A Comparative Study on Developing High-Resolution Total Precipitable Water Vapor Data Using a Physics-Based Solar Attenuation Model and Deep Learning Reconstruction Model

Precipitable water vapor (PWV) is a key atmospheric variable governing cloud microphysics and strongly influencing weather conditions. However, accurate PWV observations are limited to point-based measurements such as radiosondes and GPS-GNSS, resulting in sparse spatial coverage, particularly over oceanic regions where deep convection frequently develops. To address this limitation, this study develops two complementary approaches to generate full-grid PWV fields over western Japan. The first is a physics-based empirical method by deriving an all-sky solar attenuation model applied to Himawari-8/9 Global Horizontal Irradiance (GHI). The second is a deep learning reconstruction approach using a Convolutional LSTM–Inpainting framework trained on 512 sparse GPS-GNSS 3-hourly PWV observations and Himawari-8/9 water vapor–sensitive brightness temperature bands (Band 08, 09, and 10), with topographic constraints. The estimated PWV fields are validated against the Japanese Reanalysis for Three Quarters of a Century (JRA-3Q) from the Japan Meteorological Agency. Due to differing temporal resolutions among datasets, evaluation is conducted at 09:00 and 15:00 JST. Results show that the deep learning approach achieves substantially lower spatial-mean root mean square error (sRMSE) than the empirical method, with temporal-average values of 11.79 mm and 21.65 mm, respectively. Despite its higher error, the empirical approach better captures localized high-intensity PWV variability in cloudy regions, whereas the deep learning method produces smoother fields. These findings highlight the complementary strengths of both approaches, suggesting their respective potential to enhance atmospheric moisture representation for improved numerical weather prediction.