结合Tri-training和CV-CNN的半监督PolSAR图像分类

现有深度学习算法应用于PolSAR图像分类时,较少考虑该图像数据的复数特点,使得数据的复数域信息不能被充分利用;同时,深度学习需要大量的标签样本作为模型的训练样本,但是PolSAR图像可获取的标签样本十分有限.针对上述问题,结合Tri-training算法和复值卷积神经网络(CV-CNN)提出了半监督PolSAR图像分类算法.首先通过Wishart分类器和Tri-training算法获取一些可靠性较高的伪标签样本,然后将其加入到复值卷积神经网络的训练样本中并用于模型训练,最终完成图像分类任务.通过四幅PolSAR图像分类的仿真实验表明,该算法不仅能够有效提升伪标签样本的可靠性,同时还可提高模型的分类准确率.

Semi-supervised PolSAR image classification with Tri-training and CV-CNN model

When PolSAR image using the existing deep learning model for classification, it rarely considers the character of PolSAR data which is complex-valued. Besides, deep learning needs a large number of labeled samples, while the obtained labeled samples of PolSAR image are very limited. In view of these problems, this paper proposed a semi-supervised PolSAR image classification which combined the Tri-training and complex valued convolutional neural network(CV-CNN). Firstly, it obtained some pseudo-label samples with high reliability through the Wishart classifier and Tri-training. Then, the CV-CNN model training process used the obtained pseudo-label samples with some labeled samples. Finally, the image classification task was completed. Simulation experiments of four PolSAR image show that the proposed method not only can effectively improve the reliability of pseudo-label samples, but also can improve the classification accuracy.