Optimized Approach for Near-Real-Time Three-Dimensional Water Vapor in the GNSS Based on the Informer Model

Abstract

The integration of near-real-time three-dimensional (3D) water vapor data into numerical weather prediction is crucial for monitoring and forecasting extreme weather events but faces various challenges. We focus on reconstructing the 3D water vapor field using Global Navigation Satellite System (GNSS) water vapor tomography techniques, emphasizing two primary concerns: achieving high-precision initial 3D water vapor values and effectively partitioning the vertical tomography grid. We introduce a novel real-time, high-precision water vapor prediction model, namely, the Informer-WV model, based on the Informer framework, whose predictions serve as the initial values for tomography. Furthermore, we propose an innovative method for nonuniform vertical delineation of the tomography grid in which the upper boundary height of the 3D tomography grid is determined by the vertical prediction accuracy of the model. For practical application purposes, Hong Kong, China, was chosen as the study area. The Informer-WV model, utilizing ERA5 reanalysis data, successfully predicted the regional water vapor density for 2022. The model demonstrated a remarkable prediction accuracy, with an annual root mean square error (RMSE) better than 0.80 g/m³ compared to the actual ERA5 values. Building on this high-precision prediction, we adjusted the upper boundary altitude of the tomography grid to 5.2 km, specifically for Hong Kong. By benchmarking against radiosonde-derived water vapor density data, we analyzed the near-real-time tomography inversion results for the two weakest prediction periods of the model. The RMSE of the water vapor inversion values derived from our optimized method was reduced to 1.26 g/m³. This approach not only improved the accuracy by 19% relative to the initial predictions but also significantly outperformed the traditional tomography method.