【作者】 李琳；

【导师】 袁春； 郭子祺；

【作者基本信息】 中国地质大学（北京） ， 自然地理学， 2009， 硕士

【摘要】 遥感技术自二十世纪六时年代提出以来,已经为很多领域的具体应用提供了强大的信息支持,为了更好的利用遥感技术,人们在遥感各个方向上的研究都有了长足的进步。遥感影像需要首先通过解译,获取相关的专题信息后才能应用于地学以及其他工作中,计算机自动解译虽然在精度上与人工解译存在差距,但具有先天的速度优势。随着分类算法研究的不断深入,使计算机自动解译的精度不断提高,已能满足一定具体工作的需要。但在如今的技术飞速发展的时代,随着遥感数据的海量化以及某些领域对于信息迅速获取的迫切需求,分类的速度已经越来越得到人们的重视,这也就催生了提升分类速度算法的研究。本文则从分类效率入手,对分类算法进行相关研究。旨在满足精度要求的前提下,寻找提高分类效率,节省分类时间的改进算法。计算机自动分类方法应用较多的传统的模式识别方法,如非监督分类的ISODATA分类法,监督分类的最大似然法等。但这些传统方法受到遥感影像分辨率以及“同物异谱”、“异物同谱”现象的影响,出现较多错分、漏分,导致分类精度不高。目前也提出了一些改进算法,使其分类精度有了大幅的提高。随着遥感技术的发展,近年来出现了一些新的倾向于句法模式的分类方法,如人工神经网络方法、模糊数学方法、决策树方法、支持向量机方法等。本文选取了两类方法中的典型,包括ISODATA法、最大似然法、决策树法以及支持向量机(SVM)法进行对比研究,并最终提出决策树与SVM相结合的分类方法。经过研究,四种方法在分类精度上存在较大差异。支持向量机方法精度最高;决策树与最大似然法精度相当,位居次席;ISODATA法精度最低。但在分类效率上决策树法消耗时间最少,最大似然法与ISODATA法相当,支持向量机法消耗时间最多。本文提出的决策树与SVM相结合的方法精度上基本与单独使用支持向量机法相同,但消耗时间却大幅减少,略少于单独使用最大似然法或ISODATA法。达到了在精度基本不变的前提下,提高分类效率的目的,显示出此方法具有分类效率与精度上的综合优势。更多还原

【Abstract】 Since remote sensing technology was brought forward in 1960s, it has provided information to many areas powerfully. In order to using remote sensing technology better, people has had significant progress in all directions of the remote sensing research. Remote sensing image need to be interpreted at first, to get the thematic information, then it can be used in Geography and other works. Automatic interpretation of computer is lower in accuracy than manual interpretation, but it has advantage in efficiency innately. With the development of classification research, the accuracy of automatic interpretation of computer has improved continuously, and it can meet the demand of certain specific works. Because of the magnanimity of the remote sensing data, and the urgent need of getting information rapidly in some areas, the speed of has been paid more and more attention, and that gives the birth to the research in algorithm of improving the speed of classification. In this paper, it starts from the classification efficiency to study the classification algorithm. It designed to meet the requirements of accuracy, finding algorithm that can improve classification efficiency.Traditional methods of pattern recognition are widely used in automatic interpretation of computer, such as the ISODATA of unsupervised classification, the Maximum likelihood of supervised classification. But these methods are affected by the resolution of remote sensing images and the phenomenon of "the same thing different features" and "different things the same feature", so it makes wrong classification with low accuracy. With the development of remote sensing technology, classification of syntactic patterns has occurred in years, such as ANN, Fuzzy Models, Decision tree, Support Vector Machine (SVM). In this paper, typical methods of tows are selected to execute the comparative study, including the ISODATA, Maximum likelihood, Decision tree and Support Vector Machine (SVM). And classification method of combining Decision tree and Support Vector Machine (SVM) together is brought forward ultimately.After research, the classification accuracy of four methods are of great differences. The accuracy of SVM is the highest, Decision tree and Maximum likelihood get the second place, ISODATA is the lowest. However, in the efficiency of classification the Decision tree needs the lest time, Maximum likelihood and ISODATA get the second place, SVM need the most. The method of combining Decision tree and Support Vector Machine (SVM) together in this paper gets the same accuracy as SVM, but the time needing is reduced rapidly, slightly less than ISODATA or Maximum likelihood. It gets the purpose that in premise of accuracy unchanged basically to improve the classification efficiency. It has advantage in both accuracy and efficiency.更多还原