基于多特征融合的卫星遥感图像分类研究
发布时间：2018-11-23 13:23
【摘要】：随着遥感事业的蓬勃发展,卫星遥感图像受到人们越来越多的关注。高光谱图像作为卫星遥感图像的一个重要分支,其本身具有的高维数据蕴含了丰富的信息待我们深入挖掘。高光谱图像分类问题是现阶段遥感图像研究领域的一个热门问题,该问题涉及计算机图像学、数理统计学、矩阵论等多个学科理论。在高光谱图像分类领域,目前较为流行的分类方法是利用基于统计学习的机器学习分类算法(监督学习方法,无监督学习方法),通过建立分类模型来预测预测图像中每个像素点的类别。当使用有监督分类算法时,由于高光谱图像每个像素点对应的特征向量(光谱特征)维度较高,并且一般可供使用的训练样本数量稀少,会导致很严重的“Hughes”现象,这种特性导致了高光谱数据对分类算法的选择颇为敏感。根据前人的大量实验可知,支持向量机分类算法(SVM)比较适合高光谱图像的分类问题,此算法能够很好地克服训练集少、特征维度高等问题。在高光谱图像分类时,现有方法一般仅依赖于光谱特征,从而忽略了高光谱图像所蕴含的空间特征(即空间地理信息)。因此,为了提高图像分类精度,在分类过程中如何有效地融合这两种特征所提供的信息成为亟待解决的问题。本文中使用属性特征方法(Attribute Profiles,APs)提取空间特征,该方法利用属性过滤器(Attribute Filters,AFs)扫描高光谱图像的每个通道,计算AF覆盖区域的属性值,与一系列阈值比较后,得到对应像素点的空间特征;将得到的空间特征和光谱特征加权相加从而实现多特征融合,利用融合后的特征建立SVM分类模型。在特征融合之前,需要对数据中夹杂的噪声进行处理。近些年来,基于字典学习的矩阵稀疏表达方法越来越多地被应用于数据特征处理。在图像领域比较有代表性的稀疏表达方法是矩阵低秩分解,并结合字典学习后转换为矩阵低秩表达算法(LRR)。该方法通过矩阵分解还原矩阵的低秩性,使得噪声数据与原来的数据分解开来,从而得到高质量的特征。结合卫星遥感图像的特点,根据相邻元素具有相同类别的假设,对整幅图像分块进行低秩表达,本文提出使用基于区域划分的LRR对高光谱图像的光谱特征和空间特征进行去噪处理,对得到的新特征进行特征融合,最后再利用SVM分类算法建立分类模型;通过实验证明,本文提出的方法对于高光谱图像分类问题可以得到较高的分类精度,另与基于核函数的特征融合方法对比具有明显的优势。
[Abstract]:With the rapid development of remote sensing, people pay more and more attention to satellite remote sensing image. As an important branch of satellite remote sensing images, hyperspectral images contain abundant information for us to mine deeply. The problem of hyperspectral image classification is a hot issue in the field of remote sensing image research at present. This problem involves many disciplines such as computer graphics mathematical statistics matrix theory and so on. In the field of hyperspectral image classification, the most popular classification methods are machine learning (supervised learning, unsupervised learning), which is based on statistical learning. The classification model is established to predict the classification of each pixel in the image. When the supervised classification algorithm is used, because of the high dimension of the feature vector (spectral feature) corresponding to each pixel of hyperspectral image and the small number of training samples generally available for use, the phenomenon of "Hughes" will be very serious. This property leads to hyperspectral data being sensitive to the selection of classification algorithms. According to a large number of previous experiments, support vector machine (SVM) classification algorithm (SVM) is more suitable for hyperspectral image classification. This algorithm can overcome the problems of less training set and higher feature dimension. In the classification of hyperspectral images, the existing methods generally only depend on spectral features, thus ignoring the spatial features (i.e. spatial geographic information) contained in hyperspectral images. Therefore, in order to improve the accuracy of image classification, how to effectively fuse the information provided by these two features in the process of classification has become an urgent problem. In this paper, attribute feature method (Attribute Profiles,APs) is used to extract spatial features. In this method, attribute filter (Attribute Filters,AFs) is used to scan each channel of hyperspectral image, and the attribute value of AF coverage area is calculated, which is compared with a series of thresholds. The spatial features of the corresponding pixels are obtained. The spatial features and spectral features are weighted together to achieve multi-feature fusion, and the SVM classification model is established by using the fused features. It is necessary to deal with the noise in the data before feature fusion. In recent years, the sparse representation of matrix based on dictionary learning has been applied to data feature processing more and more. The sparse representation method in image field is matrix low rank decomposition, which is converted to matrix low rank representation algorithm (LRR). After dictionary learning. By decomposing the low rank of the matrix, the noise data is decomposed from the original data, and the high quality characteristic is obtained. Considering the characteristics of satellite remote sensing images and the assumption that adjacent elements have the same category, the block representation of the whole image is carried out in low rank. In this paper, LRR based on region division is used to de-noise the spectral and spatial features of hyperspectral images, and the new features are fused. Finally, the classification model is established by using the SVM classification algorithm. The experimental results show that the proposed method can achieve high classification accuracy for hyperspectral image classification and has obvious advantages compared with the kernel-based feature fusion method.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751
