基于最近正则子空间模型的高光谱遥感图像分类及异常检测研究

Hyperspectral image classification and anomaly detection have always been hot research topics in the field of remote sensing. With the development of remote sensing instruments, high spatial and spectral resolution images are available, nevertheless, further research on improving the performance of object classification/recognition and feature extraction is necessary. Nearest regularized subspace, recently proposed for hyperspectral imagery analysis, is a supervised classification model which has been proved to be superior to other state-of-the-art methods, such as support vector machine and sparse representation based classification. In this proposal, based on the previous work, we mainly focus on how to polish up the process of distance-weighted learning, and make it more discriminative; how to add constraints to the weight vector, such as nonnegative or sum-to-one; how to implement the classifier based on a kernel-induced feature space; how to reasonably utilize the spatial information as post-processing for the pixel-wise classifier which is only based on wealthy spectral signatures; for anomaly detection, how to modify the supervised classifier with training samples into an unsupervised detector without a priori information. The novel pattern classification approaches proposed by this research, are expected to assist the therotical analysis and applications in hyperspectral imagery.

高光谱遥感图像的分类及异常检测一直是遥感技术研究的热点。随着遥感技术的不断发展，人们可以获得更高空间分辨率和光谱分辨率的图像数据，而在发展更为精确的地物分类、地物识别、地物信息提取方面的研究还有待进一步深入。申请者在前期研究中提出一种高光谱遥感图像分类模型- - -最近正则子空间，该技术已经被证明优于目前常用的支持向量机和基于稀疏表示的分类算法。本课题拟在此前期研究的基础上，改善加权距离度量学习过程，产生更具分类性能的权重向量；增加权重向量限制使其具有丰度值非负或者和为一的约束；引入核子空间样本选择技术；利用高光谱遥感图像的空间信息，对基于像元光谱信息的分类结果进行智能修正，提高分类效果；在进行异常检测时，转化该技术使之成为不需要任何先验信息的无监督学习过程。本研究提出新的模式识别方法，可望有助于高光谱遥感图像的分析理论及应用。

在此青年基金“基于最近正则子空间模型的高光谱遥感图像分类及异常检测研究”项目的支持下，分析了最近正则子空间模型的不足及改进措施，重点研究了如何将该模型更好的用于高光谱遥感图像的分类和异常检测中。项目执行期间共发表SCI收录期刊论文21篇，其中项目ESI高被引论文2篇。这些理论成果很好地阐明了最近正则子空间和高光谱数据的相互作用关系和特点，设计和实现了包括基于核空间的正则化模型，基于空间谱间特征的表示模型分类和基于协同表示的异常检测算法，为高光谱遥感图像分析提供新技术，也为促进高光谱图像在更多领域得到合理应用和发展做出贡献。