高光谱遥感影像SVM分类中训练样本选择的研究

Use of mixed pixels as training samples for hyperspectral remote sensing image classification by SVM

支持向量机(SVM)分类的关键是发现分类最优超平面及类别间隔,而混合像元比纯净像元更接近类别边界,更容易找出最优超平面。本文针对SVM分类器的特点,在高光谱数据分类中采用混合像元作为训练样本对SVM进行训练,试验表明采用类别边界上的混合像元作为训练样本是可行的,能够获得与纯净训练样本接近的分类精度,进一步验证了SVM分类对训练样本空间分布依赖度较低的特点。

The key idea of Support Vector Machine（SVM） classification is locating an optimal separating hyper-plane and maximizing the margin between two classes.It is obvious that mixed pixels are much closer to the boundary of classes than pure pixels and much easier to locate the optimal separating hyper-plane.The paper used mixed pixels as training samples for SVM classifier in hyperspectral image classification.Experimental results showed that hyperspectral remote sensing image classification by SVM using mixed pixels was feasible,and its accuracy was similar to the accuracy derived from the use of a conventional pure pixel training set.The characteristic of the SVM classifier was demonstrated further that it has low dependence on the spatial distribution of training samples.