Measuring And Evaluating Seasonal Predictability Of The Asian Summer Monsoon

In this thesis, two types of predictability metrics, the observation-free potential measures based on information theory and signal-to-noise ratio (SNR) and the observation-involved actual measures based on verification statistics, were used to evaluate the seasonal predictability of the Asian summer monsoon (ASM). Using the hindcasts of the state-of-the-art coupled models from the National Centers for Environmental Prediction (NCEP) and the EU-funded ENSEMBLES project, we first explored the seasonal predictability of the ASM in general and then investigated specifically the seasonal predictability of one subsystem of the ASM-the East Asian summer monsoon (EASM). Several important issues related to the seasonal predictability of the ASM were addressed, such as the ability of the potential predictability measures in characterizing model forecast skills.Using the hindcasts of the single NCEP climate forecast system (CFS), we found that mutual information (MI), the information-based average potential predictability measure, can explain, to a large extent, the variation in overall deterministic prediction skill for the ASM seasonal prediction, especially anomaly correlation (AC) skill. Theoretical analysis, based on the assumption of Gaussian forecast and climatological probability distributions, revealed that the MI can measure more potential predictability than the SNR if there is a nonlinear statistical dependence between the ensemble mean (a prediction) and an ensemble member (a hypothetical observation), specifically manifested by the nonzero case-to-case variability of ensemble spread. This is because that MI essentially measures the general statistical dependence between the ensemble mean and an ensemble member, potentially nonlinear, whereas SNR only measures the linear correlation between them. Further, the practical comparison showed that MI is significantly at odds with SNR in two areas, Middle East and Northern Australia. However, within the EASM domain, MI almost coincides with SNR, suggesting the suitability of the constant spread assumption over there. As the information-based potential predictability measure for an individual prediction, relative entropy (RE) has a good relationship with C, the contribution of an individual forecast to overall AC skill, and a poor relationship with the absolute error (AE). Further, the poor RE-AE relationship was attributed to the weak case-to-case variability of ensemble spread.Using the retrospective forecasts of the ENSEMBLES coupled models, we found that average SNR can well explain the large scale spatial distributions and lead-time variations of deterministic and probabilistic prediction skills of the multi-model ensemble (MME) for the EASM seasonal prediction, confirming the applicability of the linear statistical theory responsible for the ideal SNR-forecast skill relationship. The predictability pattern over the EASM domain is mainly dominated by the signal pattern. For individual predictions, SNR also has a reasonably good relationship with prediction skills, suggesting the good power of this simple potential predictability measure in indicating the yearly variations of model forecast skills. Further analysis revealed that the variation of predictability from year to year is dominated by the signal, which has a significant year-to-year variation. In contrast, noise represented by ensemble spread shows a very weak case-to-case variability, which again confirms the great suitability of constant spread assumption.It was found that the ENSEMBLES MME outperforms the participating single model ensembles (SMEs) in terms of both deterministic and probabilistic prediction skills for the EASM seasonal prediction, while the SNR estimated using the MME is lower than the estimates using the participating SMEs, which forms a potential predictability-forecast skill contradiction. Furthermore, a statistical model taking into account the remaining stochastic component in the ensemble means of the SMEs, arising mainly from model parameterization uncertainty, was used to explain this contradiction with moderate success.It was found that seasonal predictability of the ASM shows significant lead-time dependence. In general, the longer the lead time, the worse the potential predictability and the forecast skill. Seasonal predictability of the ASM is also dependent on the variable and the region considered. Usually, Low-level zonal wind possesses better overall predictability compared with precipitation. Tropical portion of the EASM is more predictable on average than the Indian summer monsoon and the subtropical part of the EASM. It was found that great potential predictability usually appears in El Nino-Southern Oscillation (ENSO) years, whether using the NCEP SME or the ENSEMBLES MME. Using the NCEP CFS hindcasts, we revealed that RE is highly related to sea surface temperature (SST) and SST-ASM correlation patterns in the model resemble the typical ENSO structure. The different impacts of ENSO on the South Asian summer monsoon and the EASM, the two main components of the ASM, well explain the different potential predictability features in the two regions, in particular, in terms of their interannual variability. Thus, ENSO is the main source of ASM seasonal predictability.In this thesis, we explored the individual potential predictability for the seasonal predictions of the ASM using RE and SNR, and intensively examined their relationships with individual prediction skills, deterministic and probabilistic. The identified relationships undoubtedly encourage us to use the state-of-the-art coupled climate models to not only do the seasonal prediction of the ASM but also estimate a priori the confidence level of a future prediction.