高光谱图像地物分类中的特征提取方法研究

CASIA OpenIR > 毕业生 > 博士学位论文

	高光谱图像地物分类中的特征提取方法研究
	钟子沙
	2017-01
学位类型	工学博士
中文摘要	蓬勃发展的高光谱遥感对地观测技术为地物探测提供精细化的影像数据，这些数据不但包含了丰富的地物光谱信息，还包含了分辨率越来越高的空间结构信息。作为高光谱遥感的最重要应用之一，地物分类旨在为影像中的每个像素分配地物类别标签。当前，随着模式识别、计算机视觉、机器学习等学科领域的发展，高光谱遥感图像的地物分类研究取得了快速进展。然而，高光谱数据本身的高维特性、波段之间的高度相关性和混合光谱等问题，再加上数据容量的爆炸性增长，给地物分类带来了新的难题和挑战。因此，有必要提出一些有效的方法和可行的解决方案，解决高光谱图像地物分类中的这些难题。本文基于模式识别与机器学习领域内的最新理论与方法，以高光谱图像地物分类中的特征提取方法作为研究切入点，结合张量表示、哈希学习等新方法，开展高光谱图像地物分类中的特征提取方法研究。本文的主要研究贡献如下： 1. 提出了一种基于局部张量判别分析的光谱—空间特征提取方法。该方法首先将光谱—空间特征表示为二维特征张量，然后使用了一种局部张量判别分析的方法,来去除张量特征表示中的冗余信息，这样能得到更具鉴别性的低维特征，最后将提取到的低维特征输入支持向量机进行分类。在局部张量判别分析中，最优的降低维数通过优化方法自动获得，这样就减少了一部分参数选择的计算代价。论文将该方法与三种最常用的光谱—空间特征提取方法进行结合，详细的比较实验验证了该方法的有效性。 2. 提出了一个基于多特征融合的哈希二值特征提取框架。论文第一次将哈希技术引入到高光谱图像分类中来，提出一个基于哈希方法的多特征融合框架，并将它作为一种二值特征提取方法对融合的多特征进行降维与压缩，然后基于得到的二值编码特征进行高效的地物分类。由于二值编码特征的优点，使得特征在存储空间上降低了几百倍，同时在地物分类的时间上降低了至少两个数量级。详细的对比实验验证了该方法的有效性。此外，论文还细致地分析了基于哈希方法的二值特征提取在高光谱图像分类任务上的优缺点及应用潜力。 3. 提出了一种基于结构化学习的二值特征提取方法。该方法的目标主要是针对多特征组合而成的高维特征，生成紧致而具有鉴别性的二值特征。论文首先将哈希二值编码问题建模为一个约束的线性分类问题使得二值特征能最大化分类精度，然后考虑在哈希变换矩阵上应用结构化的稀疏约束来更好地处理特征冗余问题，同时应用结构化的组稀疏约束建模不同模态特征之间的结构化信息。实验结果表明该方法不仅能生成非常紧致的二值特征，还能极大地提高这些二值特征的鉴别能力，从而获得较高分类效果。
英文摘要	Nowadays, hyperspectral imaging technology on Earth observation can provide us a large amount of finer image data for land-cover analysis. These data can not only contain abundant spectral information, but also structural and spatial information with high resolution. As a vital application in hyperspectral remote sensing, land-cover classification aims at classifying each pixel into one certain known land-cover class label. Currently, with the development of pattern recognition, computer vision and machine learning, land-cover classification in hyperspectral images has achieved a lot of advances in remote sensing society. Feature Extraction is a very important issue in hyperspectral imagery classification. On one hand, hyperspectral image data provides intrinsic finer spectral information and spatial structures. How to effectively and efficiently conduct feature extraction is key to improve the performance for hyperspectral imagery classification. On the other hand, with the advent of the "big data" era, how to handle a huge amount of remote sensing data is very challenging. Thus, it is very urgent to develop some novel technologies and techniques to effectively and efficiently handle these problems. With the help of latest theories and technologies in pattern recognition and machine learning, this thesis aims to develop feature extraction methods in hyperspectral imagery classification. Specifically, we actively adopt several novel methods, such as tensor representation and feature hashing technique, to effectively and efficiently extract discriminative yet descriptive features for hyperspectral imagery classification. The main contributions and novelties can be summarized as follows. 1. We propose a spectral-spatial feature extraction method based on local tensor discriminant analysis for hyperspectral imagery classification. In this method, the spectral-spatial features are first represented into 2-D tensor representation, and then a local tensor discriminant analysis method is further adopted in order to extract discriminative features within these feature tensors, finally the reduced features are given as inputs to support vector machine for classification. With the help of local tensor discriminant analysis, this method can automatically find the optimal dimensions of the reduced feature representation, which can help alleviate the computational cost during classification. In the experiments, three commonly used spectral-spatial feature extraction methods are adopted to construct the feature tensors, detail experimental results and analysis are presented. The comparative results validate the proposed method. By using the proposed 2-D spectral-spatial tensor representation, the tensor counterparts of traditional methods can help improve the performance of classification. 2. An efficient multiple feature fusion framework with hashing technique is proposed. In this framework, the concatenated long vectors based on multiple features are reduced to compact binary codes using hashing technique. Based on four benchmark hyperspectral data sets, we systematically evaluate the classification performance of the proposed framework with different hashing methods and also compare it to five subspace-based dimension reduction methods and other six different strategies for multiple feature fusion. Experimental results are quite encouraging, demonstrating that hashing methods could achieve the comparable or better performance. Based on the extensive experimental evaluations, we further discuss advantages and disadvantages of using hashing for fusing multiple features. 3. A learning-to-hash based structured binary feature extraction method based on multiple features is proposed for hyperspectral image classification. The goal is to generate compact and discriminative binary codes. First, a linear classification model is adopted to maximize the performance of binary features. Then two structured regularization terms are further integrated to better handle the problem of feature redundancy and to simultaneously catch the structural information among different modalities of features. Detail comparative experiments demonstrate that the proposed method can generate very compact yet discriminative binary features for classification.
关键词	高光谱遥感图像特征提取地物分类张量表示哈希学习二值编码
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/13011
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所模式识别国家重点实验室
推荐引用方式 GB/T 7714	钟子沙. 高光谱图像地物分类中的特征提取方法研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
Thesis(最终版).pdf（14829KB）	学位论文		限制开放	CC BY-NC-SA