Machine learning-based estimation of hourly GNSS precipitable water vapour

Abstract

Water vapour plays a key role in long-term climate studies and short-term weather forecasting. Therefore, to understand atmospheric variations, it is crucial to observe water vapour and its spatial distribution. In the current era, Global Navigation Satellite Systems (GNSS) are widely used to monitor this critical atmospheric component because GNSS signals pass through the atmosphere, allowing us to estimate water vapour at various locations and times. The amount of precipitable water vapour (PWV) is one of the most fascinating quantities, which provides meteorologists and climate scientists with valuable information. However, calculating PWV accurately from processing GNSS observations usually requires the input of further observed meteorological parameters with adequate quality and latency. To bypass this problem, hourly PWVs without meteorological parameters are computed using the Random Forest and Artificial Neural Network algorithms in this research. The first step towards this objective is establishing a regional weighted mean temperature model for Austria. To achieve this, measurements of radiosondes launched from different locations in Austria are employed. The results indicate that Random Forest is the most accurate method compared to regression (linear and polynomial), Artificial Neural Network, and empirical methods. PWV models are then developed using data from 39 GNSS stations that cover Austria’s entire territory. The models are afterwards tested under different atmospheric conditions with four radiosonde stations. Based on the obtained results, the Artificial Neural Network model with a single hidden layer slightly outperforms other investigated models, with only a 5% difference in mean absolute error. As a result, the hourly PWV can be estimated without relying on measured meteorological parameters with an average mean absolute error of less than 2.5 mm in Austria.