Data-driven improvements to GPROF-based satellite snowfall retrievals with a focus on mountain snowfall

Abstract

Snowfall is a critical component of Earth's hydrological and climate system despite only 5% of Earth's annual precipitation falling as snow. Satellite-based snowfall estimates, particularly those obtained from the Global Precipitation Measurement (GPM) Microwave Imager (GMI), struggle to accurately estimate the total annual snowfall accumulations, especially in mountainous regions of the world. Part of the challenge is due to the reference precipitation used in the GMI-based algorithms, while radiometers struggle to distinguish between the microwave signatures of surface snowpack and snowfall. The aim of this dissertation is to evaluate the impact machine learning-based GMI retrievals have on snowfall estimates, explore how temperature and climatological adjustments to the reference precipitation can provide additional information to the retrieval, and asses if these changes lead to improved snowfall accumulations required for modeling the lifecycle of snow. A key objective of this study is to improve snowfall accumulation estimates in mountainous areas, where snowpack is a critical component of water storage. First, snowfall rates estimated from the Goddard Profiling Algorithm (GPROF) for GMI are compared using three types of GPROF algorithms: one Bayesian (GPROF V7) and two neural network versions (GPROF-NN 1D and GPROF-NN 3D). The highest detection and quantitative statistics are observed using GPROF-NN 3D with both neural network retrieval algorithms outperforming the Bayesian version. It is shown that artificial biases in the retrieval statistics can result from the selected threshold for snow/no-snow classification. Coincident in-situ snowfall and radar data are also used to evaluate the temperature dependency of the reflectivity-snowfall (Z-S) relationship and how it impacts the GPROF retrievals. Second, an evaluation of the three GPROF algorithms is conducted in the mountains of the western United States. Using data from a snow reanalysis dataset, water year snowfall accumulations from the Multi-Radar Multi-Sensor (MRMS) are adjusted to produce more realistic snowfall magnitudes and spatial patterns. These adjustments were found to decrease errors in snowfall accumulation estimates for all three retrieval algorithms, resulting in significant improvements when compared to independent SNOTEL observations. These results provide a positive outlook for snowfall retrievals in mountainous regions by incorporating additional information to the retrieval algorithm. Finally, a framework for incorporating satellite precipitation estimates into a snow evolution model in the western United States is presented that offers a flexible design to account for different study domains. The objective of this framework is to present an approach for deriving snow water equivalent (SWE) from satellite precipitation estimates given the difficulties of directly measuring SWE from passive microwave sensors. A UNet-based retrieval model is used to estimate precipitation at 30 minute time resolution across the currently available passive microwave and infrared sensors. The initial precipitation estimates were found to have a systematic bias across the study period, which, after correction, produced realistic spatial patterns of snow depth and snow water equivalent, but underestimated the magnitudes compared to two reference snow model simulations.