Seasonal prediction skills in the CAMS-CSM climate forecast system

The seasonal prediction skills in the CAMS-CSM (the acronym stands for the Chinese Academy of Meteorological Sciences Climate System Model) climate forecast system is evaluated with a set of retrospective forecast experiments during the period of 1981–2019. The CAMS-CSM, which has been registered for the sixth phase of the coupled model intercomparison project (CMIP6), is an atmosphere–ocean–land–sea ice fully coupled general circulation model. The assimilation scheme used in the forecast system is the 3-dimentional nudging, including both the atmospheric and oceanic components. The analyses mainly focus on the seasonal predictable skill of sea surface temperature, 2-m air temperature, and precipitation anomalies. The analyses revealed that the model shows a good prediction skill for the SST anomalies, especially in the tropical Pacific, in association with El Niño-Southern Oscillation (ENSO) events. The anomaly correlation coefficient (ACC) score for ENSO can reach 0.75 at 6-month lead time. Furthermore, the extreme warm/cold Indian Ocean dipole (IOD) events are successfully predicted at 3- and even 6-month lead times. The whole ACC of IOD events between the observation and the prediction can reach 0.51 at 2-month lead time. There are reliable seasonal prediction skills for 2-m air temperature anomalies over most of the Northern Hemisphere, where the correlation is mainly above 0.4 at 2-month lead time, especially over the East Asia, North America and South America. However, the seasonal prediction for precipitation still faces a big challenge. The source of precipitation predictability over the East Asia can be partly related to strong ENSO events. Additionally, the anomalous anticyclone over the western North Pacific (WPAC) which connects the ENSO events and the East Asian summer monsoon (EASM) can be well predicted at 6-month lead time.


Introduction
Climate events on seasonal time scales are essentially related to societal economy and people's living life. Hence, how to get accurate seasonal forecast has always been an important issue. The prediction methods can be generally divided into two types (Graham 1987;Jin et al. 2008;Tang, 2018): the statistical models (Graham 1987;Ren et al. 2018) and the dynamic models (Kirtman 2003;Zhang et al. 2003;Luo et al. 2005;Zheng et al. 2006;Jin et al. 2008;Zhang and Gao 2016;Zhu et al. 2017). Over the past decades, benefited from advances in climate system models (CSMs) and the improvements in observation network, such as the Tropical Atmosphere Ocean (TAO) program conducted since the 1980s (McPhaden 1995) and the Tropical Ocean-Global Atmosphere (TOGA) program during 1985(McPhaden 1998, the dynamical seasonal prediction has achieved great progresses (Zhang, 2020). Many operational and scientific centers have established seasonal climate prediction system based on dynamical methods (Saha 2006;Molteni et al. 2011;Becker et al. 2014;Liu et al. 2015;Johnson et al. 2019).
In addition to the inherent performance of CSMs, the dynamical seasonal prediction skill, on a certain extent, is dependent on the data assimilation system (Moore and Kleeman 1996;Chen et al. 1997;Moore et al. 2006;Jin et al. 2008). Unlike the predictable source of the traditional weather forecast within about 10 days, which mainly comes from the atmospheric initial conditions, the predictable source of seasonal, or even longer time period, forecasts mainly is attributed to the slow-varying processes of the climate system, which is largely resided in the ocean memory. Thus, the ocean initialization plays a vital role in the seasonal forecast system.
To obtain accurate ocean initiation conditions (ICs), a variety of data assimilation methods have been developed to assimilate the ocean observations, including the real-time in situ observations and the satellite products (Behringer and Xue 2004;Balmaseda et al. 2013). The relatively advanced and sophisticated data assimilation schemes include the 3-dimensional variational method (3D-Var) (Lorenc 1986;Barker et al. 2004;Cazes Boezio and Ortelli 2019), the 4-dimensional variational method (4D-Var) (Sugiura et al. 2008;Gao et al. 2016Gao et al. , 2018, and the ensemble Kalman filter (EnKF) (Evensen 1994;Liu et al. 2004;Zhang et al. 2007;Anderson et al. 2009;Eliashiv et al. 2019). By assimilating the sea surface temperature (SST) data into a hybrid coupled model (HCM) using a 3D-Var method, significant improvement could be obtained for ENSO prediction, especially at the 4-month lead time (Tang et al. 2003). The 4D-Var is considered to be more physically consistent. Sugiura et al. (2008) applied the 4D-Var initialization approach to obtain a successful prediction of the 1997/98 El Niño event up to 1.5year lead time. However, since the physical processes and parameterizations are different among the CSMs, the assimilation scheme used in a CSM cannot be directly applied to the others (Tang et al. 2003). Furthermore, although the complex assimilation scheme might obtain a better initial fields, the computational cost could be a major concern when these approaches are applied to high-resolution coupled models, and thus further scientific and technical efforts are still required (Small 2014).
