We are addressing the probabilistic QPE challenge in two stages. In Stage I we identify precipitation types from ABI sensor onboard GOES-16 using insights from GV-MRMS. In Stage II  probabilistic QPE modeling will be applied using these insights.  

I. Classifying precipitation: Prognostic and Diagnostic Modeling

A machine learning-based classification model is developed by deriving a comprehensive set of features using five ABI channels and numerical weather prediction observations. ​The developed prognostic model shows promising results in identifying the occurrence/non-occurrence of precipitation as well as precipitation processes from ground radars  with overall accuracy of around 75%. It is suggested to utilize probabilities instead of deterministically separating precipitation types especially in regions with uncertain classifications. Ultimately, this effort will aid towards improved precipitation characterization and retrievals from space.

Peering into the model shows satellite observations are important in separating Rain and No-Rain areas. For stratiform precipitation types, predictors related to atmospheric moisture content, such as relative humidity and precipitable water, are the most important predictors, while for convective types, predictors such as 850-500hPa lapse-rate and Convective Available Potential Energy (CAPE) are more important. The diagnostic analysis confirms the benefit of  spatial textures derived from ABI observations to improve the classification accuracy. It is recommended to combine  the heritage water vapor channel T6.2 with the IR T11.2 channel for improved precipitation classification. There is more than 10% improvement in detection of stratiform and tropical precipitation types compared to using T11.2 alone.