Precipitation Nowcasting using Data Augmentation

○ This paper proposes a simple data augmentation technique specifically designed to mitigate the data unbalancing problem in precipitation nowcasting. We consider the existence of one or more observational systems, each one comprised of a set of (either weather or rain gauge) stations. We use simulated data coming from the ERA5 numerical model to complement precipitation observations made by rain gauge stations, and use the resulting synthetic observations to augment data for a given weather station. We present preliminary results training a machine learning model using this data augmentation technique. These results show that the technique can be useful to improve the predictive performance of the resulting forecasting model.