APPLICATIONS OF BAYESIAN NEURAL NETWORKS TO GLOBAL PRECIPITATION MEASUREMENT MISSION DATA

Abstract

Meteorological remote sensing efforts have advanced operational decision making and scientific research over the last half-century by providing high-quality global observations of the land, atmosphere, and ocean. The continued development of convolutional neural networks (CNNs) and Bayesian neural networks shows potential for allowing some of these datasets to be synthetically produced where they cannot be directly observed. In this thesis, global precipitation measurement mission (GPM) data is used to train a rain-type classification Bayesian CNN (BCNN) using passive microwave data. Additionally, regression CNNs and BCNNs are trained to predict precipitation using GOES-16 multispectral infrared data over a tropical maritime region. The rain-type classification BCNN shows a 17% improvement in accuracy over existing literature, and the regression models demonstrate a proof of concept in using GPM radar data and geostationary radiances to train skillful CNNs and BCNNs to predict radar reflectivity and rain rate. The experiments demonstrate both the promise of using these data sources to train accurate models and the possible advantages of using BCNNs to quantify and better understand prediction uncertainty for these applications.