基于多特征融合的高光谱遥感图像分类研究

发布时间：2018-03-14 07:08

本文选题：遥感图像分类　切入点：特征提取　出处：《北方民族大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着遥感影像技术和信息技术的迅速发展,遥感影像数据量呈现快速增长趋势。面对海量的遥感数据,如何利用计算机按照一定规则自动对影像进行分类成为具有挑战性的研究课题。传统的方法是目视解译,该方法需要丰富的专业经验知识,又需要充足的户外实地调查资料,而且这种识别方法建立在具有一定的先验知识基础上,所以识别难度大,效率较低。高光谱遥感技术又称为成像光谱遥感技术,具有图像和光谱信息融合的特点,高光谱图像上的每一个像元分别对应一条独特的光谱曲线,因此利用这一性质可以根据地物的光谱反射率来识别遥感影像中的地物类型。同时高光谱图像兼具光学特性和光谱识别能力,已经成为遥感影像领域广泛研究与应用的焦点。遥感图像分类的应用在遥感图像研究中具有重要意义。基于单个特征分类算法对于遥感图像的分类能力是比较有限的,因此为了提高高光谱遥感图像分类的精度,本文提出了基于多特征融合的高光谱遥感分类方法,主要研究工作如下:1.高光谱遥感图像分类方法相关概述。分析了高光谱遥感图像的特征,对前人的高光谱图像特征分类方法进行分析和总结,比较单特征分类方法的不足,提出了基于多特征融合的高光谱遥感图像分类研究。2.经过分析研究发现,遥感图像分类可以从两方面改进:一方面,改进特征;另一方面,改进分类算法。因此本文提出了基于这两方面改进的算法。基于特征改进方法是以现有的特征作为基础,详细介绍纹理特征,直方图特征,降维特征并且利用各类特征的优势,对上述三类特征进行归一化融合,获得融合特征,接着使用AdaBoost集成的三种分类算法Real AdaBoost,Gentle AdaBoost,Modest AdaBoost分别对高光谱遥感数据单特征和融合特征进行分类,并且对分类结果进行对比。基于分类算法改进方法是在RVM算法基础上,将局部二值化对比算法(LBPC)和形态学算法(Morphologic Algorithm)进行改进获得L-M Algorithm算法,接着将改进的算法嵌入到RVM算法,然后获得融合特征分类结果。为了使实验更加充分,本文还利用SVM和AdaBoost算法进行分类对比实验。3.分类算法选择。分类算法性能的优劣直接决定着分类效果的好坏,因此,分类算法的选择是高光谱图像分类中的关键一步。实验结果表明,相比于单特征分类精度,融合算法获得的分类精度更高。因此可以得出本文提出的融合算法对于特征分类效果更可靠,更精准的结论。
[Abstract]:With the rapid development of remote sensing image technology and information technology, the amount of remote sensing image data is increasing rapidly. How to use the computer to classify images automatically according to certain rules has become a challenging research topic. The traditional method is visual interpretation, which requires rich professional experience and sufficient outdoor field investigation data. And this recognition method is based on a certain priori knowledge, so the recognition is difficult and inefficient. The hyperspectral remote sensing technology, also called imaging spectral remote sensing technology, has the characteristics of image and spectral information fusion. Each pixel on a hyperspectral image corresponds to a unique spectral curve, Therefore, this property can be used to identify the types of ground objects in remote sensing images according to the spectral reflectivity of ground objects. At the same time, hyperspectral images have both optical properties and spectral recognition capabilities. The application of remote sensing image classification is of great significance in remote sensing image research. The classification ability of remote sensing image based on single feature classification algorithm is relatively limited. Therefore, in order to improve the accuracy of hyperspectral remote sensing image classification, a hyperspectral remote sensing classification method based on multi-feature fusion is proposed in this paper. The main research work is as follows: 1. Overview of hyperspectral remote sensing image classification methods. The features of hyperspectral remote sensing images are analyzed, and the former hyperspectral image feature classification methods are analyzed and summarized, and the shortcomings of single feature classification methods are compared. The hyperspectral remote sensing image classification based on multi-feature fusion is proposed. 2. Through analysis and research, it is found that remote sensing image classification can be improved from two aspects: on the one hand, improved feature; on the other hand, Therefore, this paper proposes an improved algorithm based on these two aspects. The improved method is based on the existing features, including texture features, histogram features, dimensionality reduction features and the advantages of all kinds of features. The above three kinds of features are normalized fused to obtain the fusion features. Then the hyperspectral remote sensing data single feature and fusion feature are classified by using Real boost Gentle boost Modest AdaBoost, three classification algorithms of AdaBoost ensemble. The improved method based on RVM algorithm is to improve the local binary contrast algorithm and Morphologic algorithm to obtain L-M Algorithm algorithm, and then embed the improved algorithm into RVM algorithm. Then the fusion feature classification results are obtained. In order to make the experiment more fully, this paper also uses SVM and AdaBoost algorithm to carry on the classification contrast experiment .3. the classification algorithm choice. The performance of the classification algorithm directly determines the classification effect, therefore, The selection of classification algorithm is a key step in hyperspectral image classification. Experimental results show that, compared with the accuracy of single feature classification, Therefore, the fusion algorithm proposed in this paper is more reliable and accurate for feature classification.
【学位授予单位】：北方民族大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】