Forecasting Low-Visibility Conditions at Vienna Airport with Tree-Based Statistical Models

Low-visibility conditions at airports can lead to capacity problems and therefore to delays or cancelation of arriving and departing airplanes. To keep the capacity as high as possible, accurate visibility forecasts are required. Therefore tree-based statistical nowcasting models were developed, which derive decision rules by recursive partitioning. These trees produce probabilistic forecasts of the low-visibility procedure (LVP) state by recursively splitting the data with respect to the predictor variables. In each step, the predictor that has the highest association with the LVP state will be selected, until the model cannot be significantly improved. Benefits of these models are fast update cycles and short computation times. Highly-resolved meteorological observation data at the airport form the large pool of input variables for the models. For low-visibility forecasts with lead times shorter than one hour, only visibility and ceiling are choosen as important input variables by the model. With longer lead times, variables such as wind speed, relative humidity and temperature become important. Forecasts of the tree-based models are more accurate than persistence climatology forecasts for each lead time, whereby the gain of the tree-based models becomes better with increasing lead times compared to the persistence climatology forecasts.