Using spatiotemporal relational random forests (SRRFs) to predict convectively induced turbulence

Convectively induced turbulence is a common threat for airlines on a daily basis. Caused by various atmospheric conditions, turbulence can result in damage to aircraft, private property, as well as risks to passenger health.  Current FAA guidelines call for multiple safety precautions in regards to areas of potential turbulence; the most common being the avoidance of the region. However, as a seemingly random, and in the case of convectively induced and clear air turbulence, invisible event, prediction of such scenarios is challenging. Large areas of airspace may become off-limits in an attempt to avoid these possibly dangerous areas. Alternatively, turbulence forecasts or detections may be missed, leading planes to fly into “safe” areas only to experience turbulence. In addition to the physical hazards of turbulence, a new problem of airspace economics is evident as areas safe for flight may be off limits in correspondence with the FAA guidelines. For this reason, prediction of turbulence is a valuable endeavor. 
   

Currently, there are several turbulence prediction methods, each with their own limitations. Pilot report methods are useful in that they give some lead time for other aircraft. However, turbulence is a short-term event, and such reports may be invalid by the time new aircraft enter the area. In addition, pilot interpretation of such events introduces the human error on both a temporal and spatial scale. Next is the graphical turbulence guidance system, which sends out hourly forecasts of possible turbulent areas. With much to learn about the genesis of turbulence, the grid is much too coarse in relation to aircraft size. Finally, there is the NCAR Turbulence Detection Algorithm (NTDA). While this is a useful tool in the prediction of in-cloud turbulence, the NTDA fails to predict in areas outside of clouds, where CIT is most prevalent.
   

Using Spatiotemporal Relational Random Forests (SRRFs), a probabilistic prediction of turbulence is acquired. SRRFs are powerful probability forests with the ability to ask questions based on both spatial, and temporal relations. Using a bootstrap resampled subset for the training of each tree in the forest, questions based upon chi squared are used to split instances in a recursive manner. The resulting tree gives the probability of whether there is turbulence or not for each instance in the testing set. Once all trees give a probability, the mean is calculated, resulting in a corresponding label of turbulent or non-turbulent.
     

Multiple sources of data are employed to train the SRRFs, including WRF co-located model data, in-situ aircraft observations, GTG forecasts, and NTDN lighting data. Initial results show a mean Gerrity Skill Score of 0.60, suggesting this approach has significant skill in the forecasting of turbulence. In addition to the prediction of turbulent cases, the SRRF algorithm is used to find important atmospheric variables in the creation of such scenarios. Using these results, the researchers hope to aid in a stronger understanding of prediction of convectively induced turbulence and may lead to improved diagnoses and forecasts.