基于改进三重训练算法的高光谱图像半监督分类

Semi-supervised classification for hyperspectral image based on improved tri-training method

高光谱数据维数高,有标签的样本数量少,给高光谱图像分类带来困难。本文针对传统三重训练(tri-training)算法在初始有标签样本数量较少的情况下分类器间差异性不足的问题提出了一种基于改进三重训练算法的半监督分类框架。该方法首先通过边缘采样策略(margin Sampling,MS)选取最富含信息量的无标签样本,然后在训练每个分类器之前通过差分进化算法(differential evolution,DE)利用所选取的无标签样本产生新的样本。这些新产生的样本将被标记并且加入训练样本集来帮助初始化分类器。实验结果表明,该方法不仅能够有效地利用无标签样本,而且在有标签数据很少的情况下能够有效地提高分类精度。

The classification of hyperspectral images is difficult due to their highly dimensional features and limited number of training samples. Tri?training learning is a widely used semi?supervised classification method that addres?ses the problem of the deficiency of labeled examples. In this paper, we propose a novel semi?supervised learning algorithm based on an improved tri?training method. The proposed algorithm first uses a margin sampling ( MS) technique to select the most informative samples, and then uses a differential evolution ( DE) algorithm to generate new samples within the selected unlabeled samples. The newly generated samples are then labeled and added to the training set to help initialize the classifiers. We experimentally validated the proposed method using real hyperspec?tral data sets, and the results indicate that the proposed method can significantly reduce the need for labeled sam?ples and can achieve high accuracy compared with state?of?the?art algorithms.