Flexible discriminant techniques for forecasting clear-air turbulence

Forecasting aircraft (clear-air) turbulence is currently based on a system of observations by pilots combined with a mostly subjective evaluation of turbulence indices derived from numerical weather prediction models. We address the issue of improving the forecasting capability of the single indices by combining them in a nonparametric multidimensional regression model, and applying discriminant analysis to the resulting predicted values. Thus we enhance the predictive skills of the indices considered in isolation and provide a more robust algorithm. We adopt the paradigm of flexible discriminant analysis (FDA), and use multivariate adaptive regression splines (MARS) and neural networks (NN) in the regression stage. The data for this case study covers the period 12-15 March 1999, for the United States. Results of the analyses suggest that our statistical approach improves upon current practice to the point that it holds promise for operational forecasts.