Speaker
Description
In this study, we exploit a 10-year archive of composite radar images over the Swiss Alpine region to improve precipitation nowcasting methods based on persistence. The underpinning hypothesis is that growth and decay is more predictable in mountainous regions due to the presence of systematic orographic effects. To this end, the variables related to geographical location, mesoscale motion, freezing level height, and time of the day are used as input to train machine-learning algorithms to predict the temporal evolution of precipitation, that is, its growth and decay patterns. The flow direction and speed are described by the motion of radar precipitation echoes, computed through a variational optical flow technique, while the freezing-level height is extracted from the analyses of the numerical weather prediction model COSMO. The average long-term growth and decay patterns are effectively reproduced, but the case-to-case variability is found to be underestimated by the predictions. This result nicely illustrates the limitations of deterministic machine learning predictions and highlights the importance of modelling the predictive uncertainty, too.