A more economic and practical approach in assimilation is the Newtonian relaxation (nudging) approach, in which the model initial fields are nudged towards the observation or reanalysis with restore coefficients. One of the most widely used nudging schemes is the SST nudging, namely nudging towards the observed SST (Luo et al. 2005(Luo et al. , 2008Keenlyside et al. 2016;Zhu et al. 2017) or SST anomalies (Chen et al. 2004;Ren and Nie 2020). The mechanism for SST-nudging scheme is that not only the realistic temperature in the mixed layer can be reproduced, but also the observed variability of the subsurface temperature can be partially obtained owing to the air-sea interaction and coupling (Keenlyside et al. 2005;Luo et al. 2005;Kumar and Zhu 2018;Wang, 2019). Therefore, this scheme is more suitable for the area where air-sea interaction is strong, such as the tropical Pacific. By implementing the SST nudging scheme into the SINTEX-F, Luo et al. (2005) achieved a remarkable extended ENSO prediction skill up to 2-year lead time. Besides, the prediction on Indian Ocean Dipole (IOD) event, which has important impact on the global climate and is characterized by an opposite SST anomaly between the western and eastern Indian Ocean (Saji et al. 1999), can also be achieved using nudging assimilation in the dynamic climate model. Luo et al. (2007) have shown the extreme positive IOD event can be predicated at a 2-3-season lead using only SST nudging scheme into the SINTEX-F, but the model failed to predict some cold ones owing to model biases in the subsurface conditions. Generally, the IOD skillful prediction is only about 1-2 seasons ahead (Saji et al. 1999;Wajsowicz 2004;Luo et al. 2007;Liu et al. 2016), which is significantly shorter than that of ENSO events. Due to the more complicated physical processes concerning the IOD, such as the Asian/Indian monsoon, the internal dynamics of the Indian Ocean, and the external influences of ENSO, the dynamical coupled model is hardly beat the statistical model (Shi et al. 2012). In order to have substantial skill on other atmospheric variables, such as air temperature and precipitation, more observed data might need to be assimilated as well. Baehr (2014) reproduced the observed surface temperature anomalies at 2-4-month lead time, especially in the tropics, based on the initialization of atmospheric, oceanic and sea ice components of the MPI-ESM climate model with a full field nudging technology. Dunstone (2016) demonstrated the winter North Atlantic Oscillation can be predicted extending near one year ahead which used the 3-dimensional (3D) nudging scheme.
Recently, a state-of-the-art model, called CAMS-CSM, is developed by the Chinese Academy of Meteorological Sciences (CAMS), which is fully coupled with atmosphere-ocean-land-sea ice components (Rong, 2018). The CAMS-CSM has been registered for the sixth phase of the coupled model intercomparison project (CMIP6) (Eyring et al. 2016;Rong et al. 2020) and shown good simulation performance on climatological mean states, seasonal cycles, and major climate variability modes including Madden-Julian Oscillation, IOD and ENSO in pre-industrial and historical experiments Hua et al. 2019;Lu and Ren 2019). The fundamental features of ENSO, including its dominant patterns and phase-locking, and its impact on the East Asian climate are well reproduced in historical experiments. Moreover, the relationship between IOD and ENSO events in the observations is well captured by CAMS-CSM model (Lu and Ren 2019). The purpose of this study is to document a new climate forecast system based on the CAMS-CSM model, which is initialized with both the atmospheric and oceanic components. For the analysis of predictive skill, the analyses mainly focus on the SST, but other variables will be briefly discussed as well.
The rest of this paper is arranged as follows. The climate model, experimental designs, data and methods are introduced in Sect. 2. The performance of the seasonal climate forecast system is evaluated in Sect. 3, including the SST, 2-m air temperature and precipitation anomalies, and climate anomalies related to ENSO over the East Asia-Western North Pacific (EA-WNP) in the ENSO decaying summer. Conclusion and discussion are included in the Sect. 4.

