Improvement of decadal predictions of monthly extreme Mei-yu rainfall via a causality guided approach

While the improved performance of climate prediction systems has allowed better predictions of the East Asian Summer Monsoon rainfall to be made, the ability to predict extreme Mei-yu rainfall (MYR) remains a challenge. Given that large scale climate modes (LSCMs) tend to be better predicted by climate prediction systems than local extremes, one useful approach is to employ causality-guided statistical models (CGSMs), which link known LSCMs to improve MYR prediction. However, previous work suggests that CGSMs trained with data from 1979-2018 might struggle to model MYR in the pre-1978 period. One hypothesis is that this is due to potential changes in causal processes, which modulate MYR in different phases of the multidecadal variability, such as the Pacific Decadal Oscillation (PDO). In this study, we explore this hypothesis by constructing CGSMs for different PDO phases, which reflect the different phases of specific causal process, and examine the difference in quality as well as with respect to difference drivers and thus causal links between CGSMs of different PDO phases as well as the non-PDO phase specific CGSMs. Our results show that the set of predictors of CGSMs is PDO phase specific. Furthermore, the performance of PDO phase specific CGSMs are better than the non-PDO phase specific CGSMs. To demonstrate the added value of CGSMs, the PDO phase specific versions are applied to the latest UK Met Office decadal prediction system, DePreSys4, and it is shown that the root-mean squared errors of MYR prediction based on PDO phase specific CGSMs is consistently smaller than the MYR predicted based on the direct DePreSys4 extreme rainfall simulations. We conclude that the use of a causality approach improves the prediction of extreme precipitation based solely on known LSCMs because of the change in the main drivers of extreme rainfall during different PDO-phases.