联合丰度信息与空谱特征的高光谱影像分类研究

发布时间：2018-06-27 21:10

本文选题：高光谱遥感 + 空谱特征　；参考：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年博士论文

【摘要】：高光谱遥感图谱合一的特性将由物质成分决定的光谱信息与反映地物纹理、形状的空间信息有机结合,从而实现对地物的精准分类、识别以及属性分析等,极大地提高了人类对客观世界的认知能力。高光谱遥感技术已被广泛应用于地质制图、环境监测、植被调查、农业遥感、海洋遥感、大气研究等诸多领域。然而,高光谱遥感数据在带来大量有效信息的同时,也给高光谱影像分类问题带来了诸多挑战。高光谱影像分类主要存在以下问题:第一,由于地物的组成及分布情况的复杂性和成像条件的影响,高光谱影像上依然存在“同物异谱”与“异物同谱”的现象,导致误分的情况很容易出现;第二,高光谱影像上混合像元的普遍存在,进一步增大了地物分类的难度;第三,随着图像的波段增加,对监督分类中训练样本的需求也急剧增加,造成小样本分类时的病态问题。近年来发展起来的空谱特征分类、合成核分类、半监督分类算法以及光谱解混技术,对以上高光谱图像分类问题提供了可能的解决途径。本文通过开展空谱特征和丰度信息提取方法,从监督分类与半监督分类两个角度研究联合丰度信息和空谱特征的高光谱图像分类方法。并采用两组具有不同光谱、空间分辨率和不同地面复杂度的航空高光谱数据进行分类实验,验证本文所提算法的有效性。论文的主要研究成果和结论如下:1)研究了基于形态学的空谱特征提取方法,梳理了形态学属性轮廓特征的提取流程。通过对形态学属性轮廓特征的提取,既充分利用了地物光谱信息,又考虑了影像上地物的空间结构特征,进而降低“同物异谱”与“异物同谱”对分类的影响,大大提高了高光谱图像分类的精度。2)提出了一种基于类别的端元提取与稀疏解混算法。针对光谱解混的端元和分类的类别往往不一致,导致由解混得到的端元丰度信息无法有效地应用到分类中的问题,该算法首先针对训练样本的每一种类别进行端元提取,得到分别对应于每种类别的端元,所有类别提取的端元一起组成端元集,利用该端元集进行稀疏解混,进而得到与类别相对应的丰度信息。实验结果表明,该算法为地物分类中普遍存在的混合像元问题提供了一种有效的解决方法,能够有效提高图像分类的精度。3)研究了一种基于马尔科夫随机场的空间优化算法。考虑到自然界中地物的分布通常具有一定的空间连续性,反映在遥感影像中,即相邻像元是同种地物类型的概率最大。该算法采用一种马尔科夫随机场先验模型来促使邻近像元属于同一类别,通过对空间邻域信息的利用达到空间优化的目的。实验结果表明,该空间优化算法能够有效的去除椒盐噪声,提高分类精度与效果。4)提出了一种结合丰度信息与空谱特征的合成核分类方法(CKSMLRSSA)。该方法通过在核结构层面上将空谱特征和丰度信息特征进行融合,从而更加有效地用于图像分类。根据合成核理论,首先将丰度信息组成的核与空谱信息组成的核以合成核的方式进行组合,进而用这个合成核替代原来的单核函数进行地物分类。实验结果表明,通过合成核的方法能够有效地集成空间、光谱与丰度等不同的特征空间,获得与基于特征选择的堆叠组合分类相当的分类精度,而不需要进行特征选择的过程,提高了高光谱图像分类效率。5)提出了一种结合丰度信息与空谱特征的半监督分类方法(BTSSA-SMLR)。针对高光谱遥感小样本分类问题,该算法通过结合丰度信息与后验概率进行主动学习,从类别隶属度和混合像元的角度选取信息量最丰富的样本用于半监督分类,进而提高半监督分类的效率。实验结果表明,结合丰度信息和后验概率的主动学习方法,可以利用少量的标记样本实现较高的总体分类精度,有效地减轻样本标记工作量,减少分类器样本训练所需时间,提高半监督分类的效率。
[Abstract]:The characteristic of hyperspectral remote sensing atlas combines the spectral information determined by the material components and the spatial information reflecting the texture and shape of the ground objects, so as to realize the accurate classification, recognition and attribute analysis of the ground objects, which greatly improves the human cognitive ability to the objective world. Hyperspectral remote sensing technology has been widely used in geology. Cartography, environmental monitoring, vegetation survey, agricultural remote sensing, ocean remote sensing, atmospheric research and many other fields. However, hyperspectral remote sensing data bring a lot of effective information, but also bring many challenges to the classification of hyperspectral images. The main problems of hyperspectral image classification are as follows: first, the composition and distribution of ground objects With the influence of complexity and imaging conditions, the phenomenon of "heterophysics" and "ISO spectrum" still exist on hyperspectral images, which lead to the occurrence of misclassification. Second, the general existence of mixed pixels on hyperspectral images further increases the difficulty of the classification of ground objects; third, with the increase of the image band, the training of the supervised classification The demand for training samples has also increased dramatically, causing the morbid problems in the classification of small samples. In recent years, the classification of space spectrum features, synthetic kernel classification, semi supervised classification algorithm and spectral unmixing technology have provided a possible solution to the above hyperspectral image classification problems. This paper extracts the characteristics of space spectrum and the extraction of abundance information. The method is to study the hyperspectral image classification method of joint abundance information and space spectrum features from two angles of supervised classification and semi supervised classification. Two groups of Aeronautical hyperspectral data with different spectral, spatial resolution and different ground complexity