Advancing deep learning-based precipitation nowcasting model with radar and ground observation

To forecast accurate predictions of flooding in a basin, it is essential to have accurate precipitation forecasts. Recently, there have been a number of studies that utilize machine learning techniques to improve the performance of precipitation prediction models. Several radar-based precipitation prediction models have been developed and shown their capabilities, mainly for spatial precipitation prediction over basins. However, since the model is developed based on radar data, which has limitations in accurate rainfall estimation, it is challenging to accurately predict the actual precipitation that falls in a basin. In this study, we proposed a strategy to improve the prediction performance of radar-based models using machine learning by incorporating observations from gauge stations during the training process. To develop a model that produces predicted rainfall up to 6 hours for the Yeongsangang River basin, a conditional generative adversarial neural network (cGAN) was applied. Along with the radar data, the model was trained using the rainfall data from gauge stations (Automated Synoptic Observing System, ASOS) within the domain of the Yeongsangang River basin. Through the results, we showed that the model that considers point rainfall data as an additional input performs better in predicting the spatial pattern of rainfall than the model that only uses radar data.