Variable selection for tropical cyclogenesis predictive modeling

© 2016 American Meteorological Society. Variable selection for short-term forecasting (up to 72 h) of tropical cyclone (TC) genesis has been investigated. IBTrACS data (1979-2014) are used to identify the genesis time and position of over 2500 TCs between 30°N and 30°S. Tracks are extended using a tropical cloud cluster (TCC) dataset, which is also used to identify over 28 000 nondeveloping TCCs. Subsequently, corresponding local environment states at various atmospheric pressure levels are retrieved from ERA-Interim data. An initial selection of potentially favorable variables for TC genesis is made based on mutual information, which forms the set of nodes for graphical model structure learning using the Peter-Clark (PC) algorithm. Structure learning identifies the variables with the strongest influence on TC genesis, while taking into account the interrelationship with other variables. Variables are ranked based on the maximum observed p value in all (conditional) independence tests of the variable with the TC genesis node. The results indicate that potential vorticity (600 hPa), relative vorticity (925 hPa), and (vector) vertical wind shear (200-700 hPa) are the highest ranked variables for forecasting up to 72 h. These are followed by the basin and zonal wind speed (200 hPa), and for very short lead-time divergence (925 hPa), air temperature (300 hPa), and average vertical velocity. Predictive modeling with logistic regression confirms the superior performance of the top-ranked variables. The presented variable ranking (methodology) can be used as a building block for the creation of genesis indices or predictive models in the future.