A curve clustering algorithm to increase skill in extratropical transition forecasts

Extratropical transition of tropical cyclones is currently a challenge for forecasters in both time and space (e.g. Jones et al. 2003). Forecasts of track, intensity and storm structure impact assessments of potential societal impacts due to the evolving storm. In this study, a new statistical method is developed to assess the spread of model ensemble member forecasts of track and storm structure. This approach is based upon a method known as an Expectation-Maximization (E-M) curve clustering model (Gaffney 2004). The E-M model is a non-parametric clustering algorithm that reaches a final solution by maximizing the derived log-likelihood function using posterior probabilities that each curve belongs to a particular cluster. A previous analysis of operational forecast models utilized K-means clustering of individual time period forecasts of storm structure to evaluate storm evolution and model skill (e.g. Arnott et al. 2004). In the methodology presented here, curve clustering will be based around two-dimensional (latitude and longitude) and three-dimensional curves. The 3D curves represent storm structure evolution using the cyclone phase space (Hart 2003). The clustering will reduce the numerous member ensemble solutions to two to four outcomes that are most probable. The hypothesis is that the E-M model will improve insights into alternative forecast scenarios and so potentially contribute to increases in the skill of extratropical transition forecasts.