【作者】 李奇峰；

【导师】 郭同德；

【作者基本信息】 郑州大学 ， 水利信息技术， 2015， 硕士

【摘要】 遥感影像是水利信息中的重要信息源,从遥感影像中提取所需信息是利用遥感影像的关键步骤。随着遥感数据获取手段的增多,遥感影像数据量飞速增长,如何高质、高效地进行遥感影像的分类显得至关重要。本文主要探讨遥感影像的自动分类问题,在对国内外相关文献进行阅读、归纳的基础上,做了以下研究。分析了目前该类研究中存在的考虑因素单一、研究方法综合性不足等问题。提出了遥感影像分类研究应结合智能算法和多特征描述来展开的观点。同时,根据支持向量机(SVM)在遥感影像分类领域的研究现状,提出了将多种方法描述的纹理特征和影像的光谱特征相结合,并利用SVM分类器进行分类的方法。介绍了SVM的基本理论、基本算法,详细讨论了SVM参数选择算法。根据SVM的泛化误差界,分析了SVM的小样本特性及其对模型复杂程度的控制能力。同时对极大似然估计、最近距离(NN)、K近邻(K-NN)、朴素贝叶斯等分类算法,就精度、效率、适用条件做了分析对比。在对灰度直方图、Gabor小波、离散傅里叶环状采样和离散小波分解四种纹理描述方法进行介绍和比较的基础上,根据Gabor小波滤波器的导出过程,提出了尺度参数选择的基本指导原则,对离散傅里叶环状采样方法进行了改进,进一步提出了DFT平均环状采样直方图方法。结合多特征描述以及SVM遥感影像分类算法,基于Lib SVM、Open CV、Free Image、SQLite、QT等开源工具和C++语言开发了一套实验系统,并以郑州市西北方向某一区域的Landsat8 OLI影像为例进行了一系列实验。实验表明,本研究所提出的SVM分类算法,其分类精度远高于最大似然估计、K近邻、朴素贝叶斯等分类算法的精度;所选用的四种纹理描述算法均具有一定区分能力,其中Gabor小波和DFT平均环状采样直方图方法区分能力最强;结合纹理特征和光谱特征进行SVM影像分类,可以将分类精度提高10%,总体分类精度最高可达96.2%;结合多种纹理描述算法可进一步提高SVM的影像分类精度。实验中还发现,若将区分度高的纹理描述算法和区分度低的纹理描述算法进行组合,其分类精度反而高于多种区分度均较高的纹理描述算法的组合,本文从模型复杂度控制的角度对这一现象进行了分析和解释。更多还原

【Abstract】 Remote sensing image is an important data source of water conservancy information. Extracting the required information from the remote sensing image is a key step. With the diversification of the means to acquire image data, the number of remote sensing image data is increasing. How to make the classification of remote sensing image with more efficiency is critical. This paper focuses on automatic classification of remote sensing image. O n the basis of reading the domestic and foreign literature, several aspects concerning the classification was conducted.It was shown that, only a few facts were taken in account in the classification of remote sensing image, and the methods used by scientists were not comprehensive enough. We think that the intelligent algorithms and multi- feature description should be combined with each other. At the same time, based on research status of the support vector machine(SVM) in this field, combining of the spectral characteristics of images and texture features described by a variety of methods as the input vectors, the SVM classifier was proposed.The primary theories and algorithms were introduced, and the parameters selection algorithms were discussed in detail. According to the generalization error bounds of SVM, the small sample properties of SVM and its effect on the complexity control of learning model was analyzed. At the same time, the maximum likelihood estimation, nearest neighbor(NN), K-nearest neighbor(K-NN), naive bayesian, support vector machine classification of remote sensing image, ware researched and compared on their accuracy, efficiency and applicable conditions.As texture description method, basic introduction and comparison of the gray histogram, Gabor wavelet, discrete Fourier circle sampling and discrete wavelet decomposition were presented. According to the derived process of Gabor wavelet filter, the basic guiding principle of the scale parameter selection was proposed. Through the improvement of DFT average circle sampling histogram method, the discrete Fourier circle sampling method was improved.Combining with multi- feature description and SVM classification algorithm, based on C++ language, Lib SVM, Open CV, Free Image, SQLite, Q T and other open source tools, a set of experimental system has been developed. Based on a piece of Landsat8 OLI image(a small part of northwest of Zhengzhou) as an example data, a series of experiments were conducted.Our experiments showed that, SVM classification algorithm proposed by this study has a much higher precision than the MLE, K-NN and naive Bayes. Each of the four kinds of texture description algorithm has distinguishing ability, especially the Gabor wavelet and DFT average circle sampling histogram methods. Compared with image classification by spectral characteristics and multi- features combined in SVM could improve the classification accuracy by 10% and the overall classification accuracy up to 96.2%. Combined with a variety of texture description algorithm, the classification accuracy of SVM image could be improved further. The experiments also showed that, combining high distinction texture description algorithm with low texture description algorithm, the classification accuracy would be higher than that of the combination of a variety of higher distinction description algorithm. The analysis and explanation of this phenomenon was carried on from the perspective of the model complexity control.更多还原