Machine learning cloud top height detection based on GNSS radio occultations: a step ahead towards an operational use

The Global Navigation Satellite Systems (GNSS) Radio Occultation (RO) technique allows the sounding of the atmosphere with a vertical resolution of about 100 m in the upper troposphere. It has already been demonstrated that the RO bending angle, by showing clear anomalies at the cloud top heights, is an efficient parameter to highlight the presence of dense clouds in the atmosphere. The objective of this work is to use the bending angle anomaly technique to systematically detect the presence of dense clouds in the atmosphere as well as their altitude and type. Several studies demonstrated the detection efficiency of the bending angle on tropical cyclones, severe convection and volcanic clouds altitude with high accuracy. However, the clouds type differentiation remains a challenge. One of the main issue on this regard, is the lack of volcanic cloud case studies, due to the low number of eruptions in comparisons to the extreme weather events, and to the large uncertainties on volcanic clouds detection techniques.

In this work we collected all the RO collocate in a short time range with tropical cyclones and volcanic clouds, and we collocate them with the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) backscatter. The bending angle anomaly profile is given in input to a machine learning algorithm to retrieve the presence of the cloud and its height. The CALIOP backscatter has 30-meter vertical resolution in the troposphere and 60-meter in the upper troposphere/lower stratosphere. We manually constrain the cloud edges, compute the cloud top height from each cloud and use this value as target for the algorithm output. To get a balanced training of the algorithm, we add to the dataset an equal number of clear sky samples.

The algorithm aims at quickly providing the cloud top height to be used for aviation and nowcast issues and to be included in early warning systems.