面向遥感图像分类的稀疏流形降维方法研究

Taking into account the continuous accumulation, remote sensing images inherit the massive, high-dimensional external features and the internal characteristics of strong intra-class differences, low inter-class similarity, and complex disturbance factors, which pose many new problems and challenges for their classification. With this, the past several decades have witnessed a significance development in using manifold learning to study the problems of remote sensing image dimensional reduction. This is due to the fact that the manifold learning could be able to eliminate redundant information, reduce dimension, and extract intrinsic information, so that it is surely a valuable and promising preprocessing method for remote sensing image classification. However, there still exist a lot of problems for the divisibility and the uniform of low-dimensional representation when using current manifold learning methods to analyze the remote sensing image. Based on the above considerations, the purpose of this project is to address the new characteristics of remote sensing images with complex feature information and large background noise. Firstly, we use the sparse representation-based high-dimensional spatial neighbor distribution to investigate the accurately represent complex manifold distribution by comprehensively considering their local and global features. Secondly, we discuss a KL divergence-based similarity measurement method by embedding the high-dimensional spatial manifold structure to obtain a uniform, separable low-dimensional representation. Moreover, we study the convolutional neural network-based light-weight generalization model to adaptively construct a direct mapping from high dimension to low dimension. With above considerations, this project is hoped to achieve a low-dimensional representation with inter-class separable and intra-class consistency so as to offer some fundamentally support to improve the accuracy of remote sensing image classification.

随着遥感图像的不断积累，其呈现出海量、高维度的外部特征与类内差异性强、类间相似度低、干扰因素复杂的内部特点，对遥感图像分类提出新的挑战。流形学习方法因具有剔除冗余信息、约简数据维度、提取本征信息的能力被引入到遥感领域，是一种有价值和前景的遥感图像分类预处理方法。现有的面向遥感图像分析的流形学习方法仍然存在数据低维表示可分性与均匀性不足的问题。本项目拟针对具有复杂地物信息、繁杂背景噪声的遥感图像呈现出的新特点，首先，研究基于稀疏表征的高维空间近邻分布模型，综合考虑地物的局部和全局特征，准确刻画其在高维空间中的流形分布；然后，研究基于KL散度的相似性度量方法，实现高维空间流形结构有效的嵌入，获得均匀、可分的低维表示；最后，研究轻量级基于卷积神经网络的泛化模型，自适应地构建高维空间与低维空间的直接映射。本课题可实现遥感图像类间可分性和类内一致性的低维表示，为遥感图像分类精度的提高提供底层支持。

遥感大数据呈现出海量、高维度的外部特征与类内差异性强、类间相似度低、干扰因素复杂的内部特点，流形学习方法因具有剔除冗余信息、约简数据维度、提取本征信息的能力被引入到遥感领域，是一种有价值和前景的遥感图像分类预处理方法。本项目提出了一种基于混合高斯模型的高维空间近邻模型，实现了遥感大数据高维空间分布的有效刻画；针对遥感数据波段间的空间冗余问题，提出了协同耦合的稀疏表征方法，实现了局部特征与全局特征的兼顾；研究了基于卷积神经网络的泛化模型，实现了遥感数据的类间可分性核类内一致性的有效表征，为遥感图像的分类问题提供有效支撑。