A Neural Network-Based Method for Time Series Modeling of 3-D Atmospheric Refractivity Using Radio Occultation Measurements

Abstract

Satellite measurements of the atmosphere by remote sensing techniques are becoming the trend for atmospheric probing. The problem is that such measurements are usually not continuous in time series for given Earth locations. This article presents a neural network-based method for time series modeling of the atmospheric refractivity in 3-D space using radio occultation measurements from the COSMIC mission. The method offers an opportunity to obtain continuous refractivity values in time series for any given Earth location, even if there are no measurements. Time series inputs of year, day of year, and hour of day are used to train the 3-D space measurements of refractivity from both COSMIC-1 and COSMIC-2 missions. The data cover periods from April 2006 to September 2020, and the data for the Nigerian region (and from altitudes 0.1 to 39.9 km) are illustrated. The effectiveness of solar activity indicator [sunspot number (SSN)], as an input for the neural network process, is investigated. The result shows that the SSN was an insignificant input for improving the prediction accuracy of the model. A comparison of the model predictions and the COSMIC measurements shows that there is very good correlation (correlation coefficients are approximately 1.0) between the model predictions and the COSMIC measurements. Typical root-mean-square errors between the model predictions and the COSMIC measurements are about 7.3 N-units on ground, 0.6 N-units at 20 km, and 0.3 N-units at 40 km. The model predictions are demonstrated to vary (in time and space) in patterns that agree with physical measurements of refractivity.