Bias Assessment and Correction of Atmospheric Temperature and Vapor Density Profiles from Six Ground-Based Microwave Radiometers over Wuhan

This is the authors' abstract. We don't add key points for this paper.

Ground-based microwave radiometers (MWRs) continuously retrieve atmospheric temperature and vapor density profiles, but systematic biases relative to radiosonde observations limit their quantitative application in nowcasting and data assimilation. In this study, we matched temperature and vapor density profiles from six ground-based MWRs with L-band radiosonde observations at Wuhan to evaluate data quality for 2025 and analyze the sources of station-to-station bias differences. We established linear regression, random forest (RF), and artificial neural network (ANN) models to correct biases by height group and sky condition, respectively, and compared the corrected data with the original data. The results show that, compared with radiosonde data, the MWR temperature was generally cold-biased under clear (−0.895 °C), cloudy (−0.729 °C), and rainy (−0.188 °C) skies, whereas the vapor density was moist-biased under all three conditions (+0.417, +0.209, and +0.47 g/m3). The per-station biases differed markedly and could not be explained by spatial separation alone, indicating station-specific environmental and representativeness differences rather than instrument differences. All three correction methods significantly reduced the deviations; taking the RMSE together with the residual bias as the joint criterion, the ANN was selected as the best method, reducing the test set temperature RMSE from 2.802 °C to 2.284 °C (−18.5%) and the vapor density RMSE from 2.764 g/m3 to 1.746 g/m3 (−36.8%), followed by random forest (−17% and −35.9%) and linear regression (−6.8% and −6.8%). After removing the small constant residual bias, the ANN is both the most accurate and effectively unbiased. The corrected profiles brought the thermodynamic instability parameters to a usable level for severe-convection nowcasting and provided an early-warning signal in the pre-onset period of gale and short-duration heavy rain, which also provide a reference for the quantitative application of microwave radiometer data.