Assimilation of POES Radiance Observations and NCEP Conventional Observations in GSI for Tornado Outbreak Prediction

The use of numerical weather prediction (NWP) has brought significant improvements to the forecasting of severe weather outbreaks. Tornado outbreaks can now be predicted up to four to eight days in advance; nonetheless, the precise location and timing of a tornado outbreak are often missed in forecasts prior to the onset of the outbreak. Assimilation of atmospheric and oceanic data has been performed for several years as a means of improving NWP forecasts. In particular, it has been shown that WRF-ARW forecasts of severe weather outbreaks are quantitatively improved by the assimilation of conventional wind, dewpoint, pressure, and temperature observations as compared to model forecasts with no data assimilation performed.

This research examines the effect of different types of assimilated data on tornado outbreak forecast accuracy. Infrared radiance observations from the polar-orbiting Advanced Microwave Sounding Unit-A (AMSU-A) and High Resolution Infrared Radiation Sounder-4 (HIRS-4) were assimilated, as well as conventional surface and upper-air observations. A 3DVAR configuration of the Gridpoint Statistical Interpolation (GSI) data assimilation software was used, and all possible combinations of the radiance and conventional observations were assimilated against a background field. Twenty-one synoptically driven United States tornado outbreaks from 2007 to 2011 were then simulated at 4 km in the WRF-ARW model. These model runs each encompassed 42 hours, from 1800Z on the day before the outbreak day to 1200Z on the day after. Each of the twenty-one cases contained an eight-member ensemble consisting of the seven distinct combinations of assimilated data and a single control member with no data assimilated. Ten synoptically driven United States hail and wind severe events, none of which qualified as tornado outbreaks, were also simulated in the same manner.

A learning machine technique known as a support vector machine (SVM) was used to predict the probability of a tornado outbreak for each ensemble member. This algorithm classifies data into binary categories based upon characteristics of predictor values given as input. Meteorological parameters determined to be statistically significant in association with tornadoes were extracted from the WRF-ARW simulations at 0000Z on the day after the outbreak, or the 30th hour of the WRF forecast. This was a time in which severe weather was usually ongoing in a given outbreak. The covariates were input to the SVM, which then issued a yes/no prediction of a tornado outbreak for each ensemble member. Contingency statistics were calculated for each of the ensemble member types over the 21 tornado outbreaks and 10 non-tornadic outbreaks.

The results indicate that assimilating any one type of data, whether infrared radiance observations or conventional surface and upper-air observations, improves tornado outbreak forecasts above a forecast without data assimilation. However, assimilating both conventional and satellite observations provides the most accurate outcome for prediction of a tornado outbreak. When both sources of satellite data were assimilated in conjunction with conventional observations, the SVM failed to predict only one of the 21 tornado outbreaks.

Given these results, further research is planned for quantifying the forecast value of different sources of assimilated data. Local National Weather Service offices, many of which now run NWP models for their own forecast regions at high resolutions, could benefit from statistics regarding the forecasting usefulness of assimilating different sources of data for different types of weather phenomena.