基于稀疏表示的高光谱图像分类和解混方法研究

发布时间：2018-01-23 14:21

本文关键词： 高光谱图像稀疏表示空间信息图像分类光谱解混　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：高光谱图像分类技术和光谱解混技术在高光谱遥感技术领域占有非常重要的地位,一直是国内外学者的重点研究方向。近年来,随着稀疏表示在图像处理方面的广泛应用,一些学者开始将其应用到了高光谱遥感图像处理方面,尤其是高光谱图像分类和光谱解混方向,已经取得了一些成就。但是现有的基于稀疏表示的图像分类和光谱解混方法只考虑了高光谱图像的光谱信息,没有考虑到图像的空间信息和高光谱图像本身的特征信息。在高光谱图像中,存在这样一种现象,相邻像元可能包含相似或相同的物质,并且物质的含量也是相似的,基于上述特征,可以利用中心像元的邻居像元在分类和解混模型中添加空间约束,将高光谱图像的空间信息和光谱信息结合起来,从而提高分类和解混的精度。论文的研究方向主要有以下两方面:1.在高光谱图像中,存在这样一种现象,相邻像元可能包含相似或相同的物质,这样,它们很可能在分类过程中归为同一类。基于高光谱图像的这种特征,本文提出了一阶邻域系统加权约束,即用中心像元周围的4个像元在分类模型中去约束该像元,并且与中心像元越相近的邻居像元在约束中占的比重越大,该约束使得中心像元和周围的4个像元在分类过程中包含同样的光谱信息。然后将此约束添加到稀疏分类模型中,提出了一种基于一阶邻域系统加权约束的新的分类算法。为测试新算法的分类性能,利用常见的AVIRIS和ROSIS传感器搜集的高光谱图像进行实验,采用总分类精度、平均分类精度和kappa系数3种评价标准评价算法性能。实验结果表明,一阶邻域系统加权约束充分利用了空间信息和图像本身的特征,分类精度有了大幅提高,分类性能优于现有分类算法。2.高光谱图像中的像元是由光谱信息和空间信息共同组成的,光谱信息是独立的,而空间信息是相关的。由于马尔科夫随机场是一个模拟空间相关性的强大工具,它不仅考虑到图像中相邻像元的相关性,同时也考虑到了图像本身的特征,所以本文采用马尔科夫随机场在稀疏解混模型中添加空间相关性约束,提出了一种基于自适应的马尔科夫随机场的稀疏解混算法。为测试该算法的解混性能,本文提供了模拟图像数据和真实的AVIRIS图像数据进行实验,并利用SRE(信号噪声比)分析实验结果。实验结果表明,对于模拟图像,基于自适应的马尔科夫随机场的稀疏解混算法取得了更高的SRE值,对于真实图像,新算法解混后得到的丰度图像比现有算法更加光滑,丰度图像细节也展现的更加全面。
[Abstract]:The technology of hyperspectral image classification and spectral deconvolution plays a very important role in the field of hyperspectral remote sensing, and has been a key research direction of domestic and foreign scholars in recent years. With the wide application of sparse representation in image processing, some scholars begin to apply it to hyperspectral remote sensing image processing, especially hyperspectral image classification and spectral deconvolution. Some achievements have been made, but the existing image classification and spectral de-mixing methods based on sparse representation only consider the spectral information of hyperspectral images. The spatial information of the image and the characteristic information of the hyperspectral image are not taken into account. In the hyperspectral image, there is a phenomenon that adjacent pixels may contain similar or identical substances. And the content of matter is similar, based on the above characteristics, we can use the neighbor pixel of the center pixel to add spatial constraints to the classification and the mixed model, and combine the spatial information and spectral information of hyperspectral image. In order to improve the accuracy of classification and mixing. The main research direction of this paper is as follows: 1. In hyperspectral images, there is a phenomenon that adjacent pixels may contain similar or identical substances. It is very likely that they can be classified into the same class in the classification process. Based on this feature of hyperspectral images, a first-order neighborhood system weighted constraint is proposed in this paper. That is to say, four pixels around the center pixel are used to constrain the pixel in the classification model, and the neighbor pixel which is close to the center pixel occupies a larger proportion in the constraint. The constraint makes the central pixel and the surrounding pixel contain the same spectral information in the classification process. Then the constraint is added to the sparse classification model. A new classification algorithm based on weighted constraints of first-order neighborhood system is proposed. In order to test the classification performance of the new algorithm, hyperspectral images collected by common AVIRIS and ROSIS sensors are tested. The performance of the algorithm is evaluated by three evaluation criteria: general classification accuracy, average classification accuracy and kappa coefficient. The experimental results show that the weighted constraints of the first-order neighborhood system make full use of the spatial information and the features of the image itself. Classification accuracy has been greatly improved, classification performance is better than the existing classification algorithm .2.The pixel in hyperspectral image is composed of spectral information and spatial information, and spectral information is independent. Because Markov random field is a powerful tool to simulate spatial correlation, it not only considers the correlation of adjacent pixels in the image, but also takes into account the characteristics of the image itself. In this paper, a sparse demultiplexing algorithm based on adaptive Markov random field is proposed by adding spatial correlation constraints to the sparse demultiplexing model. In this paper, the simulated image data and the real AVIRIS image data are provided, and the experimental results are analyzed by using the SRE (signal-noise ratio). The experimental results show that, for the simulated images. The sparse demultiplexing algorithm based on adaptive Markov random field achieves a higher SRE value. For real images, the abundance images obtained by the new algorithm are more smooth than the existing algorithms. Abundance image details are also more comprehensive.
【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】：2014
【分类号】：TP751

【参考文献】