Initialized near-term regional climate change prediction

Abstract

Climate models are seen by many to be unverifiable. However, near-term climate predictions up to 10 years into the future carried out recently with these models can be rigorously verified against observations. Near-term climate prediction is a new information tool for the climate adaptation and service communities, which often make decisions on near-term time scales, and for which the most basic information is unfortunately very scarce. The Fifth Coupled Model Intercomparison Project set of co-ordinated climate-model experiments includes a set of near-term predictions in which several modelling groups participated and whose forecast quality we illustrate here. We show that climate forecast systems have skill in predicting the Earth's temperature at regional scales over the past 50 years and illustrate the trustworthiness of their predictions. Most of the skill can be attributed to changes in atmospheric composition, but also partly to the initialization of the predictions.

Figure 1. Forecast quality of several climate indices.

(a–c) Time series of the ensemble-mean forecast anomalies averaged over the forecast years 2–5 (solid, Init) and the accompanying non-initialized (dashed, NoInit) experiments of the global-mean near-surface air temperature (SAT) (a), the AMV (b) and IPO (c) indices. The observational time series, GISS global-mean near-surface air temperature and ERSST for the AMV and IPO, are represented with dark (positive anomalies) and light (negative anomalies) grey vertical bars, where a 4-year running mean has been applied for consistency with the time averaging of the predictions. The box-and-whisker represents the multi-model ensemble range (anomalies with respect to the multi-model ensemble mean) of Init (solid) and NoInit (dashed), where the whiskers correspond to the maximum and minimum, the box to the interquartile range and the horizontal bar to the median. The predictions have been initialized once every year over the period 1961–2006. (d–f): Correlation of the ensemble mean with the observational reference along the forecast time for 4-year averages. The one-sided 95% confidence level with a t-distribution is represented in grey, where the number of degrees of freedom has been computed taking into account the autocorrelation of the observational time series, which are different for each forecast time. A two-sided t-test (with the number of degrees of freedom computed taking into account the autocorrelation of the observational time series) for the differences between the Init and NoInit correlation found no significant results with confidence ≥90%. (g–i): RMSE of the ensemble mean along the forecast time for 4-year forecast averages. Squares are used where the Init skill is significantly better than the NoInit skill with 95% confidence using a two-sided F-test where the number of degrees of freedom takes into account the autocorrelation of the observation minus prediction time series. (j–l) Ensemble spread estimated as the s.d. of the anomalies around the multi-model ensemble mean.

Figure 2. Near-surface air-temperature forecast quality.

(a,b) RMSSS (multiplied by 100) of the ensemble mean of the Init multi-model for predictions averaged over the forecast years 2–5 (a) and 6–9 (b). A combination of temperatures from GHCN/CAMS air temperature over land, ERSST and GISTEMP 1200 (ref. 49) over the polar areas is used as a reference. Black dots correspond to the points where the skill score is statistically significant with 95% confidence using a one-sided F-test taking into account the autocorrelation of the observation minus prediction time series. (c,d) Ratio of RMSEs between the Init and NoInit multi-model experiments for predictions averaged over the forecast years 2–5 (c) and 6–9 (d). Contours are used for areas where the ratio of at least 75% of the individual forecast systems has a value above or below 1 in agreement with the multi-model ensemble-mean result. Dots are used for the points where the ratio is statistically significantly above or below 1 with 90% confidence using a two-sided F-test that takes into account the autocorrelation of the observation minus prediction time series. Poorly observationally sampled areas are masked in grey.

Figure 3. Near-surface temperature and precipitation relative linear trends.

Ratio between the slope of the linear trend and the residual variability (1 per year) over 1961–2010 for (a) near-surface temperature and (b) GPCC precipitation. A combination of temperatures from GHCN/CAMS air temperature over land, ERSST and GISTEMP 1200 over the polar areas is used as a reference. Monthly values have been smoothed with a 4-year running average before estimating the trend and the residual variance. Poorly observationally sampled areas are masked in grey.

Figure 4. Precipitation forecast quality.

(a,b) RMSSS (multiplied by 100) of the ensemble mean of the Init multi-model for predictions averaged over forecast years 2–5 (a) and 6–9 (b). GPCC precipitation is used as a reference. Black dots correspond to the points where the skill score is statistically significant with 95% confidence using a one-sided F-test taking into account the autocorrelation of the observation minus prediction time series. (c,d) Ratio of RMSEs between the Init and NoInit multi-model experiments for predictions averaged over forecast years 2–5 (c) and 6–9 (d). Contours are used for areas where the ratio of at least 75% of the individual forecast systems has a value above or below 1 in agreement with the multi-model ensemble-mean result. An inference tests at the grid point level was applied to assess if the ratio is statistically significantly above or below 1 with 90% confidence using a two-sided F-test that takes into account the autocorrelation of the observation minus prediction time series, but no point was found significant. Both predictions and the observational reference were smoothed to a 5° grid to reduce the spatial variability of the results.

Figure 5. Multi-model ensemble spread for the near-surface temperature.

Ratio between the spread and the RMSE of the ensemble mean for Init (a) and NoInit (b) for the predictions averaged over forecast years 2–5. A combination of temperatures from GHCN/CAMS air temperature over land, ERSST and GISTEMP 1200 (ref. 49) over the polar areas is used as a reference.