Support vector machines for regional clear-air turbulence prediction

Clear-air Turbulence (CAT) is a significant safety issue for aviation at upper levels in the atmosphere. Since CAT is not observable with traditional remote sensing techniques, it is particularly difficult to avoid. The current FAA-sponsored CAT forecasting product, the Graphical Turbulence Guidance System (GTG), calculates many indicators or diagnostics of CAT potential from larger-scale numerical weather prediction model output and compares them to current turbulence observations from pilots (PIREPs). It then combines the diagnostics using a global optimization technique to provide the final CAT forecast. Theory suggests that many CAT diagnostics may vary in their predictive skill depending on the geographic region, but GTG is unable to exploit these regional dependencies due to an insufficient number of timely PIREPs. Recently, more plentiful and objective observation data have become available from the In-situ Turbulence Observation System. This system is currently installed on about 200 United Airlines' aircraft and provides data at one minute intervals. This high-resolution data now allows the development of CAT forecasts on a regional scale. For each region of the continental U.S. (determined by CAT climatology), we have used the machine learning classification technique of Support Vector Machines to determine the best subset of CAT diagnostics that together have the highest forecasting performance, regardless of the diagnostics' individual performances. To search efficiently through the state space of all feature subsets, we used a forward selection algorithm; the search is guided by a five-fold cross validation method on test sets of in-situ observation data. The results of the regional CAT forecasts determined in this manner are shown to provide better skill than the current GTG algorithm. This approach will ultimately be used to replace the current operational GTG CAT forecasting system.