Deep Learning Short-term Heavy Rainfall Forecasting Using Pseudo Data

Heavy rainfalls cause great disasters all over the world. Predicting such rainfalls is challenging, and many studies have attempted to predict them for a long time. Recently, deep learning (DL) approaches have outperformed previously existing ones in other fields. Some studies also have begun to use it for rainfall prediction. However, few studies applied the method to heavy rainfall prediction. Since the accuracy of the DL depends on the training sample size, relatively rare heavy rain would be unsuitable for the DL approaches due to its less sample size than general rainfall one. One of the ways to overcome the data imbalance issue is using pseudo data generated by data augmentation (DA). DA is a powerful tool for imbalanced datasets in the field of image recognition, which is a method to increase the sample size of minor data with a transformation of original data.

In this study, we constructed a DL rainfall prediction model specialized for heavy rains using the DA method. Radar AMeDAS Precipitation (RAP) data, provided by Japan Meteorological Agency, were used. The spatiotemporal resolution is 1x1 km and 30 minutes. We divided the data into 12 squared tiles, which side is 256 km. We split the data as follows: from 2006 to 2012 as a training set, from 2013 to 2015 as a validation set, and from 2016 to 2018 as a testing set.

We used U-Net model. The input data were the previous 6 hours' precipitation distribution (every 30 minutes, continuous values), and the teacher data was the 6 hours ahead precipitation distribution (every 60 minutes, categorical values).

We prepared 2 cases to evaluate the impact of pseudo rainfall data. CASE1: we trained the model on the only original RAP data. CASE2: we trained the model on the RAP data plus pseudo one. We generated pseudo data using rotating and zooming transformation and got various advection and areas of precipitation.

As a result, it was found that CASE2 was more accurate than CASE1 when we predicted heavy rainfall. For example, "Northern Kyushu heavy rainfall in July 2017", which resulted in over forty victims, the CASE2 model was able to predict, particularly the two characteristic precipitation bands 6 hours before they occurred.

We successfully applied the DA method to DL heavy rainfall prediction. In the future, we will investigate how much we should appropriately rotate and zoom up the data.