Emulating Tropical Marine Radar Reflectivity by Applying a Convolutional Neural Network to Geostationary Satellite Radiances

Abstract:

Ground-based radar, while used widely for characterizing structures of precipitating systems and estimating rainfall, is constrained by its limited spatial coverage when compared to satellite-based systems. Data coverage is especially sparse over oceans where few radars exist. Artificial Intelligence (AI) applied to geostationary satellite data can be utilized to overcome this limitation by using geostationary radiances in multiple spectral bands to emulate passive microwave or radar data over large areas at high spatial and temporal resolution. Existing literature demonstrates the capability of convolutional neural networks (CNNs) to emulate WSR-88D composite radar reflectivity over the continental U.S using radiances from Geostationary Operational Environmental Satellites (GOES). However, any relationship between geostationary radiances may be highly dependent on the spatial distribution of the radiance as well as properties of the satellite data (e.g. spectral, spatial resolution) itself. This is the type of problem at which CNNs excel. In this research, various CNNs were trained using two infrared METEOSAT satellite bands and data from a high-quality tropical marine radar dataset captured during the DYNAMO field campaign over the Indian Ocean in 2011. We will demonstrate the ability of different CNN architectures to accurately emulate the radar-observed structure of cloud systems using a small field campaign dataset as a first step toward assessing the use of CNNs on larger radar datasets elsewhere in the tropics that may require more quality control before use. A successful model can then be used to supplement ground-based radar and polar orbiting active sensors over the largely data-sparse tropical oceans.