Radar-based precipitation nowcasting using task-segmented recurrent generative adversarial networks

Abstract

With the objective of making a high-resolution and precise forecast of regional precipitation, precipitation nowcasting has become an important technique to give the early public warning on potential landslides and floods caused by heavy precipitation. Many deep learning models have recently been shown to outperform traditional Numerical Weather Prediction (NWP) models, extrapolation-based methods for precipitation nowcasting. However, current deep learning-based precipitation nowcasting is seriously limited by the following two perspectives. Firstly, the predicted precipitation images from current deep learning models will become blurry with increasing lead-time. Secondly, most of the current deep learning models can only provide precipitation nowcasting within 2 hours. To overcome these problems, we proposed a novel task-segmented architecture called TS-RainGAN built with recurrent generative adversarial networks for precipitation nowcasting. In our experiments of 6-hour precipitation nowcasting using radar images at 1.3 km resolution, TS-RainGAN can produce photo-realistic images. At the same time, other deep learning models such as ConvLSTM, MIM, and PredRNN++ will result in considerable blurry images. In addition, TS-RainGAN can attain the highest scores evaluated by common skill metrics (CSI, FAR, and POD) and maintains a good sharpness mark throughout the 6-hour nowcast. To our knowledge, TS-RainGAN is the first deep learning model that can handle 6-hour precipitation nowcasting using radar data and produces radar-based rainfall nowcast without obvious blurriness.

Date of Award	2021
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Jimmy Chi Hung FUNG (Supervisor)