TJ37.4 Convective Initiation Forecasts Through the Use of Machine Learning Methods

Wednesday, 9 January 2013: 2:15 PM
Room 18A (Austin Convention Center)
Mark S. Veillette, MIT Lincoln Laboratory, Lexington, MA; and H. Iskenderian, P. M. Lamey, and L. J. Bickmeier

Storm initiation is a very challenging aspect of nowcasting. Rapidly forming storms that appear in areas of little to no pre-existing convection can pose a danger to aircraft, and have the potential to cause unforeseen delays in the national airspace system (NAS). As such, detection and prediction of the initial development of convective storms is critical to NAS operations and planning. The Corridor Integrated Weather System (CIWS) currently provides deterministic 0-2 hour storm forecasts over the NAS, and represents the 0-2 hour portion of the 0-8 hour deterministic CoSPA storm forecasts. CIWS includes a convective initiation (CI) module, however this module has difficulty initiating convection in areas of little or no pre-existing convection. In this study, we seek to improve the capabilities of the CI module using machine learning methods to detect regions of imminent convection and enhance the storm initiation to the 0-2 hour forecast. Improvements to the current CI detection capabilities will prove to be a benefit in the short term, as well in the longer term plans of the Federal Aviation Administration's (FAA) Next Generation Air Transportation System (NextGen).

In order to improve the capabilities of the CI module in CIWS, data from a variety of sources are fused together to produce a forecast of CI. Data incorporated into the CI algorithm include: Satellite fields from NASA's Satellite Convective Analysis and Tracking (SATCAST), convective instability fields, and a collection of numerical models which includes NOAA's North America Rapid Refresh Ensemble Time Lag System (NARRE-TL), the Localized Aviation MOS Program (LAMP), Short Range Ensemble Forecasts (SREF), and High Resolution Rapid Refresh (HRRR) model forecasts. These fields are brought together in a machine learning framework to create a probabilistic model which is used to initiate new growth in the deterministic CIWS 0-2 hour forecast. A variety of machine learning classifiers, including logistic regression, neural networks, support vector machines, and random forests, are used to investigate which technique works best with the data available. The skill of this updated CI capability is being assessed over the summer of 2012 using multiple skill metrics including CSI, bias and fraction skill score.

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