基于多视角特征和空间信息的极化合成孔径雷达图像分类方法研究

Polarimetric synthetic aperture radar (PolSAR) is an advanced instrument for remote sensing and it is widely applied in military and civil fields, therefore the accurate interpretation of PolSAR images is crucial, and PolSAR image classification is the key step of PolSAR image interpretation. Some main problems of existing PolSAR image classification methods include the inefficient usage of multi-view features and the spatial information, the uncertainty of the initialization for unsupervised classification methods and so on. Specific to these problems, based on multi-view features and spatial information, this project aims to propose accurate PolSAR image classification methods from the perspectives of machine learning and cognitive learning. The concrete research of this project includes: 1) proposing two multi-view subspace learning methods based on deep local discriminant canonical correlation analysis and adaptive view weights with the tensor representation for accurate supervised classification, 2) using local density peaks and efficient multi-view co-training spectral clustering for accurate unsupervised classification, 3) based on the mechanism of the ventral visual pathway, building a multi-channel two-way hierarchical model to learn superior features, which can provide new ideas for improving the accuracy of PolSAR image classification.

极化合成孔径雷达（SAR）是一种先进的微波遥感工具，广泛应用于军事和民用领域，因此对极化SAR图像的精确解译至关重要，而极化SAR图像分类是极化SAR图像解译的重要环节。本项目针对现有极化SAR图像分类方法未充分利用多视角特征和空间信息、无监督分类时初始状态无法自动确定等问题，从机器学习和认知学习两个角度出发，充分利用多视角特征和空间信息，研究极化SAR图像的精确分类方法。具体包括：1）提出一种基于深度局部判别典型相关分析的多视角子空间学习方法和一种基于张量自适应视角权重的多视角子空间学习方法，提取特征，进而实现极化SAR图像的精确有监督分类；2）通过精确定义和快速寻找密度峰值自动确定类别数目和初始类别中心，并提出高效多视角协同谱聚类方法，实现极化SAR图像的精确无监督分类；3）基于人类腹侧视觉通路认知机理，建立多通路双向层级模型，学习高层特征，为提高极化SAR图像分类的精度提供新的思路。

极化合成孔径雷达（SAR）是一种先进的微波遥感工具，广泛应用于军事和民用领域，因此对极化SAR图像的精确解译至关重要，而极化SAR图像分类是极化SAR图像解译的重要环节。本项目充分利用极化SAR图像的多视角特征，从机器学习的角度，针对有监督分类，提出了基于自适应视角权重多视角子空间学习的极化SAR图像分类方法和基于多视角在线学习的极化SAR图像实时分类方法；针对无监督分类，提出了基于协同图嵌入的极化SAR分类方法和基于稀疏和流形正则化的极化SAR图像多特征选择方法。从认知学习的角度，提出了基于轻量双流卷积神经网络的SAR目标识别方法。以上成果充分利用了极化SAR数据的多视角特征，实现极化SAR数据的精确解译。