Mapping RO observations using different ML algorithms

Abstract

Through Global Navigation Satellite Systems (GNSS) Radio occultation (RO), the bending of radio signals due to their refractions while passing through earth’s atmosphere can be measured using Low Earth Orbit satellites. The refractivities determined with such measurements are crucial in the creation of Numerical Weather Models (NWM), which are used to display the atmosphere’s temperature, pressure, and humidity. Due to the sparse and non-uniform distribution of RO observations, this data cannot resolve synoptic variabilities, and improvements in horizontal resolution are necessary. This thesis compares the abilities of various Machine Learning algorithms - namely, Random Forest Regression (RFR), Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN) - to predict and map RO refractivity. Data from COSMIC-2 and ECMWF forecasts interpolated at 2 km height are used to train and test the models. Using only the COSMIC data, the RFR achieved the most accurate prediction with a posterior standard deviation of 8.72 N-units. Incorporating the ECMWF forecasts the CNN model was able to predict the refractivity with a standard deviation of 5.64 N-units. Additionally, the thesis assessed the impact of various factors such as different forecast times, temporal splitting of the training data, and the use of the refractivity residuals for training the models.