Remotely profiling the tropospheric absolute humidity using a single ground-based GNSS meteorological station based on an artificial neural network

Abstract

Near-real-time profiling of the vertical distribution of tropospheric water vapor is important for a wide range of weather monitoring and forecasting applications. In this study, a GNSS-based Profiling Tropospheric Humidity Network (GPTHNet) was developed to retrieve tropospheric absolute humidity using observations from a single ground-based GNSS meteorological station. The input variables of GPTHNet include observation time, precipitable water vapor (PWV), ground-level meteorological parameters from the GNSS meteorological station, and the nearest radiosonde humidity profile prior to the observation epoch. Historical radiosonde observations are used to provide initialization information for GPTHNet. The training data for the artificial neural network (ANN) models were mainly obtained from the Atmospheric Radiation Measurement (ARM) User Facility and covered the period from 2002 to 2011. Absolute humidity was estimated at 40 height levels ranging from 250 m to 10,000 m, with a vertical interval of 250 m, using two ANN profiling models. The performance of the two ANN profiling systems was evaluated by comparing their predictions with independent humidity profiles obtained from radiosondes and Raman lidar. The ANN_Init generally outperformed the ANN_NoInit, particularly in the middle and upper troposphere, where the relative error of the ANN_Init was noticeably smaller than that of the ANN_NoInit. For the independent GNSS–radiosonde-combined validation, the RMSE of the ANN_Init reached a maximum of 1.90 g/m 3 at 500 m and decreased to below 0.1 g/m 3 at the highest levels; against the Raman lidar observations, the relative RMSE increased from about 12% at the lowest level to about 75% at the upper levels during the monsoon season. The results indicate that the accuracy of the ANN profiling algorithm is comparable to that of tomographic techniques.