Using GOES Brightness Temperatures to Assess the Accuracy of Short-Range Forecasts from the High-Resolution Rapid Refresh (HRRR) Model

Infrared brightness temperatures from geostationary satellite sensors are very valuable for numerical weather prediction model validation because they provide detailed information about the cloud and water vapor distributions over large geographic domains with high spatial and temporal resolution not available with other observing systems. To promote the routine use of these observations for this purpose, we have developed a near real-time verification system for the U.S. High Resolution Rapid Refresh (HRRR) model that uses observations from the GOES imager. The satellite-based verification system can be used by forecasters to rank the accuracy of current and prior HRRR model forecasts, which are produced on an hourly basis across the contiguous U.S. with 3-km horizontal resolution. The ability to quickly identify which of the many overlapping HRRR forecast cycles is most accurate at the current time is an important feature of the system that improves the efficiency of the forecaster model evaluation process.

To support the real-time verification system, we are also using the simulated brightness temperatures to assess the HRRR forecast accuracy over longer time periods. Simulated GOES infrared brightness temperatures sensitive to clouds and upper-level water vapor are generated during each forecast cycle using the Community Radiative Transfer Model (CRTM) and are then compared to real GOES observations using a variety of statistical methods to assess the model accuracy. In addition to conventional statistics, such as root mean square error, we are also using neighborhood and object-based verification methods such as the Fractions Skill Score and the Method for Object-based Diagnostic Evaluation (MODE) that are less sensitive to spatial displacement errors and can also provide useful information about cloud object attributes (e.g. orientation, size, etc.). In this presentation, we will discuss results from our ongoing model validation efforts using these analysis tools, with the primary emphasis placed on assessing the accuracy of the cloud field. The results show that the HRRR model is often deficient in upper level clouds at the start of a given forecast cycle; however, it then contains too many upper level clouds after a few hours, presumably because the model spin-up process is too vigorous. Results using the Fractions Skill Score for different modes of convection show that skillful forecasts can be obtained for model grid lengths greater than 50 km in most situations, with the greatest skill obtained for strongly forced convection. The MODE object-based verification tool is being used to assess the accuracy of the cloud objects containing the coldest brightness temperatures.