2-5-year SST Prediction Based On CMIP5 And CMIP6 Return Data

Accurate predictions of the Earth's climate for the next few years will benefit agriculture production and the living conditions of the world's population,as well as the formulation of short-term policies.Using the yearly initialized decadal predictions of the sea surface temperature(SST)from the CMIP5 and CMIP6 datasets for the period of 1960-2009 and 1960-2014,respectively,the systematic errors of SST predictions for forecast lead years 2-5 are analyzed.Then three kinds of drift correction methods are adopted to reduce the systematic errors.The prediction skills of CMIP5 and CMIP6 datasets in the North Pacific,the North Atlantic and tropical Indian Ocean,tropical Pacific Ocean are evaluated and intercompared.Two statistical post-processing methods are implemented to improve the prediction skills over the North Atlantic for CMIP5 and CMIP6 datasets for 2-5 forecast lead years.The main conclusions are listed as follows:(1)The sources of systematic errors for SST predictions in the North Pacific and the North Atlantic are analyzed from three aspects: trend error,initial condition error and climatology error.In the North Pacific,each model shows significant trend error and climatology error.The SST trend predicted by climate models in the middle latitudes of the North Pacific is larger than the observed one.The climatology error of colder SST is offset by the positive trend error to a certain extent.In the northwest of the North Atlantic,the SST prediction error for 2-5 forecast lead years includes trend error,initial condition error and climatology error.All models have negative trend error and initial condition error in the northwest of the North Atlantic.The systematic error in the eastern half of the North Atlantic is mainly due to the climatology error.Based on these three error sources,the mean drift correction(MDC),trend-based drift correction(Tr DC)and initial condition-based correction(ICDC)methods are used to remove the relevant errors,respectively.The results show that Tr DC can effectively remove the trend error in the middle latitudes of the North Pacific and the northwest of the North Atlantic.Taking root mean square error(RMSE)as the evaluation metric of the prediction skill,the optimal method of drift correction is found.The results show that Tr DC as the optimal correction method covers the largest area in both the North Pacific and the North Atlantic,MDC follows,and ICDC is the least.The prediction skill improvement percentage of the optimal drift correction method relative to MDC is the highest in the North Pacific region.(2)The evaluation of prediction skills for CMIP5 and CMIP6 climate models for 2-5forecast lead years show that,the prediction skills of CMIP5 and CMIP6 models are very close in the North Pacific and North Atlantic,and the prediction skills of CMIP6 models are higher in the tropical Indian Ocean and tropical Pacific Ocean.In terms of the spatial patterns,only Can CM4 in the CMIP5 dataset is capable of predict the Pacific Decadal Oscillation(PDO)pattern that in agreement with the observations.Although most of the models cannot predict the spatial pattern of the PDO precisely,the accurate prediction of the SST in the mid-latitudes of the North Pacific ensures reliable prediction skills in the North Pacific.In the North Atlantic,all the models except the Can ESM5 in the CMIP6 dataset can reproduce the Atlantic multidecadal oscillation(AMO)mode that in agreement with the observations.The dipole mode index(DMI)and Ni(?)o 3.4 index are used to assess the prediction skill over the tropical Indian Ocean and tropical Pacific Ocean at the first and 2-3th forecast lead years.For the first forecast lead year,the Indian Ocean Dipole(IOD)prediction skill of CMIP6 is higher than that of CMIP5 in terms of correlation coefficient,RMSE and IOD intensity prediction.The prediction skill of ENSO for CMIP6 models are higher than that of CMIP5 during forecast lead years 1and 2-3 in ENSO mode prediction and hit rate,especially the La Ni(?)a hit rate.The reason why CMIP6 has higher prediction skills than CMIP5 in the tropical Pacific Ocean and tropical Indian Ocean is that the DMI and Ni(?)o 3.4 index are correlated in CMIP6 climate models,while no obvious correlation between two indexes in CMIP5 models.(3)The stepwise pattern projection method(SPPM)and the multi-model super-ensemble(MMSE)method are implemented to improve the SST prediction skills averaged over 2–5forecast lead years in an independent training period over the North Atlantic for CMIP5 and CMIP6 hindcasts.The results show that the RMSE can be effectively reduced by using SPPM correction.Based on the comprehensive analysis of anomaly correlation coefficient(ACC)and mean square error skill score(MSESS)for SPPM prediction,SPPM-MIROC5 and SPPMCMCC are the optimal SPPM predictions in CMIP5 and CMIP6,respectively.In addition,the performance of SPPM correction is unstable for certain models,for example,the regional averaged ACC of SPPM-Can CM4 is smaller than that of Can CM4 prediction without SPPM correction.Comparing the prediction skills of multi-model ensemble mean(MME),optimal SPPM prediction and MMSE prediction,the regional averaged ACC of optimal SPPM prediction is the highest among three post-processing predictions.The running training periodbased stepwise pattern projection method(RTSPPM)can effectively solve the problem of unstable prediction skill of traditional SPPM.The regional averaged ACC of RTSPPM prediction of all models is higher than or equal to the corresponding SPPM prediction.Compared with the MMSE with fixed training period,the running training period-based multimodel super-ensemble(RTMMSE)prediction has no significant improvement in prediction skills.