Application of a Radar Echo Extrapolation-Based Deep Learning Method in Strong Convection Nowcasting

Abstract

Strong convection nowcasting has been gaining importance in operational weather forecasting. Recently, deep learning methods have been used to meet the increasing requirement for precise and timely nowcasting. One of the promising deep learning models is the convolutional gated recurrent unit (ConvGRU), which has been proven to perform better than traditional methods in strong convection nowcasting. Despite its encouraging performance, ConvGRU tends to produce blurry radar echo images and fails to model radar echo intensities that have multi-modal and skewed distributions. To overcome these disadvantages, we tested the structural similarity (SSIM) and multiscale structural similarity (MS-SSIM) indexes as loss functions. The SSIM and MS-SSIM loss functions are composed of luminance, contrast, and structure and provide more information about the intensity, grade, and shape of the radar echo, which can reduce blurring. Due to multi-layer downscaling, MS-SSIM extracted more radar echo characteristics, and its extrapolation was the most realistic and accurate among all of the loss function schemes. Only the MS-SSIM scheme successfully predicted strong radar echoes after 2 h, especially those at the rainstorm level.