压缩感知域高光谱图像稀疏解混方法研究

Hyperspectral unmixing technology is the key technology of follow-up analysis and quantitative applications such as hyperspectral remote sensing image classification, feature identification, anomaly target detection, et al. The existing spectral unmixing methods for hyperspectral data in Nyquist sampling domain are not suitable for hyperspectral compressive imaging(HCI) data based on the compressive sensing (CS) theory. HCI data is first reconstructed to obtain high-dimensional reconstruction data and then is unmixed to obtain endmembers and abundance by the existing spectral unmixing method, which has low processing efficiency and complex computing. Furthermore, the reconstruction error of high-dimensional reconstruction data affect the accuracy of the unmixing process. Therefore a novel hyperspectral image unmixing method in compressive sensing domain is necessary. To deal with such problems, according to the property of HCI data, a new high accuracy hyperspectral unmixing method in CS domian is proposed by the establishment of hyperspectral unmixing model in CS domain, which directly obtain endmembers and abundance from HCI data. The research contents are as follows: 1) dimensionality reduction method of hyperspectral data in CS domain; 2) hierarchical backward orthogonal matching pursuit-based endmember extraction algorithm; 3) abundance estimation algorithm based on nonlinear adaptive directional lifting sparse representation. Finally a new high accuracy hyperspectral unmixing method in CS domain is implemented, which provide the theoretical foundation and technical support for the development of hyperspectral remote sensing information processing and quantitative remote sensing applications.

高光谱解混是高光谱遥感图像分类、地物识别、异常目标检测等后续分析和定量化应用的关键。现有的解混方法均针对奈奎斯特采样域的高光谱数据，而对于以压缩感知（CS）理论为基础的高光谱压缩成像技术所获取的CS域高光谱压缩成像数据进行解混，则必须先对压缩成像数据进行重构后再进行解混，导致处理效率低、计算量大，而且重构的高光谱重建数据存在重构误差问题影响解混处理的精度，因此有必要研究新的CS域高光谱压缩成像数据解混方法。本项目将针对高光谱压缩成像数据特性，通过建立CS域高光谱图像稀疏解混模型，直接从压缩成像数据和给定光谱库中求解端元光谱和丰度系数，以获得更高精度的解混性能，包括：1）CS域高光谱数据降维方法；2）基于分层后退型正交匹配跟踪的端元提取；3）基于非线性自适应方向提升稀疏表示的丰度估计算法。项目最终实现高精度CS域高光谱图像解混方法，为高光谱信息处理的发展和定量遥感应用提供理论基础和技术支撑。

本项目分析了高光谱压缩成像数据的特性，并设计了低复杂度的压缩感知域高光谱稀疏解混模型，该模型包含端元提取模型和丰度估计模型；针对已知光谱库中的光谱曲线数远大于实际端元个数，采用贪婪迭代的方式提取光谱库的端元作为丰度估计的端元子集，提出了基于贪婪算法的高光谱图像稀疏解混方法，并设计了快速优化求解方法，具有解混精度高、重构效果好、耗时短、效率高的优点；针对高光谱图像数据的丰度系数存在行稀疏特性和空间相关特性，提出了联合稀疏贝叶斯学习、非凸稀疏低秩约束、非凸l2,q-l2,p约束等丰度估计极小化模型，有效提高了高光谱丰度系数的重构精度。在本项目的资助下，在国内外期刊和国际会议上发表论文14篇，录用1篇，申请国家发明专利2项。