Proposal of PolSAR Land Classification Using Quaternion Convolutional Neural Networks

Yuya Matsumoto, Ryo Natsuaki, Akira Hirose

Abstract: This paper proposes a quaternion convolutional neural network (QCNN) for PolSAR land classification. The QCNN learns spatial features of polarization with quaternion convo-lutional layers. The QCNN learns relationship between components of input three-dimensional vectors. This property also makes the QCNN map inputs to space of higher dimension with the same number of parameters than a real-valued CNN (RVCNN). In our experiments, the QCNN shows better classification performance than conventional networks. We also present visually that the quaternion kernels extract spatial features by quaternionic convolution.