Deep Learning Imaging for 1-D Aperture Synthesis Radiometers

Abstract

For one-dimensional aperture synthesis (1-D AS) radiometers, truncated sampling occurs in the frequency domain due to the system baseline limitation. Therefore, there is an obvious Gibbs oscillation in the reconstructed image. To solve this problem, an imaging method based on a one-dimensional convolutional neural network (1-D CNN) is proposed in this paper. Compared with deep learning methods based on two-dimensional convolutions, the 1-D convolution not only reduces the amount of computation but also produces further performance improvements. The input data of the network are the 1-D visibility function samples, and the output data are the 1-D brightness temperature (BT) samples. The network learns the mapping relationship from the training of the 1-D visibility function samples and 1-D BT samples to complete 1-D AS imaging without any prior knowledge. To verify the performance of this imaging method, simulations and experiments based on the airborne C-band one-dimensional microwave interferometric radiometer (ACMIR) system are implemented. The simulation and experimental results demonstrate that the proposed AS-CNN method achieves higher performance than the IFFT method in terms of image quality and Gibbs phenomenon suppression. In the case of an antenna failure and missing baseline, the AS-CNN method proposed in this paper can still obtain a BT image with high imaging quality, which shows that the robustness of the network is better than that of the IFFT method.