摘要
摘要
高光谱图像分类是高光谱图像处理技术的重要组成部分。通过计算机等设备来对高光谱图像中属于不同地物特有的空谱信息进行分析，并采取有效的分类手段将所有像元划分各自独立、相互不重叠的区域中去。但在分类过程中样本标签的标定需要特定的仪器以及大量的人力、物力，导致可获的能够使用的有标签样本的数目少之又少，为了解决这一问题一系列的研究方法被提出来：半监督学习，直推式学习，主动学习以及通过特定方法生成另外的虚拟样本。但是在传统的方法中，大部分研究方法都是从全局的特征空间或者标准矩形邻域出发，往往忽视了高光谱图像自身拥有的特定的空间信息和光谱信息，缺乏针对性的解决方法。基于以上的分析，本文提出了基于自适应邻域的方法来解决传统算法中的缺陷，尽最大可能的利用高光谱图像自身特有的空间信息和光谱信息来提高分类精度，本文主要研究内容如下：
pines
（1）提出了一种结合自适应邻域和主动学习的高光谱图像分类算法。本模型旨在改变传统主动学习选择样本中的按着在全局特征空间上依照SVM决策面来选择样本导致的被选择样本空间分布不均，且过程缓慢的缺陷。在选择样本进行标记的时候只选择自适应邻域中距离支持向量最远且在图像空间上互相不靠近的样本进行标记，选择出自适应邻域中含有更多信息量的样本加入到训练集中，扩大训练集的规模。在Indian Pines和Pavia University数据图像上验证得出选择这样的样本可以获得更好的分类结果。
（2）提出了一种结合自适应邻域和半监督学习的高光谱图像分类算法。本模型旨在改变传统半监督学习基于全局特征空间和空间矩形邻域的方法利用无标签样本来训练分类器中因错误利用样本的方法导致的分类精度不高，甚至在一定次数以后精度下降的情况。在选择无标签样本进行利用时只在训练样本的自适应邻域中选择出和训练样本在类标上分类结果相同的无标签样本加以利用，并给定训练样本的相同类标。通过分类类标一样且同在自适应邻域同时约束，选择的样本给定正确伪类标的可能性大大提高。并且在Indian Pines和Pavia University数据图像上验证得到这样利用无标签样本可以提高分类精度。
（3）提出了一种结合半监督学习和堆栈自编码的高光谱图像分类模型。本模型旨在改变深度学习中传统线性组合增加虚拟样本的方法在特征空间线性可分的情况会因组合权值全为正数导致生成大量冗余样本，且在线性不可分的的情况下生成大量的错误伪类标样本。本模型在选择线性组合“母体”样本时，只选择训练样本自适应邻域中的无标签样本半监督的进行线性组合，且组合权值在数值选择上有正有负，并给
西安电子科技大学硕士学位论文
本，使用这样的样本来训练堆栈自编码提取分类特征。并且在Indian Pines和Pavia University数据图像上验证得到这样生成的虚拟样本可以提升分类精度。
关键词：高光谱图像分类，半监督学习，主动学习，自适应邻域，堆栈自编码
ABSTRACT
ABSTRACT
Hyperspectral image classification is an important technology in hyperspectral image processing.The space spectrum information which belongs to different objects in hyperspectral images is analyzed by computer and other equipment, and all pixels are divided into independent and non overlapping regions by effective classification, but the number of the available labeled samples is few, because the specific instruments and a lot of manpower and material resources are needed to label the samples. A series of research methods are proposed to solve this problem, such as semi-supervised learning, active learning, and generating the virtual samples by many specific methods. But most of these research methods, in the traditional way, are using the global feature space and the standard rectangle neighborhood. They often neglect the specific spatial information and spectral information of the hyperspectral images, and result in the lack of the specific solutions. Based on the above analysis, in this paper, three hyperspectral image classification methods based on adaptive neighborhood are proposed. The spatial and spectral information of the hyperspectral image are use
d by the maximum possible to improve the classification accuracy of hyperspectral image. The main contents are as follows:
Firstly, a hyperspectral image classification model based on adaptive neighborhood and active learning is proposed. The labeled sample in hyperspectral image are limited. so , selecting training set with the most effective samples can improve the classification accuracy. The purpose of this model is to change the situation in the traditional active learning on how to select the most effective samples, which are often based on the decision surface of SVM in the feature space, and the method often results in uneven spatial distribution of the selected samples and the process of selecting samples is slow. In the model proposed in Section 3, which selecting the points farthest from the support vector in the adaptive neighborhood to marker the true class label. The adaptive neighborhood representates the similarity of characteristic and the support vector external representates contain more information, which are constrained at the same time. Selecting samples from the adaptive neighborhood with more information for the training set which can expanding the scale of the training set. A better classification result can be obtained by using these samples. And verify the model by the Indian Pines and Pavia University data sets.