[Abstract]:With the rapid development of remote sensing, people pay more and more attention to satellite remote sensing image. As an important branch of satellite remote sensing images, hyperspectral images contain abundant information for us to mine deeply. The problem of hyperspectral image classification is a hot issue in the field of remote sensing image research at present. This problem involves many disciplines such as computer graphics mathematical statistics matrix theory and so on. In the field of hyperspectral image classification, the most popular classification methods are machine learning (supervised learning, unsupervised learning), which is based on statistical learning. The classification model is established to predict the classification of each pixel in the image. When the supervised classification algorithm is used, because of the high dimension of the feature vector (spectral feature) corresponding to each pixel of hyperspectral image and the small number of training samples generally available for use, the phenomenon of "Hughes" will be very serious. This property leads to hyperspectral data being sensitive to the selection of classification algorithms. According to a large number of previous experiments, support vector machine (SVM) classification algorithm (SVM) is more suitable for hyperspectral image classification. This algorithm can overcome the problems of less training set and higher feature dimension. In the classification of hyperspectral images, the existing methods generally only depend on spectral features, thus ignoring the spatial features (i.e. spatial geographic information) contained in hyperspectral images. Therefore, in order to improve the accuracy of image classification, how to effectively fuse the information provided by these two features in the process of classification has become an urgent problem. In this paper, attribute feature method (Attribute Profiles,APs) is used to extract spatial features. In this method, attribute filter (Attribute Filters,AFs) is used to scan each channel of hyperspectral image, and the attribute value of AF coverage area is calculated, which is compared with a series of thresholds. The spatial features of the corresponding pixels are obtained. The spatial features and spectral features are weighted together to achieve multi-feature fusion, and the SVM classification model is established by using the fused features. It is necessary to deal with the noise in the data before feature fusion. In recent years, the sparse representation of matrix based on dictionary learning has been applied to data feature processing more and more. The sparse representation method in image field is matrix low rank decomposition, which is converted to matrix low rank representation algorithm (LRR). After dictionary learning. By decomposing the low rank of the matrix, the noise data is decomposed from the original data, and the high quality characteristic is obtained. Considering the characteristics of satellite remote sensing images and the assumption that adjacent elements have the same category, the block representation of the whole image is carried out in low rank. In this paper, LRR based on region division is used to de-noise the spectral and spatial features of hyperspectral images, and the new features are fused. Finally, the classification model is established by using the SVM classification algorithm. The experimental results show that the proposed method can achieve high classification accuracy for hyperspectral image classification and has obvious advantages compared with the kernel-based feature fusion method.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751
【相似文献】
相关期刊论文 前10条
1 张帆;陈俊杰;郭浩;;基于脑功能超网络的多特征融合分类方法[J];计算机工程与应用;年期
2 刘杰;;一种基于多特征融合的自适应目标跟踪策略[J];大众科技;2016年11期
3 刘红;王晔;雷长海;;基于多特征融合的中医舌像检索研究[J];计算机应用研究;2010年02期
4 王益尧;;基于多特征融合问句理解问答系统研究[J];数据通信;2018年02期
5 李志欣;李艳红;张灿龙;;一种多特征融合的场景分类方法[J];小型微型计算机系统;2018年05期
6 霍富功;王诗琴;;基于多特征融合的转移粒子滤波红外小目标跟踪算法研究[J];信息与电脑(理论版);2018年12期
7 蒋晓玲;章慧;;基于多特征融合的激光医学图像拼接[J];激光杂志;2015年12期
8 孙汝星;范荣双;;基于支持向量机的多特征融合影像云检测[J];测绘与空间地理信息;2018年06期
9 乔瑞萍;孙贺;董员臣;王方;;基于多特征融合的井盖检测系统实现[J];电子技术应用;2018年06期
10 顾鑫;李U，
本文编号：2351757
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2351757.html
