Seasonal rainfall predictability over the Huaihe River basin is evaluated in this paper on the basis of 23-year (1981–2003) retrospective forecasts by 10 climate models from the Asia-Pacific Economic Cooperation (APEC) Climate Center (APCC) multi-model ensemble (MME) prediction system. It is found that the summer rainfall variance in this basin is largely internal, which leads to lower rainfall predictability for most individual climate models. By dividing the 10 models into three cat-egories according to their sea surface temperature (SST) boundary conditions including observed, predicted, and persistent SSTs, the MME deterministic predictive skill of summer rainfall over Huaihe River basin is investigated. It is shown that the MME is effective for increasing the current seasonal forecast skill. Further analysis shows that the MME averaged over predicted SST models has the highest rainfall prediction skill, which is closely related to model’s capability in reproducing the observed dominant modes of the summer rainfall anomalies in Huaihe River basin. This result can be further ascribed to the fact that the predicted SST MME is the most effective model en-semble for capturing the relationship between the summer rainfall anomalies over Huaihe River basin and the SST anomalies (SSTAs) in equatorial oceans.
It is commonly accepted that the atmosphere behaves as a chaotic system ( Palmer, 1993), and, as a consequence, deterministic predictability associated with initial conditions is limited. An accurate weather forecast beyond approximately two weeks is impossible due to internal dynamic instability and nonlinear interactions in the atmosphere ( Lorenz, 1982). However, atmospheric predictability on a seasonal time scale is strongly influenced by the evolution of lower boundary conditions, particularly sea surface temperature (SST) (e.g., Shukla, 1981). Previous studies have reported that seasonal predictability is greater in the tropics than that in the extratropical regions and that predictability in the Pacific-Atlantic Ocean sector is greater than in the Indian Ocean-Asian monsoon region (e.g., Quan et al., 2004).
While several dynamical seasonal prediction systems have been developed in various operational and research centers in the past two decades, their predictive skills of precipitation and temperature differ significantly among regions due to the uncertainties in the representation of physical processes in climate system models ( Lin et al., 1998; Saha et al., 2006). Under this context, multi-model ensemble (MME) techniques have been developed, and a number of international projects have been organized to reduce model errors associated with the model parameterization schemes ( Krishnamurti et al., 1999; Palmer et al., 2004). Among these, the Climate Prediction and its Application to Society (CliPAS) project is one of the most comprehensive and includes support from the Asian- Pacific Economic Cooperation (APEC) Climate Center (APCC) and participation of the National Hydro-Meteorological Services from APEC member countries ( Wang et al., 2008a). Systematical evaluation of the CliPAS retrospective seasonal forecast results indicate that although the current state-of-the-art climate models exhibit better prediction skill for large-scale circulation and climate anomalies, the seasonal rainfall predictions over land have little skill ( Wang et al., 2008a). In addition, the atmosphere-ocean interaction is of great importance for the improvement of seasonal prediction skill ( Wang et al., 2008b).
With the development of seasonal climate prediction, the increasing requirements for extending the seasonal climate prediction to agriculture and hydrological sectors have been recognized, and the evaluation of the rainfall prediction skill over various river basins has drawn increasing attention. Located in the East Asian monsoon region, the Huaihe River basin (30°N-37.5°N, 110°E- 122.5°E) has experienced frequent floods and droughts leading to severe damages to the local economy ( Shi, 2007). To mitigate the impact of such disasters, skillful hydrological prediction of soil moisture and streamflow on seasonal scales are needed, which requires skillful prediction of seasonal rainfall.
In this study, the current forecast skill of summer rainfall over the Huaihe River basin is been evaluated by using the APCC multi-model retrospective seasonal forecast results. In addition, the dependence of model forecast skill on ensemble groups with various SST specifications is investigated.
The retrospective forecast datasets from 10 climate models used in this study were obtained from the APCC MME prediction project ( Wang et al., 2008a). Table 1 gives a brief summary of the model description and experimental design for the retrospective forecast in which “CMIP” represents a coupled model hindcast with an SST field generated by the coupled model itself; “AMIP” represents an SST field specified as the observed global distribution during the hindcast period; “SMIP2” and “SMIP2/HFP” represent the second phase of Seasonal Prediction Model Intercomparison Project (SMIP 2) as described by Kang and Lee (2004). AMIP examines one- and two-season potential predictability with observed SST as surface boundary condition, SMIP2 and SMIP2/HFP examines actual predictability with the boundary conditions containing forecasted information in the seasonal forecast.
