Learning from Precipitation Events in the Wider Domain to Improve the Performance of a Deep Learning–Based Precipitation Nowcasting Model

Abstract

Abstract Deep learning models have been shown to perform well in terms of the radar-based precipitation nowcasting problem when compared with advection-based models. Yet, most of the existing literature has used locally trained models, and relatively little is known about how to construct deep learning–based nowcasting models applicable in wider domains. We conduct experiments for precipitation nowcasting using deep learning models spanning the region around Japan with the Japan Meteorological Agency radar precipitation dataset with 1-km spatial resolution and 5-min temporal resolution. We trained the model with radar data sampled from all over Japan, then applied transfer learning with regional data. Through this experiment, it is found that combining data from all over Japan is effective in improving the forecast performance of heavy precipitation with deep learning models. Significance Statement This paper proposes a methodological improvement for the problem of short-term rainfall forecasting (nowcasting) using deep learning. We trained deep learning models with data sampled from all over Japan, then applied transfer learning with regional data. We compared the performance metrics of several modeling approaches and demonstrated that learning from heavy precipitation examples collected from wider domains improves the performance significantly. This finding suggests that there are a lot of commonalities in the time evolution of rainfall patterns at different geographical locations, which could be exploited to improve the model’s performance. A future research topic could be to learn from a larger dataset, including precipitation datasets from other countries.