On statistical nowcasting of road surface temperature

In this study, multiple linear regression (MLR) and the generalized additive model (GAM) approaches were used to build statistical models for 6 hr nowcasts of road surface temperature (RST) in the northeast of Vienna, Austria. GAMs were more suitable for historical analysis, particularly for decomposing the terms to identify the different influences of the meteorological covariates on RST. By contrast, for RST nowcasting, the simpler and more robust MLR models are recommended, with better applicability for real-time operational runs. In MLR models, the forecasted air temperature was the most prominent predictor, followed by the measured RST. In independent testing, the MLR models showed better prediction skill, with daily root-mean-square error (RMSE) around 1 degrees C. In accordance with the linear correlativity, the MLR models were built with more predictors for daytime than at night but still generated a larger RMSE at midday. Furthermore, the MLR models could reproduce the correct diurnal variation and could forecast RST below freezing point better than above 0 degrees C. Four case studies, i.e. snowy, cold front, cloudy and sunny, were diagnosed in detail. The predicted RSTs were close to the measurements and depicted the trend well, including the persistent and rapid cooling (warming) and correct diurnal variation.