Microwave Radiometer Calibration Using Deep Learning with Reduced Reference Information and Two-Dimensional Spectral Features

The accuracy of geophysical retrievals from radiometers relies on the quality of calibrations, which encompasses both absolute radiometric accuracy and spectral consistency. Radiometers have employed various calibration techniques, which include the utilization of external calibration targets, vicarious sources, and internal calibrators like noise diodes or matched reference loads. Calibration techniques pose several significant challenges such as frequency dependence, instrumental effects, environmental influences, drift, aging, and radio frequency interference. Recent advancements in hardware and processing units have enabled passive radiometers to collect raw samples of the observed scene that contain both temporal and spectral information. Leveraging advanced modeling techniques such as deep learning (DL) architecture can detect subtle correlations, non-linear dependencies, and higher-order interactions within the data. This capability allows them to extract valuable information that may have been difficult to capture using conventional methods. This study will utilize NASA's Soil Moisture Active Passive (SMAP) satellite's level 1A and level 1B data products to develop a DL-based radiometer calibrator to estimate antenna temperature. Spectrograms of second raw moments equivalent to power carrying the two-dimensional spectral features will be the primary input in a supervised convolutional neural network-based architecture. DL-based calibrator has demonstrated high R ² and low root mean square error when incorporating spectral information from both reference and noise diodes and when not considering this information. The findings from this analysis will suggest that the ancillary features in DL-based calibrators such as internal thermistor temperature and loss elements exhibit sufficient accuracy in estimating antenna temperature to compensate for variations in receiver noise temperature and short-term gain fluctuations in the absence of the reference load and noise diode power. The proposed calibration technique with reduced reference information might enable radiometers for a higher number of antenna scene observations within a footprint.