4:30 PM - 4:45 PM
[MGI26-11] Downscaling Precipitation Data using Convolutional Neural Network Coupled with Wasserstein Generative Adversarial Networks
★Invited Papers
Keywords:Deep-learning, Precipitation, Super-resolution, Optimal transport
In recent years, Japan has experienced an increasing trend in disasters caused by heavy rainfall. Accurate weather forecasting is crucial to mitigate the damage from such disasters. In particular, the prediction of precipitation fields in a high spatial resolution is essential to reduce the damage caused by localized heavy rain events. In the field of image processing, super-resolution techniques using deep learning have been progressing, and it is proposed that such techniques outperform conventional interpolations.
This study proposes a method that employs the optimal transport cost (Wasserstein Distance) as the loss function for super-resolving precipitation fields. Because the Wasserstein Distance can quantitatively assess the distance between specific distributions as the loss function, it is expected to produce spatially finer precipitation images than the conventional deep learning models. However, direct calculation of the Wasserstein Distance for two-dimensional rainfall data is numerically difficult; therefore, the framework of Wasserstein Generative Adversarial Networks (WGAN) is adopted in this study. We train two networks concurrently: a generator that conducts super-resolution, and a discriminator that provides the Wasserstein Distance. A well-trained discriminator efficiently estimates the Wasserstein Distance, and feeds this output back to the generator so that the generator enables super-resolution based on the optimal transport cost.
In this study, we used the Radar/Raingauge-Analyzed Precipitation (RRAP) provided by the Japan Meteorological Agency. By training a model with pairs of coarsened RRAP data and original data, we trained learning models for restoring the high-resolution precipitation fields from the coarsened rain fields.
Deep learning models trained by the commonly used Mean Square Error (MSE) for the loss function tended to produce spatially smooth precipitation images because MSE fails to evaluate shape or structure effectively. In contrast, precipitation images generated by the WGAN have spatially fine precipitation fields as original images. It was found that the CNN was not able to restore high-frequency precipitation patterns sufficiently, whereas the WGAN successfully restored almost all high-frequency precipitation patterns.
This study proposes a method that employs the optimal transport cost (Wasserstein Distance) as the loss function for super-resolving precipitation fields. Because the Wasserstein Distance can quantitatively assess the distance between specific distributions as the loss function, it is expected to produce spatially finer precipitation images than the conventional deep learning models. However, direct calculation of the Wasserstein Distance for two-dimensional rainfall data is numerically difficult; therefore, the framework of Wasserstein Generative Adversarial Networks (WGAN) is adopted in this study. We train two networks concurrently: a generator that conducts super-resolution, and a discriminator that provides the Wasserstein Distance. A well-trained discriminator efficiently estimates the Wasserstein Distance, and feeds this output back to the generator so that the generator enables super-resolution based on the optimal transport cost.
In this study, we used the Radar/Raingauge-Analyzed Precipitation (RRAP) provided by the Japan Meteorological Agency. By training a model with pairs of coarsened RRAP data and original data, we trained learning models for restoring the high-resolution precipitation fields from the coarsened rain fields.
Deep learning models trained by the commonly used Mean Square Error (MSE) for the loss function tended to produce spatially smooth precipitation images because MSE fails to evaluate shape or structure effectively. In contrast, precipitation images generated by the WGAN have spatially fine precipitation fields as original images. It was found that the CNN was not able to restore high-frequency precipitation patterns sufficiently, whereas the WGAN successfully restored almost all high-frequency precipitation patterns.