262 Support vector machine techniques to predict tropical cyclone re-intensification following extratropical transition

Monday, 24 January 2011
Washington State Convention Center
Israel Vaughn, University of Arizona, Tucson, AZ; and J. S. Tyo and E. A. Ritchie

Handout (1.3 MB)

The extra-tropical transitions (ET) of tropical cyclones are significant contributors to weather-related disasters globally. One way to reduce the societal impact of these disasters is to provide early warning of these events, which can potentially be accomplished via full numerical simulation. However, using full numerical modeling has proven to be difficult due to the apparent chaotic nature of the underlying system dynamics. Early warning can also be accomplished via machine learning techniques and classification of ET events. Our previously published work used support vector machines (SVM) to attain a probability of detection (PD) of ~76% with a corresponding false alarm rate (FAR) of ~27% on a subset (a single pressure surface of potential temperature) of back-fitted full-physics numerical prediction model data.

In this study we extend this subset to the full volume of the western North Pacific for both potential temperature and equivalent potential temperature for input into the support vector machine (SVM) classifier, attaining ~80% PD with a ~26% FAR. In addition, we apply SVM to the full model data for the western North Pacific, a data set of ~600 million points, which a computationally intensive problem. One of the limitations of SVM is that it is a supervised learning technique, and does not elucidate the natural manifold of physically meaningful data embedded in the model data. For this reason we also apply Riemannian Manifold Learning (RML) to the model data for all variables and the volume of the western North Pacific region. The implementation of RML is being accomplished utilizing the CUDA framework, which uses consumer grade graphics cards (NVidia GTX 465s) for fast parallel computing. RML will potentially reduce the dimension of the relevant data, and give us a physically-meaningful manifold. The resulting manifold can then be submitted to a supervised learning algorithm for binary classification. We present our SVM results and introduce RML in this presentation.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner