Forecast Correction Models for Systematic Errors in Different Atmospheric Cloud Regimes.

Given the diversity of cloud forcing mechanisms, numerical weather prediction (NWP) models still face many challenges to forecast all different cloud types through the depth of the troposphere. The Naval Research Laboratory (NRL) is developing forecast correction models ALFS (Advanced Learning Forecast System) to remove systematic errors in atmospheric cloud produced by NWP models. ALFS was built for general application to all NWP models and its statistical corrections aid in the understanding NWP model error trends. The framework implements a Unet-Convolution Neural Network with features extracted from NWP forecast output and satellite cloud observations. The same framework was applied to specific cloud families to increase the correlations between the predictor variables and the predicted clouds. Six cloud types were selected based on their general physical characteristics: 1) Lower tropospheric stable clouds, 2) lower tropospheric unstable clouds, 3) mid-tropospheric clouds, 4) upper tropospheric clouds, 5) deep precipitating clouds, and 6) intense convective clouds.

ALFS was initially applied to merge cloud observations from the Geostationary Operational Environmental Satellite (GOES 16) with output from the Coupled Ocean/Atmosphere Mesoscale Prediction System (COAMPS)^#. A quantitative assessment and evaluation based on an independent data set show that that AFLS corrected geographical cloud coverage errors in COAMPS and improved skill score for all cloud types and forecast lead times (3-12 hours). Building on that success, ALFS has been expanded to ingest feature input from global NWP models like the NAVY Global Environmental Model (NAVGEM) with different predictors and a larger grid size. The latest results also show that ALFS is able to greatly improve the equitable threat scores (ETS) for NAVGEM forecasts of each of the cloud types for the 3 to 12 hour forecasts.

An overview of framework and comparative assessments of results for the two NWP models: COAMPS and NAVGEM will be presented.