Cloud Mesoscale Cellular Classification and Diurnal Cycle Using a Convolutional Neural Network (CNN)

Abstract

Marine stratocumulus (MSC) clouds are important to the climate as they cover vast areas of the ocean’s surface, greatly affecting radiation balance of the Earth. Satellite imagery shows that MSC clouds exhibit different morphologies of closed or open mesoscale cellular convection (MCC) but many limitations still exist in studying MCC dynamics. Here, we present a convolutional neural network algorithm to classify pixel-level closed and open MCC cloud types, trained by either visible or infrared channels from a geostationary SEVIRI satellite to allow, for the first time, their diurnal detection, with a 30 min. temporal resolution. Our probability of detection was 91% and 92% for closed and open MCC, respectively, which is in line with day-only detection schemes. We focused on the South-East Atlantic Ocean during months of biomass burning season, between 2016 and 2018. Our resulting MCC type area coverage, cloud effective radii, and cloud optical depth probability distributions over the research domain compare well with monthly and daily averages from MODIS. We further applied our algorithm on GOES-16 imagery over the South-East Pacific (SEP), another semi-permanent MCC domain, and were able to show good prediction skills, thereby representing the SEP diurnal cycle and the feasibility of our method to be applied globally on different satellite platforms.