11:00 〜 13:00
[AAS04-P02] Convective storm nowcasting using a deep GAN from high-resolution radar datasets
キーワード：Nowcasting, Convective storm, Convolutional neural network, LSTM, Generative Adversarial Network
Accurate nowcasting of precipitation has been a challenging task in meteorological societies to mitigate damages from various hazardous weather phenomena. Nowcasting techniques utilizing deep neural networks such as Convolutional Neural Network (CNN) or Long Short-Term Memory (LSTM) which can capture spatiotemporal features from sequences of observations have been advanced in the last decade. In recent years, a generative adversarial network (GAN) is drawing more attention in the area of nowcasts of convective precipitation. The GAN is comprised of two sub networks which are called a generator and a discriminator. The latter is a classifier to determine whether the image is real or fake. The GAN can learn through competing processes between them in a way to decrease losses of the generator, which means that it becomes harder for the discriminator to distinguish generated images from real images. This learning continues until the discriminator cannot distinguish the generated images from the real images any longer and then the generator makes predictions of future frames. In this study, a deep GAN model is developed to nowcast precipitation by targeting localized convective storms over a short lead time period (< 20 min). Events of meso-γ-scale convective systems that occurred over the Tokyo metropolitan area on 1 June, 2019 and 24 July, 2018 were chosen and radar datasets observed by the X-band Multi-Parameter Phased Array Weather Radar (MP-PAWR) at Saitama city were used. The MP-PAWR has the merit to advance convective nowcasting at very short-term range since it can finish a 3-D volume scan at a very high resolution of 30 s for the internal structure of convective rain systems.
In order to detect 2-D flows in successive radar images, we use an optical flow method to obtain echo motion vectors of the precipitation systems. Radar reflectivity and Doppler velocity datasets are used as input to the deep GAN and convolutional LSTM (ConvLSTM) models in this study. The datasets are converted to the Cartesian coordinates and the x and y pixel resolution is 300 m. We perform supervised training of the GAN and ConvLSTM models using sequential radar images at consecutive time steps and test them using the unseen test data at the same time steps in order to predict multiple future radar images (i.e., many-to-many sequence prediction). The generator uses ConvLSTM and the discriminator uses CNN for learning the input data. Also, the dual frame and sequence discriminators are used to help resemble the distributions of the generated images to be closer to the real radar images and enhance the performance of the GAN model. For the convective events above, the performance results of the GAN model will be presented as we evaluate it by comparing with predicted results from the ConvLSTM models using statistics and forecast skill measures like Threat Score (Critical Success Index) in this study.
In order to detect 2-D flows in successive radar images, we use an optical flow method to obtain echo motion vectors of the precipitation systems. Radar reflectivity and Doppler velocity datasets are used as input to the deep GAN and convolutional LSTM (ConvLSTM) models in this study. The datasets are converted to the Cartesian coordinates and the x and y pixel resolution is 300 m. We perform supervised training of the GAN and ConvLSTM models using sequential radar images at consecutive time steps and test them using the unseen test data at the same time steps in order to predict multiple future radar images (i.e., many-to-many sequence prediction). The generator uses ConvLSTM and the discriminator uses CNN for learning the input data. Also, the dual frame and sequence discriminators are used to help resemble the distributions of the generated images to be closer to the real radar images and enhance the performance of the GAN model. For the convective events above, the performance results of the GAN model will be presented as we evaluate it by comparing with predicted results from the ConvLSTM models using statistics and forecast skill measures like Threat Score (Critical Success Index) in this study.