The Met Office is conducting research into
the effects of sea surface temperatures and other climatic variables on
tropical rainfall. As part of this research, experimental forecasts have been
made of seasonal rainfall for the Sahel (region 1) for each year from 1986 onwards.
Since 1992, forecasts of seasonal rainfall have also been made for a slightly
redefined rectangular Sahel (region 2, 15W to 37.5E and 12.5N to 17.5N), for an
area south of the Sahel (region 3, 7.5W to 33.75E, 10N to 12.5N), and for an
area extending further south to the coast (region 4, approximately 7.5W to
7.5E, 5N to 10N). The four regions are labelled in figure 1a.
The statistical forecasting techniques are
based on May and June sea surface temperature (SST) anomaly patterns. Further
details can be found in Folland et al, 1991. Several forecasts have been made
using different versions of each technique, and they have been averaged
together with dynamical forecasts produced using the Met Office GLObal SEAsonal
(GLOSEA) coupled ocean-atmosphere circulation model and persistence forecasts
(observed rainfall for last year’s season) to obtain the forecasts shown below
in figure 1. The GLOSEA model has
replaced the atmosphere-only model forced with persisted SST which was used for
these forecasts in previous years.
The forecast period for regions 2-4 is
July-September. For region 1 annual rainfall is predicted, though most of the
rain in this region falls during July- September. For forecasting purposes, the
predicted rainfall indices are categorised into quints which are equi-probable
over 1961-1990. The 5 quints are referred to as Very Dry, Dry, Average, Wet and
Very Wet. In table 1 the quints are defined as percentages of 1961-1990 average
SEA SURFACE TEMPERATURE ANOMALIES
The SST indices used to predict rainfall in
N Africa represent regional and global scale anomaly patterns. Most important
are tropical Pacific and Atlantic anomalies, and interhemispheric differences
in anomalies.
SST in the northern hemisphere is
predominantly warmer relative to normal than SST in the Southern hemisphere.
This interhemispheric contrast and below average SST in the tropical East
Pacific favours above average rainfall in regions 1,2 and 3. In the Gulf of
Guinea region, SST is close to average which favours near average rainfall in
region 4.
Note. At the time of the forecast we only
have SST anomalies for the first half of June 2003. For forecasting purposes,
we assume the anomalies observed during
the first half of June will persist for the remainder of the month. Since SST
persistence is quite high on this timescale, this is unlikely to have a large
impact on the forecast.
THE
PREDICTION SYSTEM
The forecasts are weighted combinations of
statistical forecasts (table 2), dynamical forecasts and persistence (last year’s
observed seasonal rainfall). The statistical best estimate forecasts are
produced by linear regression with SST indices as predictors. Statistical
probability forecasts are calculated from the same SST indices using linear
discriminant analysis.
Prior to 2001, only predictors (a) and (b)
were used. In 2001 predictors (a) were replaced by predictors (c) which use
more up to date SST analyses. Predictors
(d) were added for areas 1,2 and 3. The new predictors (c and d) were found to improve trial forecast
skill over 1951-2000 (Table 3). In particular adding predictor (d) improved
skill in predicting region 1 and 2 rainfall variability between 1981-2000.
Predictors (b) and (c) are poor at predicting variability over this period. The
trial forecasts referred to in table 3 were produced using the jackknife method
in which data for the forecast year and the next two subsequent years are
excluded when calculating prediction equations.
The statistical forecast is a correlation
skill weighted combination of methods b,c and d. Predictors b,c and d are
approximately weighted 0.25, 0.25 and 0.5 respectively. Predictor d has a
higher weight than predictors b and c since this predictor is
much better at predicting the 1981-2000 seasons than predictors b and c
and since forecasts from predictors b and c are quite highly correlated with
each other.
The dynamical forecast was produced using
the new Met Office GLObal SEAsonal (GLOSEA)
coupled model which is a combination of the 19 level HADAM3 version of
the Met Office Atmosphere global circulation model and the Met Office Ocean
Model. The forecast is based on an ensemble of 40 GLOSEA runs each initialised
with slightly different perturbations of ocean surface conditions observed in
early June. Further information about
dynamical ensemble forecasts at the Met Office can be found on our
website at www.metoffice.com/research/seasonal.
