J4.5
Spatiotemporal relational random forest (SRRF) prediction of convectively-induced turbulence: a severe encounter case study

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 26 January 2011: 9:30 AM
Spatiotemporal relational random forest (SRRF) prediction of convectively-induced turbulence: a severe encounter case study
2A (Washington State Convention Center)
Timothy Sliwinski, Florida State University, Tallahassee, FL; and J. Trueblood, D. J. Gagne II, A. McGovern, J. K. Williams, and J. Abernethy

The purpose of our research has been to create a forecast of convectively-induced turbulence (CIT) for the Contiguous U.S. (CONUS). Most CIT is encountered as aircraft attempt to avoid the well-documented hazards of in-cloud convective turbulence. Although current FAA guidelines address CIT, these guidelines cover broad regions and make large areas of airspace off-limits. With our research, we aim to efficiently label localized areas of turbulence and decrease the area of off-limits airspace.

Throughout the course of our investigation, we've utilized Spatiotemporal Relational Random Forests (SRRFs) to create a graphical turbulence forecast across the CONUS. To examine the SRRF's performance, we've chosen to focus on the recent accident involving severe turbulence encountered by United Airlines flight 967 on July 20, 2010. The SRRF approach was created at the University of Oklahoma and is based on the principles of random forests. Random forests are collections of decision trees, each trained on a subset of the training data, which vote based on how their decision nodes distinguished turbulent cases from non-turbulent cases. These votes give a probability that turbulence will be encountered at a given point and time. Turbulence is of high variability in spatial and temporal extent, and SRRFs are well suited as a prediction method since spatial and temporal changes in fields of variables have been added to the distinctions generated by regular random forests. Once we've generated a forest with high predictive skill, we can take the distinctions it generated and quickly focus it on each point over CONUS to generate a probabilistic prediction for turbulence on an update cycle efficient enough to be of use to pilots in the air. The distinctions the trees make are also significant in that they give us an idea of how important each variable is in distinguishing factors that promote CIT formation. In this sense, SRRFs can advance our knowledge of how turbulence relates to changes in the environment. This aspect is demonstrated through an examination of changes in conditions prior to, during, and after the July 20th event and the associated evolution of the turbulence prediction field.