RFI Mitigation via Auto-Encoder Framework in SAR Data

features of targets in the images and further be employed to classify different targets. Previous methods, such as the deep convolutional neural network (DCNN), were used to mitigate both narrowband and wideband interferences in SAR images (W. Fan, F. Zhou, et. al., Remote Sens., 11(14), 1654, 2019). For the unsupervised learning frameworks, such as auto-encoder (Y. Wang, H. Yao, et. al., Neurocomputing, 184, 232-242, 2016) and generative adversarial network (GAN) (I. Goodfellow, J. Pouget-Abadi, M. Mirza, et al, Advances in neural information processing systems, 27, 2014), they can discover naturally intrinsic property of the data without any pre-assigned labels and are widely leveraged in the image denoising area. Also, the PCA-based unsupervised learning methods (M. Tao, J. Su, et. al, Remote Sens., 11, 2438, 2019) are widely used to mitigate RFIs in the corrupted SAR data. Therefore, we come out an idea that the unsupervised learning frameworks have great potential on separate the RFI and the useful data. In this paper, we use an auto-encoder framework, combining real and imaginary parts of complex radar data as two branches, to mitigate strong RFIs from corrupted SAR data. It brings another perspective for RFI mitigation via deep unsupervised learning approach.