During spring season in the autonomous province of South Tyrol frost events can cause major damage to fruit crops, which are essential to the regional agriculture and economy. Water sprinklers may prevent damage under particular local atmospheric conditions determined by air temperature (Tair), wet-bulb temperature (Twb) and wind speed (v). The aim of this thesis is to create statistical models that provide binary and ordered probabilistic forecasts of these variables for the forthcoming 24 hours to assist in keeping the cost of such events to a minimum.
The data for the three predictands comes from the automatic weather station in Schlanders covering the months Feb-Apr from 2010 to 2014. The frequency of Tair below 0◦C strongly decreases towards April, especially in the afternoon. The small number of such events exposes the estimated models to overfitting.
Numerical weather prediction (NWP) data in combination with observed data on site are post-processed to standard(SLR) and ordinal logistic regression (OLR) models. These are calculated for every third hour. To reduce overfitting, the predictors are chosen by either stepwise backwards selection or the Least Absolute Shrinkage and Selection Operator (LASSO). The ranked probability skill score (RPSS) and the Brier skill score (BSS) are applied for the verification of the models.
Using the total set of predictors, the smallest median RPSS with -0.20 for Tair and 0.30 for Twb occur during evening hours. Applying the LASSO effectively prevents overfitting for both parameters, while the stepwise backwards selection shows little to no effect. By creating one model for all hours combined, the forecasting skill for all the models is increased, especially for the evening hours. For the wind speed, the combined model only improves the forecasting skill for these single hour models which use fewer predictors.