面向光谱-空间特征集合的高光谱遥感影像度量学习与分类研究

Classification is an important and hot research issue in the field of hyperspectral remote sensing image processing. The traditional spectral-based classification methods perform poor mainly because they don’t exploit the spatial information and haven’t considered the inter-pixel correlations. Combining the spatial and spectral information for hyperspectral image (HSI) processing is an efficient way to improve the classification performance and becomes the development trend of hyperspectral analysis, where the use of spatial information and construction of spatial and spectral similarity metrics are crucial. Therefore, by combining the spectral pixel and its spatial neighboring pixels, the project builds set-to-set similarity metrics for the HSI classification based on the local homogeneous pixel sets of HSI, which aims to obtain desirable spectral-spatial classifiers. Based on the spectrally point-to-point similarity metric, the project focuses on: 1) convex-hull-based set-to-set similarity metric learning and classification model, including the description of the set-to-set structural similarity, regularized set-to-set distance metric learning, set-based sparse representation classification, kernel classification and collaborative representation classification algorithms; 2) set-to-set kernel similarity metric learning and classification model, including the multi-scale representation of local homogeneous regions, definition of set-to-set kernel, ideal regularization for the set-to-set kernels, composite spatial and spectral set-to-set kernels framework. The scientific research achievements help to improve the accuracy of HSI classification and to promote the remote sensing information processing level, which have important academic value and large application prospect.

分类是高光谱遥感图像处理领域的研究热点。传统的基于光谱像素点的分类方法，因未考虑空间邻域像素的相关性，而效果欠佳。结合空间域和光谱域信息（空谱信息），进行高光谱图像分析，是提升分类性能的有效途径。其中，关键问题在于如何有效地利用空谱信息，建立合理的光谱域和空间域相似性度量。为此，本项目将光谱像素点与其空域邻域像素相结合，以高光谱图像局部同源区域或像素集为对象，建立区域或集合相似性度量，以期实现准确的地物目标分类。项目研究内容如下：1）集合凸包距离度量学习与分类，具体包括像素集结构相似性描述、正则化集合度量学习、像素集稀疏表示、核表示及协同表示分类；2）集合核度量学习与分类，包括局部同源区域的多尺度表示框架、集合核定义、集合核理想正则化提升算法、空谱组合集合核分类框架。项目研究成果有助于提升高光谱图像分类精度、推进高光谱遥感信息处理水平，具有重要的学术价值和较大的应用前景。

度量距离反映样本间的相似关系。通过学习得到准确的度量关系是分类成功的关键。本项目面向光谱—空间特征集合，开展像素集到像素集的集合凸包距离度量和集合核度量学习与分类研究，有助于改善高光谱图像像素/像素集之间的相似性度量关系、提升高光谱图像分类性能、推进高光谱遥感信息处理水平，具有重要的学术价值。项目研究成果能够应用于农作物遥感分类、湖泊湿地分类、土地利用/覆盖分类等方面，具有较大的应用前景。按照项目研究计划，项目组成员圆满完成了预期研究目标，取得了一些有意义的研究结果。分析了像素集内各像素表示系数的非一致近邻正则化关系，建立了自适应像素集以及基于像素集合的距离度量学习与分类模型。开展了集合核度量学习与分类模型研究，建立了基于组合集合核的支持向量机和极限学习机器算法，并进一步地建立了用于提升集合核性能的理想正则化一般框架。考虑像素集中不同像素之间差异性，建立了近邻加权联合稀疏表示模型和基于最大相关熵的稳健联合稀疏表示模型。