结合波段选择的差分进化高光谱图像分类

Differential evolution hyperspectral image classification combined with band selection

高光谱图像具有数据维数高、有标签样本少的特点,影响了现有分类方法的效果。针对这一情况,提出一种结合波段选择的半监督分类算法。该算法首先通过波段选择方法,去除高光谱图像中的冗余信息,进而降低复杂度和提高泛化能力;然后通过差分进化算法交叉变异无标记样本,选取置信度高的样本扩充入标记样本群以提高分类精度。实验结果表明,该算法能够有效地提升在标记样本有限的情况下分类器的分类精度与分类速度。

Hyperspectral images have the characteristics of high dimension of data and few labeled samples, which affects the effect of existing classification methods. Aiming at this situation, we propose a semi-supervised classification algorithm combining band selection. Firstly, the band selection method is used to remove the redundant information in the hyperspectral image, reducing the complexity and improving the generalization ability. Then, the differential evolution algorithm is used to cross-mutates the unlabeled samples, and expand the labeled sample group by selecting the samples with high confidence. The experimental results show that the proposed method can effectively improve the classification accuracy and speed of the classifier under the condition of limited label samples.