentation efficiency, information maintenance and segmentation precision. These results validate that the proposed method can effectively reduce the speckle noise, and improve efficiency, thereby it can provides a more accurate segmentation.