Conditional Generative Adversarial Networks Enhance Atmospheric Thermodynamic Profile Retrieval from Ground-Based Microwave Radiometer Measurements

Accurate retrieval of atmospheric temperature and humidity profiles is critical for applications such as weather forecasting, climate monitoring, and atmospheric research. Ground-based microwave radiometers (MWRs) are widely employed for these retrievals due to their capability to provide continuous, high-resolution observations under various weather conditions. However, traditional statistical and machine learning-based retrieval algorithms often face challenges in accuracy and robustness, especially in complex atmospheric conditions. This study presents a novel deep-learning approach using conditional generative adversarial networks (cGANs) to enhance the retrieval of temperature and humidity profiles. By employing adversarial learning, cGANs improve the quality of data generation and reconstruction. The network is conditioned on brightness temperatures from MWRs, enabling it to learn the nonlinear relationships between observed radiances and atmospheric profiles effectively. The proposed method achieves remarkable performance, with R² values of 0.99 for temperature and 0.96 for humidity, and root mean square error (RMSE) of 2.39 K and 0.54 g m⁻³, respectively. Notably, cGANs significantly enhance relative humidity (RH) retrievals, achieving R² of 0.55 and RMSE of 16.93%, outperforming both traditional optimal estimation (OE) and several established machine learning methods. Importantly, the cGANs model is trained and validated by using datasets that include both clear and cloudy skies, and the results demonstrate that the model maintains high accuracy across both conditions. These findings highlight the potential of advanced deep-learning methods, such as cGANs, to significantly improve MWR-based retrieval of atmospheric temperature and humidity profiles.