Microwave SAIR Imaging Approach Based on Deep Convolutional Neural Network

Microwave synthetic aperture interferometric radiometers (SAIRs) are very powerful instruments for high-resolution remote sensing of the atmosphere and the earth surfaces at microwave frequencies. Microwave SAIR imaging reconstruction from interferometric measurements suffers from hardware non-identities, limited prior information, and noise interference, and consequently often requires expert calibration strategies to reduce imaging error and improve the accuracy of the reconstruction. In this article, we propose a new SAIR imaging approach with a deep convolutional neural network (CNN) learning framework to optimize the reconstruction performance. We interpret interferometric measurements of SAIR as a signal encoding representation and SAIR imaging as the corresponding decoding representation. A deep CNN framework with additional fully connected layers is utilized to autonomously learn the decoding representation from interferometric measurement samples and perform SAIR imaging. The supervised learning forward model with hyperparameters makes that the proposed approach could accurately obtain the SAIR imaging representation involving multiple systematic features for real applications. We demonstrate the performance of the proposed imaging approach through extensive numerical experiments. Compared with conventional handcrafted Fourier transform and sparse regularization reconstruction imaging approaches, the proposed imaging approach based on deep learning is superior in terms of image quality, computing efficiency, and noise suppression.