基于进化多目标优化的高光谱稀疏混合像元分解技术研究

发布时间：2018-06-28 07:36

本文选题：高光谱遥感 + 线性混合模型　；参考：《西安电子科技大学》2015年硕士论文

【摘要】：随着高光谱遥感技术飞速发展,高光谱遥感影像数据已经被广泛应用到环境污染监测、矿产勘探等方面。传感器技术的发展瓶颈以及地表地物分布复杂多样,导致在高光谱遥感数据中普遍存在混合像元。混合像元分解精度直接影响着高光谱遥感影像数据的后续应用,因此,混合像元分解技术已经成为高光谱遥感技术中一个亟待解决的问题。随着混合像元分解技术的重要性日益突出,越来越多的研究学者加入到了混合像元分解技术这个领域的研究中,这又进一步促进了混合像元分解技术的发展。其中,基于稀疏约束的混合像元分解技术已经成为混合像元分解技术的潮流,其旨在从已知光谱库中寻找能够表示观测像元光谱向量的最佳线性组合。传统的基于稀疏约束的混合像元分解算法,使用凸的L1范数替代非凸的L0范数,然后通过引入权重系数将两个或者多个代价函数整合成一个目标函数进行优化。但是L1范数不一定在所有情况下都等价于L0范数。此外,传统的基于稀疏约束的混合像元分解算法引入的权重系数都比较敏感,而且对于不同的数据集难以确定一个最佳的参数。因此,为了解决传统的基于稀疏约束的混合像元分解算法中出现的一些问题,在前面学者研究的基础上,本文做了一些关于稀疏约束混合像元分解问题的研究,主要研究内容如下:首先,对当前一些高光谱混合像元分解技术做了介绍。根据混合尺度的情况,混合像元分解模型主要分为线性模型和非线性模型。线性模型是当前使用最为广泛的模型,本文的研究也是基于线性模型进行的。当前的线性模型算法主要分为四大类:基于信号子空间的、基于空间几何的、基于统计的和基于稀疏约束的算法。其次,针对当前基于稀疏约束的混合像元分解算法中存在的一些问题,本文将该算法建立成一个进化多目标优化模型。为了解决进化多目标模型中遇到的大规模优化问题,本文提出了协同进化多目标算法MOSU。由于高光谱遥感图像中每个像元的丰度向量是稀疏的,因此我们提出了基于稀疏性的分组策略,以便对目标函数进行更为高效的优化。此外,本文还提出了基于拐点的协作机制,避免了复杂的计算,更加适用于混合像元分解问题。通过仿真实验对比,证明了本文提出的多目标稀疏约束混合像元分解算法的有效性。最后,为了有效利用图像的空间上下文信息,本文在MOSU的基础上增加了一个有关空间邻域信息的目标函数。这将高光谱影像数据的光谱信息和空间上下文信息结合起来,充分利用每个像元与其周围邻域像元之间的关系。仿真实验证明,引入空间上下文信息提高了算法对噪声的鲁棒性。
[Abstract]:With the rapid development of hyperspectral remote sensing technology, hyperspectral remote sensing image data have been widely used in environmental pollution monitoring, mineral exploration and other aspects. The bottleneck of sensor technology and the complex distribution of surface and ground objects lead to the widespread existence of mixed pixels in hyperspectral remote sensing data. The accuracy of mixed pixel decomposition directly affects the subsequent application of hyperspectral remote sensing image data. Therefore, hybrid pixel decomposition technology has become an urgent problem in hyperspectral remote sensing technology. With the increasing importance of hybrid pixel decomposition, more and more researchers have joined in the field of hybrid pixel decomposition, which has further promoted the development of hybrid pixel decomposition. Among them, the mixed pixel decomposition technique based on sparse constraints has become the power flow of hybrid pixel decomposition technology, which aims to find the best linear combination to represent the spectral vector of observed pixel from the known spectral database. The traditional mixed pixel decomposition algorithm based on sparse constraints uses convex L1 norm instead of non-convex L0 norm, and then integrates two or more cost functions into one objective function by introducing weight coefficients. But L1 norm is not necessarily equivalent to L 0 norm in all cases. In addition, the traditional mixed pixel decomposition algorithm based on sparse constraints is sensitive to the weight coefficients, and it is difficult to determine an optimal parameter for different data sets. Therefore, in order to solve some problems in the traditional mixed pixel decomposition algorithm based on sparse constraints, based on the previous research, this paper does some research on sparse constraint mixed pixel decomposition problem. The main research contents are as follows: firstly, some techniques of hyperspectral mixed pixel decomposition are introduced. According to the mixed scale, the mixed pixel decomposition model is mainly divided into linear model and nonlinear model. Linear model is the most widely used model at present, and the research in this paper is also based on linear model. The current linear model algorithms are divided into four main categories: signal subspace-based, spatio-geometric, statistical and sparse-constrained algorithms. Secondly, aiming at some problems in the current mixed pixel decomposition algorithm based on sparse constraints, this paper presents an evolutionary multi-objective optimization model. In order to solve the large-scale optimization problem in evolutionary multi-objective model, a co-evolutionary multi-objective algorithm MOSU is proposed in this paper. Because the abundance vectors of each pixel in hyperspectral remote sensing images are sparse, we propose a sparse grouping strategy to optimize the objective function more efficiently. In addition, this paper proposes a cooperation mechanism based on inflection point, which avoids complex computation and is more suitable for mixed pixel decomposition. The simulation results show the effectiveness of the proposed algorithm. Finally, in order to effectively utilize the spatial context information of the image, this paper adds an objective function about spatial neighborhood information based on MOSU. It combines the spectral information of hyperspectral image data with the spatial context information and makes full use of the relationship between each pixel and its neighboring pixel. Simulation results show that the robustness of the algorithm to noise is improved by introducing spatial context information.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP751

【相似文献】