基于峰值密度聚类的高光谱图像分析方法研究

发布时间：2018-03-27 17:41

本文选题：图像分析　切入点：高光谱图像　出处：《郑州大学》2017年硕士论文

【摘要】：高光谱图像包含丰富的空间信息和光谱信息,具有图谱合一的特点。空间信息中的每个像元,在数百个波段上形成的连续光谱信息特征,是完成图像分析的重要依据。高光谱图像整合物体空间的几何特征与光谱的类别属性对地物探测、目标识别等领域的研究具有重要价值。研究高光谱图像分析方法,不论是筛选高质量波段图像,还是精细探测地物类型,以及地物属性分类,都为高光谱图像信息的深入挖掘和量化分析提供有效途径。高光谱图像的分析方法主要包括:波段选择、端元提取、地物分类。波段选择是高光谱图像的预处理步骤,选择高信噪比波段图像,去除冗余信息是精确处理的重要保证。端元提取确定高光谱图像中纯净的地物类别光谱,是混合像元在亚像元完成目标探测的重要步骤,所提取的端元光谱对分类及识别精度有重要影响。地物分类完成同类别地物的区域划分,为决策规划提供依据。传统高光谱图像分析方法欠缺空间信息和光谱信息间的有机结合,缺乏确定高光谱图像分析时波段数目、端元数量、地物类别的参数化指标。针对以上缺点,本文创新性地提出基于峰值密度聚类算法的高光谱图像分析方法,利用算法得出的参数化指标,依次完成波段选择、端元提取、地物分类的高光谱图像分析。引入超像元分割算法,提升高光谱图像空间信息与光谱信息的利用率,降低算法运算负担,增强全局优化能力,系统性研究高光谱图像分析方法。本文在第一章高光谱遥感研究概述的基础上,围绕高光谱图像分析技术,开展波段选择、端元提取、地物分类三方面论述。第二章讨论经典聚类方法的不足,阐明峰值密度聚类与超像元分割算法的优势。借鉴国内外研究进展,提出创新性的技术路线和研究思路。从本文的第三至第五章开始,各章节以算法原理,实验数据,评价指标,分析讨论的整体框架,分别对应波段选择、端元提取和地物分类的高光谱图像分析方法展开研究。总体来说,本文在高光谱图像分析中取得如下进展:(1)针对高光谱相邻波段图像之间信息相关性强,冗余度大的问题。提出一种基于峰值密度聚类的波段选择算法。该方法在选择波段的数量方面,提出辅助性参数化方案。通过确定的波段数量,利用高光谱图像波段之间的相关性,划分出信息量相关的波段图像子集,提取以上子集的峰值密度聚类特征,构建组合特征的高光谱图像,并采用支持向量机完成地物分类。该方法通过参数指标确定选择波段数量,成功去除高光谱相邻波段图像的冗余信息对分类效果的干扰,克服算法收敛慢、全局优化能力弱的缺陷,提升波段选择效果。(2)本文在综合经典端元提取算法思想的基础上,提出基于峰值密度聚类的端元提取算法。该方法利用线性光谱混合模型,结合混合像元光谱在高维空间上的分布特点,估计端元数目筛选纯像元光谱,解决端元提取算法中端元数目估计问题。(3)在地物分类方面,针对目前大多数算法利用高光谱图像空间信息不足的缺点,提出整合超像元分割和峰值密度的高光谱图像聚类算法。使超像元作为峰值密度算法的最小聚类单位,利用图像形态学特征减少高光谱图像数据量,排除异常像元的干扰。该方法不仅减轻算法运算负担,同时确定地物的类别数量,提升聚类精度。最后,第六章对全文工作进行总结,并指明下一步研究方向。
[Abstract]:Hyperspectral image contains abundant spectral information and spatial information, has the characteristics of spectral image. Each pixel in the spatial information and spectral information of continuous features formed in hundreds of bands, is an important basis for image analysis. The class attribute of geometric features of the hyperspectral image integration object space and spectrum of feature detection study, object recognition has important value. Analysis method of hyperspectral image, whether screening of high quality band image, or fine detection of the type of features, and features of attribute classification for hyperspectral image mining in-depth and quantitative information analysis provides the effective way. The analysis method of hyperspectral image mainly includes: the band selection of endmember extraction, classification. Band selection is a preprocessing step of hyperspectral image, high SNR band image, remove redundant information is accurate processing An important guarantee. Endmember extraction determine objects spectral pure hyperspectral image, is an important step to complete the mixed pixel target detection in subpixel endmember extraction, which have an important impact on the accuracy of classification and recognition. The classification of objects with complete division of categories, provide the basis for decision-making. The traditional hyperspectral the image analysis method is lack of spatial information and spectral information of the organic combination, lack of the hyperspectral image analysis when the band number, the number of endmembers parametric index classes. In view of the above shortcomings, this paper proposes the analysis method of hyperspectral image peak density based clustering algorithm, using parametric index algorithm is derived. In order to complete the band selection, endmember extraction of hyperspectral image classification. The introduction of super pixel segmentation algorithm, hyperspectral image spatial and spectral information promotion The utilization rate, reduce the computation burden, enhance the ability of global optimization and system analysis method of hyperspectral image. In this paper, the first chapter outlines the hyperspectral remote sensing research, on the hyperspectral image analysis technology, carry out band selection, endmember extraction, feature points on the three aspects. The second chapter discusses the shortcomings of the classic clustering method, clarify the peak density clustering and super pixel segmentation algorithm. The advantages of using domestic and foreign research, put forward the technical route and research innovation. From the beginning of the third to fifth chapters, each chapter in the algorithm principle, experimental data, evaluation index, analysis and discussion of the overall framework, corresponding to band selection, expansion study on hyperspectral image endmember extraction and classification analysis method. In general, this paper made the following progress in hyperspectral image analysis: (1) for high spectral image of the adjacent band The information redundancy problem of strong correlation. Proposed a band selection algorithm based on peak density clustering. In the number of band selection methods, put forward the scheme of auxiliary parameters. By determining the number of bands, using the correlation between bands of hyperspectral image, divides the band image information related subset. The peak density clustering feature extraction above subset of the hyperspectral image to construct feature combination, and the support vector machine to complete the classification. This method through the parameters determine the choice of band number, successful removal of interference of redundant information in hyperspectral image classification results of the adjacent band, to overcome the slow convergence of the algorithm, global optimization is weak. Enhance the band selection effect. (2) based on the comprehensive classic endmember extraction algorithm based on the idea, put forward the endmember extraction algorithm based on density clustering. The peak By using the method of linear spectral mixture model, combined with the characteristics of spectral distribution of mixed pixels in high-dimensional space, estimate the number of endmembers were pure pixel spectra, solve the number of endmembers endmember extraction algorithm in the estimation problem. (3) in the classification, for most of the current algorithms insufficient use of hyperspectral image spatial information shortcomings, put forward hyperspectral image clustering algorithm integrating the super pixel segmentation and peak density. The super pixel as the smallest unit of peak density clustering algorithm, using morphological image reduction of hyperspectral image data, excluding abnormal like yuan interference. This method not only reduces the computational burden, and determine the number of object categories, improve the clustering accuracy. Finally, the sixth chapter summarizes the work and points out the next research direction.

【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】