Synergizing GNSS, MODIS, and ERA5 for high-resolution PWV retrieval: a two-stage machine learning approach over Hong Kong

Abstract

Atmospheric water vapor is a fundamental driver of the Earth’s energy balance and climate system, making precipitable water vapor (PWV) a key indicator for meteorological monitoring and extreme weather forecasting. This study introduces an innovative two-stage machine learning framework, named the Decision Tree and Random Forest PWV Fusion (DTRP) framework, to generate a high-precision, spatiotemporally continuous PWV product over Hong Kong by fusing multi-source data, including GNSS, Moderate Resolution Imaging Spectroradiometer (MODIS), and ERA5. GNSS PWV derived from 18 Continuously Operating Reference Stations (CORS) served as the benchmark. In the initial stage of the DTRP framework, a regression tree model is employed to correct systematic deviations in MODIS Near-Infrared (NIR) PWV. The model exploits the differences between ERA5 and MYD NIR PWV, together with spatiotemporal features, to estimate GNSS–MYD deviations, thereby yielding an enhanced PWV product (EMYD). Building on this, the second stage of DTRP applies a random forest framework to integrate EMYD with original ERA5 PWV along with meteorological covariates, generating the final fused product (RF PWV). Statistical evaluation confirms the initial correction successfully eliminated the systematic negative bias in MYD NIR PWV. The RMSE decreased to 3.45 mm with an R² of 0.95, corresponding to improvements of 55.8% in accuracy and 0.17 in explanatory power over the original data. The subsequent fusion stage achieved a further refined RMSE of 3.29 mm, outperforming the MYD, ERA5, and EMYD products by 57.9%, 30.4%, and 4.6%. The final RF PWV product demonstrates nearly unbiased estimation capability and maintains strong agreement with reference data, achieving an R² of 0.96. Spatiotemporal and correlation analyses confirmed the superior consistency and reliability of the RF PWV under diverse conditions and its effectiveness in decoupling errors from geographic and atmospheric influences. The proposed DTRP algorithm presents a valuable framework for predicting severe convective weather.