Process-Based Cloud Cover Verification for Improved Understanding of Physical and Statistical Forecasts

The Naval Research Laboratory (NRL) is developing a machine learning-based cloud nowcasting system consisting of merged data from GOES 16 and the Coupled Ocean/Atmosphere Mesoscale Prediction System (COAMPS^®#). GOES 16 predictors and ground truth observations consist of reflectance and brightness temperature as well as retrievals of cloud top height (CTH), total condensed water path (CWP), cloud top phase, and cloud type. COAMPS predictors are drawn from forecasts of microphysical, thermodynamic, kinematic, variables as well as synthetic brightness temperature forecasts at 5 km grid spacing.

Brightness temperature prediction proved to be difficult due to uncertainties in cloud location. However, the machine learning algorithms were more adept at forecasting cloud probability. The performance accuracy was noted to be dependent on cloud type. For that reason separate machine learning algorithms are being developed for specific sets of atmospheric processes. Deep convection, shallow convection, advection, stable stratocumulus, and synoptic lift have been initially identified. In an effort to improve the transparency of the machine learning predictions, the COAMPS forecast errors have been characterized for each set of processes. Deterministic errors are often quite large, especially for poorly resolved convective clouds, while deeper, more strongly forced clouds are better predicted. This information will help determine the positive effects of including other predictors in the machine learning algorithms such as atmospheric stability, surface heating rates and turbulent kinetic energy.

^#COAMPS^® is a registered trademark of the Naval Research Laboratory.