A Random Forest–Based Global Gridded Model for Vertical Correction of Precipitable Water Vapor Accounting for Nonlinear Vertical Variation

Abstract

Abstract Vertical correction of precipitable water vapor (PWV) is a critical step in PWV fusion and atmospheric research. To address the limited accuracy of traditional global‐scale PWV vertical correction models that do not explicitly account for nonlinear vertical variations, we developed RF‐PWV, a global 1° × 1° PWV vertical correction model based on random forest. The model was trained using pressure‐level PWV data derived from the fifth‐generation European Center for Medium‐Range Weather Forecasts reanalysis (ERA5) monthly average hourly data from 2008 to 2017. RF‐PWV was compared with two empirical models, EPWV‐H and GPWV‐H. Validation results indicate that: (a) when evaluated using 1‐hr ERA5 PWV profiles for 2018, RF‐PWV achieves a mean bias of 0.00 mm and an root mean square error (RMSE) of 0.75 mm, corresponding to reductions of 95.24% and 32.35% relative to EPWV‐H, and 98.02% and 27.77% relative to GPWV‐H, respectively; (b) when validated against observations from 788 global radiosonde stations in 2018, RF‐PWV exhibits reductions in bias and RMSE of 90.39% and 56.74% relative to EPWV‐H, and 39.30% and 31.10% relative to GPWV‐H, respectively. Analyses across height‐difference intervals and seasons demonstrate that RF‐PWV exhibits superior vertical correction performance within altitude ranges where PWV is primarily concentrated, while maintaining stable accuracy under diverse climatic conditions. The proposed model is independent of meteorological variables; instead, PWV differences for arbitrary height differences at a given grid point can be obtained using only temporal information and elevation differences. This provides a reliable solution for high‐precision global‐scale PWV vertical correction.

The authors' abstract, as published at the source. Journal of Geophysical Research Machine Learning and Computation, 2026 · DOI ↗