Statistical decadal predictions for sea surface temperatures: a benchmark for dynamical GCM predictions
- First Online:
- Received:
- Accepted:
DOI: 10.1007/s00382-012-1531-9
- Cite this article as:
- Ho, C.K., Hawkins, E., Shaffrey, L. et al. Clim Dyn (2013) 41: 917. doi:10.1007/s00382-012-1531-9
Abstract
Accurate decadal climate predictions could be used to inform adaptation actions to a changing climate. The skill of such predictions from initialised dynamical global climate models (GCMs) may be assessed by comparing with predictions from statistical models which are based solely on historical observations. This paper presents two benchmark statistical models for predicting both the radiatively forced trend and internal variability of annual mean sea surface temperatures (SSTs) on a decadal timescale based on the gridded observation data set HadISST. For both statistical models, the trend related to radiative forcing is modelled using a linear regression of SST time series at each grid box on the time series of equivalent global mean atmospheric CO_{2} concentration. The residual internal variability is then modelled by (1) a first-order autoregressive model (AR1) and (2) a constructed analogue model (CA). From the verification of 46 retrospective forecasts with start years from 1960 to 2005, the correlation coefficient for anomaly forecasts using trend with AR1 is greater than 0.7 over parts of extra-tropical North Atlantic, the Indian Ocean and western Pacific. This is primarily related to the prediction of the forced trend. More importantly, both CA and AR1 give skillful predictions of the internal variability of SSTs in the subpolar gyre region over the far North Atlantic for lead time of 2–5 years, with correlation coefficients greater than 0.5. For the subpolar gyre and parts of the South Atlantic, CA is superior to AR1 for lead time of 6–9 years. These statistical forecasts are also compared with ensemble mean retrospective forecasts by DePreSys, an initialised GCM. DePreSys is found to outperform the statistical models over large parts of North Atlantic for lead times of 2–5 years and 6–9 years, however trend with AR1 is generally superior to DePreSys in the North Atlantic Current region, while trend with CA is superior to DePreSys in parts of South Atlantic for lead time of 6–9 years. These findings encourage further development of benchmark statistical decadal prediction models, and methods to combine different predictions.
Keywords
Decadal prediction Statistical Sea surface temperatures Global climate model1 Introduction
Climate predictions for the near-term (up to about 30 years), especially on regional scales, can be used to inform adaptation actions to a changing climate, for example infrastructure planning and hazard preparedness (Adger et al. 2005; Challinor 2009). In recent years, there has been major progress in the development of such predictions based on global climate models (GCMs) initialised with atmospheric and oceanic observations (e.g. Smith et al. 2007; Keenlyside et al. 2008; Fyfe et al. 2011). As part of the Fifth Coupled Model Intercomparison Project (CMIP5; Taylor et al. 2011), experimental decadal predictions produced by initialised GCMs from different modelling groups will be compared (e.g. Smith et al. 2012). Proper evaluation of these dynamical decadal predictions can aid efforts in improving future GCM simulations, for example their initialisation schemes.
On decadal timescales changes in climate are caused by both its response to radiative forcing and its internal variability (Keenlyside and Ba 2010; Solomon et al. 2011; Goddard et al. 2012b). With a non-stationary climate, an initialised GCM can show significant prediction skill if it predicts a long-term forced trend consistent with observations (e.g. Lee et al. 2006; van Oldenborgh et al. 2012), but often users are also interested in the magnitude of the internal variability. The skill of a particular initialised GCM in predicting such internal variability may be assessed by comparing its retrospective forecasts (also known as hindcasts in the decadal prediction literature) and those produced by the identical GCM without assimilation of observations (e.g. Smith et al. 2007). Alternatively, Laepple et al. (2008) proposed the use of bias-corrected ensemble mean projections from multiple uninitialised GCMs as benchmarks to evaluate the skill of decadal retrospective forecasts by initialised GCMs.
