融合深度可分离卷积的多尺度残差UNet在PolSAR地物分类中的研究

极化合成孔径雷达(Polarimetric Synthetic Aperture Radar, PolSAR)地物分类作为合成孔径雷达(Synthetic Aperture Radar, SAR)图像解译的重要研究内容之一，越来越受到国内外学者的广泛关注。不同于自然图像，PolSAR数据集不仅具有独特的数据属性同时还属于小样本数据集，因此如何更充分地利用数据特性以及标签样本是需要重点考虑的内容。基于以上问题，该文在UNet基础上提出了一种新的用于PolSAR地物分类的网络架构——多尺度可分离残差UNet (Multiscale Separable Residual Unet, MSR-Unet)。该网络结构首先利用深度可分离卷积替代普通2D卷积，分别提取输入数据的空间特征和通道特征，降低特征的冗余度；其次提出改进的多尺度残差结构，该结构以残差结构为基础，通过设置不同大小的卷积核获得不同尺度的特征，同时采用密集连接对特征进行复用，使用该结构不仅能在一定程度上增加网络深度，获取更优特征，还能使网络充分利用标签样本，增强特征传递效率，从而提高PolSAR地物的分类精度。在3个标准数据集上的实验结果表明：与传统分类方法及其它主流深度学习网络模型如UNet相比，MSR-Unet网络结构能够在不同程度上提高平均准确率、总体准确率和Kappa系数且具有更好的鲁棒性。

Research on Multi-scale Residual UNet Fused with Depthwise Separable Convolution in PolSAR Terrain Classification

As one of the important research contents of Synthetic Aperture Radar(SAR) image interpretation, Polarimetric Synthetic Aperture Radar(PolSAR) terrain classification has attracted more and more attention from scholars at home and abroad. Different from natural images, the PolSAR dataset not only has unique data attributes but also belongs to a small sample dataset. Therefore, how to make full use of the data characteristics and label samples is a key consideration. Based on the above problems, a new network on the basis of UNet for PolSAR terrain classification—Multiscale Separable Residual Unet(MSR-Unet) is proposed in this paper. In order to extract separately the spatial and channel features of the input data while reducing the redundancy of features, the ordinary 2D convolution is replaced by the depthwise separable convolution in MSR-Unet. Then, an improved multi-scale residual structure based on the residual structure is proposed. This structure obtains features of different scales by setting convolution kernels of different sizes, and at the same time the feature is reused by ​​dense connection, using the structure can not only increase the depth of the network to a certain extent and obtain better features, but also enable the network to make full use of label samples and enhance the transmission efficiency of features information, thereby improving the classification accuracy of PolSAR terrain. The experimental results on three standard datasets show that compared with the traditional classification methods and other mainstream deep learning network models such as UNet, the MSR-Unet can improve average accuracy, overall accuracy and Kappa coefficient in different degrees and has better robustness.