LightGBM-Based Bias Correction of ERA5 Upper-Tropospheric Temperature Profiles and Its Differential Impacts on Cloud-Top-Height Retrievals Over East Asia

Abstract

The quality of atmospheric background temperature profiles is a critical factor governing the accuracy of passive-satellite cloud-top-height (CTH) retrievals. However, existing CTH retrieval studies generally treat profiles from the Fifth-Generation European Centre for Medium-Range Weather Forecasts (ECMWF) atmospheric reanalysis (ERA5) as fixed inputs, without accounting for their systematic biases. Moreover, the extent to which background field quality quantitatively affects CTH retrieval accuracy across different retrieval paradigms has not been systematically characterized. To address these gaps, we developed a LightGBM-based bias-correction scheme that incorporates physically informed structural features of the vertical temperature profile. The scheme was applied to ERA5 upper-tropospheric temperature profiles over East Asia using radiosonde observations from 136 stations collected during 2019–2020. The ERA5 profiles were interpolated onto 91 pressure levels with a vertical spacing of 10 hPa, and biases within the 150–500 hPa layer were corrected using a globally trained model, with correction performance assessed across seven climatic zones. The proposed scheme reduced the overall root-mean-square error (RMSE) by 9.59% and substantially alleviated the systematic cold bias of approximately −0.2 K. In addition, a latitude-dependent reversal in the sign of the bias near 35°N was identified and physically interpreted. Corrected profiles were applied to Himawari-8 Advanced Himawari Imager (AHI) CTH retrievals and validated against spatially independent Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO). For high clouds (6.0–13.5 km), the physical infrared window (IRW) method exhibits monotonic and continuous sensitivity to background field quality. Specifically, RMSE and mean absolute error (MAE) improve by 3.41% and 8.06%, respectively, for correction magnitudes $\geq 2.0$ K. In the summer deep-convection case study, the retrieval error decreases from 5.05 to 1.43 km, corresponding to a 71.7% reduction. In contrast, the data-driven eXtreme Gradient Boosting (XGBoost) model exhibits a threshold-type response, with consistent gains emerging only when ERA5 biases exceed 1.5 K. These results demonstrate that correction benefits scale directly with background field bias magnitude, rather than producing uniform improvements across all CTH retrievals. The most pronounced gains occur in the upper troposphere–lower stratosphere (UTLS), where ERA5 biases are concentrated, and under deep convective cloud scenarios, as illustrated by the representative cases examined in this study. We provide a systematic quantitative characterization of how background field correction magnitude modulates CTH retrieval accuracy across physical and data-driven methods, offering actionable guidance for retrieval algorithm design and background field quality standards.