A Dynamical One-Month Lead Seasonal Rainfall Prediction for March

to May 1999 for the North-Eastern Area of South America

contributed by Tony Evans and Richard Graham



Room 329, NWP Division UK Meteorological Office, Bracknell, United Kingdom





The UK Meteorological Office (UKMO) is attempting to estimate potential dynamical seasonal predictability given ideal surface boundary conditions on a global scale as part of the European Union funded PROVOST (PRediction Of climate Variations On Seasonal and interannual Timescales) project (Graham et al., 1999). In addition, an updated version of the UKMO Unified Model (UM), HADAM3, has been integrated in 9-member ensembles to a range of four months for each season over 19 years from 1979 to 1997. Initial conditions are obtained from the ECMWF reanalysis and SST anomalies from the UKMO GISST and Reynold's OI datasets. The ensemble members were initialized at 24 hour intervals using 0000Z analyses finishing on the day prior to the start of the season. The north-eastern region of South America has been identified as an area of relatively high predictability for the UM and based on this assessment we present a real-time seasonal forecast at one month lead. The area with higher predictability broadly stretches through much of eastern Brazil, the Amazon Basin, French Guiana, Surinam and Guyana, even covering the southeastern parts of the Caribbean.

In north-eastern South America, correlations over the 19 years between ensemble-mean rainfall anomalies and anomaly values obtained from the rainfall data set from the CPC Merged Analysis of Precipitation (CMAP - Xie and Arkin, 1997), a combination of satellite- and rain-gauge derived data, exceed 0.5 with correlations in excess of 0.8 in the vicinity of the Nordeste region of Brazil (Fig. 1). The observed data set is gridded to the same resolution as the model (2.5ox3.75o), but in order to reduce noise four by four gridpoints have been joined together to produce Figure 1. Time series of ensemble mean simulated rainfall and the CMAP data, expressed as a percentage of climatology (Fig. 2), illustrates the fact that the dynamical model is able to capture the interannual variability reasonably well. Moreover, comparison of observed and ensemble mean quintiles (well-below, below, normal, above and well-above) shows that the simulation mean is in error by at most one quintile in 17 out of 19 years.

Here we attempt to give an estimate of probabilistic predictability using the Relative Operating Characteristic (ROC) approach (Stanski et al., 1989). The ROC for a specific event (e.g. well-above average) is expressed in the form of a curve plotting hit rates against false alarm rates for a specific event over a range of forecast probability thresholds. Note that both the hit and false alarm rates are defined as the proportion of the observed frequencies of the event and the non-event respectively. ROC evaluations of the simulations for all gridpoints in each region over all 19 seasons for five quintile events have been produced (e.g. ROC for region D, Fig. 3). Note that the greater the skill of the ensemble, the more the ROC curve must bow up towards the top left corner; the point (0,1) corresponding to perfect deterministic skill and points on the diagonal corresponding no skill (i.e. performance available from a climate or random forecast). Thus the area under the ROC curve provides a useful overall index of skill; a value of 0.5 (the area under the diagonal) or less indicating no skill, and a value of one indicating perfect deterministic skill. ROC areas for each region and for each event quintile are shown in Figure 4. Probabilistic skill exceeds the chance level for every event type in all regions and is apparently highest for the more extreme events (i.e. well-below and well-above average), a result seen frequently in categorical forecasts e.g. Van den Dool and Toth (1991), who state that low skill near the mean may be a feature of categorical definition rather than having a physical explanation.

Forecasts as produced for the 1999 March to May season are derived from nine-member ensemble runs, but with two major design differences from the predictability experiments outlined above. First, initialization for the predictions is from 2-4 February rather than late February as in the simulations. Secondly the real-time experiments use persisted SST anomalies (from January) throughout. Neither is thought likely to have significant negative impacts on the model's ability to provide real-time predictions. Skill over the region as deduced from the simulations remains high throughout the year, whether for months 1 to 3 or 2 to 4 of the simulations. Indeed equivalent levels of skill tend to be present on a monthly time scale, although with some drop-off into the fourth month. Hence the shift in start date is considered unlikely to affect potential predictability for the region from that deduced for the 'standard' seasons. Experiments for 12 boreal spring seasons have been carried out to test the impact of using persisted anomalies. While there is some inevitable loss of predictability associated with the use of persisted anomalies this appears to be minimal in areas of relatively high predictability such as considered here, and certainly does not eliminate predictability in terms of the levels normally associated with seasonal forecasts (Evans et al., 1998). Use of persisted anomalies fails, of course, during seasons in which there is a substantial readjustment of SST anomalies over ocean areas related to a given region's rainfall; experience has been gained of such failed forecasts for the Nordeste in preliminary work with the model. Currently there is no solution to this problem of rapid intraseasonal SST anomaly distribution changes: the forecasts given below are conditional on the continuity of the January anomalies.

