A fog forecasting method in a deeply embanked valley

This paper presents a statistical model used to forecast fog in the Meuse Valley in Belgium. The method is a bootstrap discriminant analysis using eight predictors: river surface temperature, air pressure, air temperature at two elevations, wind speed and relative humidity at the same two locations. These data are measured from November 1989 to April 1990. Tests are done to determine the number of resampling needed for this data set and the optimum projection delay for prediction from the meteorological data. The best results are obtained for the prediction at 0700 h UT using meteorological data at 0400 h UT. The reliability of the model is given by a probability α = 0.16 of clear weather forecasting when it is foggy and a probability β = 0.26 of fog forecasting when there is clear weather. These results are finally checked on 27 new observations in November 1990: the 6 foggy days are perfectly predicted and 24% of the clear days are badly predicted.