Mon. Wea. Rev., 143, 933-954.

Probabilistic Prediction of Tropical Cyclone Intensity from a Multiple-Linear Regression Model


Chia-Ying Lee
International Research Institute for Climate and Society, Columbia University, Palisades, NY

Michael K. Tippett
Department of Applied Physics and Applied Mathematics, Columbia University, New York, NY

Suzana J. Camargo
Lamont-Doherty Earth Observatory of Columbia University, Palisades, NY.

Adam H. Sobel
Department of Applied Physics and Applied Mathematics, and Lamont-Doherty Earth Observatory, Columbia University, New York, NY


Abstract

The authors describe the development and verification of a statistical model relating tropical cyclone intensity to the local large-scale environment. A multiple linear regression framework is used to estimate the expected intensity of a tropical cyclone given the environmental and storm conditions. The uncertainty of the estimate is constructed from the empirical distribution of model errors. NCEP-NCAR reanalysis fields and historical hurricane data from 1981 to 1999 are used for model development, and data from 2000 to 2012 are used to evaluate model performance. Seven predictors are selected: initial storm intensity, the change of storm intensity in the past 12 hours, the storm translation speed, the difference between initial storm intensity and its corresponding potential intensity, deep-layer (850-200 hPa) vertical shear, atmospheric stability, and 200 hPa divergence. The system developed here predicts storm intensity changes in response to changes in the surrounding environment with skill comparable to existing operational tools. The probabilistic intensity predictions are shown to be reliable and skillful. Since one application of such a model is to predict changes in TC activities in response to natural or anthropogenic climate change, we examine the performance of the model using data that is readily available from global climate models, i.e., limited variables and monthly averages. We find that statistical models based on monthly data (as opposed to daily) with only a few essential predictors, e.g., the difference between storm intensity and PI, perform nearly as well at short leads as when daily predictors are used.