A machine learning framework to map Atmospheric Rivers and Planetary Boundary Layer height from GNSS radio occultation observations

We use machine learning (ML) to map column-integrated water vapor (IWV) to characterize atmospheric rivers (ARs) and to map planetary boundary layer (PBL) height given Global Navigation Satellite Systems (GNSS) radio occultation (RO) data. GNSS RO of the Earth’s atmosphere obtains vertical profiles of microwave refractivity with vertical resolution approaching 100 meters. RO is effectively a water vapor sounder in the lower troposphere and is especially sensitive to vertical gradients associated with the top of the PBL. RO soundings undersample synoptic variability of the atmosphere with severe nonuniformity in spatial and solar angle distribution, making the creation of atmospheric model-agnostic level 3 climatologies a complicated task. ML methods have already shown great promise for mapping RO retrieved quantities in the horizontal and time. We present two applications of ML to RO data, the first to characterize ARs, and the second to map PBL height.

In a mission architecture trade study, we use an ML approach to determine what type of small-satellite constellation would be appropriate to map ARs with detail sufficient for atmospheric process studies and for the prediction of the severe weather on the U.S. Pacific coast that results from ARs. Because ARs are high-volume flows of water vapor in filaments within the PBL and RO sounds water vapor, RO data are ideally suited as input to ML algorithms for the study of ARs. How many low-Earth orbiting RO sounders would be needed to gain desired information on ARs remains an open question, as does our ability to map ARs with existing program-of-record RO data. We answer these questions by formulating ML algorithms to map ARs in the North Pacific Ocean from simulated and real RO data. Simulated RO sounding geolocations are defined by various sizes and types of Walker constellations, realistic GNSS orbits, and the interpolation of refractivity profiles from the ECMWF operational forecast system. We develop two neural networks, one to convert refractivity soundings between 0 and 10 km to IWV and another to map IWV in the horizontal in 1-hr time windows. We find that optimal performance is obtained with Walker constellations of 36 or more RO satellites in near-polar orbits, appropriate orbits for temporal uniformity of the RO’s sampling density. The advent of GNSS RO satellites in micro-satellite and nano-satellite form factors makes such constellations feasible and affordable in the very near future.

We then use ML to map PBL height. The PBL is the part of the atmosphere closest to the Earth’s surface. In the PBL, turbulent processes often affect the vertical redistribution of heat and moisture and their exchange influences cloud evolution and large- to meso-scale circulation. The PBL height can be used to describe climatological processes in a specific region, including cloud characterization. The high vertical resolution of RO observations is suitable to model PBL height in individual profiles. Initially, we use the changes in refractivity to model the PBL height at RO locations and times. Then, we apply ML to produce global PBL heights with a high temporal resolution.