are used for classification experiments to verify the effectiveness of the proposed algorithm. The results are as follows: 1) the method of spatial spectral feature extraction based on morphology is studied, and the extraction process of morphological features is combed. By extracting the features of morphological attributes, it takes full advantage of the spectral information of the ground objects and takes into account the spatial structure features of the objects on the image, and then reduces the "heterophysics" and "foreign objects". The effect of the same spectrum on classification greatly improves the accuracy of the classification of hyperspectral images (.2). A classification based end extraction and sparse solution mixing algorithm is proposed. The end element and classification of spectral solutions are often inconsistent, resulting in the uneffective application of the unmixed endpoint abundance information to the classification problem. The end element is extracted for each category of the training sample, and the end element corresponding to each category is obtained. The endelements extracted from all categories are composed of endelements together, and the abundances are mixed with the end set. The results show that the algorithm is a common mixture in the classification of ground objects. The pixel problem provides an effective solution and can effectively improve the accuracy of the image classification. A spatial optimization algorithm based on Markov random field is studied. It is considered that the distribution of the ground objects in the nature usually has a certain spatial continuity, which is reflected in the image of remote sensing, that is, the adjacent pixels are the same types of objects. The algorithm uses a Markov random field prior model to make the adjacent pixels belong to the same category, and achieves the purpose of spatial optimization by using the spatial neighborhood information. The experimental results show that the spatial optimization algorithm can effectively remove salt and pepper noise, improve the classification accuracy and effect.4), and put forward a combination abundance. A synthetic kernel classification method (CKSMLRSSA) for the characteristics of information and space spectrum. This method combines the characteristics of the spatial spectrum and the abundance information at the nuclear structure level, so that it can be used more effectively for image classification. The experimental results show that the method can effectively integrate space, spectrum and abundance, and obtain the classification precision equivalent to the stacked combination classification based on feature selection without the process of feature selection. A semi supervised classification method (BTSSA-SMLR) combining abundance information and space spectrum features is proposed for hyperspectral image classification efficiency (.5). In view of the small sample classification problem of hyperspectral remote sensing, the algorithm is active learning by combining the abundance information and the posterior probability, and selects the most abundant information from the angle of category membership and mixed pixel. The sample is used for semi supervised classification to improve the efficiency of semi supervised classification. The experimental results show that a small number of labeled samples can be used to achieve higher overall classification accuracy with a small number of labeled samples, which can effectively reduce the workload of the sample marking, reduce the time required for the training of the classifier sample and improve the semi supervised classification. The efficiency of classification.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：博士
【学位授予年份】：2017
【分类号】：P237

【相似文献】