SAR图像方向性上下文协方差矩阵构建方法及地物分类应用

SAR image directional context covariance matrix:Construction and its application in terrain classification

合成孔径雷达(SAR)是对地观测等领域的重要传感器,得到了广泛应用。然而,单通道SAR图像每个像元只包含一个复散射值,有限的信息量在一定程度上限制了其应用性能。针对这一问题,本文提出了一种SAR图像空间纹理信息表征方法:方向性上下文协方差矩阵(DCCM)。DCCM通过提取邻域中不同方向上的散射强度变化来拓展利用图像上下文信息,将单个像元从一个标量拓展为一组矩阵,从而极大提升了像元信息容量,实现了信息增广。在此基础上提取的DCCM纹理特征能够更完备地表征地物空间纹理特性,有望应用于地物分类等领域。本文将DCCM纹理特征分别与传统分类器和卷积神经网络结合,构建了基于方向性上下文协方差矩阵的SAR图像地物分类方法。利用AIRSAR和UAVSAR数据开展的对比实验表明,相比基于灰度共生矩阵、Gabor滤波和多级局部模式直方图3种典型纹理特征的分类方法,本文方法结合传统分类器时对总体分类精度提升达到7%以上,结合卷积神经网络时也展现出了优异的分类性能以及更好的稳健性。

Synthetic Aperture Radar(SAR)is a kind of high-resolution imaging sensor,which is able to work under nearly all weather and illumination conditions.SAR plays an important role in earth observation.For single-band,single-polarization SAR image,however,there’s only one complex scalar in each pixel.So that the information contained in such single-channel SAR image could be quite limited,which limits its performance in various applications.Since terrain classification is one of the typical tasks in SAR image interpretation,this paper takes it as an example to demonstrate the problem and gives our solution.To address the above problem,this paper proposes a representation for the spatial information on SAR image—the Directional Context Covariance Matrix(DCCM).DCCM obtains the variance of pixel intensity in several orientations inside the neighborhood in order to make use of the context information.During such process,the target pixel in extended from a complex scalar to a group of matrices,so that its information content is increased.Besides,the matrix form also enables some of the advanced matrix algorithms to be applied to single-channel SAR image.On the basis of it,the DCCM texture feature is derived,which can better represent the texture properties on SAR image and shows better discriminability for different land covers.Then,the texture feature is combined with two traditional classifiers as well as the Convolutional Neural Network(CNN),respectively.Thereafter,a SAR image classification scheme is established.To illustrate the performance of proposed method,terrain classification experiments are carried out on AIRSAR and UAVSAR datasets.Methods based on three commonly used texture features,the gray level co-occurrence matrix(GLCM),Gabor filters and Multilevel Local Pattern Histogram(MLPH)are taken into comparison.On traditional classifiers,the overall classification accuracies are increased by 7%on both datasets.While combining with CNN,the overall accuracies and kappa coefficients are significantly improved with DCCM texture feature than the original SAR data.The proposed feature also shows nice efficiency and better robustness when compared to other texture features.The experiment results indicate that DCCM is an effective representation that is suitable for SAR image.DCCM is efficient,robust and easy-to-use.The proposed DCCM based classification method can improve the classification performance of single-channel SAR image by increasing the pixel information content.Beyond that,DCCM could be a promising method for many other SAR image interpretation tasks.