Using machine learning to improve multi-wavelength spaceborne radar precipitation retrievals

[14-Dec-2022]
Abstract: We employ a new physically-based machine learning retrieval approach at Ku- and Ka-band frequencies using neural networks in ice and liquid phase retrievals. The retrieval is trained and tested with inputs of observed particle size distributions from aircraft microphysical probes collected during various NASA field campaigns (OLYMPEX, GCPEX, MC3E, CAMP2EX) and state-of-the-art forward simulations of hydrometeor scattering properties at multiple frequencies. In the ice phase retrievals, attenuation is ignored, and the retrieval provides outputs of normalized number concentration, ice water content, and melted and unmelted mass-weighted mean diameter. In the liquid phase, path integrated attenuation is used to constrain the retrieval in a column-iterative approach; the number concentration, liquid water content, mass-weighted mean diameter, and Gamma particle size distribution shape parameter are retrieved. Evaluation of these retrievals is performed against a separate validation set of airborne microphysical data directly and statistically, as well as from independent data from ground-based sensors (scanning radars, ground-based in situ measurements) depending on the field campaign, and shows decreases in bias and random error from current spaceborne radar approaches. These retrievals are used evaluate operational retrievals from the NASA Global Precipitation Measurement and CloudSat missions. I will discuss using new techniques for physical understanding of the retrievals, as well as other potential machine learning approaches for spaceborne precipitation retrieval, and has implications for the upcoming NASA Atmosphere Observing System multi-frequency radars.