As shown in Table 1, three methods are used to specify the SST field for SMIP2 and SMIP2/HFP; thus, the above- mentioned 10 models have been further categorized into three groups for analysis. The “obs_SST” ensemble indicates that the observed evolution of monthly SST anomalies (SSTAs) has been specified during the hindcast period; the “persist_SST” ensemble indicates that the observed SSTAs field at the starting month has been adopted and has remained unchanged during the hindcast period; and the “predict_SST” ensemble indicates that model-predicted SSTAs have been used during the hindcast period.
The present study focuses on the predictability of summer rainfall in the Huaihe River basin. The 23-year monthly retrospective forecast data during the common period of 1981-2003 was chosen and consists of several sizes of ensemble members for various climate models. All of the hindcast experiments were initiated in June, and results for June, July, and August were analyzed.
The observational rainfall datasets used in this study were the Climate Prediction Center Merged Analysis of Precipitation (CMAP) monthly datasets ( Xie and Arkin, 1997) recorded from 1979 to 2006 with 2.5°× 2.5° horizontal resolution. The SST datasets used were the improved Extended Reconstructed SST Version 2 (ERSST V2) monthly data ( Smith and Reynolds, 2004) from 1854 to 2009, with 1° × 2° resolution in latitude and longitude. For comparison, all of the retrospective forecast data were regridded into a common 2.5° × 2.5° grid.
In this study, potential predictability was assessed with the method proposed by Koster et al. (2004). This approach is based on the perfect model assumption in which one member of the ensemble forecast is assumed to be the "truth’’ while the rest of the ensembles are considered as model forecasts. Because observed climate is in fact one of many possible realizations of the climate system, this approach produces an analog to the real world in which the forecast model and the underlying "climate system’’ share the same exact physics ( Luo and Wood, 2006). The potential predictability shows how well the climate system can predict itself. Because each forecast member is assumed to be the truth, each square of the correlation coefficient ( R-square) of the truth and ensemble mean forecast is calculated. The potential predictability is then defined as the averaged R-square.
The metrics used to measure prediction skill of deterministic forecast included the anomaly temporal correlation coefficient (TCC), anomaly pattern correlation coefficient (PCC), and root mean square error (RMSE). The distribution of the predominant precipitation patterns in the Huaihe River basin was analyzed by the Empirical Orthogonal Function (EOF) method.
The scattering plot shown in Fig. 1 gives the results for potential predictability of the various models for June, July, August, and summer mean (JJA), which represents an average of JJA. The averaged R-square values were calculated by using the 23-year hindcast data initiated in June with a lead time of 0-2 months. It was determined that the potential predictability of summer rainfall by most climate models was relatively low with an R-square lower than 0.1; however, the predictability of HMC and MSC_GM3 in June to August were higher than 0.1. In addition, the R-square reached approximately 0.4 in summer, which indicates high predictability. Moreover, the R-square of MGO in August was also higher than 0.1. These results are consistent with those obtained by the analysis of variance method ( Rowell, 1998).
Because the prediction of summer rainfall in the Huaihe River basin is of great significance to policymakers in emergency management and water resources sectors, the deterministic prediction skill of climate models is evaluated in this section. The TCC and RMSE against the target month between the hindcasts and observation are shown in Fig. 2. The TCC of June rainfall forecast for the predict_SST MME was nearly 0.3, which is higher than the other three MMEs. The TCCs of June rainfall forecast for 10-model MME, persist_SST MME, and obs_SST MME were 0.2, 0.1, and -0.1 respectively. The higher TCC for the predict_SST MME was also observed for August and the summer season. For July, the TCC for the predict_SST MME was also relatively high, although its skill was slightly lower than that of the 10-model MME (Fig. 2a).
Figure 2b shows the RMSE of the hindcasts for June, July, August, and JJA respectively. The predict_SST MME had the minimum RMSE compared with the observation in all lead times among the four MMEs, while the RMSE of the obs_SST MME was generally the largest.
