Generalization characteristics of complex-valued reservoir computing for interferometric synthetic aperture radar applications

This paper discusses the principles and properties of complex-valued reservoir computing (CVRC), a proposed method for adaptive processing of interferograms with low computational cost. The topographic information of elevation, slope and aspect, obtained from InSAR data, can be utilized for hazard mapping and land use planning. Many recent studies use convolutional neural networks (CNNs) for image recognition tasks. However, CNNs have problems of high computational cost and resolution degradation of classification results. CVRC, a reservoir computing-based approach, provides a solution to these problems. We compare between CVRC and real-imaginary parts separated reservoir computing, and show that CVRC has a high generalization ability through an experiment on aspect classification.