Machine Learning Techniques to Assign Observation Error for Airborne Radio Occultation Profiles in Atmospheric Rivers

Airborne radio occultation (ARO) is a remote sensing method that measures the properties of the atmosphere through the refractive delay and bending of Global Navigation Satellite System (GNSS) signals from satellites near the horizon as observed by a receiver on an aircraft. These measurements are used to recover thermodynamic profiles. ARO is used in reconnaissance campaigns to observe Atmospheric Rivers (ARs), which are narrow filaments of water vapor transport in the lower troposphere, that drive much of the precipitation on the US West Coast. To utilize the observations in Numerical Weather Prediction (NWP), an estimate of the accuracy is necessary. We hypothesize the observation error of an ARO profile can be impacted by the atmospheric conditions sampled by the raypath of the GNSS signal due to a variety of factors, including sharp gradients in refractivity that may induce superrefraction. ARO profiles have been validated in comparisons against collocated dropsonde observations and reanalysis products within the AR environment. In refractivity, ARO remains unbiased with respect to the European Centre for Medium-range Weather Forecasting Reanalysis-5 (ERA5) above 5 km altitude, with a standard deviation around 1%. Standard deviation increases to ~4% near the surface, and bias increases as well. However, this estimate for the entire AR environment does not consider the different canonical features of the AR and their associated effects on observation error. Currently, ARO data are assimilated using an observation error estimate that is a function of latitude and season. No method has yet been developed to characterize and assign observation errors dynamically for individual ARO profiles specifically within ARs. We present several potential observation error assignment methods that use machine learning clustering techniques to divide ARO profiles into unique clusters representative of different atmospheric regimes. We show that ARO profiles can be separated into clusters, and calculate the error characteristics within each cluster. The significance of the differences in error among different clusters is assessed. This represents an important step towards more realistic observation error estimates for data assimilation in NWP. Improvements in observation error estimates will improve the robustness of forecasting for AR events when ARO is assimilated. Increased forecast skill will allow stakeholders to make better-informed water management and disaster preparedness decisions in an increasingly arid and unpredictable climate in the western US.