An improved dual polarimetric sar quad-pol image reconstruction method based on full convolutional end-to-end neural network

Compared with quad polarization, dual polarization (DP) not only has twice wide-swath of observation but also decreases the synthetic aperture radar (SAR) system energy budget. In this paper, an end-to-end full convolutional neural network is proposed to achieve full polarimetric SAR image reconstruction based on dual polarimetric SAR data. Firstly, the feature extraction (FE) network is utilized to extract the multi-scale features of the dual-pol SAR data. Then, a feature translation (FT) network is proposed to achieve the stacked multi-scale features fusion and the quad-pol SAR image space mapping. The weighted cross-entropy loss function is designed to resolve the unbalanced reconstruction of different polarimetric channels. The measured ALOS/PALSAR data is utilized to validate the superiority of the proposed method.