基于高光谱数据的抗噪广义模型及解混算法研究

Spectral unmixing is a key technology in the hyperspectral data processing. Multiplicative noises are brought about in the process of hyperspectral imaging, which are caused by the changes in atmospheric conditions and the relative positions of the sensor, surface, and illumination source. However, the traditional unmixing models consider the reconstruction random errors as the additive noise, which neglect the influences of the multiplicative and mixed noises. Moreover, the linear mixture model (LMM) does not perform well for the effects of nonlinearity. In order to solve the interferences of diverse noises, our project studies linear anti-noise unmixing model and corresponding algorithms. Furthermore, the nonlinear influences on the linear anti-noise model is considered. As a result, the accuracies and stabilities of the unmixing results can be improved..The contents of this project are listed as follows: (1) The improved unmixing algorithms are presented, which are utilized to design the dictionaries. (2) Study the noises evaluation algorithms. Construct an appropriate divergence to overcome the diverse noises. The relationship between the errors of the linear anti-noise model and the nonlinear effects is analyzed. Segment the spatial space by the clustering methods, and detect the nonlinear term by the theory of the markov random field. Construct a generalized anti-noise unmixing model. (3) The theories of the compressed sensing and the sparse representation are stutied, and a series of fast hyperspectral unmixing algorithms are proposed. Experiments are executeded to demonstrate the achievements of our project. In order to guarantee the effectiveness of our project, the experiments results to the real distributions of the ground are compared. .The achievements of the projects will provide new unmixing models for spectral unmixing. The theories and the algorithms of our project can be directly used in the remote sensing image classification, interpreted. In addition, the research on the multiplicative noise will provide new ideas for medical imaging, video surveillance, radar and other remote sensing image processing.

混合像元解混是高光谱数据处理的关键技术。传统的解混模型忽略了乘性噪声及混合噪声，且线性解混模型忽略了端元之间非线性作用。本项目通过抗噪模型及算法研究解决解混过程中多种噪声干扰，在线性抗噪模型基础上考虑非线性作用，并构建字典获得合适的初始值，提高解混结果精确性、稳定性。.研究内容：1）改进传统的混合像元解混算法，提取端元，设计端元字典库；2）研究噪声评估算法，构造抗噪分离度准则；分析抗噪线性模型误差与非线性项的关系，对图像进行聚类分析，根据马尔科夫随机场理论检测非线性项，构建抗噪广义解混模型；3）研究压缩感知理论及稀疏重建算法，提出解决抗噪广义模型的快速解混算法；实验验证本项目研究成果，并结合实地考察数据评价其实效性。.本项目将为混合像元分解提供新的解混模型及算法，研究成果可以直接应用于遥感图像分类、解译。此外，处理噪声的研究将为医学图像、雷达遥感图像及视频监控等处理提供新思路。

由于高光谱数据的混合像元解混是个不适定问题，且高光谱数据成像过程中受到外界多种环境因素干扰，在高光谱数据的海量信息中提取有效信息提高混合像元解混结果精确性及稳定性面临着诸多挑战。本项目组针对复杂的成像机制带来的多样化噪声、非线性因素使得解混结果更加不精确、不稳定问题，我们分析线性抗噪解混模型的信号重构误差，解决解混过程中的非线性影响；评估高光谱图像中多种噪声，构造有别于常用的欧几里德距离的分离度准则，构建广义抗噪解混模型，抑制多种噪声对解混结果的干扰，同时考虑非线性因素导致的误差。在基于本项目构建的新模型基础上，我们提出了新的解混算法。本课题理论贡献在于：(1)所提出的解混算法能够适用于不同传感器拍摄的高光谱数据的解混问题；(2)通过改进的解混算法求得的端元字典保证了初始值的选择，降低了解混结果对初始值依赖性，且Block Coordinate Descent算法保证了迭代效率；（3）虽然压缩感知理论已经被广泛应用于混合像元解混领域，但压缩感知问题求解是NP-hard的，且高光谱数据中端元谱线与地面实测谱线之间的差异性导致了压缩感知算法求得的结果误差较大。本项目组利用高光谱数据提取字典，由压缩感知及稀疏重建理论设计解混算法，避免了上述问题。项目研究成果将直接促进定量遥感技术的发展。特别地，在军事领域，本项目“同谱异物、同物异谱”的研究成果可以实现辅助伪装规划，使得我方目标更好地与“背景”融合，达到隐真示假目的。此外，项目研究成果可以推广到视频理解、环境监测、医学图像处理、海洋探测等领域。在项目资助下，项目组在计算机领域较有影响力的期刊上发表科研论文5篇，获得授权的发明专利8项。在本项目研究成果辐射下，项目负责人入选“湖州市1112人才工程培养人选后备人选”，项目组成员获得浙江省自然科学基金2项，湖州市科技计划项目一项，项目组培养2名青年教师晋升为副教授，毕业博士生2名。