Being much less computationally expensive to run, predictions from statistical models can also serve as benchmarks when assessing the skill of dynamical GCM predictions. A good benchmark statistical model should be able to capture the basic characteristics of the climate system which we want to predict. Ideally it should also be trained solely by historical observations without information from physical climate models, but this is limited by the length and quality of available observational record. Simple benchmarks such as persistence and climatology have been extensively used in the evaluation of seasonal climate predictions (e.g. Barnston et al 1994; Colman and Davey 2003). As for decadal climate predictions, since both radiative forcing and internal variability can be important, more advanced benchmark statistical models which also incorporate these effects are more appropriate. For example, Lean and Rind (2009) projected global annual mean air temperatures using a multiple linear regression of temperature on anthropogenic influence, solar radiation, ENSO variability and volcanic aerosols. Krueger and von Storch (2011) produced decadal retrospective forecasts of global annual mean temperatures by separating the observed temperature time series into two components representing forced trend and internal variability respectively. The forced trend was modelled by a linear regression between multi-model ensemble mean temperatures and atmospheric CO_{2} concentration. After removing this forced trend component, the residual internal variability was modelled as a first-order autoregressive process. Fildes and Kourentzes (2011) compared retrospective forecasts of global mean temperatures for up to 20 years ahead by the UK Met Office Decadal Prediction System DePreSys (Smith et al. 2007; which will also be considered in this study, see Sect. 2.2) with those by various time series models, ranging from a local linear trend model to a multivariate neural network. DePreSys was found to be more skillful than the statistical models for lead time of 1–4 years, while some of the more complex statistical models, such as the neural network model, have better prediction skill for longer lead times.
Some other studies have produced benchmark statistical models for climate variables other than air temperatures. With a perfect model approach using control integrations of two GCMs, Hawkins et al. (2011) showed that two empirical methods, namely Linear Inverse Modelling (LIM) and Constructed Analogue (CA), have significant skill in predicting the internal variability of sea surface temperatures (SSTs) over parts of the Atlantic for up to a decade. Zanna (2012) applied LIM to gridded observed Atlantic SST anomalies, with the forced trend removed using cubic splines. The retrospective forecasts were found to be skillful relative to climatology for a lead time of up to around 5 years.
In this paper we present two benchmark statistical models for predicting annual mean SSTs with a lead time of up to a decade based on the gridded observed data set HadISST. Variations of SSTs have important implications for atmospheric conditions, such as precipitation patterns and tropical cyclone activity (e.g. Sutton and Hodson 2005; Zhang and Delworth 2006; Smith et al. 2010). We will assess the predictive skill of the two statistical models by verifying a set of retrospective forecasts with start times from mid-1960s to mid-2000s, with a focus on examining the regional skill for different lead times. In addition, a brief comparison will be made with the corresponding forecasts by DePreSys to assess the relative strengths of dynamical and statistical predictions for different regions. Predictions for the Atlantic Ocean are particularly interesting and our discussion will put greater emphasis on this sector.
This paper is structured as follows. Section 2 gives a summary of HadISST, DePreSys and the external forcing time series used in this study. This is followed by descriptions of our benchmark statistical models and verification measures in Sect. 3. In Sect. 4 we assess the prediction skill of the statistical models, while in Sect. 5 we compare the skill of these models with DePreSys. The statistical forecast of Atlantic SSTs for the years 2012 to 2021 is presented in Sect. 6. Further discussion on our findings and concluding remarks are given in Sect. 7.
2 Observational and GCM data
2.1 HadISST
We use the HadISST data set (Rayner et al. 2003) to train our statistical prediction models, correct biases in the DePreSys predictions and verify retrospective forecasts from both the statistical models and DePreSys. This data set contains global monthly interpolated fields of SSTs on a 1° × 1° grid from 1870 to 2011. For the rest of this paper, we consider the anomalies of annual means of SSTs. To allow direct comparison between the statistical and DePreSys retrospective forecasts, here the annual mean for a certain year is defined to be the 12-month average from the December of the previous year to the November of the year concerned. The anomalies for each grid box are calculated by removing the corresponding mean for the years 1986 to 2005. It should be noted that the amount of data (e.g. in-situ SST observations, satellite data) used to construct this gridded data set varies with space and time. Particularly, because of the sparseness of observations over the southern oceans and near the polar regions, we perform our analyses only from 30°S to 70°N. Grid boxes covered with sea-ice are also omitted in our analyses.