The TCC scores and RMSE taken as an index for model’s predictive skill suggest that the prediction skill for the predict_SST MME is the highest, the 10-model MME has lower skill than the predict_SST MME and the prediction skill of the persist_SST MME ranks third. The obs_SST MME has the poorest prediction skill.
 | Table 1 Description of model information. |
 | Figure 1 The potential predictability of 10 models of forecasted June, July, August, and summer season mean (JJA) precipitation. Model color representation is indicated at the top left corner of the plot. |
 | Figure 2 (a) Temporal Correlation Coefficient (TCC) and (b) Root Mean Square Error (RMSE) of the anomaly percentage of regional averaged summer precipitation between the multi-model ensembles (MMEs) and the observations. The x axis represents the forecasted June, July, and August and JJA. |
More detailed spatial distribution of rainfall within a basin is often necessary for better streamflow prediction. Moreover, the capability of climate models in reproducing the observed dominant rainfall modes over a region is a key factor in determining the model’s predictive skill. In this section, we evaluate the abilities of the models in reproducing the principal modes of summer precipitation anomaly variability over the Huaihe River basin during the hindcast period. The spatial distribution of July mean precipitation in the Huaihe River basin generally follows two major patterns including 1st and 2nd EOF modes. The former, as shown in Fig. 3a1, is characterized by the contrast of rainfall distribution between the southeastern and northwestern regions, while the latter, as shown in Fig. 3a2, is characterized by the north-south contrast of rainfall anomalies over the region. The 1st and 2nd EOF modes account for 40.9% and 16.6% of the total variance, respectively.
Figure 3a shows the first two EOF modes for the observed July precipitation, while Figs. 3b-e indicate the hindcasted counterparts for the predict_SST MME, per- sist_SST MME, obs_SST MME, and 10-model MME. Left and right columns of Fig. 3 show the first and second eigenvectors of July precipitation, respectively. The observed 1st EOF rainfall pattern was accurately reproduced by the predict_SST MME, as shown in Fig. 3b1 with a PCC of 0.9 between the hindcast and observation. However, the rainfall patterns from other MMEs differed significantly from the observations. The PCCs of the 1st EOF pattern between observation and persist_SST MME, obs_SST MME, and 10-model MME were 0.24, 0.81, and 0.71 respectively. The time series associated with the eigenvectors, shown in Fig. 3f1, varied in a similar manner. The TCC for the predict_SST MME was 0.3, whereas those for the persist_SST MME, obs_SST MME, and 10-model MME were 0.04, 0.14, and 0.21, respectively. The predict_SST MME most accurately reproduced the first eigenvector of the observed July precipitation in Huaihe River basin, while the performance of persist_SST MME was the poorest.
Generally, the second eigenvector of the July precipitation in the Huaihe River basin rather than the first was better reproduced by the 10 models. As shown in the right column of Fig. 3, predict_SST MME, obs_SST MME, and all-model MME reproduced the observed rainfall spatial pattern with high accuracy. The PCCs of the above three MMEs were 0.92, 0.89, and 0.92, and the TCCs were 0.25, -0.02, and 0.28, respectively. However, the rainfall pattern of the 2nd EOF mode for the persist_SST MME showed a northwest-southeast contrast, which differs significantly from the observation. The PCC for the persist_SST MME was 0.62, and the TCC was 0.24 during the hindcast period.
It is worth noting that the time series in 1988 and 1998 of Principal Component 1 (PC1) agree quite well between observation and retrospective forecast; the same is true for the time series in 1987, 1988, and 1997 of Principal Component 2 (PC2). Because 1987-88 and 1997-98 were typical ENSO years, these results suggest that the model performance in the Huaihe River basin may be closely related to El Niño-Southern Oscillation (ENSO)-related SSTAs.
 | Figure 3 1st (left column) and 2nd (right column) EOF modes of the observed precipitation and four hindcasted MMEs’ precipitation of July and the (f1)-(f2) time series associated with the eigenvectors. Solid and dashed lines indicate the observed and hindcasted time series of the MMEs, respectively. |
A significant relationship has been reported between
SSTAs in the equatorial Pacific Ocean and summer pre- cipitation anomalies in the Huaihe River basin ( Qian et al., 2009). To clarify the possible reasons for the predict_SST MME exhibiting a relatively higher predictive skill, the capability of climate models in reproducing the above- mentioned summer rainfall and SST relationship is further analyzed in this section.
The correlation between observed July precipitation over the Huaihe River basin and observed SSTs in previous winters over the equatorial oceans is shown in Fig. 4a, and the counterparts for the four different MME hindcast results are shown in Figs. 4b-e, respectively.
