A New Deep Learning Nowcast Model of Radar Imagery using Generative Adversarial Network for Operational Rainfall Nowcasting

The Hong Kong Observatory (HKO) has been operating an in-house developed nowcasting system, namely the Short-range Warning of Intense Rainstorms in Localized Systems (SWIRLS), for supporting forecast operation and warning services of rainstorm and associated severe convective weather in Hong Kong. With the advent of machine learning (ML) techniques including the Convolutional Long Short Term Memory (ConvLSTM) (Shi et al., 2015) and the Trajectory Gated Recurrent Units (TrajGRU) (Shi et al., 2017) developed in collaboration with a local university, these ML models have been utilized in experimental or real-time trials, providing additional reference on predicting heavy rain for the next couple of hours. Comparing to extrapolation of radar reflectivity based on motion field retrieved from optical flow technique, the above ML models demonstrate improvements in capturing location, growth and decay of rain bands.

Similar to other ML methods, the above deep learning nowcasts have a problem that blurrier images are generated beyond one to two hours of nowcasts, making them ineffective in capturing heavy rain and significant convection. In this study, a new ML framework, namely ResConvLSTM-GAN, enhances ConvLSTM technique using the generative adversarial network (GAN) pipeline, style transfer technique, and a dynamically balanced loss function for generating more reliable and realistic radar nowcasts. From verification, it is shown that ResConvLSTM-GAN is more capable of maintaining and generating small-scale features when compared with the traditional ML nowcast methods. Case studies on recent rainstorm events in Hong Kong using ResConvLSTM-GAN will be illustrated. This study also illustrates the application of ResConvLSTM-GAN to satellite imagery to enhance rainfall or significant convection nowcasts over a larger geographical coverage. Future work on enhancing the ResConvLSTM-GAN framework and extending its applications will also be presented.