基于稀疏表示的高光谱图像解混算法研究

发布时间：2018-06-29 05:11

本文选题：高光谱遥感 + 稀疏解混　；参考：《北方民族大学》2017年硕士论文

【摘要】：高光谱遥感图像既含有大量的空间信息,还含有充裕的光谱信息,是遥感领域近年来的研究热点。然而,在实际的高光谱图像中,由于传感器空间分辨率的限制和自然界地物的复杂性,单个像元通常聚集了多种特征地物,它们依据某种比例混合而成,形成混合像元。混合像元的存在阻碍了高光谱图像解释、目标识别以及分类,这就要求解混技术的出现。稀疏解混是高光谱遥感数据分解中常用的线性光谱解混工具,它通过利用预先得到的光谱库完成解混,属于一种半监督的方式。过去,大多数稀疏回归方法都是基于凸松弛的,其试图得到明确定义的优化问题的全局解。近来,由于对低计算复杂度的需求,越来越多的人开始关注稀疏约束的贪婪解混算法,其中,子空间匹配追踪(SMP)依据原始图像的不同列迭代地提取最佳端元,是目前表现较好的算法。本文研究与总结了混合像元分解的相关技术,针对真实光谱图像受噪声所影响严重的现象,在已有稀疏解混算法的基础上,借鉴了几种成熟相似度计算方法并把它们的优缺点进行简单分析对比,提出了SMP稀疏解混的改进算法,用Dice系数法替代内积法作为新的匹配准则,通过计算所有光谱信号的算术平均值,考虑了端元光谱自身的信息,而不仅仅是残差与光谱库的相关度信息,此外,本文还添加一个预分块策略,规避了端元数量很大时算法陷入局部最优的问题,同时也更好的利用了空间信息。
[Abstract]:Hyperspectral remote sensing images not only contain a large amount of spatial information, but also contain abundant spectral information, which is a research hotspot in the field of remote sensing in recent years. However, in the actual hyperspectral images, due to the limitation of sensor spatial resolution and the complexity of natural features, a single pixel usually gathers a variety of feature features, which are mixed according to a certain proportion to form mixed pixels. The existence of mixed pixels hinders the emergence of hyperspectral image interpretation, target recognition and classification. Sparse demultiplexing is a commonly used linear spectral descrambling tool in hyperspectral remote sensing data decomposition. It is a semi-supervised method by using the pre-acquired spectral database. In the past, most sparse regression methods were based on convex relaxation, which attempted to obtain the global solution of the well-defined optimization problem. Recently, due to the demand for low computational complexity, more and more people begin to pay attention to the greedy demultiplexing algorithm with sparse constraints, in which subspace matching tracking (SMP) iteratively extracts the best endpoints according to the different columns of the original image. Is the current performance of the better algorithm. In this paper, the related techniques of mixed pixel decomposition are studied and summarized. Aiming at the phenomenon that the real spectral image is seriously affected by noise, based on the existing sparse demultiplexing algorithms, Several mature similarity calculation methods are used for reference and their advantages and disadvantages are simply analyzed and compared. An improved algorithm for sparse demultiplexing of SMP is proposed. The Dice coefficient method is used instead of the inner product method as a new matching criterion. By calculating the arithmetic mean of all spectral signals, the information of endmember spectrum itself, not only the correlation information between residual and spectral database, is considered. In addition, a preblocking strategy is added in this paper. It avoids the problem of local optimization when the number of endelements is large, and makes better use of spatial information.
【学位授予单位】：北方民族大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】