|
363526-191001-021000
NA17OAR4590140
N/A
7/1/2017
2020-6-30 0:0:0
Completed
$203,373.00
Transition of Machine-Learning Based Rapid Intensification Forecasts to Operations
Mercer
Andrew
MSU
Coastal Hazards CH
OAR
A machine learning-based forecast for tropical cyclone rapid intensification (RI) was built as a first effort to implement an operational RI forecast task. This novel unsupervised learning technique employed Global Forecast System analyses to identify features that are most helpful in discriminating RI and non-RI environments. To develop a new classification predictor set, researchers identified predictors to emphasize outer band structures versus inner-core structures of tropical cyclones, and then they retained and added optimal discriminating fields to the existing SHIPS-RII predictors (currently used operationally to make RI forecasts). A fully cross-validated support vector machine (SVM) classifier was built from these predictors to predict Atlantic RI on 3,605 tropical cyclone timesteps. They tested this scheme in the Joint Hurricane Testbed experiments at the National Hurricane Center, with results indicating skill improvements of up to 18% better than climatology (roughly a 20% improvement over the performance baseline seen in the SHIPS-RII). Afterwards, the classifier was tested on cases from 2017-2019 Atlantic Hurricane seasons to identify its performance in true forecast mode. Testing and evaluation will continue.
|