Spatial resolution of complex-valued reservoir computing in aspect classification using InSAR data

Convolutional neural networks (CNNs) have been utilized for SAR image analyses such as aspect classification and land cover grouping. However, reservoir computing-based approach was also proposed for aspect classification because CNNs have serious problems of resolution degradation and large computational cost. In this paper, we conduct an experiment of aspect classification by using complex-valued reservoir computing (CVRC) and complex-valued CNNs (CVCNNs). We also discuss the tradeoff between the spatial resolution of the CVRC results and noise robustness. In addition, we compare these approaches in terms of segmented image resolution and computational cost. Experimental results show that CVRC realizes one-hundredth learning time as well as one-fifth classification time, compared with CVCNNs, with the same resolution as that of original interferogram.