[M-GI26] Data-driven approaches for weather and hydrological predictions

[MGI26-10] Deep Learning for Boreal Summer Intraseasonal Oscillation (BSISO) Prediction and Explainability

*Yuki Maeda¹, Masaki Satoh¹ (1.Atmosphere and Ocean Research Institute, The University of Tokyo)

Keywords:Boreal Summer Intraseasonal Oscillation, Deep Learning, Explainable AI

This study uses deep learning to investigate the monthly scale predictability of BSISO (Boreal Summer Intraseasonal Oscillation) regarding what factors contribute to the predictability. The BSISO is one of the most pronounced sub-seasonal variability in the global monsoon system in summer. Compared with the Madden-Julian Oscillation (MJO) in winter, the structure of BSISO is more complicated, with northward propagation over the northern Indian Ocean and western North Pacific as well as eastward extension along the equator. The BSISO profoundly influences various space-time scale phenomena in the tropics and the extratropics. However, the numerical models have large gaps between their forecasting skills and potential predictability, which is a problem for sub-seasonal predictability. In this research, we use a convolutional neural network (CNN), one of the deep learning approaches for statistical forecasting of the BSISO. Reanalysis and satellite observation data are used as input data, with training data from 1979 to 2014 and validation data from 2015 to 2021. Our results show that the correlation skill of the BSISO indices of the CNN model is higher than that of existing numerical models. To extract the interpretable predictive information of the CNN model, we perform an analysis using "eXplainable AI (XAI)". Due to their complexity, we show this is a powerful tool for analyzing deep learning models considered "black box" models. We will further extend this method to study the predictability of the (western) North Pacific High and the relationship between the North Pacific High and BSISO variability.