Tracking Marine Boundary Layer Cloud Transitions Using Machine Learning

Abstract:

Stratocumulus clouds are considered the air conditioners of the climate system due to their vast coverage and brightness. A relatively small change in their albedo or extent (on order of a few percent) would be sufficient to offset all of the radiative warming due to anthropogenic activities. The mesoscale organization of cloud fields has been shown to be influenced by sea surface temperature, precipitation, cloud-to-surface coupling and possibly aerosols. As part of the Frontier Development Laboratory (FDL) summer project we set out to determine whether changes in aerosol concentrations affect the transitions between cloud regimes (e.g. open cell to closed cell transitions) using unsupervised learning techniques.

A convolutional neural network-based approach using a mixture of experts model has been developed to predict cloud regime transitions over three-day periods during both day- and night-time along several tens of thousands of Lagrangian trajectories off the coast of Namibia. This model utilizes multiple visible and infrared channels from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) as well as cloud property retrievals from NASA. Meteorological data from the European Centre for Medium Range Forecasting (ECMWF) ERA 5 reanalysis product are used to drive the Lagrangian trajectories generated using the Hybrid Single Particle Lagrangian Integrated Trajectory (HYSPLIT) model. Half-hour gridded precipitation retrievals from the Integrated Multi-satellitE Retrievals for GPM (IMERG) product is used to quantify the role of precipitation on the changes in cloud morphology.

We compare our approach to two baseline unsupervised learning methods (Tile2Vec and InfoGAN), all trained using thermal IR imagery from SEVIRI. All methods demonstrate that consistent classification of mesoscale cloud morphology is possible without utilizing visible, daytime only observations. In particular, the mixture of experts model provides semantic segmentation of cloud structure along Lagrangian trajectories, and displays statistically robust differences in the distributions of retrieved cloud properties between different cloud types. An analysis of the cloud regime transitions over multiple days and the relative role of aerosol on these transitions will be discussed.