Improving Active Remote Sensing of Snow through the Use of Multiple Frequencies, In Situ Data, and Neural Networks

The observed spatial distribution of ice water path (IWP) and near surface snowfall (S) remain a priority for constraining global circulation models and quantifying the global water cycle. Current estimates of IWP and S from the operational algorithms of CloudSat (2C-SNOW-PROFILE) and the Global Precipitation Measurement mission Dual-Frequency Precipitation Radar (GPM-DPR) have large discrepancies. To quantify and reduce uncertainty in the true spatial distribution of IWP and S, a new multi-frequency retrieval is constructed from an amalgamation of several ice scattering property databases, in-situ measurements and implemented using deep neural networks. The new retrieval of ice water content (IWC) derived for the same frequencies of the GPM-DPR compares well with independent in-situ observations from two NASA ground validation campaigns, GPM Cold Season Experiment (GCPEX) and the Olympic Mountains Experiment (OLYMPEX), and outperforms the single-frequency empirical power law fits by 8-10%. Single-frequency power laws from the literature estimate a larger IWC for the same reflectivity (Z) when compared to the new retrieval, but are partially explained by the choice of particle forward model and the choice of particle size distribution (PSD). The GPM-DPR derived IWC-Z relationship shows a large low bias, underestimating IWC by a factor of 7 for Ku-band and 11 for Ka-band. Further comparisons between the new retrieval and other published multi-frequency retrievals using Bayesian statistics and optimal estimation are also investigated.