MACHINE LEARNING ENHANCED SNOWFALL RATE RETRIEVALS FROM PASSIVE MICROWAVE SENSORS

Y. Fan^1,2, Y. Hong¹, J. Dong^1,2, H. Meng², C. Kongoli¹, Y. You³, R. Ferraro⁴, T. Ren⁵, P. Yang⁵

¹CISESS, University of Maryland, ²NOAA/NESDIS/STAR, ³Department of Earth and Ocean Sciences, University of North Carolina Wilmington, ⁴ESSIC, University of Maryland, ⁵Department of Atmospheric Sciences, Texas A&M University

Snowfall accounts for a large fraction of winter precipitation in mid- and high latitudes. The NOAA/NESDIS snowfall rate (SFR) dataset is a global, liquid equivalent snowfall rate estimation. It is derived from observations taken by passive microwave radiometers, including ATMS, MHS, GMI and SSMIS sensors, aboard nine polar-orbiting satellites. The SFR algorithm consists of two main components: snowfall detection (SD) and rate estimation. Recently, the SD component has been updated to machine learning (ML) models, which was trained against a large collection of in-situ observations. The XGBoost ML model shows the best performance, hence it’s implemented in the SFR operational retrieval system. The core of the second component, rate estimation, is a 1DVAR-based inversion model. It has been improved by using ML-based initial conditions. Lastly, the retrieved SFR values are further biased-corrected with a neural network (NN) model. The ML SD and SFR algorithms have been successfully applied to cross-scan instruments such as ATMS and MHS sensors. We will present the development of the ML SD and SFR algorithms for the conical scanning GMI and SSMIS sensors and their global validation results. Case study compared with Stage IV radar and gauge combined precipitation rate, CPR near surface snowfall rate, NOHRSC hourly snowfall analysis and ERA5 hourly snowfall rate reanalysis will also be provided. Another notable improvement to the snowfall rate estimation is the development of a ML multi-ice-habit algorithm. Our studies show that the snowfall rate derived from space-borne passive microwave (PMW) observations is highly sensitive to ice habit assumptions in the retrieval algorithm and the currently used sphere-like particles tend to overestimate SFR. We have selected several non-spherical ice/snow particles that perform well globally and developed a ML multi-ice-habit algorithm to predict the probability of each ice habit to optimize the snowfall rate estimation. This multi-ice-habit approach is superior to the single-ice habit method and achieves an overall improvement of approximately 10% (up to 40% for deep cloud cases) in statistical metrics.