2.2 DePreSys
The UK Met Office Decadal Prediction System, DePreSys (Smith et al. 2007), is based on the third Hadley Centre coupled GCM HadCM3 (Gordon et al. 2000). Its atmosphere component has a horizontal resolution of 2.5° × 3.75° and 19 vertical levels, while its ocean component has a 1.25° × 1.25° resolution with 20 vertical levels. We use a perturbed physics ensemble of DePreSys with nine variants designed to sample climate model uncertainty. One of these nine variants uses the standard HadCM3 settings of physical parameters, while for the other eight, simultaneous perturbations of 29 atmospheric parameters are employed. These variants are chosen from previous experiments with larger perturbed physics ensembles (Collins et al. 2006, 2011), with the aim of spanning a wide range of parameter values while simulating physically plausible climate variability. The equilibrium climate sensitivity of the ensemble ranges from 2.6 to 7.1 °C.
For each model variant, atmospheric and oceanic analyses were assimilated from December 1958 to November 2007 as anomalies with respect to the model climate to create the initial conditions. Time-varying radiative forcings derived from observed changes in greenhouse gases, aerosol and solar irradiance were used up to the year 2000, after which the forcing based on the SRES A1B scenario (Nakićenović and Swart 2000) was applied. A total of 46 retrospective forecasts of global SSTs, starting on 1 November of each year from 1960 to 2005 with each extending to 10 years ahead, are available for each ensemble variant. We consider only the ensemble mean of annual mean SST forecasts from December to November.
2.3 Global mean equivalent CO_{2} concentration
In the modelling of the forced trend of observed SSTs, we use the global mean equivalent CO_{2} concentration from the Representative Concentration Pathways (RCPs) greenhouse gases concentration historical and projection data set (Meinshausen et al. 2011). These data have been used to drive the CMIP5 climate simulations. The equivalent CO_{2} concentration incorporates the global net effects of all anthropogenic forcing agents, including greenhouse gases and aerosols. Observed concentrations from 1760 to 2005 are available. For subsequent years, the RCP4.5 concentration scenario is used for the forecasts we present in this paper, but the sensitivity of our results to alternative scenarios have also been tested.
3 Statistical prediction models for SST hindcasts and forecasts
In order to compare with DePreSys, our statistical retrospective forecasts for annual mean SSTs cover the same time period, i.e. we produce 46 sets of retrospective forecasts starting from 1960 to 2005, each of 10 years in length. In addition, we produce a forecast starting from 2011 for the years 2012–2021.
3.1 Modelling the forced trend (Trend)
3.2 Comparing the forced trend and internal variability
A related variable, the time at which the climate change signal emerges from the interannual variability, is often of interest in climate change risk assessment. A number of studies (e.g. Giorgi and Bi 2009; Mahlstein et al. 2011; Diffenbaugh and Scherer 2011; Hawkins and Sutton 2012) have examined this ‘time of emergence’ for future climate change using climate model projections, but here we can assess when and where the signal has already emerged in the historical SST observations. Figure 2c shows the time of emergence defined as the year when the ratio of the magnitude in estimated forced change from the year 1871 (pre-industrial time) to σ last crossed and exceeded the threshold of one in the time series.^{1} The climate signal emerged over parts of the Indian Ocean in the 1960s, while most of the Atlantic and western Pacific saw emergence in the 1970s and 1980s. As the forced trend was small over the far North Atlantic and eastern Pacific, the climate signal remained smaller than the interannual variability throughout the 141 years. Figure 2d also shows the time of emergence but considers the forced change from the 1986–2005 mean. Even with this more recent baseline, the climate change signal has already emerged in tropical Indian Ocean and tropical Atlantic in the 2000s.
Another feature worth exploring is the importance of longer timescale variability relative to interannual variability. This is because the former is related to slow ocean processes and is considered to be, at least potentially, more predictable. We consider a simple measure of the potential predictability (Boer 2004, 2011; Boer and Lambert 2008) of X^{v} for a timescale of N years, σ_{N}/σ, where σ_{N} is the standard deviation of the running N-year mean of estimated X^{v}. Figure 2e, f show this measure for N = 5 and 10 years. Larger potential predictability can be found over the most of the North Atlantic. Previous observational and modelling studies have suggested that this is related to decadal variations in the North Atlantic thermohaline circulation and the North Atlantic Oscillation (e.g. Latif et al. 2006). The potential predictability is smaller over the tropical Pacific where ENSO-related interannual variability dominates.
