Statistical Characterization of Precipitation Events Collected from the GPM DPR

Utilizing precipitation data observed from the GPM Dual-frequency Precipitation Radar over 65N-65S, globally collected precipitation events (PEs) are categorized by their 3D characteristics and connected to their large-scale environments by machine learning methods. From eight years (2014-2022) of accumulation, over 580,000 moderate-to-large scale PEs are collected. Utilizing their 3D characteristics, maximum and mean precipitation intensities, precipitation area, maximum echo-top heights, and stratiform rainfall ratio, PEs are classified into 5 categories; aggregated/small area deep convections, intense/moderate organized systems, and mid-latitude stratiform systems, by the K-means++ cluster analysis. Examining their global distributions, diurnal variations, and histogram characteristics, as well as their environments it is confirmed that these 5 categories represent well-distinguished PE categories. Secondly, we examined a potential to construct a neural network model to statistically distinguish dominant PE categories from the large-scale environmental variables in 5deg x 5deg latitude-longitude grids. Thirty-three values from 9 variables obtained from the ERA5 data are connected to each PE and used for input data to the statistical model. To build the neural network model, 3/4 of paired datasets out of 580,000, are utilized for the study data, and other 1/4 are utilized for the evaluation. After some optimizations to minimize the loss function, a statistical model is successfully constructed to retrieve the dominant PE category in each 5deg x 5deg grid from the environment. It is suggested that an adequate selection of the optimization scheme is essential for the successful construction of the statistical model.