【作者】 刘瑶；

【导师】 谢红；

【作者基本信息】 哈尔滨工程大学 ， 信息与通信工程， 2017， 博士

【摘要】 高光谱图像以极高的光谱分辨率而著称,应用领域广泛。然而,它所具有的波段数量多、数据冗余严重的特点,给其在分类和识别方面带来一定的困难。因此,有必要研究一种既高效又快速地处理高光谱数据的方法。对高光谱数据进行波段选择就是从所有波段中选择起主要作用的波段子集,既能比较完整地保留有用的信息,又能使得高光谱图像数据的维数大大降低。粗糙集理论是一种研究不精确和不确定信息的处理的方法,它能够在不损失有效信息的情况下获得数据核心知识,对数据进行约简,实现数据的分类和预测。本文把粗糙集理论中属性约简的思想应用到研究高光谱图像数据的波段选择方法中,主要研究内容如下:(1)提出基于可变精度邻域粗糙集的高光谱波段选择算法。将邻域粗糙集理论中属性约简的方法融合到波段选择中,为了增强算法对噪声数据的适应能力,引入可变精度因子β,允许上、下近似存在一定的分类误差。邻域半径δ和可变精度因子β的选取是算法的重要影响因素,不同的δ和β将会得到不同的属性重要度,最终影响属性约简子集。本文研究了δ和β对波段子集中波段数量和分类准确率的影响,通过实验确定它们的取值范围,进行最优波段子集的搜索。(2)提出基于邻域一致性准则的高光谱波段选择算法。针对属性依赖度度量中没有考虑决策边界中的样本这一问题,将一致性的思想引入到邻域粗糙集中,根据邻域决策误差最小化准则选择波段子集,达到去除冗余波段的目的,可以在邻域的全部取值范围[0,1]区间内选择出大小合适的波段子集,对邻域的变化有较好的适应性和不敏感性。针对高光谱波段选择算法中存在的过拟合问题,提出前向修剪和后向修剪两种策略,通过实验对比两种策略产生的最终波段子集中所包含的波段数量和分类效果,结果表明Filter与Wrapper相结合的后向修剪策略更有效,可以得到满意的分类效果。(3)提出基于邻域互信息的高光谱波段选择算法。从信息观的角度出发,将信息熵引入到属性约简,把在波段子集中添加某个波段所引起的邻域互信息的变化的大小作为该波段重要度的度量。邻域粗糙集中基于信息观的属性约简算法是以属性依赖度为度量的代数观属性约简算法的互补,能选择出基于代数观的属性约简算法无法选择出的属性。(4)提出基于最大相关最小冗余准则的高光谱波段选择算法。在基于邻域互信息的高光谱波段选择算法基础上,将波段与类别之间的相关性和波段与波段之间冗余性相结合进行研究,分别定义了最大相关最小冗余差(MRMRD)和最大相关最小冗余商(MRMRQ)两种结合算子。通过该算法可以取得最优的波段子集,利用该波段子集进行分类能达到令人满意的分类效果,并且在稳定性上,MRMRQ算法和MRMRD算法比其他算法也有较大的提高。(5)给出除分类性能外的另一种衡量波段选择算法优劣的重要指标——稳定性。对于同一个高光谱数据集,很多波段子集有相同或者近似的分类性能,本研究选择Jaccard系数为稳定性度量指标,研究样本子集之间重叠率、样本扰动变化、样本子集大小和邻域大小对算法的稳定性的影响。提出了兼顾分类性能、稳定性和波段子集大小的综合评价函数,可根据实际情况灵活设置参数来调整分类性能、稳定性和子集大小的重要性。同时,还分析了各算法对噪声的敏感性和对训练集大小的敏感性,讨论了各算法的适用性。更多还原

【Abstract】 Hyperspectral imagery(HSI)is known as its high spectral resolution,which can provide us with rich spectral information and has extensive application value.HSI has a number of bands,a huge amount of data and redundant information.These characteristics bring it certain difficulty in classification and identification.Therefore,it is necessary to explore a method for processing and analyzing data fast and utilizing information effectively.Band selection of hyperspectral data means to select a subset which contains important bands.The subset can retain useful information entirely,and can reduce the dimension greatly.Rough sets theory is a new mathematic approach to process the uncertain and vague data.It can reduce the data and obtain the core knowledge of data without loss of effective information.From the perspective of attribute reduction in rough sets theory,this study explores the band selection methods.The main contents are as follows:(1)The hyperspectral band selection method based on variable precision neighborhood rough set is put forward.By fusing the methods of attribute reduction in neighborhood rough set to band selection,this method enhances the ability to adapt to the noise data by introducing a variable precision factor β,that is allowing the approximation error of the upper and lower approximation.The selection of neighborhood radius δ and a variable precision factor β is the important factor of algorithm.The different δ and β may produce different attribute importance,and affect the attribute reduction subsets ultimately.This paper studies the effect of δ and β on the size of subset and classification accuracy,and determines the scope of δ and β through the experiment,and searches the optimal band subsets.(2)The hyperspectral band selection method based on consistency-measure of neighborhood rough set is proposed.Dependence measure does not take into account the samples in decision boundary.Concerning this issue,the consistency is introduced to neighborhood rough set.The algorithm selects band subsets based on neighborhood decision error minimization criterion for removing redundant band.And it can select appropriate size of subset in the full range of neighborhood,and has good adaptability and insensitivity to the change of neighborhood.For over-fitting problem of hyperspectral band selection algorithm,pre-pruning strategy and post-pruning strategy are proposed.The sizes of subset and classification performance of the two strategies are compared through the experiments.The result shows that the post-pruning strategy which combines the Filter method with the Wrapper method is more effective,and can obtain satisfactory classification performance.(3)The hyperspectral band selection method based on neighborhood mutual information is proposed.From the viewpoint of information view,entropy is introduced into the attribute reduction.The proposed algorithm regards the change value of neighborhood mutual information when add a band to the subset as the importance measure of the band.The attribute reduction algorithm based on information view is the complementation of that based on algebraic view,and can select the attributes which the algorithm based on algebraic view is unable to select.(4)The maximum relevance minimum redundancy(MRMR)band selection based on neighborhood rough set is proposed.Based on the neighborhood mutual information,redundancy among bands is considered fully,and the relevance between bands and categories and the redundancy among bands are combined.Two measures MRMR difference and MRMR quotient are defined.By the algorithm the optimal band subsets can be obtained,and satisfactory classification performance can achieved.In term of the stability,MRMRQ algorithm and MRMRD algorithm are better than other algorithms.(5)In addition to the classification performance,the paper provides another important indicator for measuring the performance of band selection algorithm-the stability.For the same hyperspectral datasets,several band subsets with the same or similar classification performance can be got.Selecting Jaccard coefficient as stability measure,the paper researches how overlap rate of sample subset,disturbance changes of sample subset,the size of sample subsets,and the size of neighborhood effect on the stability of the algorithm.Finally,the paper puts forward the comprehensive evaluation function which can take into account classification performance,stability,and the size of band subsets.According to actual condition,the parameters can be set flexibly to adjust the importance of classification performance,stability,and the subset size.The sensitivity of algorithms to noise and training samples is analysed,and the applicability of the algorithms is discussed.更多还原