For the rest of this paper we will put greater emphasis on the Atlantic sector for two reasons. First, the gridded historical observations in the Atlantic are more reliable because there were more available in-situ observations in the region (Rayner et al. 2003). Second, over the Atlantic there is spatial variation in the relative importance of forced long-term trend and internal variability. In the far North Atlantic where the internal variability is more important, the longer timescale variability is also more potentially predictable. This constrasts with the Indian Ocean and western Pacific where the effect of long-term forced trend dominates and the forced trend prediction described in Sect. 3.1 is perhaps adequate for achieving reasonable predictive skill.
3.3 First-order autoregressive model (AR1)
Figure 1b shows estimates of γ_{1} in the AR1 model fitted to data from all available years. The highest autoregression coefficients can be found in the North Atlantic subpolar gyre, indicating stronger persistence in the SST time series (or memory at a longer timescale; see also Zhu et al. 2010). The effect of persistence is generally weaker in the equatorial regions. Similar results can be seen for the AR1 models fitted for the individual retrospective forecasts (not shown).
3.4 Constructed analogue (CA)
The use of a second and more complex model for predicting X^{v}, constructed analogue (CA; van den Dool 2007, Chap. 7), is motivated by Hawkins et al. (2011) which employed this model for decadal predictions of SSTs using control integrations of two GCMs. It was found to be skillful over the far North Atlantic. The CA method has also been employed in operational seasonal SST forecasts for the Pacific (Barnston et al. 1994; Landsea and Knaff 2000) and seasonal predictions of soil moisture (van den Dool et al. 2003). The rationale behind CA is to develop a weighted, linear combination of historical spatial patterns of observations which is closest to the initial spatial pattern. If the future evolution of such patterns resembles the historical evolution, the CA can make predictions by carrying forward the estimated weights. An analogue needs to be constructed because there is only a remote chance of finding a ‘natural’ analogue where the error between a state in the historical record and the desired initial state is within observational errors (van den Dool 1994).^{2}
3.5 Verification for retrospective forecasts
We have seen how the forced trend (X^{f}) and the internal variability (X^{v}) components of SSTs are modelled and predicted. The predictions of SST anomalies (\(\hat{X}\)) are either the sum of predicted X^{f} and predicted X^{v} using AR1 [i.e. (3) + (5)] or the sum of predicted X^{f} and predicted X^{v} using CA [i.e. (3) + (8)]. We term these two anomaly predictions as ‘Trend+AR1’ and ‘Trend+CA’ respectively.
The three key questions in the evaluation of prediction skill of statistical SST retrospective forecasts are: (1) Are the retrospective statistical forecasts of the forced trend and internal variability skillful? (2) How does the skill vary spatially and with prediction lead times? (3) Where and for what lead times are initialised dynamical retrospective forecasts by DePreSys more skillful than the corresponding benchmark statistical forecasts? We now describe how these questions are to be addressed.
We mainly consider two skill measures, anomaly correlation coefficient (ACC) and mean squared skill score (MSSS), to verify up to 46 sets of statistical retrospective forecasts by Trend+AR1 and Trend+CA.^{3} The ACC considers the prediction and observation of the anomaly (\(\hat{X}\) and X). Since the ACC does not inform about whether the skill comes from the forced trend or internal variability (or both), we also consider the correlation coefficient between the prediction of the forced component (\(\hat{X}^{f}\); ‘Trend’ prediction) and the observed anomaly (X) and the correlation coefficient between the predicted internal variability component (\(\hat{X}^{v}\)) and the estimated internal variability component of the verifying observations (X^{v}). Note that there are two major limitations of correlation coefficients in verification. They measure only the linear association between predictions and observations, and any conditional biases in the predictions are ignored (Murphy and Epstein 1989).
The statistical significance of the correlation measures and MSSS is assessed using a non-parametric bootstrapping approach. In particular, a block bootstrapping technique is used to account for the temporal autocorrelation among successive observations (or forecasts). Details are given in the section Appendix.