The Ensemble Mean forecast (Table 1) is for near-normal conditions in the southern Carribean (region A), above-normal conditions in Guyana, Surinam, French Guiana, eastern Venezuela (region B), well-above normal in Eastern Amazonia (region C) and above-normal in north eastern Brazil (region D). This prediction for region D is in agreement with the UKMO statistical forecast of Colman and Davey (1999).

Expressed probabilistically (Fig. 5), well-above average conditions are very likely in region C (apparently a very skillful category), but spread is higher in region A where the middle three categories are given a near or higher than climatological probability of occurring. Wet or Very Wet (relatively skillful categories) conditions appear equally likely in region D.



References

Colman, A. and Davey, M., 1999: Statistical prediction of March-May 1999 rainfall in NE Brazil using input from Multiple Regression and Discriminant Analysis, this volume.

Evans, A., Harrison, M., Graham, R.,Davey, M. and Colman, A., 1998: A Dynamical One-Month Lead Seasonal Rainfall Prediction for March to May 1998 for the North-Eastern Area of South America. COLA Long-Lead Forecast Bulletin, March 1998, 7(3), 33-37.

Graham, R., Evans, A., Mylne, K., Harrison, M. and Robertson, K. 1999: An assessment of seasonal predictability using Atmospheric General Circulation Models. Submitted to Quarterly Journal of the Royal Meteorological Society.

Stanski, H.R., Wilson, L.R. and Burrows, W.R., 1989: Survey of common verification methods in Meteorology. World Weather Watch Tecnical Report. No. 8, WMO/TD 358, 114pp.

Van den Dool, H.M. and Toth, Z., 1991: Why do Forecasts for "Near Normal" Often Fail?. Weather and Forecasting 6, 76-85.

Xie, P. and Arkin, P.A., 1997. Global Precipitation: a 17-year monthly analysis based on gauge, observations, satellite observations and numerical model outputs Bulletin of the American Meteorological Society. Vol 78,2539-2558.

Forecast % A B C D
E Mean 97 (Q3) 109 (Q4) 125 (Q5) 115 (Q4)
Highest Member 150 (Q5) 140 (Q5) 134 (Q5) 125 (Q5)
E Mean 60 (Q2) 89 (Q3) 118 (Q5) 106 (Q3)

Table 1. March to May 1999 seasonal forecast rainfall, as percentages of normal, for the Ensemble Mean (E Mean - with respect to the model 1979-1997 climate) and for the highest and lowest ensemble members for each of the four areas depicted in Figure 1. The forecasts are expressed as quintiles (five equi-probable categories relative to normals): Q1:well-below, Q2:below, Q3:normal, Q4:above, Q5: well-above.

Figure 1. Correlations for simulations over March to May 1979-1997 between ensemble mean rainfall and the Observed CPC Merged Analyses of Precipitation Anomalies over 10 latitude x 15 longitude degree blocks.

Figure 2. Timeseries for Area D of Ensemble Mean (EM) rainfall (triangles joined with solid line) over March to May 1979-97 from simulations and CMAP observations (crosses joined with dashed lines) expressed as (a) percentage of normal and (b) in 5 quintile categories.

Figure 3. Relative Operating Characteristics (ROC) curves for March to May 1979-97 simulations for 5-equi-probable rainfall events relative to normal over region D (NE Brazil): Well-below (quintile 1), below (quintile 2), normal (quintile 3), above (quintile 4) and well-above (quintile 5). The curves are constructed from hit and false alarm rates at four thresholds on the forecast probability of the event (20%, 40%, 60% and 80%). The curve is bounded by the points (0,0) and (1,1) which correspond respectively to the false alarm and hit rates achieved through never and always forecasting the event. The areas under the ROC curve are provided and give an overall measure of skill, 0.50 or less representing the area below the diagonal and that obtained with a random or climatology forecast. ROC areas for other regions are illustrated in Fig. 4.

Figure 4. ROC areas for March to May 1979-97 simulations for 5-equi-probable rainfall events relative to normal over region D (NE Brazil): Well-below (quintile 1), below (quintile 2), normal (quintile 3), above (quintile 4) and well-above (quintile 5). Areas of 0.50 and below represent skill associated with a random or climatology forecast.

Figure 5. March to May 1999 forecast for probabilities of rainfall occurring in 5-equi-probable categories, or quintiles, calculated from the number of ensemble members in each category. Climatological quintile probability is 20%.