RELIABILITY ESTIMATION OF 6-MONTH PRECIPITATION PREDICTION IN ASIAN MONSOON REGION USING DEEP LEARNING

Abstract

 Precipitation forecasting covering the entire rainy season is a key to water resource management in the Asian monsoon region. This study developed two models, one for 6-month rainfall prediction using convolutional neural networks and the other for predicting its reliability. 6930 samples from 42 different models participating in CMIP6 were used to predict the accumulated rainfall over the Asian monsoon region from May to October. The anomaly correlation coefficient between the predicted values and the true values was 0.89, indicating the validity of the model. On the other hand, the test results revealed that about 15% of the sample had poor prediction performance. This study proposes a “reliability index” to identify this sample group. A new model was trained with discrete values of rainfall, and the “reliability index” of each sample was calculated from the predicted probability distribution of rainfall. The results show that the reliability index can discriminate sample groups with poor prediction performance in advance.