A059-0012
Machine Learning Approach to Classify Precipitation Type from A Passive Microwave Sensor

Wednesday, 9 December 2020
Poster
Spandan Das1, Jie Gong1,2, Chenxi Wang3,4, Dong Liang Wu1,5, Stephen Joseph Munchak6,7 and William S Olson8, (1)NASA Goddard Space Flight Center, Greenbelt, MD, United States, (2)Universities Space Research Association, GESTAR, Columbia, MD, United States, (3)University of Maryland College Park, College Park, MD, United States, (4)NASA GSFC, Greenbelt, MD, United States, (5)NASA/Goddard Space Flight Cent, Greenbelt, MD, United States, (6)NASA Goddard Space Flight Center, Greenbelt, United States, (7)University of Wisconsin Madison, Space Science and Engineering Center, Madison, WI, United States, (8)Joint Center for Earth Systems Technology, Baltimore, MD, United States
Abstract:
Precipitation flag (precipitating or not; stratiform or convective) is a key parameter for us to make better
retrieval of precipitation characteristics as well as to understand the cloud-precipitation physical
processes. The Global Precipitation Measurement (GPM) Core Observatory's Microwave Imager (GMI)
and Dual-Frequency Precipitation Radar (DPR) together provide ample information on global
precipitation characteristics. As an active sensor in particular, DPR provides an accurate precipitation
flag assignment, while passive sensors like GMI were traditionally believed not to be able to tell apart
precipitation types.


Using collocated precipitation flag assignment from DPR as the “truth”, this project employs machine
learning models to train and test the predictability and accuracy of using passive GMI-only observations
together with ancillary atmosphere information from reanalysis. Precipitation types are classified into
the following classes: convective, stratiform, convective-stratiform mixed, no precipitation, and other
precipitation. Sub-sampling with different probabilities is employed to construct a balanced training
dataset. A variety of classification algorithms are tested, including Support Vector Machines, Naive
Bayes, Random Forests, Gradient Boosting, and Neural Networks (Multilayer Perceptron Network), and
their results are evaluated and compared. The trained model has ~ 85% of prediction accuracy for every
type of precipitation. High-frequency channels (166 GHz and 183 GHz channels) and 166 GHz
polarization difference are found among the most important factors that contribute to the model
performance, which shed light on future instrument channel selection.

<< Previous Abstract | Next Abstract >>