11:00 〜 13:00
[MGI35-P14] Nowcasting localized torrential rain using multi-parameter weather radar and 3D convolutional recurrent neural networks
キーワード：Machine learning, Nowcast, Radar, Precipitation, ConvGRU
Localized torrential rains are sudden convective storms occurring on a small area of about 5x5 km2 in a few tens of minutes. They are responsible of severe flooding causing infrastructure damages, often with fatalities, and their frequency is expected to increase with climate change. Their prediction at least a few tens of minutes in advance with a high spatial resolution is called nowcast. The best applicable real-time nowcasts do not exceed 10 minutes lead-times with conventional observing and forecasting methods due to the fast evolution and narrowness of these storms, as well as because of strong non-linear processes involved in their formation and dissipation.
Under the Japanese Cross Strategic Innovation Promotion Program (SIP), a new Multi-Parameter Phased Array Weather Radar (MP-PAWR) operating in the X-band has been developed by Toshiba Co., Osaka University and NICT. The instrument is exploited in Saitama University (35.86N, 139.60E) since 2018. Combining mechanical horizontal scanning with electronic vertical one, the instrument is able to scan the whole surrounding atmosphere up to of 60 km from the instrument every 30 sec with a high spatial resolution. Hence the evolution of the thunderstorms can be described in details from the early phase of the cloud formation detectable above the altitude of 3 km. This information is essential to properly initiate torrential rain nowcasting models.
In this presentation we will study nowcasts of localized thunderstorms within a radius of 30 km (250 m resolution) from the MP-PAWR with lead-times of up to 20 min. The 3D spatial information of the MP-PAWR observations (radar reflectivity, Doppler velocity and differential reflectivity up to the altitude of 10 km) are analyzed to extrapolate in time the radar reflectivity at a given altitude, a proxy of the rainfall intensity. This is accomplished using a multi-layer encoder-decoder type supervised neural network [1]. The model uses a 3D version of the 2D convolutional recurrent network unit proposed in [2] and is composed of 4 layers related to features at different spatial resolution. The model is trained with wide range of summer and fall rainfall types observed between 2018 and 2020 which corresponds to more than 40000 sequences. Reduced area of 8x8 km2 are used for the training. When optimizing the model parameters, a penalty from a discriminator is added to the loss function using a technique taken from Generative Adversarial Networks (GANs).
[1] Baron et al. (2021): “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. https://doi.org/10.1117/12.2598915
[2] Shi, X. et al. (2015): “Convolutional LSTM network: A machine learning approach for precipitation nowcasting," in [NIPS].
Under the Japanese Cross Strategic Innovation Promotion Program (SIP), a new Multi-Parameter Phased Array Weather Radar (MP-PAWR) operating in the X-band has been developed by Toshiba Co., Osaka University and NICT. The instrument is exploited in Saitama University (35.86N, 139.60E) since 2018. Combining mechanical horizontal scanning with electronic vertical one, the instrument is able to scan the whole surrounding atmosphere up to of 60 km from the instrument every 30 sec with a high spatial resolution. Hence the evolution of the thunderstorms can be described in details from the early phase of the cloud formation detectable above the altitude of 3 km. This information is essential to properly initiate torrential rain nowcasting models.
In this presentation we will study nowcasts of localized thunderstorms within a radius of 30 km (250 m resolution) from the MP-PAWR with lead-times of up to 20 min. The 3D spatial information of the MP-PAWR observations (radar reflectivity, Doppler velocity and differential reflectivity up to the altitude of 10 km) are analyzed to extrapolate in time the radar reflectivity at a given altitude, a proxy of the rainfall intensity. This is accomplished using a multi-layer encoder-decoder type supervised neural network [1]. The model uses a 3D version of the 2D convolutional recurrent network unit proposed in [2] and is composed of 4 layers related to features at different spatial resolution. The model is trained with wide range of summer and fall rainfall types observed between 2018 and 2020 which corresponds to more than 40000 sequences. Reduced area of 8x8 km2 are used for the training. When optimizing the model parameters, a penalty from a discriminator is added to the loss function using a technique taken from Generative Adversarial Networks (GANs).
[1] Baron et al. (2021): “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. https://doi.org/10.1117/12.2598915
[2] Shi, X. et al. (2015): “Convolutional LSTM network: A machine learning approach for precipitation nowcasting," in [NIPS].