Calibration of Aperture Synthesis Radiometer Based on Deep Learning

Aperture Synthesis Radiometer (ASR) is an important instrument used in Earth remote sensing observations and has significant implications in areas such as climate change research and environmental monitoring. Nevertheless, conventional ASR calibration techniques exhibit inherent limitations, such as direction diagram correction and internal noise injection. This paper introduces ASRC-Net, a deep learning-based calibration method for ASR that can improve calibration accuracy. Firstly, the method utilizes a deep learning network (ASRC-Net-1) to conduct feature learning and extraction on spectrum data gathered by ASR, facilitating brightness temperature (BT) reconstruction and relative calibration. Subsequently, by integrating reference BT scene data, such as flat sea surfaces and employing appropriate network structure, loss functions and optimization algorithms, an absolute calibration network for ASR (ASRC-Net-2 network) is established. The network undergoes rigorous training and validation utilizing a substantial volume of measured data. Experimental findings attest that the proposed deep learning-based ASR calibration method notably enhances calibration accuracy. This method heralds a novel approach and technical avenue for precise ASR measurement, holding substantial application value and promising prospects for widespread adoption. Furthermore, it furnishes technical underpinning for high-precision in-orbit measurement of Chinese ocean salinity detection satellites.