Seasonal Prediction of Mediterranean temperature and precipitation anomalies by statistical model ensembles (already finished)

Start date: 01.10.2004
Duration: vorauss. bis Ende 2008
Funded by: DFG (Deutsche Forschungsgemeinschaft)
Local project leader: Prof. Dr. J. Jacobeit Dr. E. Hertig


The intended seasonal predictions of Mediterranean pressure, temperature, and precipitation variations comprise basic information about the probable climatic character over a longer period of time (month to season) and hence differ from forecasting individual atmospheric conditions (as done in weather forecasts). Based on extensive observational data and using different advanced statistical methods, quantitative relationships between Mediterranean climate variables and large-scale predictor fields of atmospheric, oceanic, and terrestrial variables are established in consideration of relevant time lags. In this context, sea surface temperatures and so-called centres of teleconnectivity play a decisive role in particular. The derived statistical models are verified in periods being independent from the model calibration periods. Models which pass the verification procedure are pooled into an ensemble of models. This model ensemble is subsequently used to generate probabilistic predictions of the occurrence of pressure, temperature, and precipitation anomalies in the Mediterranean area from particular parameter configurations in the predictor fields. Furthermore, models are fitted to take into account instationarities in the predictor- predictand- relationships. The application of the seasonal climate forecasts in the scope of adaptation strategies and risk management is allowed for by providing up-to-date forecast information with relevant comments on the internet.



Since understanding the predictability of the atmosphere at monthly to seasonal time scales has improved considerably, the provision of basic information about the probable climatic character over a longer period of time has become a possibility in general. Beyond the objective of scientific insight, seasonal predictions are of high economic and social importance, especially in a climatic-sensitive region like the Mediterranean area: provided that they hold an adequate performance, seasonal predictions can facilitate decisions in the range of water management, crop production, and tourism


Using Canonical Correlation Analysis with time lags from 1 to 12 months, statistical relationships between Mediterranean climate variables and atmospheric, oceanic and terrestrial predictors are established in the observational period 1950-2003. The derived statistical models are used to produce hindcasts, in such a way that the models are tested in years being excluded from the model calibration periods. Hindcast performance is used to compile ensembles of models with adequate time lags and meaningful predictor configurations for the prediction of the occurrence of temperature and precipitation anomalies in the Mediterranean area.


As an example results of ensemble hindcasts of the mean January temperature in the Mediterranean area are discussed. Preceding 500hPa- geopotential heights, sea surface temperature (SST), and/or 850hPa- air temperature anomalies are taken into consideration as predictors.
The graphs of Fig. 1 show for increasing lead times the correlation coefficients of the cross- validated hindcast ensembles of Mediterranean temperature in January, predicted from monthly, bi-monthly, and three-month means of 500hPa-geopotential heights ( a ), 850hPa-air temperature ( b ), and SSTs ( c ). The best performing hindcast ensembles are used to establish multi-type predictor hindcasts. This leads to the finding that the August to October mean of SSTs (lead time 2) combined with the November/December mean of 850hPa-air temperature (lead time 0) form an adequate predictor combination for January temperature in the Mediterranean area.



Fig. 1 : Correlation Coefficients for cross-validated hindcast ensembles of Mediterranean temperature in January, predicted with increasing lead times from monthly (red bars), two-month (blue bars), and three-month averages (green bars) of 500hPa-geopotential heights (a), 850hPa-air temperature (b), and SST anomalies (c).

The prementioned ensemble hindcast of January temperature in the Mediterranean area, predicted from 850hPa-air temperature of November/December and SSTs of August/September/October is taken to illustrate results in more detail. The categorial representation of the result (below normal, near normal, above normal, Fig. 2 ) displays that in wide parts near normal temperatures were observed as well as forecasted, whereas in the Western Mediterranean above normal values are forecasted, but near normal values occurred. Below normal anomalies in the northeastern Mediterranean area are partly reproduced. Additionally the corresponding contingency table is shown in Table 1 .

Fig. 2 : Classified forecast ensemble (left side) and corresponding observed classes (right side) of January temperature. Predictor: preceding November/ December mean of 850hPa-air temperature and August/September/October mean of SST anomalies. Threshold values for classes: limits of the confidence interval for the long-term mean at 10 % confidence level (using t-distribution).


Table 1 : 3 x 3 contingency table for the example of Fig 2. A total of 2363 grid boxes represent the Mediterranean area.

Forecast set performance is measured against ‘climatology', represented by the long-term mean. Other standard quantities, e.g. bias (for the above example see Fig. 3 ), correlation of forecasted and observed values ( Fig. 4 ), mean squared error, and amplitude error (not shown) are also used for verification. These quantities not only serve to select the best-performing models, but furthermore can give insight into systematic, phase, and amplitude errors and might therefore add to an improvement of forecast performance.

Fig. 3 : mean bias (forecast minus observation) for the example of Fig.2. Overall mean bias = 0.081 Fig. 4 : correlation of forecasted and observed values for the example of Fig. 2. Overall correlation = 0.596


The predictors-predictand-relationships are further investigated by the use of t-mode principal component analysis and composite calculations. Occuring instationarities are taken into account by means of analogue techniques, in such a way that those predictors-predictand-relationships are explicitly selected which are most similar to current parameter values. In addition long-term trends and cyclic components are incorporated through fitting procedures to further enhance forecast performance.


This work is supported by the DFG (German Research Foundation) under contract JA 831/3-1.