Thursday, 30 September 2010
ABC Pre-Function (Westin Annapolis)
Often obscured by upper level clouds in visible and infrared imagery, tropical cyclone (TC) structure can be seen in microwave imager data. Taking advantage of this capability and access to multiple years of data, an automated method to estimate TC intensity using Special Sensor Microwave Imager (SSM/I) data is developed for the Atlantic hurricane basin. From SSM/I 85 GHz data centered on 60 tropical cyclones (319 samples) and covering 11 years (1995-2005), characteristic features (including eye, banding, and other structural characteristics) are extracted for each of the historical samples. After feature selection to reduce the negative impact of redundant and irrelevant features, a feature subset and best track intensity are used to represent each sample in order to train and test (through leave-one-out cross validation) a machine learning algorithm for TC intensity estimation. Given the current data sets, features, and sole use of 85 GHz data, the root-mean-square-error (RMSE) for cross validation is 15.0 kts. Noting that the high variance of feature values for a given TC intensity can adversely affect performance, the samples are stratified by life stage and environmental shear values in order to reduce that variance and provide higher confidence in the intensity estimation. Current RMSE for mature (no samples that occur during initial intensification or final dissipation are included) TC samples with low shear (123 samples) is 13.2 kts. Further work will explore the use of 37 GHz data, additional features and samples (2006-2009) as well as the continued need for data stratification.
![](images/flag.png)
![](images/blue_ribbon.png)