A Research on Input Schemes of Polarimetric SAR Classification Through Deep Learning

Based on the input schemes in Polarimetric Synthetic Aperture Radar (PolSAR) classification through deep learning. In this paper, we propose eight polarimetric parameter input schemes through reflection symmetric decomposition (RSD) and adopt classic convolutional neural networks for the classification of GF-3 PolSAR data.The results demonstrate that the commonly used 6-parameter input scheme, which many researchers adopt, lacks the comprehensive utilization of polarization information and warrants attention. The 7 alternative methods designed in addition to this scheme are all superior to it. If the computational resources are limited, it is suggested to directly use the scheme 7, which includes all the information of the T matrix. If the device configuration allows, it is recommended to prioritize the use of the 21-parameter polarization data input scheme 8, which includes all the parameters of the T matrix and the RSD.