The Deep‐Learning‐Based Fast Efficient Nighttime Retrieval of Thermodynamic Phase From Himawari‐8 AHI Measurements.

Retrieval of the cloud thermodynamic phase (CP) is essential for satellite remote sensing and downstream applications. However, there is still a lack of efficient nighttime CP data products. A transfer‐learning‐based deep learning model, transfer‐learning‐ResUnet, is proposed to retrieve the nighttime CP of Himawari‐8 from thermal infrared channels. Cloud products of Himawari‐8 and Moderate‐resolution Imaging Spectroradiometers were selected as labels during training. A benchmark obtained by the Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observations confirmed the accuracy of the CP retrieval. During three independent months, the daytime and nighttime retrieval accuracy of the CP was 0.867 and 0.816, respectively, which was superior to that of the Himawari‐8 operational product in the daytime (0.788). Plain Language Summary: Clouds are essential for modulating global weather and climate systems and are routinely monitored by weather satellites such as Himawari‐8, a geostationary Earth‐observing satellite operated by the Japan Meteorological Agency. The cloud thermodynamic phase (CP) product is the main Level 2 product of Himawari‐8 and contains cloud properties. However, Himawari‐8 lacks a nighttime CP product, owing to its retrieval algorithm. Deep learning (DL) has proven to be effective in remote sensing. Among the various deep‐learning model structures, the fusion model of ResNet‐34 and Unet has the most accurate retrieval. Taking this model structure as the basis, the transfer‐learning‐ResUnet was established to improve the retrieval accuracy by integrating the Himawari‐8 and Moderate‐resolution Imaging Spectroradiometers cloud products as labels through a transfer‐learning method. To validate the nighttime CP retrieved by the model, data products obtained by Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observations were used. During three independent months, the daytime and nighttime retrieval accuracy of CP retrieval was 0.867 and 0.816, respectively, which were higher than that of Himawari‐8 operational product in the daytime (0.788). Key Points: The cloud thermodynamic phase obtained from Himawari‐8 is essential, but there is no accurate nighttime productA transfer‐learning‐based method provides rapidly obtained and accurate products, with an accuracy higher than that of Himawari‐8 productsIndependent comparisons with Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observations and Moderate‐resolution Imaging Spectroradiometers show that the retrieval results are reliable for both daytime and nighttime [ABSTRACT FROM AUTHOR]