Data Assimilation Research Team Events / Media

The very short-term real-time prediction of localized torrential rains is a difficult challenge that could help the development of a safer society. The typical size and lifetime of the storm events considered in this presentation are 5x5 km and 30 minutes. The use of conventional observation systems and numerical methods fail to give satisfactory predictions of such events. Under the Japanese Cross Strategic Innovation Promotion Program (SIP), a new Multi-Parameter Phased Array Weather Radar (MP-PAWR) is operated in Saitama University (35.86N, 139.61E, Japan). Combining mechanical horizontal scanning with electronic vertical one, the instrument provides polarimetric measurements of precipitations with unprecedented spatial and temporal resolution. It allows us to detect aloft the early phase of convective cells and provides a detailed description of their evolution in the 3 spatial dimensions with a time resolution of 30 sec. This information is essential to properly initiate numerical models for very short-term forecast (nowcast) of torrential rains.
The supervised deep neural network (DNN) is a promising numerical technique to realize nowcasts. It has the capability to capture the non-linear behavior of the convective cells and can be operated in real-time with a relatively small computer. In addition, the model can handle the large number of dimensions of the MP-PAWR observations and can be trained with past observations of various rain types. In this presentation I will describe the model developed in the National Institute of Information and Communications Technology (NICT, Japan) for the past 2 years. It has been designed to exploit altitude information and the high spatio-temporal resolution of the measurements. The model uses 3D spatial resolution in Gated Recurrent Units (convGRU3D), a technique inspired from the 2D version proposed by Shi et al. [2015] for the time extrapolation of 2D weather radar maps. It has been trained to make nowcasts of rainfall rates at an altitude of 0.6 km with a lead-time of 20 min and a horizontal resolution of 0.5 km.
The NICT model will be compared with another 3D-DNN developed independently at Osaka University (Ushio et al., 2022) and with a 3D optical flow based advection model studied in Riken [Otsuka et al., 2016]. The two DNNs have different architectures and training strategies but use the same MP-PAWR data. Their comparison could highlight some of the limitations of this approach and indicates ways for improvement.
Baron, P., Hanado, H., Kim, D.-K., Kawamura, S., Maesaka, T., Nakagawa, K., and Ushio, T., Very short-term prediction of torrential rains using polarimetric phased-array radar (MP-PAWR) and deep neural networks," Proc. SPIE 11859, Remote Sensing of Clouds and the Atmosphere XXVI, 1185928 (Sept 13, 2021)
Otsuka, S., Tuerhong, G., Kikuchi, R., Kitano, Y., Taniguchi, Y., Ruiz, J. J., Satoh, S., Ushio, T., and Miyoshi, T., Precipitation nowcasting with three-dimensional space-time extrapolation of dense and frequent Phased-Array Weather Radar observations," Weather and Forecasting 31(1), 329-340 (2016).
Shi, X., Chen, Z., Wang, H., Yeung, D., Wong, W., and Woo, W., Convolutional LSTM network: A machine learning approach for precipitation nowcasting," in [NIPS], (2015).
Ushio, T., Kim, D.-K., Baron, P., Wada, Y., Mega, T., Mizutani, F., Wada, M., Yoshikawa, E., Satoh, S., and Hanado, H., Recent progress on the Phased Array Weather Radar at X band," Proc. of IEEE radar conference, submitted (2022).