Pseudo‐Radar Reflectivity by Applying U‐Net Models to Global Passive Microwave Brightness Temperatures

Abstract Spaceborne radar systems such as the Global Precipitation Measurement Mission (GPM)'s core satellite Dual‐frequency Precipitation Radar (DPR) provide global insight into precipitation structure, storm morphology, and hydrological cycles. However, their limited spatial and temporal sampling and high cost constrain their ability to continuously monitor precipitation globally. This study explores the use of Convolutional Neural Networks (CNNs) to generate pseudo‐radar reflectivity profiles from Passive Microwave (PMW) Brightness Temperatures (Tbs) alone, aiming to fill gaps in spaceborne radar observations. Collocated GPM Microwave Imager (GMI) and DPR data sets from 2020 to 2021 are used to train and evaluate models with a U‐Net architecture. Two strategies are investigated: a Full Frequencies (FF) ensemble using all 13 GMI channels (10–183 GHz), and a High Frequencies (HF) ensemble restricted to 6 channels ≥89 GHz, reflecting configurations of radiometers on CubeSats. Results show that CNN models successfully reproduce large‐scale precipitation occurrence and echo top distributions, with over 50% of echo tops estimated within 1 km of DPR observations. The FF ensemble outperforms HF for warm rain and shallow systems due to the sensitivity to liquid water added by low‐frequency channels, whereas HF shows comparable skill at higher altitudes dominated by ice‐phase hydrometeors. Both models systematically underestimate reflectivity, particularly at higher values (e.g., >40 dBZ) and altitudes, highlighting challenges in capturing extreme events. The findings demonstrate the potential of AI‐driven pseudo‐radar products from abundant PMW observations to complement spaceborne radar measurements, with implications for future precipitation monitoring and nowcasting capabilities from PMW sensors on CubeSats.