Full-Learning Rotational Quaternion Convolutional Neural Networks and Confluence of Differently Represented Data for PolSAR Land Classification

Quaternion convolutional neural networks (QCNNs) expand the range of their applications in processing optical and polarimetric synthetic aperture radar (PolSAR) images. Conventional real-valued convolutional neural networks (RVCNNs) compress a three-channel input image into a single-channel feature map and ignore the relationship among the channels. In contrast, QCNNs deal with the input image as a single quaternion matrix and perform quaternion operation without the reduction of the channels. They can learn the interrelationship among the channel components. Though there exist two types of QCNNs, they have problems, respectively. One type conducts physically unclear quaternion convolution by using simple quaternionic multiplications. The other employs quaternion rotations with fixed axes, resulting in impairment of expression ability. In this article, we propose full-learning rotational QCNNs, which perform quaternion rotation in convolution, and update all the four parameters of a quaternion weight by backpropagation. They realize quaternion rotational convolution with high expression ability. We also propose using two different kinds of features, namely PolSAR pseudocolor features and Stokes vectors normalized by their total power. These two features allow neural networks to learn totally different characteristics of land surface. We train two networks with these features independently. Then, we merge their two classification results to obtain final decision to compensate for the shortcomings of the respective features. Experiments demonstrate that our proposed QCNNs show better classification performance than that of RVCNNs and the two existing QCNNs. We also find that the combination of the two features improves final classification results measured by F-scores.