进化多目标高光谱图像波段选择与分类

发布时间：2018-05-16 03:18

本文选题：高光谱图像分类 + 特征选择　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：高光谱图像是成像光谱仪在数十甚至数百个以上的连续的光谱通道上对地物持续遥感成像所成的图像。由于高光谱图像具有很高的光谱分辨率和丰富的波段信息，所以高光谱图像在很多领域都应用广泛，，例如植被、生态、大气、以及海洋等研究领域。但是正因为高光谱图像的波段多，所以它的相邻波段的相关性高，造成了波段冗余，信息重复等问题。所以对于高光谱图像来说，特征选择和分类技术已经成为高光谱图像研究的热点。本文在高光谱图像特征选择和分类上做了以下三个工作： (1)针对传统的基于像素的高光谱图像分类的方法中，只是利用了谱段信息进行分类，没有考虑图像的空间的相关性，本文提出了一种基于均值漂移和稀疏表示分类的高光谱图像空谱分类方法。采用融合机制融合由基于稀疏表示分类得到的高光谱图像分类图和利用均值漂移聚类得到的包含不同像素点的封闭区域的分割图，得到最终的分类结果图。通过该策略，将自适应空间信息融入了分类结果中，增强了区域一致性，大大提高了分类正确性。 (2)提出了基于多目标免疫克隆算法实现高光谱图像同时波段选择和分割，在聚类数未知的情况下，分割部分采用多目标免疫克隆聚类算法同时结合无监督的特征选择，在实现分割的同时实现高光谱图像的波段选择。并采用上一章的融合策略，将基于稀疏表示分类器得到的像素级分类图与得到的分割图进行融合得到最终的分类结果图。通过同时波段选择和多目标分类，在多幅高光谱图像上可实现使用尽可能少的特征维数和在少量训练样本得到较好的分类结果。 (3)提出了一种多目标粒子群分类算法。粒子群分类是一种收敛快的全局优化算法，这里我们在经典的PSO分类的目标函数上增加了另外两个目标函数，类内判据和类间判据，我们把多目标粒子群优化算法用于UCI数据集和高光谱图像上进行了分类，期望能在结合PSO和多目标的优势得到比较好的分类结果。本文的工作得到了国家自然科学基金(61272282)，“教育部新世纪优秀人才支持计划”(NCET-13-0948)和中央高校基础科研业务费(K50511020011)等项目的资助。
[Abstract]:Hyperspectral image is the image of continuous remote sensing imaging of ground objects in dozens or even more than hundreds of continuous spectral channels by imaging spectrometer. Because hyperspectral images have high spectral resolution and rich band information, hyperspectral images are widely used in many fields, such as vegetation, ecology, atmosphere and ocean. But because the hyperspectral image has many bands, the correlation of its adjacent bands is high, which leads to the redundancy of the band and the repetition of information. Therefore, for hyperspectral images, feature selection and classification techniques have become the focus of hyperspectral image research. In this paper, the following three works have been done on the feature selection and classification of hyperspectral images: In the traditional method of hyperspectral image classification based on pixels, only the spectral segment information is used to classify the image, and the spatial correlation of the image is not considered. In this paper, a space-spectrum classification method for hyperspectral images based on mean shift and sparse representation is proposed. The fusion mechanism is used to fuse the hyperspectral image classification map based on sparse representation and the segmentation map with different pixel points by means of mean shift clustering to obtain the final classification results. Through this strategy, adaptive spatial information is incorporated into the classification results, which enhances the regional consistency and greatly improves the classification accuracy. (2) A multi-objective immune clone algorithm is proposed to realize the simultaneous band selection and segmentation of hyperspectral images. In the case of unknown clustering number, the segmentation part adopts multi-objective immune clone clustering algorithm combined with unsupervised feature selection. At the same time, the band selection of hyperspectral image is realized. Using the fusion strategy in the previous chapter, the pixel level classification map based on sparse representation classifier and the segmentation graph are fused to obtain the final classification result graph. By simultaneous band selection and multi-target classification, it is possible to use as few feature dimensions as possible on multiple hyperspectral images and obtain better classification results in a small number of training samples. A multi-objective particle swarm optimization algorithm is proposed. Particle swarm optimization (PSO) is a fast convergent global optimization algorithm. Here we add two other objective functions to the objective function of the classical PSO classification, the intra-class criterion and the inter-class criterion. We apply the multi-objective particle swarm optimization algorithm to the classification of UCI data sets and hyperspectral images in order to obtain a better classification result by combining the advantages of PSO and multi-targets. The work of this paper is supported by the National Natural Science Foundation of China 61272282, NCET-13-0948) and K50511020011) of the National Natural Science Foundation of China (NCET-130948) and the National Natural Science Foundation of China (NCET-130948).
【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】