结合多特征描述和SVM的遥感影像分类研究

发布时间：2018-08-30 19:28

【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。
[Abstract]:Remote sensing image is an important information source in water conservancy information, and extracting the information needed from remote sensing image is a key step to use remote sensing image. With the increase of remote sensing data acquisition means, the data volume of remote sensing image is increasing rapidly. How to classify remote sensing image with high quality and efficiency is very important. This paper mainly discusses the automatic classification of remote sensing images. On the basis of reading and summing up the relevant literature at home and abroad, the following research is done. This paper analyzes the problems of single factor and lack of comprehensive research methods in this kind of research at present. The viewpoint that the classification of remote sensing images should be developed with intelligent algorithm and multi-feature description is put forward. At the same time, according to the research status of support vector machine (SVM) in remote sensing image classification, this paper proposes a method which combines the texture features described by many methods with the spectral features of the image, and uses SVM classifier to classify the image. This paper introduces the basic theory and algorithm of SVM, and discusses the SVM parameter selection algorithm in detail. According to the generalization error bound of SVM, the small sample characteristics of SVM and its ability to control the complexity of the model are analyzed. At the same time, the maximum likelihood estimation, nearest distance (NN) KNN (K-NN) and naive Bayes classification algorithms are analyzed and compared in terms of accuracy, efficiency and applicable conditions. On the basis of introducing and comparing four texture description methods, such as gray histogram Gabor wavelet, discrete Fourier ring sampling and discrete wavelet decomposition, according to the derivation process of Gabor wavelet filter, The basic guiding principle of scale parameter selection is put forward, the discrete Fourier ring sampling method is improved, and the DFT average annular sampling histogram method is further proposed. Combined with multi-feature description and SVM remote sensing image classification algorithm, an experimental system is developed based on open source tools such as Lib SVM,Open CV,Free Image,SQLite,QT and C language, and a series of experiments are carried out with the example of Landsat8 OLI image in a certain area northwest of Zhengzhou. Experimental results show that the classification accuracy of the proposed SVM classification algorithm is much higher than that of the maximum likelihood estimation (MLE) algorithm and the naive Bayes classification algorithm. Among them, Gabor wavelet and DFT mean ring sampling histogram have the strongest ability to distinguish, and combine texture feature and spectral feature to classify SVM image. The classification accuracy can be improved by 10%, and the overall classification accuracy can be up to 96.2.The SVM image classification accuracy can be further improved by combining various texture description algorithms. It is also found in the experiment that if the combination of texture description algorithm with high classification degree and texture description algorithm with low differentiation degree is combined, the classification accuracy of the algorithm is higher than that of many texture description algorithms with high degree of differentiation. This paper analyzes and explains this phenomenon from the angle of model complexity control.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP751

【参考文献】