For decadal climate predictions, verification is commonly performed on predictions averaged over a range of lead times, but there is no clear consensus on the choice of temporal averaging period (e.g. Smith et al. 2010; Hawkins et al. 2011; van Oldenborgh et al. 2012). We adopt the framework suggested by Goddard et al. (2012a) and consider three lead time periods, year 1, years 2–5 and years 6–9. Note that a skillful time-averaged (multi-annual) prediction does not imply that the predictions for individual years within the averaging period (annual predictions) are skillful, as errors in the individual predictions with different signs may cancel out by averaging.
The verification procedures for the ensemble mean DePreSys retrospective forecasts are similar to those described above but two additional steps are required. Since the spatial resolution of DePreSys and HadISST are different, the HadISST data are first interpolated onto the grid of DePreSys using bilinear interpolation prior to verification. In addition, the mean bias of the ensemble mean DePreSys retrospective forecasts (difference between the modelled and observed climate) is removed for each lead time. In order to avoid the introduction of artificial skill, this is performed in a cross-validation manner, i.e. for each lead time, we calculate and remove the mean bias for each forecast individually using forecast and observation data from all the other start times.
4 Skill of the statistical retrospective forecasts
We now assess the skill of retrospective forecasts by our statistical models, Trend+AR1 and Trend+CA. We will also compare our results with some other previous studies on benchmark statistical models reviewed in the Introduction.
4.1 Trend+AR1 forecasts
For the tropical Pacific, there is little predictive skill for all lead times. Despite large interannual variability (Fig. 2a), the effect of persistence is generally weak (Fig. 1b) as the timescale of ENSO is typically less than 1 year. Both the forced trend and AR1 forecasts for the internal variability do not show obvious skill.
The predictive skill for the far North Atlantic is more interesting. The ACC for Trend+AR1 drops from around +0.7 to around −0.8 as we move from year 1 to years 6–9. The correlation for Trend is negative for all three periods, however the correlation for AR1 is greater than 0.5 for year 1 and years 2–5. The MSSS is also significantly greater than zero (at 10 % level; same for below unless otherwise stated) for these two periods. These suggest that the AR1 model gives skillful predictions of the internal variability over the far North Atlantic for shorter lead times, but the poor forced trend prediction offsets such skill. These results are similar to the global decadal air temperature retrospective forecasts by Krueger and von Storch (2011), where the ACC is larger than 0.7 for year 1 but is not significantly greater than zero (at 5 % level) for year 10.
4.2 Trend+CA forecasts
There is quite strong evidence that CA is skillful in predicting the internal variability in the far North Atlantic, and performs considerably better than the simple AR1 model. Significant positive correlation coefficients for the internal variability (greater than 0.5) can be seen for all three lead time periods, while the MSSS also shows that the Trend+CA forecasts are significantly superior to the Trend forecasts even for years 6–9. These explain the changes in MSSS with lead time of Trend+CA relative to Trend+AR1. Both methods are similarly skillful for year 1, but Trend+CA has a significant advantage at extended lead times. Similar to Trend+AR1, the ACC for Trend+CA in the far North Atlantic changes from positive to negative as we move from year 1 to years 6–9 because the predicted forced trend is not consistent with observations within the verifying period. However, the drop in magnitude for Trend+CA is smaller as CA better predicts the internal variability.
Another region worth noting is the South Atlantic. There is some evidence that Trend+CA gives skillful predictions at longer lead times in parts of this region, even though we have seen that the South Atlantic is less potentially predictable than the far North Atlantic (Fig. 2e, f). From Fig. 1a the dependence of SSTs on radiative forcing appears to be stronger near the coast than in the central parts of the South Atlantic. This might explain the higher correlation coefficients for the Trend forecasts nearer to the coast. As for the CA predictions of internal variability, the correlation coefficients are rather low for year 1, but increase gradually as we move to longer lead time periods. This is consistent with the positive MSSS of Trend+CA relative to both Trend and Trend+AR1 forecasts, especially for years 6–9 with a large area of significantly positive MSSS (Fig. 7). Figure 5c shows the time series of observations and retrospective forecasts for a grid box in South Atlantic (centred at 10.5°S 14.5°W; also marked with a cross in Figs. 6, 7). The positive trend is well predicted. In addition, the Trend+CA forecasts for year 8 (purple line) appear to capture the interannual variability quite well, particularly from around 1980 to 2000.
