Radar Image Extrapolation with Conditional Generative Adversarial Network

Abstract

Forecasting hazardous weather, especially severe convective weather, is critical to protecting the environment and human life. Radar image extrapolation is a widely adopted method for this purpose, including the traditional optical flow-based method and deep learning-based methods. However, the optical flow approach suffers from a significant decline in accuracy with time, and the predictions generated by ConvRNN and CNN-based models tend to be smooth and blurry. To solve these limitations, we propose a Radar Multi-scale Fusion Generative Prediction Network (Fugen-Net) for the task of radar image extrapolation. We use an enhanced U-Net as the generator of a conditional generative adversarial network (CGAN) and design a Multi-scale Feature Fusion and Transmission (MFT) structure to improve the feature extraction capacity and reduce the number of parameters. To evaluate the forecasting ability of Fugen-Net, we conduct experiments using 4 years of radar images collected in Beijing from 2014 to 2017 and compare its performance with optical flow and U-Net. The experimental results demonstrate that Fugen-Net outperforms the other two methods when both high threshold skill scores and image realism are considered.