[A-AS08] Multiple scale structure and their interactions in Asian monsoon system

[AAS08-P06] Constraining rainfall bias in tropics with inpainting deep learning

*Kiyoharu Hasegawa¹, Shinjiro Kanae¹ (1.Tokyo institute of technology)

Keywords:deep learning, raifall, CMIP6, inpainting, tropics

Thanks to their remarkable development, numerical models have shown high prediction accuracy in short-term rainfall forecasting. On the other hand, in regions where rainfall is closely related to complex meteorological phenomena such as monsoon including tropical regions, there is still a large uncertainty. Deep learning, a powerful statistical method, can learn spatial characteristics and supplement missing values from surrounding data. Applying this method, we propose a method to reproduce rainfall distribution in the tropics from surrounding rainfall information.
Using the CMIP6 rainfall dataset, rainfall in regions other than the entire northern region of Thailand is learned from 65 ensembles of global rainfall data, and the model outputs rainfall within the entire northern region.
The results showed that the deep learning model tended to overestimate the actual rainfall, but was able to accurately reproduce the rainfall distribution. This method is expected to be applied in practice because of its low computational cost and limited data requirements. It also has the potential to be applied to global regions.