空间约束半监督高斯过程下的高光谱图像分类

针对高光谱遥感图像分类中带标记训练样本较少、导致分类正确率偏低的问题,提出用于高光谱图像分类的空间约束半监督高斯过程方法.由于高光谱图像的特征空间满足流形分布假设,大量未标记样本可以使数据空间变得更加稠密,从而有助于更加准确地刻画局部空间特性,提高分类的精度和普适性.通过对高斯过程模型中的核函数施加空间近邻约束,建立未标记样本与带标记样本之间的空间联系.该半监督高斯过程分类器不仅可以提升高光谱遥感图像的分类性能,而且构造简单,实现方便.实验结果表明,在仅有少量带标记的训练样本情况下,半监督高斯过程分类方法对高光谱图像有较高的分类精度和稳定性.

Semi-supervised learning based Gaussian processes for hyperspectral image classification

A new classification method based on spatial semi-supervised Gaussian processes (SSGP) was proposed to address the problem of low hyperspectral imagery classification performance caused by a small number of labeled training samples. As the feature space of a hyperspectral imagery satisfies the assumption of manifold distribution, a lot of unlabeled samples will make the feature space denser so that the local spatial character can be exploited more precisely and the classification accuracy and generality can be improved. In SSGP, the constraint of spatial neighborhood was imposed into the kernel function of Gaussian process, so the spatial correlations of labeled and unlabeled samples can be embedded in the kernel function. SSGP not only raises the classification performance, but also is easy to build and realize. Experimental results show that SSGP method is very good at classification of hyperspectral images in terms of classification accuracy and stability in the case of small size of labeled training samples．