高光谱遥感影像三维空谱特征提取与小样本分类技术研究

发布时间：2018-10-24 12:18

【摘要】：高光谱遥感影像包含了关于地表物质的丰富空间、辐射和光谱三重信息,在遥感信息处理中的地位日益凸显,已被广泛用于城市地物分类、环境监测、军事目标识别等诸多领域。然而,传统的遥感信息处理技术在解译高光谱遥感影像时面临着诸多问题和挑战:如何有效利用高光谱影像提供的多重信息、如何解决小样本分类的性能瓶颈等问题。本文从高光谱遥感影像的自身特点出发,在总结现有遥感信息处理技术的基础上,研究如何从高光谱遥感影像中获取判别力强的三维空谱特征,并进一步将多种空谱一体化特征进行融合,实现地物的有效分类。现有的空谱特征提取方法大多是将不同波段的空间特征机械地组合起来,没有充分利用高光谱遥感影像空谱结构中的上下文信息。针对这一问题,本文提出了一种基于三维局部二值模式(Local Binary Patterns,LBP)的高光谱遥感影像空谱特征提取方法。该方法利用模式统计的方式对获得的三维LBP编码特征进行归约,建立面向三维拓扑结构的稠密编码模式来代替传统的加权编码,使三维LBP编码模型具有旋转不变性,提升特征的鉴别力。此外,通过引入松弛变量实现三维LBP编码模型中阈值操作的模糊处理,消除噪声以及同类地物由于反射值分布不匀匀带来的影响,提升了三维LBP编码特征的鲁棒性。实验结果表明,基于三维LBP编码的高光谱遥感影像空谱特征提取方法能够有效利用高光谱影像空谱结构中的上下文语义,显著提升小样本分类的性能。另一方面,不同的三维空谱特征提取方法得到的特征具有各自不同的适用特性。具体来说,三维Gabor特征对影像中的光照和阴影有较好的鲁棒性,三维形态学特征能够有效地捕捉影像中地物的形状特征,而三维局部二值模式对影像中的空间特征具有较好的表征能力。由于这三种特征分别从光照鲁棒性、地物形状、空间关系三个层面表征高光谱影像中关于地物的三种不同属性,因此它们对高光谱影像中的地物区分能力存在差异和互补性。基于此,本文提出了一种基于稀疏决策融合的高光谱遥感影像三维空谱特征分类方法。该方法通过稀疏表示对上述三类空谱特征进行决策融合,整合它们之间的优点,实现优势互补。实验表明,基于稀疏决策融合的高光谱遥感影像三维空谱特征分类方法能够充分发挥三种空谱特征的优势,进一步改善了高光谱遥感影像的小样本分类性能。
[Abstract]:Hyperspectral remote sensing images contain rich spatial, radiative and spectral information about surface matter. They play an increasingly important role in remote sensing information processing, and have been widely used in the classification of urban features and environmental monitoring. Military target recognition and many other fields. However, the traditional remote sensing information processing technology is faced with many problems and challenges in interpreting hyperspectral remote sensing images: how to effectively utilize the multiple information provided by hyperspectral images and how to solve the performance bottleneck of small sample classification. Based on the characteristics of hyperspectral remote sensing images and the existing remote sensing information processing techniques, this paper studies how to obtain the discriminant 3D spatial spectrum features from hyperspectral remote sensing images. Furthermore, various spatial spectral features are fused to realize the effective classification of ground objects. Most of the existing space-spectrum feature extraction methods combine spatial features of different bands mechanically and do not make full use of the context information in the spatial spectral structure of hyperspectral remote sensing images. In order to solve this problem, a novel spatial spectral feature extraction method for hyperspectral remote sensing images based on 3D local binary mode (Local Binary Patterns,LBP) is proposed. In this method, the obtained 3D LBP coding features are reduced by pattern statistics, and a dense coding pattern oriented to 3D topology is established to replace the traditional weighted coding, which makes the 3D LBP coding model rotation-invariant. Enhance the ability to identify features. In addition, the fuzzy processing of threshold operation in 3D LBP coding model is realized by introducing relaxation variables, which eliminates the influence of noise and the unevenness of reflection distribution of similar ground objects, and improves the robustness of 3D LBP coding features. The experimental results show that the spatial spectral feature extraction method based on 3D LBP coding can effectively utilize the context semantics of hyperspectral image space-spectrum structure and improve the performance of small sample classification. On the other hand, the features obtained by different three-dimensional space-spectrum feature extraction methods have different applicable characteristics. Specifically, 3D Gabor features are robust to illumination and shadow in images, and 3D morphological features can effectively capture the shape features of ground objects in images. The 3D local binary model has a better ability to represent the spatial features in the image. Because these three features represent the three different attributes of the hyperspectral images from the three aspects of illumination robustness, ground object shape and spatial relationship, there are differences and complementarities in their ability to distinguish ground objects in hyperspectral images. Based on this, a spatial feature classification method for hyperspectral remote sensing images based on sparse decision fusion is proposed. This method combines the advantages of the above three spatial spectral features by sparse representation and realizes the complementary advantages. The experimental results show that the 3D space-spectrum feature classification method based on sparse decision fusion can give full play to the advantages of the three space-spectrum features and further improve the classification performance of small samples of hyperspectral remote sensing images.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】