Analysis of Mesocyclone Detection Algorithm Variables

The Mesocyclone Detection Algorithm (MDA) is used in the Weather Surveillance Radar –1988 Doppler (WSR-88D) to detect rotation associated with tornadoes and other severe weather. The MDA analyzes Doppler radar radial velocity volume scans to compose a number of attributes thought to be related to mesocyclone formation. The 23 attributes of the MDA are compared to truthed tornado data in exploratory and diagnostic analyses to examine the underlying structure of the MDA. Results of these analyses indicate that the MDA is a highly correlated system with a wide variety of complexity in those correlations. This multicollinearity can hinder statistical prediction. Measured associations between the attributes vary from near zero correlation to complex correlations with values greater than 0.8, binding up to nine MDA attributes. In diagnostic analyses, linear and logistic regressions are performed on various sets of MDA attributes in an attempt to distinguish tornado events from non-tornado events. Logistic regression is found to be the most successful model due to its parsimony and ability to classify correctly tornado versus non-tornado cases. This research has shown that the number of attributes in the MDA can be decreased by projecting the 23 correlated attributes on a number of uncorrelated dimensions. Using principal component analysis (PCA), multivariate exploration of the data determines that 9 dimensions are needed to describe 85% of the variability of the MDA attributes. While the MDA is currently an improvement over older algorithms, this research shows that it is advantageous to reduce the redundancy of the MDA to make it a more useful tool.