14:15 〜 14:30
[MGI26-03] Short-term deep-learning rainfall prediction in Japan
キーワード：深層学習、豪雨、降水予測、防災
In Japan, the number of heavy rainfall disasters is increasing yearly. As an aspect of evacuation, predicting such rainfall rapidly and accurately is important. In recent years, the deep-learning precipitation forecasting method has been spotlighted. While improving the accuracy of numerical weather models is important, it is also necessary to consider a data-driven perspective. However, the study for such a deep learning method is still insufficient in Japan. Also, most deep-learning rainfall prediction studies focused on ordinal rain (i.e., light rain). Therefore, this study aims to develop a precipitation forecast model available in the Japanese region and can predict precipitation of 50 mm h-1 or more.
The data used in this study is the "Radar AMeDAS Precipitation" provided by the Japan Meteorological Agency (JMA). This data is a rainfall distribution observed by the JMA C-band radar, corrected by ground rainfall, and has a 1 km squared spatial resolution. This product is generally regarded as the ground truth in Japan. We divided this product into 13 tiles as a 256x256 grid square and trained the single model using all tiles. Also, tiles that meet the following conditions were extracted as training tiles: when one precipitation map contains 0.5% or more precipitation grids of 50 mm h-1 or higher. The data from 2006 to 2012 was used as the training data, 2013 to 2015 as the validation data, and 2016 to 2018 as the test data.
We employed the U-net and added “Attention Gate (AG)” and “Deformable Convolution V2 (DC).” AG was added after skip connections, and DC was added to deconvolution blocks. This model was trained to predict the precipitation every hour for 3 hours ahead using 3-step precipitations up to 2 hours ago, including current observation. In the testing process, the predicted three steps were recursively used as input to predict the next 3-hour precipitations. This made it possible to predict up to 6 hours in advance.
We compared our prediction results with JMA's results from operational prediction, named short-term precipitation forecast. We evaluated these results using the Precision, Success Ratio, and Threat score by extracting the days when precipitation occurred from the test period data and using the bootstrap method.
When the threshold was set at 5 mm h-1, the two models had almost the same scores up to 3 hours ahead. On the other hand, in the subsequent predictions from 4 hours to 6 hours ahead, the JMA’s prediction tended to overestimate, but our model did not show such a tendency. Furthermore, our model showed significantly higher predictive accuracy than the short-term precipitation forecast when the threshold was set at 50 mm h-1. Although there was a slight tendency to overestimate, it could predict the heavy rainfall 1 hour and 2 hours in advance.
The sample size of the winter data used in this study was relatively small compared to the sample size of summer data. However, it was found that our model also had high predictive performance in winter. This is because precipitation characteristics are partly common in summer and winter. Since precipitation can also occur in winter due to low pressure and front, as in summer, it was possible to use the learning features from summer data for winter prediction, which is called transfer learning.
In this study, our model still tends to overestimate and have a short lead time in heavy rain forecasting. Performing transfer learning and training on heavy rain data from Japan and other countries, as well as additional meteorological factors such as water vapor and wind speed, would improve the accuracy of our model.
The data used in this study is the "Radar AMeDAS Precipitation" provided by the Japan Meteorological Agency (JMA). This data is a rainfall distribution observed by the JMA C-band radar, corrected by ground rainfall, and has a 1 km squared spatial resolution. This product is generally regarded as the ground truth in Japan. We divided this product into 13 tiles as a 256x256 grid square and trained the single model using all tiles. Also, tiles that meet the following conditions were extracted as training tiles: when one precipitation map contains 0.5% or more precipitation grids of 50 mm h-1 or higher. The data from 2006 to 2012 was used as the training data, 2013 to 2015 as the validation data, and 2016 to 2018 as the test data.
We employed the U-net and added “Attention Gate (AG)” and “Deformable Convolution V2 (DC).” AG was added after skip connections, and DC was added to deconvolution blocks. This model was trained to predict the precipitation every hour for 3 hours ahead using 3-step precipitations up to 2 hours ago, including current observation. In the testing process, the predicted three steps were recursively used as input to predict the next 3-hour precipitations. This made it possible to predict up to 6 hours in advance.
We compared our prediction results with JMA's results from operational prediction, named short-term precipitation forecast. We evaluated these results using the Precision, Success Ratio, and Threat score by extracting the days when precipitation occurred from the test period data and using the bootstrap method.
When the threshold was set at 5 mm h-1, the two models had almost the same scores up to 3 hours ahead. On the other hand, in the subsequent predictions from 4 hours to 6 hours ahead, the JMA’s prediction tended to overestimate, but our model did not show such a tendency. Furthermore, our model showed significantly higher predictive accuracy than the short-term precipitation forecast when the threshold was set at 50 mm h-1. Although there was a slight tendency to overestimate, it could predict the heavy rainfall 1 hour and 2 hours in advance.
The sample size of the winter data used in this study was relatively small compared to the sample size of summer data. However, it was found that our model also had high predictive performance in winter. This is because precipitation characteristics are partly common in summer and winter. Since precipitation can also occur in winter due to low pressure and front, as in summer, it was possible to use the learning features from summer data for winter prediction, which is called transfer learning.
In this study, our model still tends to overestimate and have a short lead time in heavy rain forecasting. Performing transfer learning and training on heavy rain data from Japan and other countries, as well as additional meteorological factors such as water vapor and wind speed, would improve the accuracy of our model.