Rainfall nowcasting with Deep Neural Networks: trends and strategies to overcome limitations

초록

Deep learning models have demonstrated remarkable performance in short-term rainfall prediction by effectively capturing the complex spatiotemporal patterns of rainfall systems using vast datasets. This has helped to address the limitations of traditional Numerical Weather Prediction (NWP) models and radar echo extrapolation methods, such as computational speed and nonlinear precipitation dynamics. In this paper, we systematically review deep learning-based rainfall nowcasting models, categorizing them into three main groups: Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Generative Adversarial Networks (GANs). We also conduct an in-depth analysis of the common limitations identified in these models and explore strategies to overcome them. Innovative approaches, such as the application of diverse loss functions, transfer learning, sampling techniques, and ensemble models, are discussed in terms of their impact on enhancing model performance. Based on these findings, we propose future research directions for rainfall nowcasting. This review aims to provide a comprehensive understanding of the current state of deep learning-based technologies for rainfall nowcasting and serve as a guide for future research and technological advancements in the field. © 2024 Korea Water Resources Association. All rights reserved.

키워드