结合空域信息与显著极化特征的PolSAR深度学习分类方法研究

Polarimetric synthetic aperture radar (PolSAR) has attracted widespread attention all over the world due to its powerful capability of target information acquisition and broad prospects of application. The increasing image resolution and the advanced techniques of machine learning have brought new development opportunities to high-precision and intelligent classification methods for PolSAR images. However, the inherent scattering characteristics of PolSAR data lead to some problems when directly applying deep learning-based classification methods. These problems include the effect of speckle noise, the complicated scattering mechanisms, and the sensitivity of PolSAR data to observation conditions. This project aims to incorporate the domain knowledge of PolSAR interpretation to extract the salient polarimetric features which are with superior capability of characterizing the scattering information. Then the input of deep learning networks can be optimized, thus facilitating the model training process and also improving the classification performance. Moreover, the rich and fine spatial information of PolSAR image is incorporated into feature extraction and image classification, so that the classification result can be improved further. The researches of this project are as follows: accurate extraction of spatial information, the extraction and analysis of polarimetric features, and the integration of spatial information and salient polarimetric features for PolSAR classification. Overall, the spatial information is beneficial to extracting reliable polarimetric features, and these two kinds of information, i.e., spatial information and polarimetric information are complementary to each other. Therefore, the integration of them can greatly improve the classification performance, and preserve the detailed structural information as well.

极化合成孔径雷达（PolSAR）因其强大的目标信息获取能力和广阔的应用前景引起各国普遍关注。图像分辨率不断提升以及机器学习快速发展给高精度、智能化PolSAR分类带来了新的发展契机，但是数据固有的散射特性导致基于深度学习的分类方法面临一些突出问题，例如相干斑噪声、复杂散射机理、数据对观测条件敏感等。本项目拟结合极化SAR解译的领域知识来深入挖掘具有优越地物散射表征能力的特征（显著极化特征），优化深度学习网络的数据输入，降低模型训练难度并提升分类性能；另一方面，结合图中丰富的空域信息进行特征提取和图像分类，从而进一步提升精细分类的效果。因此，拟开展如下研究：空域信息精准提取、极化特征提取与分析、结合空域信息与显著极化特征的图像分类。其中空域信息有助于提取稳定可靠的极化特征，所获取的极化信息与空域信息具有一定的互补性，将两者共同应用于PolSAR分类能够改善分类效果，且保留图像的细致结构信息。

极化SAR强大的信息获取能力以及主动微波成像的独特优势使其成为民用遥感与军事应用中不可替代的重要手段。PolSAR数据大量获取、图像分辨率不断提升以及机器学习快速发展给高精度、智能化PolSAR分类带来了新的发展契机，然而极化SAR数据固有的散射特性导致相关研究面临一些突出问题，表现在相干斑噪声影响、散射机理复杂、训练样本有限、数据对观测条件敏感等。针对这些问题，本项目重点围绕极化SAR图像显著极化特征提取、超像素分割、深度学习网络模型构建、空域信息利用等关键技术问题展开研究。项目取得的成果如下：提出了基于有监督流行学习的极化SAR特征提取方法，实现了高维极化信息的有效压缩与去冗，同时实现了邻域空间信息的有效利用；构建了基于胶囊网络的极化SAR分类网络框架，实现了多层级特征信息的有效利用；提出了基于最小生成树的极化SAR图像多尺度超像素分割方法，构建了像素间更加精确的距离度量，实现多尺度超像素的快速精准分割；提出了基于超像素图卷积神经网络的极化SAR分类方法，在提升分类精度和效率的同时降低了对训练样本的需求。研究成果发表在IEEE TRGS，IEEE JSTARS和IEEE GRSL等高水平期刊和国际会议上，共23篇。研究成果能够进一步丰富极化SAR图像解译理论和方法，提升深度学习方法在极化SAR图像解译方面的应用水平，实现对极化SAR图像不同地物和目标的快速准确分类。