现有的高光谱图像分类方法多注意到空间信息的利用,并未考虑地物在空间分布上具有连续性的特点。基于此,提出了一种空谱加权近邻(SSWNN)高光谱图像分类算法。通过构造测试样本点的近邻空间,过滤近邻空间中与测试样本标签不一致的空间近邻点,有效解决了近邻空间内异类点对中心像元分类的干扰,改善了图像的椒盐效应。根据空间近邻点和测试像元之间的光谱相似性为空间近邻点赋予不同的权重,增大了同类像元间的相似性和异类像元间的差异性,并通过引入正则化系数,得到训练样本和测试样本近邻空间的距离,选择距离最小的训练样本标签作为测试样本的标签。该方法在Indian Pines和PaviaU高光谱数据集上的总体分类精度分别达到了96.75%和98.54%,高于文中所列的其他算法。

Existing hyperspectral image classification methods focus on using spatial information without considering the continuity of ground object in spatial distribution. Based on this, this paper proposes a space-spectrum weighted nearest neighbor hyperspectral image classification algorithm. By constructing the neighboring space of the test sample points, the spatial neighboring points in the neighboring space that are inconsistent with the test sample labels are filtered to further remove the interference of heterogeneous points in the neighboring space towards the classification of central pixels and improve salt and pepper effect. According to the spectral similarity between the spatial neighbors and the test pixels, different weights are assigned to the spatial neighboring points, which increases the similarity between similar pixels and the difference between the heterogeneous pixels. The distance between the training sample and the test sample neighboring space is obtained by introducing the regularization coefficient, and the training sample label with minimum distance is selected as the label of the test sample. The overall classification accuracies by this method on the Indian Pines and PaviaU hyperspectral datasets reach 96.75% and 98.54%, respectively, which are higher than those by other algorithms listed in the paper.