[A-HW21] Interdisciplinary approach to support climate change adaptation measures in regional scale

[AHW21-05] S2S prediction challenge with deep learning in Thailand

*Kiyoharu Hasegawa¹, Shinjiro Kanae¹ (1.Toyko Institute of Technology)

Keywords:deep learning, 1 month rainfall prediction, Thailand, CMIP5, Sea surface temperature, ECMWF

Subseasonal to seasonal (S2S) prediction still remains as a challenge task even today. S2S prediction refers to a lead time ranging from one to several months, which is known to result in rapidly decreasing forecast accuracy. Deep learning (DL), one of the most powerful statistical models, is expected to overcome current difficulty of S2S prediction because of its spatio-temporal locality. In this study, one-month rainfall prediction is carried out for the Chao Phraya River basin in Thailand during the rainy season from May to October, where the introduction of long-term rainfall prediction to dam management is urgently required for flood mitigation. The DL model is constructed using global maps of sea surface temperature and heat content as input values. 17 climate models output from CMIP5 dataset is used to expand the training data from 65 existing observations to 2500, which enables us to train the DL model proper. The prediction accuracy of the DL model is compared with that of the physical model and the linear regression model. The results show that the DL model is slightly inferior to the physical model, but has higher prediction accuracy than the linear model. In addition, the deep learning model has the highest prediction accuracy when trained only on the CMIP5 dataset, indicating that the DL model has potential to outperform the physical model in the future.