17:15 〜 18:30
[AAS04-P01] 3D Precipitation Nowcasting: RESNet applied to Highly Dense PAWR Data
キーワード:Precipitation Nowcasting, Residual Neural Networks, Weather Forecasting
Sudden heavy rain may lead to disasters like flooding and loss of life and property. To reduce the risk, predicting sudden downpours is of key importance. However, predictability of such events is limited to only for a very short range within an hour or shorter because of their abruptness. In this case nowcasting is an effective approach. Detecting sudden heavy rain even 10 minutes before it occurs can reduce the damage drastically.
Precipitation nowcasting is the process of short-range prediction based on observation data. In the case of sudden rainfalls, this process is difficult due to the fast evolution of the rain and its chaotic nature. Therefore, we need innovative techniques.
The novel Phased-Array Weather Radar (PAWR) offers dense 3D images of reflectivity every 30 seconds. We took advantage of this big data to perform nowcasting using neural networks. We use Residual Neural Networks (RESNet) to compress the images and extract information relevant for the prediction. Next, we use a Convolutional Neural Network (CNN) to make the prediction. Afterwards, we use the same RESNet to map the forecast to the original domain. The RESNet and the CNN are trained jointly for the compression to maximize the prediction accuracy. Our first results show that in most cases we can predict precipitations up to 30 minutes, with an error rate (false positives + false negatives) of 8% . The use of the RESNet allowed to alleviate the memory load and the computational complexity of the prediction. Moreover, training the RESNet and the CNN jointly reduced immensely the prediction noise in non-precipitation regions and improved the accuracy in precipitation regions.
Precipitation nowcasting is the process of short-range prediction based on observation data. In the case of sudden rainfalls, this process is difficult due to the fast evolution of the rain and its chaotic nature. Therefore, we need innovative techniques.
The novel Phased-Array Weather Radar (PAWR) offers dense 3D images of reflectivity every 30 seconds. We took advantage of this big data to perform nowcasting using neural networks. We use Residual Neural Networks (RESNet) to compress the images and extract information relevant for the prediction. Next, we use a Convolutional Neural Network (CNN) to make the prediction. Afterwards, we use the same RESNet to map the forecast to the original domain. The RESNet and the CNN are trained jointly for the compression to maximize the prediction accuracy. Our first results show that in most cases we can predict precipitations up to 30 minutes, with an error rate (false positives + false negatives) of 8% . The use of the RESNet allowed to alleviate the memory load and the computational complexity of the prediction. Moreover, training the RESNet and the CNN jointly reduced immensely the prediction noise in non-precipitation regions and improved the accuracy in precipitation regions.