西安电子科技大学硕士学位论文
Secondly, a hyperspectral image classification model based on adaptive neighborhood and semi-supervised learning is proposed. This model aims to change the traditional semi- supervised learning on utilizing the unlabeled samples from the global feature space or using space rectangular neighborhood. The classification accuracy caused by the error method of utilizing the unlabeled samples is not high, even if the accuracy is decline after a certain number of iterations. The model proposed in Section 4, which selecting the unlabeled samples in the adaptive neighborhood of the training sample with the same classification label as the label of training sample. By the same classification labels and both in the same adaptive neighborhood, the probability of the selected sample belongs to the same label as the training sample is greatly increased. Selecting such sample and adding it to the training set for training can greatly improve the classification results. And verify the model by the Indian Pines and Pavia University data sets.
Thirdly, a hyperspectral image classification model based on semi-supervised learning and Stacked Auto-Encoder is proposed. The purpose of this model is to change the traditional linear combination in the feature space of linear separable with virtual sample generated by all positive weights that most of them are redundant samples. and a large number of virtual samples with false labels are generated under the condition that the feature space is nonlinear separable. In the model proposed i
n Section 5, choosing the samples in the adaptive neighborhood of training sample for combination when select the "Parent" samples. The choice of the combined weights are positive and negative together. Giving the same label as training samples and using the SVM classifier to remove redundant samples. Using such virtual samples generated to train the SAE for extracting feature which can increases the classification performance. And verify the model by the Indian Pines and Pavia University data sets.
Keywords: hyperspectral image classification, semi-supervised learning, active learning, daptive neighborhood, Stacked AutoEncoder
插图索引
插图索引
图2.1中心点确定的自适应邻域 (6)
图2.2主动学习流程示意图 (7)
图2.3自动编码器模型示意图 (7)
图2.4堆栈自编码网络模型 (8)
图3.1相同数量的不同样本选择导致不同分类结果 (11)
图3.2某数据图像在特征空间的真实分布 (12)
图3.3基于主动学习和SVM相结合的算法示意图 (13)
图3.4利用自适应邻域组合空谱信息生成新特征 (14)
图3.5结合自适应邻域和主动学习的高光谱图像分类流程图 (15)
图3.6属于自适应邻域样本点的分类 (16)
图3.7样本点和自适应邻域样本点真实分类示意图 (16)
图3.8属于自适应邻域样本点分类及选择标准示意图 (17)
图3.9 A VIRIS Indian Pines数据图像的自适应邻域结果 (19)
图3.10 Pavia University scene数据图像的自适应邻域结果 (21)
图4.1高光谱图像像元点在特征空间真实分布 (26)
图4.2随机选择样本集以及通过Softmax分类结果 (27)
图4.3高光谱所有像元分类结果 (27)
图4.4均质区域大的情况的样本 (28)
图4.5均质区域小或处于边界的情况的样本 (28)
图4.6结合自适应邻域和半监督学习的高光谱图像分类流程图 (30)
图4.7属于自适应邻域样本点的分类以及挑选情况 (30)
图4.8样本点和自适应邻域样本点真实分类示意图 (31)
图4.9存在噪声的高光谱图像分类结果 (32)
图4.10基于投票滤波示意图 (33)
图4.11迭代得到不同过程的训练集和分类结果 (34)
图4.12 SVM、SOMP、BTC-GF和本章方法对比结果图 (36)
图4.13迭代得到不同过程的训练集和分类结果 (37)
图4.14 SVM、SOMP、BTC-GF和本章方法对比结果图 (39)
图5.1针对特征空间线性可分的线性组合增加样本示意图 (43)
图5.2针对特征空间线性不可分的线性组合增加样本示意图 (44)

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。