-
Improving the anticipation of severe weather (McGovern, Brotzge,
R. A. Brown, Droegemeier). Approximately 25% of all tornadoes are
not warned in advance, and nearly 80% of all tornado warnings turn
out to be false alarms (when no tornado actually occurs). Too many
false alarms can increase public apathy, and tornadoes without
advance public warning can cause otherwise preventable deaths and
injuries. This pro ject will use data mining and knowledge
discovery methods to extract storm type information and attributes
from weather radar data and model output to identify common
relationships among storm events associated with a false alarm or
missed tornado warning. Once these relationships are known, new
techniques can be developed to improve our tornado warning process
(McGovern et al.,2007; Hiers et al., 2008).
- Improving the automatic selection of WSR-88D radar scanning
strategies (McGovern, R. A. Brown). Within its nationwide
network of WSR-88D radars, the National Weather Service uses
different radar scanning strategies depending on the type of storms
occurring within the coverage region of a given radar. At the
present time, forecasters have to monitor the radar display and
change the scanning strategy manually. We will investigate how data
mining and ML can be used to ob jectively identify storm type and
thereby automatically select the appropriate scanning strategy
(Gagne II et al., 2008).
-
Establishing the limits of predictability of tornadic
outbreaks (Richman, McGovern). At the present, severe storm
outbreak type is forecast 24 hours in advance. Through data mining and
modeling of the differences between severe weather outbreaks and
tornadic outbreaks, it is expected that signatures of tornadic
outbreaks can be made at least several days in advance. This project
applies ML techniques, such as support vector machines, to classify
outbreak type and establish the limits of predictability in both
severe weather and tornadic events.