The CAMS-CSM climate model
The forecast system is based on the CAMS-CSM , which was developed at the CAMS. The atmospheric component of CAMS-CSM is a modified version of ECHAM5.4 (Roeckner et al. 2003). It is configured by the T106L31 resolution, with a 1° resolution in horizontal direction and 31 vertical layers extending from surface to 10 hPa. The modifications mainly include a Two-step Shape Preserving Advection Scheme (TSPAS) for passive tracer transport (Yu 1994) and k-distribution scheme for radiation transfer parameterization (Zhang et al. 2006a, b). The oceanic component is the Modular Ocean Model version 4 (MOM4) (Griffies et al. 2004), which is configured by a resolution of about 1° × 1° in zonal and meridional directions and 50 vertical layers. The atmospheric and oceanic components are coupled by the GFDL Flexible Modeling System (FMS) coupler. Moreover, the land model named CoLM (Dai 2003) is included in the atmospheric component, and a five-layer sea-ice model (SIS) is included in MOM4 (Winton 2000).
More details on the model description can be referred to Rong et al. (2018).

Design of the CAMS-CSM hindcast experiments
The nudging initialization scheme is achieved by adding a Newtonian damping term into the prognostic equations, resulting in the model fields adjusted towards observations. In this study, both the atmospheric and oceanic ICs are constructed by the 3D nudging technology. For the atmospheric initial conditions, the 6-h air temperature, relative humidity, the surface pressure, zonal and meridional winds from the 55-year Japanese Reanalysis Project (JRA55) atmospheric reanalysis data (Kobayashi et al. 2015) are assimilated with a restoring time scale of 0.5 days. For the oceanic initial conditions, the pentad temperature and salinity above 1000 m from the National Center for Environmental Prediction (NCEP) GODAS ocean reanalysis data (Behringer and Xue 2004) are assimilated in the global ocean south of 60° N. The restoring time scale is set at 5 days near the equator and gradually increased to 30 days outside of 40°N/S. Eight ensemble simulations have been conducted to calibrate and evaluate the prediction skills, starting once a month from January 1980 to December 2019. The 6-month predictions with initial conditions (ICs) at 0000, 0600, 1200, and 1800 UTC were made on the 20th and 21th of every month, which constitute the above mentioned eight samples. Daily mean data were stored for the several (8-11) days in the initial incomplete month, and they were not used to calculate monthly mean. The following prediction month is the so-called the "month lead time". For example, the prediction at 1-month lead time from January initiation conditions is the forecast for February. Due to the first prediction initiated from January 1980, the 1-6-month lead time prediction results are not completed in the first half year of 1980. Then only the 39-year prediction period from January 1981-December 2019 are used for the following analysis. These 8 runs from the hindcasts form the ensemble that is used for calibration and skill assessment of seasonal predictions.

Verification data and methods
The verification data used in this paper include the SST, precipitation, 2-m air temperature and 850-hPa wind. The observed global monthly SST dataset is the Extended Reconstructed Sea Surface Temperature (ERSST v5) dataset with 2° × 2° resolution (Huang 2017). The 2-m air temperature and 850-hPa wind data are from the 55-year Japanese Reanalysis Project (JRA-55) (Kobayashi et al. 2015). The precipitation data from the Global Precipitation Climatology Project (GPCP) (Adler et al. 2003). All the above reanalysis/ observations cover a 39-yr period from 1981-2019 and are interpolated onto the CASM-CSM model grid. The anomalies of each variable are calculated relative to 1981-2010 climatology. The statistical methods include regression, correlation and physical analysis, and the standard two-tailed Student's t test is used to estimate the significant levels. The anomaly correlation coefficient (ACC) and root-mean-square error (RMSE) between the predicted and the observations are used as measures of the prediction skill.

Impact by nudging scheme to reduce model biases
Before the evaluation of the prediction skill, the effect of the nudging assimilation scheme is investigated by comparing the results between historical experiment and nudging experiment. The differences of the climatological mean of SST and subsurface temperature related to the observations/reanalysis data are compared between the historical experiment and the nudged experiments ( Fig. 1). In the historical experiment (Fig. 1a, c), there are obvious positive temperature biases over the central and eastern tropical Pacific, Atlantic Ocean and the Southern Ocean, with maximum temperature bias above 3 degree. In the northern Pacific and Atlantic, the historical experiment shows some SST biases, which are larger than 2 degree in most regions. Meanwhile, the temperature simulations over the equatorial ocean show similar biases with those in SST biases. The 20-isothermal over the eastern (western) tropical Pacific and Atlantic Ocean is deepening (shallowing) of the observations owing to the positive (negative) temperature biases. However, the temperature is simulated much warmer in the subsurface over the Indian Ocean, corresponding to the much deeper of 20-isothermal in the historical experiment.
