A Statistically Derived Prediction Procedure for Tropical Storm Formation

Abstract

A statistical forecasting experiment was performed to test the capability of predictors derived from observational data (analysis) fields at 950, 700, 500 and 200 mb to forecast tropical storm formation (genesis). National Oceanographic and Atmospheric Administration tropical mosaic visible satellite images and the Joint (United States Navy and Air Force) Typhoon Warning Center's Post-Season Best Track analyses of tropical storms were used to select a representative collection of tropical cloud clusters, some of which became tropical storms (GO cases), others of which did not (NO GO cases). Navy Fleet Numerical Oceanography Central archived analysis fields of surface pressure, winds, sea surface temperature, and moisture were accessed at locations and times corresponding to cloud cluster positions 24, 48 and 72 hours prior to tropical storm formation/nonformation, and candidate predictors were formed from these analysis data. The number of predictor candidates was increased by also calculating a predictor candidate's Laplacian, and the magnitude of its gradient. A Special local-maximum enhancement technique was also applied to some of the candidate predictors. Stepwise discriminant analysis was applied to these candidate predictors to select subsets with greatest predictive capability for forecasting tropical storm formation at projections of 24,48 and 72 hours. The resulting statistical forecast algorithms were evaluated on independent data, against climatology, and against a basic technique derived solely from latitude and longitude. The results show our forecast technique possesses considerable skill in predicting tropical storm formation with good pre-figurance, post-agreement, threat, and Brier scores.