Over parts of the tropical Atlantic, CA appears to have some predictive skill for the internal variability as well. The correlation coefficient for the internal variability exceeds 0.4 for year 1 and years 2–5. The MSSS of Trend+CA relative to both Trend and Trend+AR1 also shows that the CA is more skillful for these two periods and perhaps even for years 6–9. However, comparing the time series of observations and year 8 forecasts for the grid box centred at 17.5°N 57.5°W in Fig. 5d, Trend+CA does not seem to have a clear advantage over Trend+AR1.
Our results are generally consistent with Hawkins et al. (2011). Their verification of retrospective forecasts by CA using the control integration of HadCM3 also gave ACC greater than 0.5 in the far North Atlantic for all lead time periods (years 1, 2, 3–5 and 6–10). However, while we have seen positive correlation coefficients for internal variability in the South Atlantic at extended lead times for our forecasts, the ACC for their CA forecasts in the region was close to zero.
4.3 Regional average retrospective forecasts
Correlation measures for predictions in four specified regions for different lead time periods
Model | Year 1 | Years 2–5 | Years 6–9 |
---|---|---|---|
Subpolar gyre (SPG; 60–66°N, 10–60°W) | |||
Trend | −0.58 [−0.73,−0.22] | −0.64 [−0.83,−0.13] | −0.39 [−0.72,0.34] |
Trend+AR1 | 0.65 [0.42,0.75] | −0.09 [−0.48,0.27] | −0.70 [−0.84,0.01] |
Trend+CA | 0.66 [0.40,0.77] | 0.27 [0.01,0.43] | 0.15 [−0.27,0.48] |
DePreSys | 0.77 [0.55,0.85] | 0.84 [0.58,0.92] | 0.80 [0.38,0.89] |
Main development region (MDR; 10–20°N, 20–80°W) | |||
Trend | 0.06 [−0.26,0.39] | 0.24 [−0.14,0.58] | 0.40 [−0.05,0.73] |
Trend+AR1 | 0.49 [0.20,0.67] | 0.30 [−0.18,0.56] | 0.42 [−0.03,0.73] |
Trend+CA | 0.44 [0.12,0.64] | 0.53 [0.16,0.74] | 0.47 [0.15,0.65] |
DePreSys | 0.63 [0.37,0.79] | 0.61 [0.16,0.79] | 0.54 [0.09,0.76] |
North Atlantic (NATL; 10–60°N, 0–75°W) | |||
Trend | 0.43 [0.10,0.69] | 0.59 [0.34,0.80] | 0.68 [0.45,0.86] |
Trend+AR1 | 0.87 [0.66,0.93] | 0.78 [0.52,0.89] | 0.73 [0.51,0.90] |
Trend+CA | 0.85 [0.65,0.92] | 0.71 [0.33,0.85] | 0.57 [0.15,0.76] |
DePreSys | 0.89 [0.68,0.93] | 0.91 [0.75,0.96] | 0.89 [0.71,0.95] |
South Atlantic (SATL; 10–30°S, 50°W–20°E) | |||
Trend | 0.45 [0.23,0.65] | 0.65 [0.41,0.83] | 0.60 [0.36,0.78] |
Trend+AR1 | 0.47 [0.27,0.61] | 0.64 [0.40,0.83] | 0.60 [0.38,0.79] |
Trend+CA | 0.41 [0.18,0.57] | 0.58 [0.32,0.80] | 0.65 [0.45,0.79] |
DePreSys | 0.56 [0.35,0.72] | 0.43 [0.08,0.67] | 0.32 [0.07,0.51] |
As in Table 1 but for root-mean-squared error (RMSE; in K)
Model | Year 1 | Years 2–5 | Years 6–9 |
---|---|---|---|
Subpolar gyre (SPG; 60–66°N, 10–60°W) | |||
Trend | 0.47 [0.39,0.53] | 0.44 [0.36,0.51] | 0.46 [0.36,0.50] |
Trend+AR1 | 0.31 [0.26,0.33] | 0.38 [0.31,0.44] | 0.46 [0.36,0.50] |
Trend+CA | 0.30 [0.25,0.33] | 0.37 [0.31,0.41] | 0.42 [0.33,0.47] |
DePreSys | 0.26 [0.23,0.29] | 0.21 [0.16,0.24] | 0.22 [0.19,0.26] |
Main development region (MDR; 10–20°N, 20–80°W) | |||
Trend | 0.32 [0.28,0.36] | 0.23 [0.20,0.27] | 0.23 [0.19,0.27] |
