GC54A-04:
Using Mixture Regression to Understand and Model Tropical Cyclone Intensification in Relation to the Environment and Climate

Friday, 19 December 2014: 4:45 PM
Emmi Yonekura, Ning Lin, Yuyan Wang and Jianqing Fan, Princeton University, Princeton, NJ, United States
Abstract:
Representing the relationship between tropical cyclone intensity and the surrounding environment is key to projecting the way tropical cyclones change with the climate. However, capturing the complex relationship between tropical cyclone intensification, the surrounding environment, and storm characteristics may not be possible with simple linear regression models like SHIPS, especially for short forecast periods in risk models. Here, we move beyond simple linear modeling and apply methods that perform variable selection, determine if the data is a heterogeneous “mixture” of multiple features, and fit linear or nonlinear functions of predictors for a 6-hour forecast window. We also incorporate a new predictor, the Ventilation Index, which represents the surrounding environmental conditions in a way that is known to affect tropical cyclone intensity according to hurricane physics. The data used to construct the models comes from the IBTrACS WMO archive for TC-specific data and ERSST v3b and the NCEP-NCAR Reanalysis for environmental variables from 1970-2010 in the North Atlantic. The two observed measures of intensity, the maximum sustained wind speed and minimum central pressure, are modeled jointly.

First, we find that variable selection in both linear and nonlinear models does not significantly reduce the number of environmental predictors needed or improve the model R-squared. Second, mixture regression is applied to establish data groupings and their associated predictors. There is an increase in R-squared by 0.10 compared to the linear regression model that uses all possible variables as predictors. The number of environmental predictors decreases from 8 to 4 for wind intensification and to 2 for pressure intensification. When we further restrict the predictor pool to use only Ventilation Index to represent the environment, mixture modeling shows a 0.10 increase in R-squared. Then, allowing nonlinear relationships with predictors in a mixture model with Ventilation Index gives a further R-squared increase of 0.20. We show that mixture modeling is the best approach to create a climate-variant model for tropical cyclone intensity that can be used as part of a tropical cyclone risk model. The techniques may also be applied to other aspects of tropical cyclones, such as storm genesis and size.