星载环境下高光谱遥感图像感知压缩方法研究

With the widely application of the higher spectral resolution remote sensing images, and the current condition of the limited satellite resources, the research on remote sensing images compression and reconstruction, which has less calculation resource consumption, higher compression ratio and higher reconstruction precision, has becoming the key. Around the three critical questions: the learning of adaptive dictionary and the reconstruction of the hyperspectral remote sensing image, this study focus on using the feature difference of hyperspectral remote sensing image from different regions during different time to train the adaptive learning dictionary and set up the optimal remote sensing image sparse representation model. After the analysis of sparse vector coefficients features of hyperspectral remote sensing image multi-scale geometric transformation, this paper proposes a method, which combines discrete chaotic system and the rotation matrix, to construct observation matrix. Finally the study on the high-precision nonlinear reconstruction of hyperspectral remote sensing image is carried out. Strive to reduce the transmission data volume and computing resources consumption through removing redundant information and compressive sensing, improve the quality of the reconstruction image data through the nonlinear and high precision to reconstruction algorithm. This paper aims at making a new breakthrough on theory and methods of acquisition, processing, network sharing and application service quality of multidimensional, multi-scale spatial information. It also provides technical support for the hyperspectral image compression under requirements of low consumption, high compression rate and high accuracy reconstruction, improves hyperspectral remote sensing image information service capabilities of spatial information network.

随着更高光谱分辨率的遥感图像的广泛应用，在星上资源受限的条件下，开展更低计算资源消耗、更高压缩比和更高重构精度的高光谱遥感图像压缩和重构的研究成为关键。围绕星载环境的高光谱遥感图像在观测矩阵构造、适应性字典学习和高精度重构这三方面存在的问题，研究利用不同地域不同时间的高光谱遥感图像数据，训练学习生成自适应字典，构建最优稀疏表达模型，针对性的提出一种结合离散混沌系统与轮换矩阵的观测矩阵的构造方法，研究高光谱遥感图像稀疏表征条件下高精度非线性重构的算法。力争通过高光谱遥感图像感知压缩减少数据量传输和计算资源消耗，通过非线性高精度重构算法提高重构图像数据的质量，争取在多维信息的获取、处理、网络化共享与应用服务质量理论、方法等方面取得新突破，为星载环境低资源消耗、高压缩率和高重构精度要求下的高分遥感图像压缩提供技术支撑，提高空间信息网络高光谱遥感图像信息服务能力。

本项目针对高光谱图像感知压缩问题，围绕观测矩阵构造、自适应字典学习和高精度重构等方面展开研究，取得了一系列重要研究成果。在感知压缩观测矩阵构造方面，构造了非相关轮换混沌观测矩阵以较高的概率精确恢复稀疏信号；在自适应字典学习方面，将基于解析字典的分块图像压缩模型分为线下解析字典学习和利用学习获得的解析字典进行图像压缩两部分来提高系统的效率；在高光谱遥感图像的非线性高精度重构算法方面，提出了一种基于结构化稀疏表示的高光谱图像压缩方法，在稀疏的基础上进一步提升了压缩率，同时，将高光谱图像视为一个三维的张量信号，提出了一种基于稀疏张量字典学习的高光谱图像压缩方法，采用线性的张量块稀疏重构算法缩短了重构时间；基于传统的无监督稀疏子空间聚类算法在高光谱图像中聚类精度受到限制，尤其是在一些高光谱地物聚类过程中会出现许多地物像素错分的情况，无法进行正确有效的聚类。因此在高光谱聚类中，我们提出了基于稀疏子空间算法监督信息的类概率传递模型研究；为了解决低秩图块对角结构下的高光谱遥感影像分类问题，本文提出了空谱块对角联合类概率结构表示和空谱块对角联合模糊类概率结构表示进行高光谱遥感影像的分类。本项目所取得的研究成果，能够在一定程度上推动高光谱图像感知压缩的发展。