Trend+AR1 | 0.26 [0.23,0.30] | 0.23 [0.19,0.26] | 0.23 [0.19,0.28] |
Trend+CA | 0.27 [0.24,0.30] | 0.21 [0.18,0.23] | 0.23 [0.19,0.27] |
DePreSys | 0.25 [0.20,0.30] | 0.21 [0.17,0.25] | 0.24 [0.19,0.29] |
North Atlantic (NATL; 10–60°N, 0–75°W) | |||
Trend | 0.23 [0.19,0.26] | 0.21 [0.18,0.24] | 0.22 [0.18,0.25] |
Trend+AR1 | 0.14 [0.12,0.16] | 0.19 [0.16,0.21] | 0.22 [0.19,0.25] |
Trend+CA | 0.13 [0.11,0.15] | 0.19 [0.16,0.21] | 0.23 [0.18,0.28] |
DepreSys | 0.13 [0.11,0.14] | 0.12 [0.10,0.14] | 0.14 [0.11,0.16] |
South Atlantic (SATL; 10–30°S, 50°W–20°E) | |||
Trend | 0.26 [0.20,0.32] | 0.18 [0.14,0.23] | 0.18 [0.13,0.23] |
Trend+AR1 | 0.24 [0.20,0.28] | 0.18 [0.14,0.22] | 0.19 [0.14,0.23] |
Trend+CA | 0.24 [0.20,0.29] | 0.18 [0.12,0.22] | 0.16 [0.12,0.18] |
DePreSys | 0.22 [0.18,0.26] | 0.16 [0.13,0.19] | 0.16 [0.13,0.18] |
For MDR, Trend+AR1 is the best performing model for year 1, while Trend+CA is the best for years 2–5 and 6–9, however the improvement in the verification metrics from the Trend prediction and Trend+AR1 to Trend+CA is smaller compared to SPG. Trend+AR1 appears to be the best for all lead times for NATL. For SATL, Trend+CA performs better than the other models at longer lead times.
We note that another possible method to obtain regional average forecasts is to first average, year by year, the historical observations for all grid boxes in the region and then make forecasts from the spatially averaged annual mean SSTs as described in Sects. 3.1 and 3.3. This alternative method gives generally similar results.
5 Comparison with DePreSys retrospective forecasts
For year 1, DePreSys significantly outperforms Trend+AR1 over large parts of the tropical Pacific, which is likely to be related to its skill in predicting ENSO at the seasonal timescale. It also performs better than Trend+AR1 for parts of the Indian Ocean and better than both statistical models over parts of the tropical Atlantic. However, both Trend+AR1 and Trend+CA have clear advantages over DePreSys in the NAC region.
A different pattern evolves at longer lead times. The advantages of DePreSys in the tropical Pacific diminish, and DePreSys performs worse than Trend+AR1 for most parts of the Pacific and the Indian Ocean for years 2–5. On the other hand, DePreSys is clearly superior to Trend+AR1 over large parts of north Atlantic including the SPG for years 2–5 and 6–9. Similar results can be seen for the comparison between DePreSys and Trend+CA. For parts of the MDR and South Atlantic, however, Trend+CA performs significantly better than DePreSys.
Tables 1 and 2 also show the ACC and RMSE of the regional average DePreSys retrospective forecasts for the four regions we considered in Sect. 4.3. DePreSys generally ourperforms the statistical models, especially for the SPG where for years 2–5 and 6–9, it has significantly higher ACC and lower RMSE than the best statistical model, Trend+CA. The results for SATL are different. Although the RMSE for DePreSys is the lowest for all three lead time periods, Trend and Trend+CA have higher ACC for years 2–5 and years 6–9 respectively.