Figure 4a indicates notably positive correlations between the observed July precipitation and SSTAs in the South Indian Ocean and equatorial eastern Pacific with the highest correlation of approximately 0.45 occurring in the Southeast Indian Ocean and equatorial eastern Pacific Ocean. Moreover, a distinct negative relationship can be found in the eastern South Pacific Ocean with a correlation coefficient of -0.45.
Figures 4b-e indicate that the predict_SST MME, persist_SST MME, and 10-model MME have all captured the relationship between SST and July-mean rainfall anomalies over the Huaihe River basin, although the correlation from the hindcast results is much stronger in the Indian Ocean and equatorial eastern Pacific Ocean and weaker in the eastern South Pacific Ocean. However, for the obs_SST MME, no correlation was detected in the Indian and Pacific oceans, which may suggest that the models using observed monthly mean SST during the hindcast period failed to capture the relationship between July precipitation in the Huaihe River basin and the SSTs in previous winters. Therefore, the prediction skill of the obs_SST models is poorer than that in the predict_SST and persist_SST ensembles. Similar results were observed for June and August (figures not shown).
 | Figure 4 Correlations between the (a) observed and (b)-(e) four hindcasted MME July precipitation of the Huaihe River basin and the observed SSTs in previous winters over the equatorial Pacific and Indian Ocean. Red solid lines denote positive correlation while blue dashed lines denote negative correlation; zero lines are not drawn. Light and dark shadings denote the correlation coefficient at 90% and 95% confidence levels, respectively. |
The present research assessed the current status of the seasonal predictability of the Huaihe River basin by using the APCC 10-model ensemble retrospective forecast data for 1981-2003. These 10 models were further categorized into three groups according to their specifications of SST field during the hindcast period, and the prediction skills from different MMEs based on these SST specifications were evaluated and compared.
A remarkable internal variance of summer rainfall in Huaihe River basin for most of the individual climate models was determined by using the ensemble retrospective forecast data, which suggest a lower potential predictability of summer rainfall for individual models. Among the 10 models, only the HMC and MSC_GM3 model showed relative higher potential predictability in the Huaihe River basin.
The predictive skill of summer rainfall anomalies was investigated by calculating the temporal correlation between the observation and multi-model hindcast results in the Huaihe River basin, together with the RMSE of the forecasted rainfall averaged over Huaihe River basin. It was determined that the predict_SST MME had a higher prediction skill than the 10-model MME, persist_SST MME, and obs_SST MME; the obs_SST MME showed the lowest prediction skill. These results indicate that the multi-model ensemble is an efficient method for increasing the current seasonal prediction skill; therefore, the methods used to generate the ensemble are also important.
Further analysis indicates that the higher prediction skill for predict_SST MME is closely related to model’s capability in reproducing the observed dominant modes of the summer rainfall anomalies in the Huaihe River basin.
The relationship between the summer rainfall anomalies with the SSTAs in equatorial oceans has been further investigated for both observation and retrospective forecasts with the various MMEs. Results show that the predict_SST MME can effectively capture the relationship between the SST signal and July-mean rainfall anomalies over the Huaihe River basin, which could explain the relatively higher predictive skill exhibited by predict_SST MME. The better performance of predict_SST MME in reproducing the observed SST-rainfall relationship could be ascribed to both the climate model itself, and the hindcast experimental design, particularly the methods of prescribing the SST field during the hindcast integration. Further studies are needed to identify the importance of air-sea interaction on the seasonal prediction skill over the East Asian monsoon region.
We thank the APCC and those institutes participating in the APCC MME operational system for providing the hindcast experiment data. We also appreciate the significant assistance of Professor TANG Youmin, Canada Research Chair in Climate Prediction and Predictability, University of Northern British Columbia. This research is supported by the National Natural Science Foundation of China (41175073), the National Science Foundation of China (NSFC)-Yunnan Province Joint Grant (U1133603), the National Basic Research Program of China (2010CB428403 and 2009CB421406), and the NOAA Climate Program Office and Michigan State University (NA10OAR4310246 and NA12OAR 4310081).
| 1 |
|
| 2 |
|
| 3 |
|
| 4 |
|
| 5 |
|
| 6 |
|
| 7 |
|
| 8 |
|
| 9 |
|
| 10 |
|
| 11 |
|
| 12 |
|
| 13 |
|
| 14 |
|
| 15 |
|
| 16 |
|
| 17 |
|
| 18 |
|