Result: Effect of Machine Learning Training Data on Water Vapor Profile Estimation Using Ground-Based Microwave Radiometer.

Several machine learning (ML) approaches have been developed to estimate vertical water vapor profiles from ground-based microwave radiometers (MWRs). Radiosonde observations, typically available only twice daily, have often served as training data. Their limited temporal coverage, however, requires long accumulation periods, and large errors remain in the lower atmosphere. This study investigates the effect of training data selection on water vapor profile estimation using the KASMI-100 MWR (FURUNO) installed at the Meteorological Research Institute. Hourly ECMWF Reanalysis v5 (ERA5) data were used for training and were compared, for the first time, with models trained on radiosonde data. The ERA5-trained models produced profiles more consistent with radiosonde measurements than those trained solely on sondes. Incorporating surface meteorological observations as supplementary training input markedly improved estimation accuracy below 1.5 km altitude. In addition, estimation accuracy decreased when cloud water was detected based on cloud-base temperatures observed by an infrared radiometer. While the bias tendencies of ERA5 may vary spatiotemporally and the same accuracy cannot be expected in all regions, this study confirms that ERA5 and surface data are beneficial for ML-based water vapor estimation. [ABSTRACT FROM AUTHOR]