基于特征重选择的高光谱图像分类

在对高光谱图像进行分类时,由于高光谱数据维度很高,通常先对其进行特征选择,进而使用分类器进行分类。在分类时,最大似然分类器因为具有多分类、概率输出等特点而经常被采用。在实际分类中,我们发现,目标的真实类别所获得的概率值通常位于所有类别的前两位,我们称该现象为混淆现象。利用这一现象,通过特征重选择,我们给出一种新颖的特征选择框架。该框架首先对数据进行特征选择,进而对目标进行分类。若分类结果满足混淆条件,则针对易混淆的两类进行特征重选择,分类并得到最终结果。理论分析和实验结果表明,该框架稳定、有效。

Feature reselection based hyperspectral image classification

In the hyperspectral image classificationliterature, since the dimension of the hyperspectral data is very high, we always take a feature selection before classification. Since the maximum likelihood is a multiclass classifier and can give probability output, it is often used in remote sensing image classification task. During the classification of hyperspectral image, we find that the real label of the target always lies in the top 2 of the classification output, and we call this phenomenon as confusion phenomenon. Take advantage of this phenomenon, and use feature reselection, we propose a novel feature selection framework. In the proposed work, we firstly take a feature selection, and then classify the given target. If the probability output satisfies the confusing condition, we reselect a feature subset to separate the two easy confusing classes. Theoretical analysis and experiment verify that proposed framework is effective and stable.