After the 3D-nudging assimilation, the results have shown that the simulation biases in the historical run, including the amplitudes and spatial structure of SST and subsurface temperature, have been largely reduced ( Fig. 1b and 1d), resulting in a sea temperature pattern that is much closer to the observations. Owing to the assimilation, the prediction skill should be better guaranteed to a large extent compared to the experiment without the nudging assimilation scheme.

Global SST anomalies
The global distributions of prediction correlation skill of SST anomalies between the observation and predictions are shown in Fig. 2 for 1-, 3-and 6-month lead times during the period 1981-2019. The result shows a high prediction skill of SST at 1-month lead time with large ACC value, with many areas over the tropics and some regions in the mid-latitudes oceans higher than 0.6 (Fig. 2a). The highest predictable regions appear in the central and eastern equatorial Pacific where correlation skill is above 0.9. The model shows high prediction skill (ACC > 0.6) in most regions of the tropical Pacific at 3-month lead time (Fig. 2b). High ACC is still seen in the western tropical Atlantic, and the ACC is higher than 0.5 in most regions in the tropical Indian Ocean. At 6-month lead time, the domain of significant SST predictabilities is mostly found in the central tropic Pacific (Fig. 2c). The predictability becomes weak over the tropical Indian and western tropical Atlantic Ocean. The prediction skills of SST anomalies at 6-month lead time in the CAMS-CSM forecast system are quite comparable with most models Similar to a and b but for the temperature bias along the equator relative to GODAS temperature in c and d, the red solid contour is the 20C isothermal in GODAS, the black solid for the historical and nudged runs, respectively. Unit: degree from the North American Multi-model Ensemble (NMME) project (Zhu et al. 2017).
In order to understand more details about the seasonal predictability of global SST, a view of seasonal difference in anomalous SST prediction skills from different initial times are examined for the lead times of 3 and 6 months ( Figs. 3 and 4), respectively. For these two lead times, the maximum prediction skill of SST anomalies is still over the tropical central and eastern Pacific where the correlation is above 0.6 for all but January ICs at 6-month lead time.
Comparing the predictability at different starting times, the most skillful prediction is that initiated from July and October, with the correlation in the tropical central Pacific up to 0.9 at 3-month lead time (Fig. 3d) and remaining above 0.7 at 6-month lead time (Fig. 4d). The highest correlation (ACC > 0.8) appears for the initiation starting from July at the 6-month lead time (Fig. 4c), indicating the high predictability of SST in winter. As the lead time increases, the corresponding correlation gradually decreased in other oceanic areas. Nonetheless, there are still some significant prediction skills in the tropical Indian and tropical Atlantic Ocean basins at 6-month lead time. Owing to the influence of ENSO, the SST warming event over these two regions tends to occur 4-5 months after the mature phases of ENSO events (Lau and Nath 2003), leading to the better predictability starting in October over the north tropical Indian Ocean and tropical Atlantic Ocean (Fig. 4d). In general, ENSO teleconnections might play an important role in the predictability of SST anomalies in the tropical Indian and Atlantic Ocean. In addition, the Southern Pacific Ocean has the best prediction skill (ACC > 0.6) in spring and summer for prediction that starts from January and July at 3-month lead times (Fig. 3a, c). Guan et al. (2014) had attributed the main reason to the effect of ENSO remote forcing.
The correlation and the RMSE are calculated for different indices from the Pacific, Indian and Atlantic Ocean at each lead month between the observation and the prediction (Fig. 5). Except for the indices for the Pacific temperature, the selection of other indices is based on Luo et al. (2005). The result shows that there is a certain degree of predictability for these three basins and with no apparent initial shock at 1-6-month lead time. The most skillful area lies in the central and eastern Pacific Ocean. The skill (either correlation or RMSE) of the index over the Pacific Ocean, especially the Niño3.4 index, is better than those in the other ocean regions, and most can significantly beat the persistence skill after 2-month lead time. In the Indian Ocean, the Eastern Indian Ocean (EIO) index (averaged SST anomalies in the domain of 90°-110° E, 10°S-0°), decreases rapidly from 0.72 at 1-month lead time to around 0.55 at 3-month lead time, with little prediction skill after 3-month lead time informed by declined correlation and increased RMSE. The prediction skill of the South Indian Ocean (SIO) index remains significant at 6-month lead time, with correlation skill (about 0.72) slightly larger than that (about 0.7) at 4-month lead time. This may be related to the lagged influence of ENSO (Alexander et al. 2002;Luo et al. 2005;Du et al. 2009). Another region which is largely affected by ENSO is the tropical Atlantic He et al. 2020), represented by the ATL index (averaged 60°-40° W, 10°-20° N) here. The model shows certain prediction skill of ATL index with ACC from 0.85 at 1-month lead time deceasing to around 0.6 at 6-month lead time, which are somewhat higher than those for the SIO index. However, the correlation and RMSE of EIO and ATL indices cannot beat their persistence skill in this 6-month lead time prediction. In other word, there are still a large challenge to predict SST anomalies in these two regions in the dynamic climate forecast than the statistical model, which may be caused by the inappropriate processes or inadequate initialization where the dynamic model adopted (Luo et al. 2005;Shi et al. 2012;Liu et al. 2016). The seasonal predictions of ENSO and IOD will be further investigated in the Sects. 3.4 and 3.5.

