Machine Learning-Based Wet Refractivity Prediction Through GNSS Troposphere Tomography for Ensemble Troposphere Conditions Forecasting

Abstract

This article introduces an innovative ensemble troposphere conditions forecasting method using wet refractivity within the context of Global Navigation Satellite System (GNSS) troposphere tomography. The current models lack coverage of diverse geographical locations and weather conditions, and they do not utilize high-spatial resolution tropospheric data to cover a large area. Moreover, their deterministic prediction mode may introduce high uncertainty into the results. This article leverages long short-term memory (LSTM) networks and genetic algorithms (GAs) to optimize hyperparameters, enabling the prediction of 3-D wet refractivity fields for ensemble forecasting under various weather conditions including rain bands sweeping in Poland and a storm in California, USA. A comparison of the 3-h predictions with Weather Research and Forecasting (WRF) model outputs at levels with a height lower than 3000 m shows root-mean-squared error (RMSE) values of 4.15 and 3.18 ppm for Poland and California, respectively. After utilizing the generative adversarial network (GAN) to produce realistic time series, ensemble forecasting is conducted. The model demonstrates exceptional accuracy in both regions, yielding an optimal threshold of 0.41, which shows a point at which the balance between true positive (TP) and true negative (TN) instances is optimized, achieving a sensitivity of 0.967 and a precision of 0.973 in Poland. Additionally, it achieves an optimal threshold of 0.52, yielding a sensitivity of 0.982 and a precision of 0.993 in California. The low false positive rate (FPR) of 0.027 in Poland and 0.011 in California underscore the adaptability and reliability of the model across diverse datasets.