ABOUT THE EQUIVALENCE BETWEEN COMPLEX-VALUED AND REAL-VALUED FULLY CONNECTED NEURAL NETWORKS - APPLICATION TO POLINSAR IMAGES

In this paper we provide an exhaustive statistical comparison between Complex-Valued MultiLayer Perceptron (CV-MLP) and Real-Valued MultiLayer Perceptron (RV-MLP) on Oberpfaffenhofen Polarimetric and Interferometric Synthetic Aperture Radar (PolInSAR) database. In order to compare both networks in a fair manner, the need to define the equivalence between the models arises. A novel definition for an equivalent Real-Valued Neural Network (RVNN) is proposed in terms of its real-valued trainable parameters that maintain the aspect ratio and analyze its dynamics. We show that CV-MLP gets a slightly better statistical performance for classification on the PolInSAR image than a capacity equivalent RV-MLP.