The 2-m air temperature and precipitation anomalies
The seasonal predictabilities of 2-m air temperature and precipitation anomalies are also evaluated for all four seasons (MAM, JJA, SON, DJF respectively represent boreal spring, boreal summer, boreal autumn and boreal winter) (Figs. 6 and 7). Although the prediction skill of the 2-m air temperature over land is much lower than that of SST, there are significant high correlations during each season in the tropics (Fig. 6), such as the South America, the southern North America, the Maritime Continent, Australia and Africa, with a correlation skill larger than 0.4 at 2-month lead time. The maximum W] indexes as a function of forecast lead months (x axis). The grey lines represent the persistence skill and RMSE for the corresponding index. All the above indexes are calculated by model ensemble mean SSTA area-averaged in the range within square brackets correlation (> 0.7) center appears in the northern South America during MAM (Fig. 6a) and DJF (Fig. 6d). Moreover, the prediction skill seems strongly dependent on regions. For instance, the prediction skill in the central Australia, especially for DJF, is obviously lower than the other places of the Southern Hemisphere, such as Brazil.
Over the mid-latitude of the Northern Hemisphere, there is significant prediction skill in the East Asia where the correlation reaches above 0.4 during JJA (Fig. 6b) and SON (Fig. 6c) at 2-month lead time. The northern North America shows significant prediction skill of the 2-m air temperature in JJA and DJF, while prediction skill is weak in SON and DJF over the high-latitude Eurasia. The prediction of precipitation over land at 2-month lead time (Fig. 7) seems much more difficult than that for SST and 2-m air temperature. The precipitation prediction skill in the Northern Hemisphere mainly lies in the Middle East, the southeast China, the western and eastern North America where the correlation skill is mainly above 0.2 during DJF (Fig. 7d). The predictability of precipitation seems better in the Southern Hemisphere. There are credible prediction skills in the southeast Australia and most regions of Brazil, where the correlation skill can reach above 0.6 in some areas during SON and DJF.

ENSO events
ENSO is a dominant mode on inter-annual timescale in the Pacific Ocean. Many previous studies have shown that the seasonal cycle of the equatorial Pacific SST plays an important role in the development of El Niño events (Latif et al. 2001;Guilyardi 2006). Hence, the seasonal cycle of SST deviation from annual mean over the equatorial Pacific in model predictions at 1-, 3-and 6-month lead times are compared with the observations (Fig. 8).
The predictions at all the three lead times generally are matched well with the observations, including the semiannual cycle and annual cycle in the western and eastern tropical Pacific, respectively. The two main discrepancies appear in the westward extension and magnitude of SST deviations. In short lead time (Fig. 8b), the evolution of annual cycle shows a closer match to the observations, with the positive (negative) SST deviations extending westward to around 160° W (130° W), although the magnitude is relatively smaller in the model prediction. The problem of "too westward propagation and too strong" occurs at both 3-and 6-month lead times (Fig. 8c, d), which could lead to the lower decaying speed of ENSO events and in turn might lead to the seasonal climate prediction biases in the ENSO decaying phase. This situation will be discussed in the Sect. 3.5.
