Seasonal prediction of African precipitation with ECHAM4–T42 ensemble simulations using a multivariate MOS re-calibration scheme

Abstract

A 15 member ensemble of 20th century simulations using the ECHAM4–T42 atmospheric GCM is utilized to investigate the potential predictability of interannual variations of seasonal rainfall over Africa. Common boundary conditions are the global sea surface temperatures (SST) and sea ice extent. A canonical correlation analysis (CCA) between observed and ensemble mean ECHAM4 precipitation over Africa is applied in order to identify the most predictable anomaly patterns of precipitation and the related SST anomalies. The CCA is then used to formulate a re-calibration approach similar to model output statistics (MOS) and to derive precipitation forecasts over Africa. Predictand is the climate research unit (CRU) gridded precipitation over Africa. As predictor we use observed SST anomalies, ensemble mean precipitation over Africa and a combined vector of mean sea level pressure, streamfunction and velocity potential at 850 hPa. The different forecast approaches are compared. Most skill for African precipitation forecasts is provided by tropical Atlantic (Gulf of Guinea) SST anomalies which mainly affect rainfall over the Guinean coast and Sahel. The El Niño/Southern Oscillation (ENSO) influences southern and East Africa, however with a lower skill. Indian Ocean SST anomalies, partly independent from ENSO, have an impact particularly on East Africa. As suggested by the large agreement between the simulated and observed precipitation, the ECHAM4 rainfall provides a skillful predictor for CRU precipitation over Africa. However, MOS re-calibration is needed in order to provide skillful forecasts. Forecasts using MOS re-calibrated model precipitation are at least as skillful as forecast using dynamical variables from the model or instantaneous SST. In many cases, MOS re-calibrated precipitation forecasts provide more skill. However, differences are not systematic for all regions and seasons, and often small.

Appendix

Over many African regions the CRU gridded data are based only on very few station data. Data sparce areas with almost no observations are the Sahara and central Africa (e.g., large parts of the DR Congo). In order to assess the sensitivity of our results to inhomogeneous data coverage, we repeated the MOS re-calibration forecast for the first and second half of the century, respectively. Figure 24 shows the correlation coefficients between the MOS re-calibrated ECHAM4 forecast and the CRU precipitation for the period from 1903 to 1952 (left column), and from 1953–2002 (middle column). For both periods, the MOS re-calibration was derived from the whole available period of data, but the forecast was performed only for the period from 1903 to 1952 or 1953 to 2002, respectively. The right column panels in Fig. 24 show the correlation coefficients for a forecast where both, the training as well as the forecast include only data from 1953 to 2002. The 95% level of significant correlations (N = 50) is now 0.28. However, for reason of comparability we kept the scale as in Fig. 12.

Fig. 24

Pointwise correlation between MOS re-calibrated ECHAM4 and CRU precipitation for (left panels) the period 1903–1952, (middle panels) the period 1953–2002, both including the period 1903–2002 into the training procedure and (right panels) for the period 1953–2002 without using the data prior to 1953

First of all, potential predictability for the JFM and AMJ season seems stronger during the first compared to the second half of the century (compare left and middle column in Fig. 24). Data are very sparce in that region, with almost no station data before the 1930s. Thus, this strong signal of potential predictability might well be an artifact of the sparce data. During the JAS and OND seasons predictability is larger during the second half of the century. Besides the data quality, this might also result from a stronger ENSO activity during the second half of the century. This follows also from Fig. 25, where the distributions of the correlation coefficients is represented.

Fig. 25

Distribution of the correlation coefficients over Africa from Fig. 24. The dotted dashed line represents the forecast for 1903–1952, the solid line the forecast for 1953–2002, both using all data during the training, the dashed line represents the forecast without using data prior to 1953

Excluding the first half of the century from the training process tends to slightly degrade the forecast (Fig. 24, right column). When comparing the distributions of the correlation coefficients (Fig. 25), the tendency to produce more negative and weaker positive correlations using the short-training period (dashed line) is particularly strong for OND, but holds for all seasons including those not shown here. Thus, no evidence is found that including the data prior to the 1950 degrades the forecast.