【作者】 杨国鹏；

【导师】 余旭初；

【作者基本信息】 解放军信息工程大学 ， 摄影测量与遥感， 2010， 博士

【摘要】 高光谱影像分类是高光谱遥感应用的关键技术之一,对于资源调查、环境监测、精细农业、军事测绘与战场环境探测等应用具有重要意义。为了满足高光谱影像分类对精度、速度和可靠性的要求,针对高光谱影像具有的高维相关特征、非线性可分的特点,本文围绕影像分类及特征提取技术,采用机器学习方法展开了深入研究。论文的主要研究内容和创新点如下:1.针对基于乘法迭代规则的高光谱影像非负矩阵分解(NMF)收敛速度慢的问题,给出了一种基于改进投影梯度的快速分解算法。采用分块坐标下降法将NMF的整体优化转化为两个子优化问题迭代求解,而每个子优化问题都通过改进的投影梯度法计算。实验分析表明,该方法能够明显提高高光谱影像NMF的收敛速度。2.针对高光谱影像广义判别分析(GDA)特征提取中核函数参数选择问题,给出了一种高斯径向基核函数参数选择方法。将训练样本数值范围规范化,按对数形式将参数空间离散化,通过交叉验证获取合适参数。实验分析显示,使用该方法选择的参数,GDA特征提取能够显著提高高光谱影像分类的精度。3.提出了一种基于简约集(RS)支持向量机(SVM)的高光谱影像分类方法。采用序列最小优化算法、交叉验证网格搜索参数选择方法构造高精度的多类SVM分类器,通过微分进化算法求解原像问题来获取简约集向量。经验证,SVM具有良好的泛化能力,RS-SVM能够保持分类精度,有效提高分类速度。4.提出了一种基于相关向量机(RVM)的高光谱影像模糊分类方法。采用序列稀疏贝叶斯学习算法提高RVM训练速度,针对一对一法构造的多类RVM分类器,将两两配对的概率输出转化为相对于地物类别的隶属度。与SVM的比较表明,RVM参数选择简单、分类速度快;利用模糊隶属度能够标识混合像元,有效提高影像分类的可靠性。5.提出了一种基于自适应提升(AdaBoost)高光谱影像集成分类方法。选择决策树树桩作为弱分类器,通过Gentle AdaBoost算法提升为分段线性的强分类器,采用一对余法解决多类分类问题。结果表明,AdaBoost分类方法的训练和分类速度快,分类精度优于一般分类方法。6.通过理论和实验分析了SVM、RVM和AdaBoost在分类精度、训练和分类速度方面的表现,指出了各自在不同高光谱影像分类应用中的潜力。SVM适用于实时性要求不高的地物精细分类,RVM适用于获取具有统计意义的地物分类,而AdaBoost适用于分类精度要求较高的快速分类。更多还原

【Abstract】 Hyperspectral imagery classification is one of the key technologies in application of hyperspectral remote sensing, and is of great significance for resource investigation, environment monitoring, fine agriculture, surveying and mapping, and battlefield detection. To fulfill the requirement of hyperspectral imagery classification for accuracy, speed and reliability, considering the characteristic of correlation in high dimension and non-linear separability, imagery classification and dimensionality reduction were studied in depth using machine learning methods. The main works and creations of this dissertation are listed as follows:1. To solve the problem of slow speed of hyperspectral imagery non-negative matrix factorization (NMF) based on multiplicative update rule, a fast factorization algorithm based on improved projection grads was given. Using the block coordinate descent method, the global optimization of NMF was transformed into two sub-optimization problems solved by iteration, and each sub-optimization problem computed by improved projection grads. Through experiments it can be seen that this method increases the convergence speed of NMF evidently.2. A radial basis function (RBF) kernel parameters selection method was given to solve the problem of kernel parameter selection in hyperspectral imagery generalized discriminant analysis (GDA) feature extraction. Firstly, the number range of training samples was standardized and parameter space was dispersed logarithmically, then parameter was obtained by cross-validation. The analysis of experiments shows that, using this approach to select kernel parameter, GDA feature extraction can improve the accuracy of hyperspectral classification evidently.3. A reduced set (RS) based on support vector machine (SVM) hyperspectral imagery classification method was brought forward. Using the sequence minimum optimized algorithm and cross validation grid searching parameter selection method, a high precision multi-class SVM classifier is constructed. Reduced set vectors are obtained by solving the pre-image problem through differential evolution algorithm. It is validated that SVM has good generalization ability and RS-SVM can keep classification precision and increase the speed of classification.4. A relevance vector machine (RVM) based on hyperspectral imagery fuzzy classification method was brought forward. In this algorithm, sequence sparse Bayesian learning algorithm was used to improve the training speed of RVM. For the multi-class RVM classifier constructed by one against one decomposition method, we transform the probability of pairwise coupling classifier into membership of the ground objects classes. Compared with SVM through experiments, RVM parameter selection is simpler, and the training and classifying speed is faster. Using fuzzy membership can label mixed pixels and improve the reliability of imagery classification effectively.5. An AdaBoost based hyperspectral imagery ensemble classification method was put forward. Firstly decision stump was selected as weak classifier, and then this classifier can be boosted into subsection linear strong classifier by gentle AdaBoost algorithm, finally multi-class classification was designed by one against rest decomposition method. The speed of training and classifying was proved fast, and the classification accuracy of this method is better than many other methods.6. Both theoretically and experimentally, classification accuracy, training and test speed of SVM, RVM and AdaBoost were analyzed, and their different potentials in different hyperspectral imagery classification applications are pointed out. SVM is suitable for fine classification of ground objects which does not require real-time processing. RVM is applicable to classification of ground objects with statistical prediction. AdaBoost can be applied in fast classification which requires high precision.更多还原