Application of the autologistic function with covariates to estimate an icing field

The continued occurrence of serious aircraft accidents attributed to in-flight icing conditions has led to great interest in the generation of reliable and accurate forecasts of this phenomenon. Traditionally, such forecasts have been verified by simple comparisons between the verification data (pilot reports) and the forecast field. However, pilot reports represent a very small proportion of the forecast grid, and they are not systematic. Thus, pilot reports by themselves can not provide a complete evaluation of in-flight icing forecasts.

In an effort to generate a more representative icing verification field, an autologistic model with covariates is applied to a variety of case studies. This model is used in the context of binary responses considering spatial correlation and covariates, in an attempt to provide inference over a large, sparsely sampled area of interest. It is motivated by the scarcity of icing observations (i.e., pilot reports of icing conditions) and the existence of easily accessible icing predictors. In this case, the covariate is the Integrated Icing Diagnostic Algorithm (IIDA), developed at the National Center for Atmospheric Research. The set-up is Bayesian, and a Gibbs sampler is implemented to estimate the model parameters. The model is evaluated based on its ability to resolve the relationship between the covariate and observations.