INVESTIGATION OF A REAL-TIME RAIN FORECASTING USING U-NET

Abstract

 A heavy rainfall prediction has long been an important goal in the field of disaster prevention. Robust and accurate predictions highly contribute to evacuating from disasters like a flood. Optical flow has become one of the standard methods for rainfall predictions, e.g. High-resolution Precipitation Nowcasts and Radar nowcasts provided by Japan meteorological agency (JMA). The aim of this study is to examine the applicability of U-Net, one of deep neural network models, for the prediction of hourly 5 categorical precipitation up to 6 hours ahead, and the effectiveness of learning pseudo rainfall data created for data augmentation to the model performance. For the model development, we used the data of Radar/raingauge Analyzed Precipitation (RAP) provided by JMA and generated the pseudo data by applying random rotation and random scaling to the original data. Using these data, we trained the model and then verified the capability of the U-Net model for rain forecasting. As a result, we successfully applied the U-Net model for a rainfall prediction and found that the model catches a rainfall intensification and weakening. The model trained with the pseudo data outperforms the one without the augmented data, particularly in the predictability of heavy rainfall for long forecast time.