【作者】 姚伏天；

【导师】 钱沄涛；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2011， 博士

【摘要】 为利用高光谱成像光谱仪,高光谱图像可通过数百个连续且细分的光谱波段对地表区域同时成像,进而获得三维图像数据。高光谱图像信息通过较窄的波段区间、较多的波段数量而提供。高光谱图像能被用于从光谱空间中对地物予以细分和识别,这是其能够在军事和民用领域得到广泛应用的重要原因。当前,在高光谱图像分类领域,核函数方法因具有可解决非线性问题和数据特征维数过多问题的能力,而受到科研人员越来越多的关注,逐步成为研究热点。最近几年,支持向量机作为一种核函数方法广泛被应用于高光谱图像分类。但是,支持向量机本身存在诸如核函数中超参数难以选择、输出结果不具有概率意义等问题,限制了其进一步的推广。高斯过程也是一种基于核函数的方法。它具有完全的贝叶斯公式化表示,能够明确的进行概率建模,使结果更易于解释。高斯过程的贝叶斯学习提供了一个范式,可以根据训练样本,实现从先验分布到后验分布的转换,以及对核函数超参数的推理。本文以基于高斯过程的高光谱图像分类技术作为主要研究内容,针对高光谱图像波段数多,波段间相关性和空间相关性强,带标记训练样本过少的特点,将高斯过程理论与亲和传播聚类算法、综合波段特征与空间特征构造空间约束核函数方法、条件随机场理论和半监督核函数理论相结合,对高光谱图像分类进行研究。主要的创新性研究成果如下：1.高光谱图像有监督分类时,若训练样本数目有限,会出现"Hughes"现象,即分类精度先随着图像波段数目的增加而增加,当到达一定极值后,分类精度又随着波段数目的增加而下降。为避免"Hughes"现象,高光谱图像分类前应先进行波段选择。本文在高光谱图像波段选择理论基础上,提出了基于波段选择的高斯过程高光谱图像分类方法,该方法先用亲和传播(Affinity Propagation)方法进行波段选择,再用高斯过程进行高光谱图像分类。实验结果表明,基于波段选择的高斯过程高光谱图像分类方法能够在较少的波段数下取得较好的分类结果。2.高光谱图像不仅拥有较高的谱段分辨率,而且像素点之间有很强的空间相邻关系。本文通过综合波段相关性和空间相关性构造空间约束核函数,提出了空间约束核函数高斯过程分类方法。实验结果表明,空间约束核函数高斯过程分类方法部分消除了同物异谱和同谱异物造成的分类错误。3.为进一步利用高光谱图像的空间结构,本文将高斯过程和条件随机场理论相结合,提出了用于高光谱图像的高斯过程与条件随机场集成分类方法。实验结果表明,高斯过程与条件随机场集成分类方法能够有效减少高光谱图像中的孤立噪声点,进而提高高光谱图像分类精度。4.高光谱图像分类中,经常遇到带标记训练样本不足的问题。半监督学习理论可有效利用少数带标记样本和大量易于获得的无标记样本,从而可以改善分类和预测精度。本文根据高光谱图像像素点之间空间局部相关性较强的特性,基于半监督流形假设,通过构造半监督核函数,创新地提出一种用于高光谱图像分类的半监督核函数高斯过程方法。一方面,该方法为非线性方法,对高维非线性的高光谱图像可以取得较好的分类效果；另一方面,该方法为非参数方法,仅需要对少数几个超参数进行学习,速度较快,也比较简单。实验结果表明,在只有少量带标记的训练样本情况下,高光谱图像分类精度有明显的提高。本文提出的基于高斯过程高光谱图像分类的多个方法,能够较好地针对高光谱图像波段数众多、光谱波段相关性和空间相关性强、以及带标记训练样本过少的特点,从而在标准高斯过程高光谱分类基础上进行改进,有较高的分类精度和较强的适应性。在本文最后,对全文做了总结并展望了未来的研究方向。更多还原

【Abstract】 Hyperspectral imagery (HSI) is a three-dimensional imagery generated by imaging spectrometer simultaneously to the same surface scenery at hundreds of bands. It contains hundreds of spectral information in a narrow spectral region. One of the main applications of HSI is to identify and recognize the materials by the rich spectral information, which is also the basic reason for widely used of HSI to military and civilian fields.Currently, kernel methods are more and more popular in HSI classification for their ability to solve nonlinear problems and less sensitive to the curse of dimensionality with respect to traditional classification techniques. As a kind of kernel methods, support vector machine(SVM) classifiers are widely used for HSI classification in recent years. However, there are some drawbacks in SVM such as difficulty of hyperparameters selection and non probabilistic outputs,which prevent their further population.Another potentially interesting kernel-based classification approach is represented by Gaussian process classifier(GPC). By contrast to SVM classifiers, GPCs are Bayesian clas-sifiers and they permit a fully Bayesian treatment of considered classification problem. GPCs have the advantage of providing output probabilities rather than discriminant func-tion values. Moreover, they can use evidence for automatic model selection and Hyperpa-rameter optimization.In this paper,research is focused on Gaussian Processes(GP) based HSI classification. Aiming at HSI features such as numerous bands, highly correlations of spectral and spatial, lack of labeled samples, We combine Gaussian Processes with Affinity Propagation(AP), kernel construction methods, conditional random fields and semi supervised learning re-spectively to propose some new Gaussian Processes based methods.The major works and contribution of this dissertation are as follows:1. The dimensionality of HSI strongly affects the performance of many supervised clas- sification methods,which is called "Hughes" phenomenon. In order to avoid it, band selection should be processed before classification of HSI. Combined with Affin-ity propagation, Band selection based Gaussian processes method is proposed in this dissertation, which means band selection by AP is followed by classification by GPC. Experimental results show that the proposed band selection based Gaussian processes method can get better classification result in a few spectral bands.2. HSI shows strong spectral and spatial correlations. By constructing a new spatial kernel function (SGK) of GP, spatial relations in HSI are included, so that classifica-tion error partially caused by "same material different spectral" and "same spectral different material" can be partially eliminated.3. In order to utilize spatial structure of HSI, we make a combination of GPC with conditional random fields(CRF) and propose GPCRF method for HSI classification. Experiments on the real world Hyperspectral images attest to the accuracy and robust of GPCRF method, because it can reduce image noise to some extent.4. In HSI classification, supervised learning methods for classification often lead to low performance because of the hard of obtaining the labeled training samples. Mean-while, there are a lot of unlabeled data in Hyperspectral images. In semi supervised learning theory, labeled samples and abundant unlabeled samples are combined to train classifiers by estimating parameters of a generative model. A new classifica-tion method of Spatial Semi-supervised gaussian processes(SSGP) is proposed in this dissertation which is based on the assumption of semi supervised manifold as-sumption. SSGP is a semi-supervised learning method, and spatial correlations of labeled samples and unlabeled samples can be build to raise the classification cor-rect rate; SSGP is a kernel method and it can deal good with the nonlinear property of HSI; SSGP is a non-parameter method and has few Hyperparameters which can be learned from the data. Experiment results show that SSGP method is very good at classification of Hyperspectral images with respect to classification accuracy and stability at the case of small percentage of labeled training samples. In this dissertation, we take the advantages of HSI features such as abundant spectral bands, highly correlations of spectral and spatial and lack of labeled samples, improve stan-dard GPC and propose several GP based HSI classification methods. The results achieved show that these methods have the potential of yielding accurate and stability.Finally, we make a conclusion and give a research Perspective.更多还原