Deep-Learning-Based Approach in Imaging Radiometry by Aperture Synthesis: Application to Real SMOS Data

Abstract

A novel image reconstruction algorithm for aperture synthesis measurements using deep learning techniques was introduced recently. This algorithm is specifically designed to retrieve brightness temperature (BT) from interferometric data, similar to those collected by the European Space Agency's (ESA) Soil Moisture and Ocean Salinity (SMOS) mission, launched in 2009. The algorithm employs a deep neural network (DNN) architecture that features a fully connected layer followed by a contracting and expansive path, enabling the network to effectively learn the relationship between simulated visibilities and BT maps. Validation with simulated data has confirmed that this approach aligns perfectly with the theoretical framework of the Van-Cittert Zernike theorem. In this study, a new DNN architecture better suited for real SMOS data is proposed. The new architecture integrates a priori information regarding the water content of each observed pixel. It also includes further enhancements to the previous DNN architecture to better accommodate real SMOS data by incorporating the effects of radiometric noise and the Faraday rotation angle, as well as selecting appropriate global BT maps for training. Finally, validation of the proposed DNN approach using large datasets of real SMOS data is presented and compared to the traditional algebraic approach. Globally, the results demonstrate a significant improvement in image quality, with a reduction in reconstruction error, better handling of residual foreign sources, such as radio frequency interference and direct solar radiation, and a notable reduction in land-sea and sea-ice contamination. Overall, the results suggest that the DNN-based approach provides substantial improvements over traditional methods, making it a promising technique for processing SMOS data.