5D.8 Use of Track Clustering to Elucidate Landfall Hazard Uncertainties of Harvey and Irma

Tuesday, 17 April 2018: 9:45 AM
Heritage Ballroom (Sawgrass Marriott)
Alex Kowaleski, Pennsylvania State Univ., Univ. Park, PA; and J. L. Evans

Although Hurricanes Harvey and Irma both made landfall on the continental United States as major hurricanes, each storm produced a substantial portion of its damage far from its landfall location, and center track. Harvey generated catastrophic freshwater flooding in the Houston area, though it made landfall 250 km from Houston, and its center never came within 120 km of the city. Irma, which landfalled in far southwest Florida, caused a record storm surge in Jacksonville, and a significant storm surge as far north as South Carolina. The effects of Harvey and Irma demonstrate that forecasting the life and property hazards from a landfalling tropical cyclone is substantially more complicated than forecasting the track of the storm center.

Kuruppumullage et al. (2016), Kowaleski and Evans (2016), and Gonzalez-Aleman et al. (2018) have demonstrated that regression model clustering (Gaffney et al. 2007) can partition forecasts from one or more ensemble prediction systems (EPSs) into a small number of distinct forecast outcomes that minimize intra-cluster spread and maximize inter-cluster separation. However, clustering of ensemble TC track forecasts is most useful as an operational tool if it elucidates how each TC-related hazard (wind, precipitation, and storm surge) varies among and within clusters.

Here, we examine how mixture-model clustering of track forecasts of Harvey and Irma from the 51-member European Centre for Medium-Range Weather Forecasts (ECMWF) EPS represents the relationship between storm track and the hazards experienced at select locations. Five-day track forecasts of Harvey and Irma from the ECMWF EPS, initialized 48-120 hours before landfall, are partitioned via regression mixture models, as in Kuruppumullage et al. (2016) and Kowaleski and Evans (2016). Optimal mixture-model specifications (polynomial order and number of clusters) are determined through analysis of Bayesian Information Criterion and Root Mean Squared Error relative to each cluster centroid’s trajectory.

After track forecasts of Harvey and Irma are clustered, we examine how the wind, precipitation, or storm surge hazards at select locations vary among track clusters and among the members that each cluster comprises. For Harvey forecasts, we examine the total precipitation and maximum 6-hour precipitation at Houston, Beaumont, and Lake Charles. For Irma forecasts, we analyze the maximum wind speed at Key West, Miami, Palm Beach, and Tampa, and the storm surge threat at Tampa, Jacksonville, and Charleston. Storm surge threat at each location is estimated from the strength and distance of the onshore flow in each ensemble member.

The results from this analysis method indicate the extent to which ensemble forecast track clustering partitions the sensible hurricane-related hazards experienced at and after landfall. This provides information about the predictability of specific hazards of Harvey and Irma at specific locations, including hazards a great distance from the tropical cyclone center. The methodology demonstrated here may be applied to forecasts of future tropical cyclones to elucidate the storm tracks that produce the greatest hazards at specific locations, and to show which hazards are relatively insensitive to variations in storm track within the ensemble.

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