Forecasting thunderstorms over a 2- to 5-h period by statistical methods

Abstract

Classical statistical techniques, such as multiple regression with variable selection and principal component analysis, were employed to define combinations of parameters from meteorological observations which optimally discriminate between the occurrence and nonoccurrence of thunderstorms. Routine observations of weather elements at five levels in the troposphere during two spring and summer seasons were analyzed objectively onto a 65-km grid which spanned much of the central United States. A thunderstorm occurrence was definced from manually digitized radar (MDR) observations with an MDR code of four or greater as the basis. The binary variable one or zero for occurrence or non-occurrence, respectively, was the predictand. Parameters which are measures of atmospheric moisture content, stability, and trigger mechanisms were calculated from gridded fields of surface and upper-air observed elements for different times each morning. These parameters were candidate predictors in the variable-selection procedures. Data from all grid points and for each day were pooled in order to provide an adequate sample of thunderstorm observations. Errors which result from usual assumptions in a regression model were quaintly analyzed.