In an ensemble forecast system, the sufficiently large samples might largely reduce the uncertainty and improve the accuracy of prediction. Previous studies have investigated the relationship between the spread and magnitude of ENSO events in an ensemble prediction (Tang et al. 2008;Hu and Kumar 2014). The Niño3.4 index, which is computed by averaging the anomalous SSTs over the domain of 5° S-5° N, 120°-170° W, is used to represent the ENSO events. The scatterplot between the multi-sample of hindcasts for the DJF Niño3.4 index (x axis) and the spread of ensemble members from different ICs (y axis) with six lead times is shown in Fig. 9. The spread (y axis) is taken the absolute value of the difference between each sample and the ensemble mean, and it can be treated as the uncertainty in Niño3.4 index. Firstly, the variance of Niño3.4 index's spread varies with the lead time ( Fig. 9 thick black line in bottom left). The spread, which represents the uncertainty in prediction, becomes bigger with increasing lead time. The spread concentrates within 0.2 °C at the 1-month lead time, but it grows up with a wide distribution between 0 and 0.8 °C at 6-month lead time. Secondly, the amplitude of spread increases as the lead time increases. This phenomenon is quite similar to the results based on CFSv2 for ENSO prediction (Hu and Kumar 2014). Compared to the value of total mean of spread (about 0.3 °C, averaged spread for all lead times), the amplitude of Niño3.4 index (about 3 °C, averaged amplitude for all lead times) is about ten times bigger. In other words, the variance of uncertainty is a smaller amount than that of amplitude, consistent with previous studies (Tang et al. 2008;Hu and Kumar 2014). Moreover, the spread to RMSE ratio, which is measured to judge whether the resulting probabilities are reliable (Johnson and Bowler 2009), is calculated as well (Figures not shown). This ratio about Niño3.4 index mostly lower than 0.6 for 1-6-month lead time, which is lower than the desired value of 1, but that value is comparable to that in the NCEP CFS Reanalysis and Reforecast (CFSRR) Project using CFSv2 model (Zhu et al. 2013). All the above analysis demonstrates a reasonable spread range of the Niño3.4 index prediction. Indeed, the ensemble size seems need to increase more, or other new perturbations should be included with respect to improve the ENSO prediction skill. The time series of Niño3.4 index during the period of 1981-2019 based on the ERSSTv5 observation and model prediction at 1-, 3-and 6-month lead times are compared to inspect the ENSO prediction skill (Fig. 10a). All of them have been smoothed by 3-month running mean. The results show that the ENSO during most periods can be well predicted at each lead time. The standard deviation of interannual Niño3.4 index at each lead time is around 0.95 ºC, which matches well with the observation (0.98 °C). The three strongest El Niño events (1982/83, 1997/98 and 2015/16) can be successfully predicted at 6-month lead time, but the amplitudes are slightly smaller than the observations. Moreover, the stronger La Niña events, such as 1988Niña events, such as /89, 1998Niña events, such as /99, 2000Niña events, such as /01, and 2011, can also be successfully predicted at 6-month lead time, and the predicted amplitudes are very close to the observations. However, the biases and uncertainties become much bigger for the weaker ENSO or neutral events, such as 1993/94, 2012/13 and 2017/18. ENSO prediction has a "spring barrier" feature both in observations and models (Latif 1994;Tang et al. 2018), which means the predicted skill of ENSO decreases rapidly during spring season. It is clear that the intrinsic "spring barrier" situation also exists in the CAMS-CSM climate forecast system (Fig. 10b, c). A significant lower center of ACC (< 0.75) lies from January to April at 3-6 months lead time (Fig. 10b), while a higher center of ACC (> 0.95) exists from July to December at 1-3 months lead time, corresponding with the higher (> 0.65 ºC) and lower (< 0.3 °C) RMSE center respectively (Fig. 10c). The predicted skill in spring at 6-month lead time can reaches 0.7, which is comparable with the performance in SINTEX-F (Luo et al. 2005). Some researches demonstrated that the spring barrier came from the phase locking of ENSO to the seasonal cycle (Wu et al. 2009). During the springtime, the phase transition of ENSO results in a low signal-to-noise ratio (Xue et al. 1994). , which the vertical length represents the standard deviation and its center point corresponds to the value on the y-axis represents the average spread. The unit is °C Some researches emphasized the lower variance of anomalous SST in spring (Fan et al. 2000). Although lots of effort have been made to reduce the spring prediction barrier, this defect still exists in the most widely used climate models (Luo et al. 2005;Becker et al. 2014;Zhu et al. 2017). However, the forecast system in CAMS-CSM still shows highly credibility to ENSO prediction at 6-month lead time, with the ACC of Niño3.4 index reaching about 0.88 and 0.75 at 3-and 6-month lead times, respectively (Fig. 10a). The RMSE (about 0.63 °C) of Niño3.4 index at 6-month lead time is much lower than one standard deviation of 0.98 °C (Fig. 10b).

