Improving Radar Remote Sensing Retrievals of Snowfall with an Emphasis on GPM-DPR
More than 50% of surface precipitation globally can be linked to ice processes aloft. Thus, in order to quantitatively describe the hydrologic cycle, the proper measurement and understanding of ice phase microphysics is required. Currently, two spaceborne radars in orbit are capable of measuring snowfall and its properties, CloudSat and the Global Precipitation Measurement mission Dual-frequency Precipitation Radar (GPM-DPR). With CloudSat’s longer record of observations and time for evaluation (est. 2006), CloudSat is currently the most confident method of quantifying the global distribution of snowfall. Comparisons of GPM-DPR to CloudSat have occurred and show that GPM-DPR is underestimating the global mean snowfall accumulation by approximately 50%.
In order to assess a potential cause of this low bias in the GPM-DPR retrieval, the microphysical assumptions are investigated here. The main retrieval assumption is that regardless of precipitation phase, measured precipitation follow an empirical relation between the precipitation rate () and the characteristic size () that was derived from tropical rainfall. Using a myriad of ground-based measurements of snow and rain particle size distributions (PSDs), the assumed relation is assessed. For rainfall, the relation is supported, and shows a good correlation (Pearson ) between and . In snowfall, there is low correlation () between and , suggesting that the relation is inappropriate.
As an alternative method to retrieve snowfall properties from GPM-DPR, machine learning is leveraged on a database of snowfall properties simulated by state-of-the-art scattering models and observed PSDs. Specifically, a neural network is used to retrieve two parameters of the particle size distribution from which other microphysical properties (e.g., ice water content, IWC) can be calculated. An evaluation on three case studies from NASA ground validation field campaigns show that the neural network statistically outperforms old methods at retrieving IWC, with a median percent error of 10% compared to in-situ observations.
As an additional evaluation of the neural network retrieval, a direct comparison to CloudSat is done. Utilizing a pre-made coincident dataset created between the two satellites, the neural network algorithm shows the best agreement to the GPM-DPR rainfall retrieval below the melting layer, suggesting the best performance. Meanwhile, CloudSat shows a decrease of about 100% in mean snowfall rate from C to C. A case study from NASA’s recent IMPACTS field campaign highlights the potential issues that CloudSat’s single frequency snowfall retrieval cannot overcome, suggesting that in moderate to heavy snowfall, GPM-DPR would likely perform better than CloudSat.
