15:30 〜 15:45
[ATT30-01] Using supervised neural networks to upscale conventional weather radar 4D-resolution
★Invited Papers
キーワード：convLSTM, Precipitation, XRAIN, Phased-Array Weather Radar, Neural Network
Localized heavy precipitations (LHP) occur suddenly in summertime and are becoming more frequent in Japan.
High 4D spatiotemporal resolutions, typically a few hundred meters and less than one minute, are essential for accurately capturing the initiation and development of LHP. The eXtended RAdar Information Network (XRAIN), operated by the Ministry of Land, Infrastructure, Transport and Tourism’s (MLIT), covers most of urban regions in Japan and has been operational for more than 10 years. Therefore, it provides a unique dataset for studying and predicting LHP, as well as assessing their long-term trends. However, XRAIN's limitations, i.e., 5-minute resolution and ~12 elevations, limit its ability to capture LHP details.
In this study, we investigate a machine learning approach to address this limitation. A supervised neural network has been developed to interpolate XRAIN data into high-resolution (HR) data (40 levels up to a height of 8 km, 1 minute time-step). To fill in the gaps in XRAIN data, we trained the model with MP-PAWR data collected in Saitama since 2018. The MP-PAWR provides 4D measurements of rain every 30 seconds with more than 100 elevation levels. The neural network architecture consists of an encoder-decoder structure with multiple LSTM layers. In the LSTM units, standard scalar operations are replaced with 3D spatial convolutions. Various loss functions (e.g., Mean Absolute Error, Structural Similarity Index, ...) are tested to train the model in order to reduce the "blur-effect" inherent with convolutional techniques.
The model will be presented and its ability to properly produce HR features will be discussed based on the analysis of LHP events.
High 4D spatiotemporal resolutions, typically a few hundred meters and less than one minute, are essential for accurately capturing the initiation and development of LHP. The eXtended RAdar Information Network (XRAIN), operated by the Ministry of Land, Infrastructure, Transport and Tourism’s (MLIT), covers most of urban regions in Japan and has been operational for more than 10 years. Therefore, it provides a unique dataset for studying and predicting LHP, as well as assessing their long-term trends. However, XRAIN's limitations, i.e., 5-minute resolution and ~12 elevations, limit its ability to capture LHP details.
In this study, we investigate a machine learning approach to address this limitation. A supervised neural network has been developed to interpolate XRAIN data into high-resolution (HR) data (40 levels up to a height of 8 km, 1 minute time-step). To fill in the gaps in XRAIN data, we trained the model with MP-PAWR data collected in Saitama since 2018. The MP-PAWR provides 4D measurements of rain every 30 seconds with more than 100 elevation levels. The neural network architecture consists of an encoder-decoder structure with multiple LSTM layers. In the LSTM units, standard scalar operations are replaced with 3D spatial convolutions. Various loss functions (e.g., Mean Absolute Error, Structural Similarity Index, ...) are tested to train the model in order to reduce the "blur-effect" inherent with convolutional techniques.
The model will be presented and its ability to properly produce HR features will be discussed based on the analysis of LHP events.