The dynamical forecast output is expressed
as both deterministic forecasts and probability forecasts for the 5 quint
categories. The model forecasts were calibrated using 9 member
ensemble hindcasts for 42 years between 1959 and 2000 produced as part of
the DEMETER project (www.ecmwf.int/research/demeter).
The deterministic forecasts are produced by correcting the ensemble mean
forecast for model bias as observed in model simulations and hindcasts for
1961-1990. To evaluate the dynamical forecast probabilities for
5 observed quint categories , 5 frequency distributions of observed
quint categories are evaluated for sets for years when the model simulates or
predicts the same category. The forecast probabilities are proportional
to the mean of these frequency distributions for the 5 categories
predicted by the 9 forecast members.
The forecasts are weighted to reflect the
reliability of the different inputs. The ratio of weights for the statistical
forecast/dynamical forecast/persistence are shown in table 4 . Persistence is
not used for the region 4 forecast, as persistence skill is negligible for this
region. Dynamical skill is somewhat
higher for region 4 than for the other regions hence the higher weights.
LAST
YEAR
Last year, the DRY category was observed in
all 4 regions.
FORECAST
SUMMARY
Forecasts for regions 1-4 are shown in figure
1. Weighted average deterministic forecasts are shown as percentages of the
1961-1990 average in figure 1a. In Figure 1b, the forecasts are expressed as
percentage standardised units (e.g. standardised values of +100 indicate
rainfalls one standard deviation above average) relative to 1961-1990 (NB.
1901-1980 for region 1 for compatibility with previous publications by the Met
Office and Nicholson (1984). Quint categories are indicated in figure 1c. The
skill of these weighted forecasts is indicated in fig. 1d by the trial forecast
correlations with observed rainfall in the period 1951-2000. The correlations
are well above the 5% significance level for all 4 regions. Probability
forecasts for the 5 quint categories are shown in figure1f-j respectively. The
Relative Operating Characteristic (ROC) skill in figure 1e is a measure of the
performance of these probability forecasts over the period 1951-2000. ROC
scores above 60% are considered to indicate significant (5% level) skill.
There are
considerable differences between the forecasts for 2003 provided by the
different methods. Persistence favours the DRY category in regions 1,2 and
3. The statistical favour the VERY WET category for regions 1,2 and 3 but
the dynamical forecast favours near or
below average rainfall in these regions. Our best estimate for these
regions is closer to the statistical
forecast because of its relatively high skill compared to the dynamical
model . For region 4 the statistical forecasts are close to average but
the dynamical forecast is VERY DRY. The best estimate forecast for this region
is VERY DRY reflecting a very strong dynamical model signal (model forecast is
2.5 standard deviations below climatology) and relatively high model skill
compared to the other regions. Confidence is LOW due to disagreement between
dynamical and statistical forecasts.
Our
best estimate forecasts are:
Region
1: WET
Region 2: WET
Region 3: WET
Region 4: VERY DRY
Hence, rainfall is expected to be greater
in 2003 than during the past 3 years in
regions 1,2 and 3 but drier than last year in region 4. There is an
above chance probability of a “VERY WET” category rainfall season in regions
1,2 and 3 (fig 1j).
REFERENCES
Folland, C.K., Owen, J., Ward, M.N and
Colman, A.W. 1991: Prediction of seasonal rainfall in the Sahel region using
empirical and dynamical methods. Journal of Forecasting, 10,
21-56.
Folland, C.K., Parker, D.E.,
Colman, A.W. and Washington,R. 1999: Large scale modes of Ocean Surface
Temperature since the late nineteenth century. In Beyond El Nino, decadal and
Interdecadal variability. Ed. A Navarra, Springer pp 75-102.
Nicholson, S.E. 1985: Sub-Saharan rainfall 1981-84. J.
Clim. Appl. Met., 24, pp 1388-1391.
