Journal of Atmospheric and Solar-Terrestrial Physics paper Mashao et al. 2026 published

Upper-air temperature is essential for assimilation in Numerical Weather Prediction systems and climate change detection, yet radiosonde (RS) measurements often contain systematic biases that compromise long-term consistency. Conventional approaches to bias characterization and correction in inverse modelling remain limited, relying heavily on model assumptions, expert judgment, and efficiency-driven simplifications. This study develops a data-driven framework to characterise and correct RS temperature biases by integrating satellite-based atmospheric soundings (COSMIC-2, Suomi-NPP, NOAA-20), ERA5 reanalysis data and machine-learning methods during the 2023/2024 seasons. The objectives are to quantify RS temperature biases at Beltsville using Sterling and satellite data as references and implement a transferable correction model for improved upper-air consistency. A Random Forest model is used to identify and predict systematic bias structures, while uncertainty characteristics are evaluated using Gaussian Mixture Model clustering based on Sterling data. The trained model is then applied to independent Beltsville RS datasets to evaluate its generalization performance. Raw RS profiles exhibited substantial mean biases of -1K to 3K and RMSE values of 1.1 K to 3.9 K across all seasons. After correction, systematic biases were substantially reduced, with mean errors constrained between 0.3 K and 1 K, standard deviations ranging from 0.19 K to 1.20 K, and RMSE improvements of 0.56–1.25 K in the mid-to upper troposphere, relative to independent RS41-GDP.1 GRUAN observations. While the current framework demonstrates robust performance at the tested stations and seasons, further evaluation across additional geographic regions, seasons, and satellite instruments is required to fully assess its potential for broader spatiotemporal application and operational integration.