[M-GI27] Data-driven approaches for weather and hydrological predictions

[MGI27-01] Global precipitation nowcasting with ConvLSTM and adversarial training

★Invited Papers

*Shigenori Otsuka¹, Takemasa Miyoshi¹ (1.RIKEN)

Keywords:precipitation nowcasting, deep learning, adversarial training

Precipitation nowcasting provides prediction of precipitation for a short time based on most recent precipitation image patterns observed by weather radars and meteorological satellites. Conventional precipitation nowcasting performs spatiotemporal extrapolation by tracking rain areas and moving them forward in time by an advection model. Recently, deep learning-based precipitation nowcasting is becoming more popular. Image processing capabilities of deep learning enable us to implement such nowcasting applications at relatively low cost. Some operational meteorological centers are already deploying their deep learning-based nowcasting systems. However, there are still issues such as a blurring effect. In this presentation, an experimental system for global precipitation nowcasting will be presented. Global Satellite Mapping of Precipitation (GSMaP) is a semi-global precipitation product by JAXA derived from mainly microwave and infrared satellite observations. Hourly updated precipitation distributions on a 0.1-deg-by-0.1-deg mesh are provided between 60S-60N. Here we trained a ConvLSTM-based model with the previous 24 images to generate the future 12 images, i.e., 12 hours lead time. To address the blurring problem, an adversarial training framework was adopted. The results indicated that the ConvLSTM-based model generally outperformed a conventional tracking-based nowcasting system.