Comparing datasets and methods to identify major severe weather outbreaks

Two techniques are proposed to distinguish major severe weather outbreaks from less significant events. The first uses gridded fields of meteorological parameters valid at the times of the outbreaks. A domain positioned at the center of the outbreak is defined, and a principal component analysis is conducted for each of the outbreaks in a training set. The matrix of principal component scores for each outbreak is used as input data for a number of statistical algorithms. The algorithms use the training data to develop statistical models to be tested on independent cases. The second technique simply calculates the number of grid points that exceed a specified threshold for various severe weather parameters. The areal coverage of the training cases is used to develop statistical models to be tested on the same set of independent cases.

The techniques are compared in two ways. First, the variability of the training models is compared by developing a number of subsets of the training data to develop multiple statistical models. The second comparison involves bootstrapping contingency statistics of the test cases, for comparison of the average and median statistics and confidence intervals to determine if one method is statistically significantly better than the alternative approach. Preliminary results indicate that the areal coverage and principal component analysis techniques perform similarly, suggesting the simpler, computationally less demanding, and easier to interpret areal coverage technique may be preferred for distinguishing major severe weather outbreaks from less significant events.

Additionally, the National Centers for Environmental Prediction (NCEP)/National Center for Atmospheric Research (NCAR) reanalysis (NNRP) dataset and the North American Regional Reanalysis (NARR) datasets are compared using the principal component analysis technique. The purpose is to determine if a higher-resolution dataset (such as the NARR) provides any additional information to distinguish outbreak types at their valid times. Preliminary findings suggest that reanalysis datasets with smaller grid spacing do not necessarily provide additional useful information. Reasons for these findings and implications on future studies will be discussed.