6 Statistical forecast for 2012–2021
The regional average forecasts by Trend+AR1 and Trend+CA are shown by thick dark red dashed and solid lines respectively in Fig. 8. To give an indication of possible errors in these forecasts, the corresponding average RMSE of the retrospective forecasts for different lead times are shown by the grey shadings. As in what we have seen in the forecast maps (Fig. 10), there is no major disagreement between forecasts by Trend+AR1 and Trend+CA. A slow cooling trend is predicted for SPG, while NATL and SATL are expected to have a warming trend. In particular, the magnitude of predicted rise in the average SST in SATL (about 0.5 K) is slightly larger than the average historical RMSE (about 0.3 K).
We note that the above forecasts are based on projected equivalent CO_{2} concentration for the RCP4.5 concentration pathway. Forecasts using alternative pathways (e.g. RCP2.6 and RCP8.5) were considered. Since the difference in equivalent CO_{2} concentration among various pathways are small (less than 10 ppm difference from RCP4.5), the forecasts are generally similar to those presented above (not shown).
7 Conclusions
Both the simple AR1 model and the more complex CA model provide skillful predictions of the internal variability in the far North Atlantic for years 2–5, and the skill of CA extends to years 6–9.
CA gives skillful predictions of the internal variability for parts of the South Atlantic for years 6–9.
Although DePreSys, an initialised GCM, performs significantly better than the statistical models for most parts of the North Atlantic at extended lead times, Trend+AR1 or Trend+CA outperforms DePreSys significantly for certain regions, such as the NAC and South Atlantic.
With a projected increase in equivalent CO_{2} concentration, both statistical models forecast a small cooling trend in the subpolar gyre region for the years 2012–2021, while most other parts of the Atlantic are expected to warm, especially in the South Atlantic.
We have also attempted to understand the source of predictive skill in our statistical forecasts. We have identified certain regions where the prediction of the forced trend plays a dominant role in having skillful forecasts, such as parts of mid-latitude and tropical Atlantic, the Indian Ocean and the western Pacific. The prediction of internal variability is more important for some other regions, such as the far North Atlantic and perhaps parts of South Atlantic. These findings can help identify areas where decadal predictions may be improved by adding more observations.
There is much scope for further work on developing benchmark statistical models for decadal predictions. While the retrospective forecasts by our two statistical models are skillful for certain regions on a decadal timescale, other modelling strategies may be considered. Specifically, the separation of the effects of radiative forcings and internal variability on SST variability is not trivial. The negative correlation for our forced trend prediction in the far North Atlantic suggests that our model is far from ideal. One aspect worth noting is the effect of tropospheric aerosols. The equivalent CO_{2} concentration data used in this study represent the total net global effects of greenhouse gases and aerosols, however the spatial distribution of aerosols is not uniform which means that our model is unlikely to capture the regional climate effects of aerosols. Meanwhile, recent work by Booth et al. (2012) suggested that aerosols have played a key role in forcing North Atlantic SST changes. Some other known radiative forcing effects are also not included in this study, for example solar activity and volcanic eruptions. Further work is required to explore the best modelling option. In addition, for the modelling of the forced trend and AR1 processes, a separate model has been fitted to each grid box. A modelling approach in which all grid boxes are fitted together and parameters for each grid box assumed to be related would be worth exploring. Finally, this study has considered SSTs because of their influence on weather systems and climate patterns. It will be also interesting to explore if our statistical models can offer skillful regional predictions of other climate variables, for example land surface temperatures.
The time of emergence may also be defined as the year when the ratio first crossed and exceeded the threshold of one. We use a more conservative measure here as the temporary reduction in equivalent CO_{2} concentration in the 1950s led to a decrease in climate change signal in some places.
For example, Hawkins et al. (2011) estimated that for Atlantic SSTs, one would need 10^{5} years of data to find a natural analogue.
Hindcasts with start year from 2002 onwards give predictions beyond year 2011, therefore less than 46 sets of retrospective forecasts can be verified for lead times longer than 6 years.
Acknowledgments
We would like to thank Nick Dunstone, Chris Ferro, Doug Smith and David Stephenson for useful comments on this work. The authors are supported by NCAS-Climate (CKH, EH, LS), the EU project THOR (CKH, EH) and Walker Institute (CKH). Research leading to this paper has received funding from the European Community’s 7th framework programme (FP7/2007–2013) under grant agreement No. GA212643 (THOR: ‘Thermohaline Overturning–at Risk’, 2008–2012) and a NERC grant (No. NE/H011420/1).
Open Access
This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.