PolSAR land classification by using quaternion-valued neural networks

We propose the use of quaternion-valued neural networks for PolSAR land classification based on Poincare sphere parameters. By using quaternion-valued feedforward algorithm, the 3-dimensional Poincare sphere parameters are treated as vectors directly instead of three independent real values. In this way, the information of the parameters are highly preserved. Then, in comparison with real-valued neural network, the proposed quaternion-valued neural network can provide higher classification accuracy. By this method, we generate successful classification images for detecting the lake, grass, forest, and town areas. The experimental results show that, by using quaternion-valued neural networks, the classification performance have been much improved.