Proposal of PolSAR Land Classification Using Full-Learning Quaternion Convolutional Neural Networks

Yuya Matsumoto, Ryo Natsuaki, Akira Hirose

Abstract: Quaternion convolutional neural networks (QCNNs) are in-herently useful for image processing and PolSAR land classification, since they can learn the relationship between the components of input vectors with quaternionic rotations. However, conventional QCNNs fix their rotation axes represented by quaternion weights, resulting in reduction of the degree of freedom (DoF) and the lose of the expression ability. In this paper, we propose QCNNs which learn all the four parameters of the quaternion weights by backpropagation. They perform learning with the maximum of DoF so that they take full advantages of quaternion learning. In addition, we use two totally different features, namely, Pauli RGB features and normalized Stokes vectors. We experimentally found that Pauli RGB features are suitable for discrimination between town and forest, and Stokes vectors between water and grass. Combining their two results complementarily improves classification results. Our proposed QCNNs show the best classification performance compared with real-valued convolutional neural networks and fixed-axis QCNN. These results demonstrate the strength of the proposed QCNN in adaptive polarization processing in multimodal data in the PolSAR field.