The Indian ocean dipole
The Indian Ocean Dipole (IOD) has significant impacts on East Asian and Western North Pacific (EA-WNP) climate (Ashok et al. 2004;Nur'utami and Hidayat 2016;Xu and Guan 2017;Jiang et al. 2019). Although it is still a great challenge to predict the IOD, the extreme warm/cold IOD events (e.g., 1995, 1998, 1999 and 2017) are well predicted at 3-month, even at 6-month lead time (Fig. 11a). Similar to the ENSO prediction, the prediction of IOD events is also difficult in spring, but it seems that the prediction barrier appears earlier. The IOD prediction skill at the 2-month lead time rapidly decreases from around 0.6 during winter time to 0.2 in spring (Fig. 11b). The RMSE at 6-month lead time has the highest value during spring (Fig. 11c). Meanwhile the prediction skill is still low in autumn and winter. However, the prediction skill of IOD index is higher starting from May to December at 1 to 3-month lead time. In particular, when the initial time starts from summer, the correlation skill of the IOD index reaches up to 0.5 at 5-month lead time. The ACC of IOD events between the prediction and the observation is 0.64, 0.51 and 0.35 at 1-, 2-and 3-month lead times, respectively, indicating a rapid decrease of prediction skill after 2-month lead time. The prediction skill of IOD is comparable with the BCC_CSM1.1 climate model .

SST anomalies in El Niño decaying summer
One of most dominant climate systems in EA-WNP is the East Asian Summer Monsoon (EASM), and its variability has great impacts on societal and economic influences. Many researches have shown that one of the most important factors affecting the EASM is the ENSO (Chen et al. 1992;Zhang et al. 1999;Wang et al. 2000). The impacts of ENSO on the EASM is not direct, and the anomalous anticyclone over the western North Pacific (WPAC) in the ENSO decaying summer (June-August; JJA (1), in which 1 represents the year after ENSO occurs) plays a bridge role connecting the ENSO and the EASM (Wu and Kirtman 2003;Yang et al. 2007;Wu et al. 2009;Xie et al. 2009). Therefore, for the successful prediction over the EASM, the WPAC and SST anomalies over the Indo-western Pacific are two important prerequisite factors.
The regressions of JJA (1) SST anomalies on the normalized previous DJF Niño3.4 index are shown in Fig. 12. In the observation (Fig. 12a), there are significant warming centers over most of the tropical Indian Ocean, the South China Sea, and the East China Sea. The distribution of SST anomalies in the Indo-western Pacific is quite similar to the observations in the decaying summer of El Niño events at all leading times. The bias at 1-month lead time lies in the overestimated amplitude in the Indian Ocean and South China Sea (Fig. 12b). In addition, the warming area in the western tropical Pacific seems to slightly shift westward compared with the observation. At 3-month lead time, the prediction is quite close to the observation in the Indian Ocean where the overestimated warming has been slightly reduced (Fig. 12c). The bias over the western North Pacific Ocean becomes bigger. When initiated with 6-month lead time, the warming pattern is much closer to the observations in the South Indian Ocean. However, the warming in the South China Sea is weaker than that in the observations (Fig. 12d). The bias seems to become even bigger with the lead time increase in the tropical western Pacific. This might be a common bias in climate models with the decaying speed of ENSO events slower than the observations, a reason that may come from the model systematic errors in the simulation of the cold togue (Tao et al. 2015;Jiang et al. 2017).

Precipitation and Circulation Anomalies in El Niño decaying summer
The successful prediction of precipitation beyond 1-or 2-month lead time is very difficult in current climate models. Since the ENSO and its decaying summer SSTA can be reasonably predicted at a quite long-term period, their relationship could be used to make the predictability of precipitation more available (Liu et al. 2017;Shi and Wang 2018). Figure 13 shows the regression of JJA (1) precipitation and 850-hPa wind anomalies at various lead times on DJF Niño3.4 index over the Indian Ocean and western North Pacific. In the observations, the WPAC is obviously found over the western North Pacific. Affected by the Pacific-Japan teleconnection, the precipitation anomalies display a distinct triple meridional structure over the EA-WNP (Fig. 13a). One positive anomalous rain band is found on its northwest and north flank of the WPAC from the South China to the South of Japan after the El Niño events, which matches with the area covered by the Mei-Yu band. Another positive rain band lies over the North Indian Ocean to the Maritime Continent, owing to anomalous winds that enhance the prevailing trade during summertime. These characters are well predicted by the CAMS-CSM forecast system (Fig. 13b, d), especially at 1-and 3-month lead times. The prediction bias at 1-month lead time is mainly because the predicted WPAC is a little bit stronger and shifts westward than the observed (Fig. 13b), resulting in much weaker moisture transport from the tropical Pacific to land over southern China and also weaker positive rainfall over South China. At the 3-month lead time, the amplitude of predicted WAPC is weakened, leading to reduced easterly anomalies on south flank of the WPAC (Fig. 13c). Besides, the positive rainfall anomalies over the South Indian Ocean are much weaker than the observations. When the prediction is initiated at 6-month lead time, the WPAC can still be successfully predicted but much weaker, resulting in weaker rainfall anomalies over the East Asia. Anyway, the south branch of positive rainfall anomalies over the North Indian Ocean to the South China Sea are successfully predicted at 6-month lead time (Fig. 13d).

