Dr. Dong-Kyun Kim Profile

Dr. Dong-Kyun Kim

at Osaka Univ

SPIE Involvement:

Author

Publications (2)

Proceedings Article | 14 November 2021 Poster + Paper

Comparative study of deep neural networks for very short-term prediction of torrential rains using polarimetric Phased-Array Weather Radar (MP-PAWR)

Philippe Baron, Dong-Kyun Kim, Kohei Kawashima, Hiroshi Hanado, Takeshi Maesaka, Tomoaki Mega, Shinsuke Satoh, Eiichi Yoshikawa, Seiji Kawamura, Katsuhiro Nakagawa, Tomoo Ushio

Proceedings Volume 11914, 1191416 (2021) https://doi.org/10.1117/12.2601830

KEYWORDS: 3D modeling, Meteorology, Radar, Convolution, Neural networks, Polarimetry, Motion models, Atmospheric modeling

Read Abstract +

Under the Japanese Cross Strategic Innovation Promotion Program (SIP), studies are conducted to perform very short-term predictions of local torrential rains based on a new multi-parameter phased-array weather radar (MP-PAWR) and deep neural networks (DNNs). The association of the two methods is expected to overcome the limitations of the conventional rains observation systems and numerical models that are not well suited to handle the rapid non-linear processes inherent in heavy convective rains. The unique spatio-temporal resolution of the observations allows us to train supervised DNNs to extrapolate the fast evolution of 3D convective cells. We compared two DNNs (CLM3D and CGRU3D) designed to fully exploit the information in the vertical dimension. Both methods use new techniques involving spatial convolutions in temporal recurrent iterations such as Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRU) ones. The core of CLM3D is a stack of convLTSM2D layers, each of which is applied to a single altitude. CGRU3D uses a multilayer encoderdecoder with convGRU3D layers, each layer is associated with a size of 3D spatial features. Forecasts with a lead-time of 10 min at an altitude of 600 m with a horizontal resolution of about 500 m are compared. The models are tested with different types of heavy precipitation: localized short-lived rains on July 24, 2018 and wide-spread ones on the 29 of the same month. The models are evaluated with respect to a 3D linear advection nowcast model (OF3D) and a persistent one. We found that the DNN and OF3D models perform better on July 24 with similar scores that are significantly higher than those of the persistent model. Considering all rain events, critical success indexes (CSI) of 0.62, 0.53, 0.55 are found for CGRU3D, CLM3D and OF3D, respectively, and 0.43 for the persistent model. Regarding only heavy precipitation, the CSIs show a great variability between 0 and 0.4 on the predictions made that day. These results clearly illustrate the great challenge of nowcasting heavy precipitation. On July 29, none of the models have significantly higher scores than those obtained with the persistent nowcast. The interesting result of this study is that the two DNNs show similar nowcasting skills whatever the intensity and the type of rain, and this despite their architectures and training strategies being different. This may indicate that optimizing the tunning of the hyperpameters and the training dataset could not bring significant improvements and, the key, could be by feeding the models with more comprehensive information on the atmospheric state.

Proceedings Article | 12 September 2021 Poster + Paper

Very short-term prediction of torrential rains using polarimetric phased-array radar (MP-PAWR) and deep neural networks

Philippe Baron, Hiroshi Hanado, Dong-Kyun Kim, Seiji Kawamura, Takeshi Maesaka, Katsuhiro Nakagawa, Tomoo Ushio

Proceedings Volume 11859, 118590T (2021) https://doi.org/10.1117/12.2598915

KEYWORDS: 3D modeling, Radar, Convolution, Motion models, Reflectivity, Polarimetry, Neural networks, Performance modeling, Volume segmentation

Read Abstract +

This study is about the development of a deep neural network to make very short-term predictions of torrential rains at the urban scale (meso-γ). The new polarimetric Phased Array Weather Radar (MP-PAWR) operating at Saitama (Japan) since 2018 is used. Thanks to the unique spatio-temporal resolution of the measurements, the precursors of torrential rains are detected aloft more than 20 minutes before the rain occurs. With this information, we aim at the prediction of surface precipitation with a lead time of 20 min, a horizontal resolution better than 500 m within a radius of 25 km around the instrument. Two supervised neural networks are considered to extrapolate radar reflectivity (Z_H) at the altitude of 600 m. The first model (model-1) is based on a technique developed for mesoscale predictions from observations at a single altitude. It uses horizontal (2D) convolutions in gated recurrent time iterations and a multilayer encoder-decoder (EC/DC) architecture. The technique is adapted to consider 3 radar parameters and 11 altitudes up to 10 km, in the same way as RBG channels in video analysis. The second model (model-2) uses similar architecture but with 3D spatial convolutions to properly describe the vertical motions between adjacent layers. The input to the models consist in 20 min long time series of ZH, Doppler velocity and differential reflectivity observations (30 sec sampling). The models are trained using all the rain events observed between August and October 2018, and are assessed using local heavy rains observed over a period of 1-hour on July, 24, 2018. The beginning of the rain is first predicted with a lead time of about 5 min, and its evolution is fairly well reproduced to lead times up to about 10 min. Results quickly degrades for longer lead times. We found that a deeper network with 4 layers EC/DC gives better 20 min predictions than a model with 3 layers, but final results were not yet obtained at the time of writing. Regarding lead-times of 10 min, model-2 gives critical success indexes (CSIs) of 0.60 and 0.40 for pixels with Z_H> 10 dBZ and 37 dBZ, which is comparable or better than results presented in other studies. For lead-times of 20 min, CSIs dropped to 0.28 and 0.10, respectively, and no other studies was found for comparison. Model-1 clearly shows poorer performance, especially for high Z_H. However, this approach demands much less calculations and the training lasts only 2 weeks long, namely half of the time spent for model-2. Therefore, it is worth further studying both approaches and potential improvements are discussed.

View contact details

UPDATE YOUR PROFILE

Is this your profile? Update it now.

Sign into your SPIE.org account

Don’t have a profile and want one?

Create an account on SPIE.org

Keywords/Phrases

Search In:

Publication Years