Evaluating the Accuracy of the High Resolution Rapid Refresh (HRRR) Model Using Satellite Infrared Brightness Temperatures

Location: FL2-1001
Jason Otkin, CIMSS/SSEC, University of Wisconsin-Madison

Infrared sensors onboard geostationary satellites provide detailed information about cloud top properties and the water vapor distribution with high spatial and temporal resolutions that make them very useful as a numerical weather prediction model validation tool.  To promote the routine use of these observations for this purpose, we developed a near real-time GOES-based verification system for the High Resolution Rapid Refresh (HRRR) model that provides operational forecasters objective tools to determine the accuracy of current and prior HRRR model forecasts when they are creating or updating short-range forecasts.  This capability has become increasingly more important in recent years due to the implementation of rapidly updating numerical models with many overlapping forecast cycles.  Besides serving as a useful forecaster model evaluation tool, long-term statistics accumulated using this system also provide an excellent means to assess the accuracy of the cloud and water vapor fields in the HRRR model forecasts.

For this presentation, we will describe the capabilities of the near real-time verification system and present results from several ongoing model validation projects.  Synthetic GOES 10.7 m infrared brightness temperatures are generated for each HRRR forecast cycle using the Community Radiative Transfer Model (CRTM) and are then compared to real GOES observations using various statistical methods to assess the model accuracy at each model forecast time.  These methods include dimensioned metrics such as root mean square error and bias, neighborhood-based metrics such as the Fractions Skill Score, and object-based verification tools using the Method for Object-Based Diagnostic Evaluation (MODE) system.  The model accuracy was assessed for two one-month periods during August 2015 and January 2016.  Overall, the results show that the simulated brightness temperatures are often too warm during the first hour of the forecast, indicating that the HRRR model initialization is deficient in upper-level clouds.  This warm bias, however, is quickly replaced by a large cold bias due to the rapid generation of upper level clouds, with the negative bias often lasting for many hours into the forecast before the excessive cloud cover dissipates.  Detailed analysis of the MODE results showed that the HRRR initialization contains too many small cloud objects, especially during August; however, the number of cloud objects becomes too low by forecast hour 2.  This behavior is consistent with the changes in the brightness temperature bias and indicates that the simulated cloud objects become too large after a few hours.