Summary
The seasonal prediction skills for the period of 1981-2019 based on the CAMS-CSM climate forecast system has been evaluated. The forecast system includes a 3-dimentional nudging initialization scheme which assimilates atmospheric and oceanic data, and eight ensemble members per month were carried out with a lead time up to 6 months. From the analysis of Niño3.4 index, all the El Niño events, including the top three strongest 1982/83, 1997/98, and 2015/16, and La Niña events can be predicted successfully at 6-month lead time. The ACC skill of Niño3.4 index can reach 0.75 at 6-month lead time, which is comparable to current numerous dynamic seasonal forecast systems from the North American Multi-model Ensemble (NMME) project (Zhu et al. 2017). However, the predicted magnitudes for some particular El Niño (La Niña) events at 6-month lead time are weaker (larger) than the observed. The largest prediction skill over the Pacific is prediction starting from July at both 3-and 6-month lead times. The results also indicate the prediction skills not only have seasonal dependence, but also significant difference among the ocean basins. As far as the IOD events, the extreme warm/cold IOD events, (e.g., 1997, 1998, 2007 and 2017) are successfully predicted at 3-month and even at 6-month lead times. The IOD events has the highest predictability for predictions starting from summertime, which remains up to 5-month lead time. Generally, a certain degree of prediction skill exists on IOD events, with a lead time up to 2 months in advance. The ACC of IOD events between the observation and the prediction can reach 0.51 at 2-month lead time.
For the prediction of 2-m air temperature over land, the prediction skill is much lower than that of SST. The seasonal prediction skills over land are largely dependent upon the initial time. The prediction starting from April and July, has reliable skill at 2-month lead time over most areas of the North Hemisphere where the correlation is mainly above 0.4, especially over the North America and the East Asia.
The prediction on the spatial patterns of SST, precipitation and circulation anomalies in ENSO decaying summer are also evaluated. The forecast system can well predict the pattern and magnitude of SSTA over the EA-WNP at 6-month lead time, especially the Indo-Pacific warming. The bias seems to become bigger over the tropical western Pacific with the lead time increase. This might be a common bias in the present coupled models, in which the speed of ENSO decaying is slower than observation. The positive rainfall related to ENSO events over the Mei-Yu region can be well predicted at 3-month lead time, owing to that (a) (b) (c) (d) Fig. 13 The regression of JJA(1) precipitation anomalies in the observation (a) and model ensemble mean predicted at 1-(b), 3-(c) and 6-month (d) lead time with DJF(0) Niño3.4 index. The areas where significant above 90% are dotted. The wind is shown when the zonal or meridional components is significant at 90% level using Student's t test the ENSO-related WPAC can be predicted at a long-term period.
The reasonable prediction skills of the results obtained in this model may mainly benefit from two parts. First is the realistic simulation of the CAMS-CSM on the climatology and ENSO evolution over the tropical Pacific . Another is the data assimilation scheme. Compared to sophisticated ones, the simple nudging scheme is competent for the seasonal prediction both in skill score and efficiency. Meanwhile, assimilating more observed data (including 3-dimentional atmospheric and oceanic variables) here may enable the extreme events, such as the unusual ENSO events in 2015/2016 (Su et al. 2016;Zhong et al. 2019), to be successfully predicted at 6-month lead time. However, the prediction skills on the Indian Ocean and Atlantic Ocean temperature still face more challenges, and the issues about the model output statistics corrections Ji et al. 1998) may should be taken into consideration. Moreover, the ocean initial conditions (OICs) could also influence the ENSO prediction skill (Zhu et al. 2012). Hence, more sets of ocean data may be considered into our assimilation system to reduce the uncertainty of prediction. By participating in the (Chinese) multi-model ensemble project, it will help to improve the seasonal climate prediction skills, as many reasonable results have already been obtained in the CAMS-CSM forecast system. The present achievements also lay a foundation for setting up a seamless weather-climate prediction system in the future.