Abstract: In view of the problem of needing a large number of training samples for the depth network, this paper proposes a classification method that deep networks are effectively trained by limited labeled data. The homogeneous region increases the number of training samples. The initial classifier is trained by the increased training samples and predicts the initial category of all unknown pixels. Each initial category is aggregated into the appropriate cluster and cluster labels as a pseudo label. Hyperspectral data together with their pseudo labels are used to pre-train a deep convolutional neural network. Finally, the pre-trained deep convolution network is fine-tuned by the limited training samples added by homogenous regions. Experimental results with real hyperspectral image data sets demonstrate that the new method can successfully train deep convolutional networks with a small number of actual labeled samples and efficiently classify hyperspectral data.

Key words: Convolutional Neural Network（CNN）, pseudo labels, semi-supervised classification, hyperspectral imagery

摘要： 研究了一种仅利用少量标记点训练深度卷积神经网络并对高光谱影像进行分类的方法。以图像分割获得的同质区增加训练样本数目；借助这些增加的样本训练初始分类器并预测所有未知点的初始类别；将每一初始类别聚集为适当的类簇，以类簇号作为伪标签对深度卷积网进行预训练；最后利用经过同质区增加的训练样本精调预训练深度卷积网。实验结果证明新方法可以在仅用少量实际标记样本的情况下成功地训练深度卷积网，对高光谱数据进行有效分类。

关键词: 卷积神经网络（CNN）, 伪标签, 半监督分类, 高光谱影像