Comparison of different deep learning architectures for the retrieval of thermodynamic profiles from GNSS-RO

Global Navigation Satellite Systems (GNSS) Radio Occultation (RO) is one of the most promising remote sensing techniques for global atmospheric sounding. RO is a limb-sounding technique that uses GNSS signals, refracted during their propagation through the Earth’s atmosphere to a receiver on a low-Earth orbit (LEO) satellite. RO data have been proven to be of enormous value for data assimilation in numerical weather prediction (NWP) as well as in climate science over the two last decades. However, retrieving products such as temperature or humidity from RO observations is not straightforward and dedicated retrieval algorithms still have limitations, such as the need for external meteorological data. On the other hand, various new RO missions are now producing over 10,000 globally distributed profiles daily. This makes the technique interesting for the application of Artificial Intelligence (AI) models to different steps of the RO retrieval chain.

This study compares an existing retrieval method entitled AROMA (Advancing the GNSS-RO retrieval of atmospheric profiles using MAchine-learning), which is based on a multi-layer perceptron (MLP), with more sophisticated deep learning (DL) architectures such as Transformers and one-dimensional convolutional neural networks (1D-CNNs). All these models are trained on multiple years of data from different RO missions, using vertical profiles of bending angle and other RO parameters as input features and operational results from a standard retrieval algorithm as targets. Validation results using both a separate test data set as well as external data will be presented, aiming to give a recommendation on the most promising type of architecture to use for the RO wet retrieval problem.