Extended Abstract (72K)J7.2
Winter hydrometeor classification using polarimetric radar and spatiotemporal relational probability trees
Andy L. Spencer, Rose-Hulman Institute of Technology, Terre Haute, IN; and A. McGovern, K. L. Elmore, and M. B. Richman
Dual polarization provides several additional radar parameters that can be used when determining precipitation types. Like reflectivity, polarimetric data points are spatially and temporally correlated and patterns in the data can be recognized and used to when generating predictions. Current classification algorithms that use polarimetric data are based on existing theories about the characteristics of various types of hydrometeors. These algorithms work well for determining the precipitation types in volume data obtained during summer storms. However, they show very little skill when tested against observations taken during the Winter Hydrometeor Classification Ground Truth Program. This is in part due to noise in the radar and ground truth data.
The large number of input parameters and the availability of ground observations provides an ideal environment for the application of supervised machine learning techniques. Many existing supervised learning algorithms demonstrate only small amounts of skill and are too inaccurate to be used successfully by forecasters. To improve accuracy, a spatial learning algorithm can be used so that relationships such as bright band can be automatically identified. This is accomplished by clustering the polarimetric data in order to identify regions of similar values. Afterwards, spatial relationships between clusters are identified and a spatiotemporal relational probability tree is used to determine which relationships correspond to different precipitation types. Initial findings suggest that this approach can increase accuracy when compared to similarly sized decision trees that do not include spatiotemporal information.
Once the learning portion of the algorithm is complete, precipitation types can be classified using data that is available at the time of the event. In addition, as new forms of data become available the probability tree can be relearned with minimal changes to the algorithm itself.
Joint Session 7, Hydrology and AI: Status and Applications—II
Tuesday, 13 January 2009, 1:30 PM-3:00 PM, Room 125A
Previous paper Next paper
Recorded presentation