【作者】 吴静；

【导师】 韩玲；

【作者基本信息】 长安大学 ， 摄影测量与遥感， 2010， 硕士

【摘要】 高光谱遥感作为一种新型的遥感方式在近几年的发展中已经广泛运用于军用和民,用的多个领域,然而如何从其产生的海量数据中快速而准确地挖掘出所需要的信息,目前仍然是一个待解决的问题。简单的支持向量机只能处理二值分类问题,不能直接处理多值分类问题。而现实世界中的大部分数据都是多类数据,所以需要对简单支持向量机作进一步扩展,使之能解决多值分类问题。本文介绍了几种用于多值分类的支持向量机分类的方法,包括“一对多”、“一对一”、基于决策树以及基于有向无环图的支持向量机分类方法,并比较了它们各自的优点和缺点。通过分析各种支持向量机分类方法的不足之处,提出了一种新的支持向量机分类的方法,即将图斑理论与多值分类的支持向量机分类方法相结合。最后通过试验比较传统的支持向量机分类技术与结合图斑特性的支持向量机的分类技术,结果表明结合图斑的支持向量机在应用于高光谱影像分类问题中取得了较好的效果。基于图斑的支持向量机方法是选择一个合适的尺度,利用光谱信息按照一定的策略将图像分割为一系列的图斑,并确保图斑内大多数像元的光谱特征相近,分别对图斑内各像元进行统计,求出各个波段的均值用该均值替换图斑内所有像元的原始亮度值。这样分类的目的是为了使各种原因产生的噪声点可以被其周围的像元同化而融合到一个图斑类别中,相当于根据其周围像元的信息对其进行了有效的恢复,不至于在分类结果上出现孤立的错分点,避免产生“椒盐现象”。试验结果表明这种方法是可行的并且分类精度和速度均较传统的支持向量机分类法有所提高。更多还原

【Abstract】 During the recent two decades, hyperspectral remote sensing has been playing an important role in both military and civil applications. It’s urgent to develop fast and accurate methods in order to discover the interested information from the huge data which were produced by hyperspectral sensors.Simple SVM can only handle binary classification problems; can not directly handle multi-value classification. In the real world most of the data is multi-class data, so the simple SVM need for further expansion, so that it can solve the multi-value classification. This paper introduces several SVMs for multi-value classification, including "One against Rest", "One against One", Decision Tree and Directed Acyclic Graph SVMs, and compares their respective advantages and disadvantages. By analyzing the deficiencies of various SVMs, a new SVM method, namely, the theory of combining spot and SVM, is put forward. Finally, comparing the traditional SVM to SVM binding spot feature, the tests show the SVM combining of spot applied to hyperspectral image classification has achieved good results.The principle of SVM based on spot is to choose an appropriate scale to split the image into a series of segmentation, according to certain strategy using spectral information. And this principle ensures the spectral features of the majority of patch pixel similar. This method gathers statistics of each pixel value in the spot and obtains the mean value of each band to replace the original value of all pixels in the spot. The purpose of this classification is that the pixel having the noise brought by various causes is assimilated by the surrounding pixels to merge into a single spot. In other words, under the information of its surrounding pixels recovering the value of the pixel having noise is to not appear the fault isolation in the classification map and to avoid the salt and pepper phenomenon. The results show that this method is feasible and the classification accuracy and speed is better than traditional support vector machine.更多还原