Experimental Forecast of 2000 Seasonal Rainfall in the Sahel and other Regions of Tropical North Africa Using Dynamical and Statistical Methods
contributed by Andrew W Colman, Mike K Davey, Richard Graham and Robin Clark
The Met Office, Bracknell , Berkshire, RG12 2SY, UK
The UK 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, irregular shaped region; bounded by thick solid line in
figure 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). Maps of these regions are shown in Figures 1 and 2.This forecast uses ocean and atmosphere information available in early May and was issued to National weather services in the affected area in May.
1. The Statistical Forecast
The statistical forecasting methods used are multiple linear regression and discriminant analysis. Predictors are indices of March and April sea surface temperature anomaly patterns which are represented by eigenvectors. Predictors for regions 1,2 and 3 include (in order of importance), a global pattern showing opposing weights in the northern and southern hemisphere oceans respectively, a global pattern with strong weights in the tropical south Atlantic, a global pattern showing ENSO related variability and regional patterns for the South Atlantic, Pacific and Indian Oceans. There is one predictor used for the Guinea region (region 4), a South Atlantic pattern with strong weights in the Gulf of Guinea. The discriminant analysis probability forecasts and linear regression forecasts are produced using the same predictors.
The linear regression output consists of deterministic "best estimate"; forecasts whilst the discriminant analysis output consists of probability forecasts for 5 climatologically equiprobable categories. The categories are quints of the observed rainfall distribution over 1961-1990; the quint boundaries are shown in table 1 as percentages of average rainfall. Further details about the forecast methods and predictors can be found in Folland et al, 1991, Journal of Forecasting, Vol 10, 21-56. A combination of forecasts are produced using different training periods (1901-1997 and 1948-1997) to capture variability on different time scales and for the regions 1,2 and 3, using 2 different sets of the predictors. The different forecasts for each region were averaged together to obtain the statistical forecasts shown in
figure 1.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.
Sea Surface Temperature Anomalies
The sea surface temperature (SST) predictor patterns focus on tropical Pacific and Atlantic anomalies, and on interhemispheric differences. SST is currently above the 1961-90 average over most of the extra-tropical Southern Ocean and below average in the extra-tropical North Pacific and North Atlantic. This interhemispheric contrast in SST anomalies favors below average rainfall in regions 1,2 and 3. Above average sea temperatures off the east coast of Southern Africa favor above average rainfall in region 4. A La Niña SST anomaly pattern in the Pacific favors above average rainfall in regions 1, 2 and 3 but this anomaly is expected to weaken or disappear by July-September according to most SST forecasts.
Statistical Forecast Skill
The skill of the forecasting methods is assessed by hindcasting rainfall for past years from 1948 to 1997 using the Jackknife method. Jackknife hindcasts are made by predicting each year in the testing period using a prediction equation based on the remaining years which provides a measure of the skill in predicting independent data.
Correlation is used as the skill measure to assess the regression forecast and LEPS is used to assess the discriminant forecast. LEPS (Linear Error in Probability Space) is similar to correlation in that a score of 1 indicates perfect skill, 0 is chance skill and -1 is perfect negative skill. LEPS is a stricter score than correlation in that LEPS is sensitive to differences between forecast and observed variance, hence LEPS scores are expected to be lower than correlation.
The Forecast
The linear regression and discriminant analysis hindcasts are shown in
figure 1. a-c and f-j respectively. Also shown are correlations between regression hindcasts (figure 1.d) and observed and LEPS skill of discriminant analysis hindcasts (figure 1.e).2. The Dynamical Forecasts
A dynamical ensemble forecast has been produced of July-September rainfall for regions 2,3 and 4 using the 19 level HADAM3 version of the Met Office's atmospheric general circulation model (AGCM). The forecast was produced from an ensemble of 9 AGCM runs each initialized with slightly different atmospheric conditions observed over the period May 3rd-4th and forced with the SST anomalies observed in April which are assumed to persist throughout the forecast period. The AGCM ensemble was run to 4 months ahead (up to early September). September was assumed to have the same standardized rainfall anomalies as forecast for August. This forecast differs slightly from the forecast submitted to PRESAO3 in that the latter was for June-August.
The dynamical forecast is expressed as a probability forecast
figure 2 for the 5 quint categories adjusted to use information about the performance of the model in simulating rainfall over the period 1961-1990. The probabilities are proportional to the frequency distribution of observed quint categories for years when the model simulated the same model rainfall quint category as predicted for 2000. (Model rainfall quints are calculated from model simulations in the same way as observed rainfall quints are calculated from observed rainfall data. Using model rainfall quints instead of observed rainfall quints to categorize the model simulations removes potential problems that could occur when the model fails to predict one of the observed categories due to model bias).3. Forecast Summary
This year the dynamical and empirical forecasts disagree. The empirical forecasts favor average or below average rainfall for regions 2 and 3 while the dynamical forecasts favor average to above average rainfall for these regions. Conversely, for region 4 the empirical forecasts favor above average rainfall whilst the dynamical forecasts show no clear preference.
In calculating the overall best estimate forecasts, two factors were considered: 1) The dynamical forecasts may be overestimating rainfall because for the dynamical forecasts, the SST anomalies observed prior to the forecast are assumed to persist throughout the forecast period. Consequently, for this year
=s forecast, the La Niña anomalies currently being observed in the Pacific are assumed to persist. However there is strong evidence that these La Niña anomalies, which are associated with above average rainfall in regions 1-3, are likely to fade or disappear.2) There is evidence that the statistical forecasts underestimate the impact of SST anomalies in the Pacific.
Taking account of these two factors, the overall best estimate forecasts are compromises between the statistical and dynamical forecasts with slight weighting towards the statistical forecasts due to their higher skill.
Our overall best estimate forecasts are:
Region 1: Dry-Average
Region 2: Average
Region 3: Average
Region 4: Wet
Confidence is LOW due to conflicting forecasts from the dynamical and statistical prediction schemes for more information about our seasonal forecasts contact
awcolman@meto.gov.ukTable 1: Quint Category Boundaries as percentages of mean rainfall for 1961-1990
|
Region |
Very-Dry/Dry |
Dry/Average |
Average/Wet |
Wet/Very Wet |
|
1 |
75 |
97 |
109 |
121 |
|
2 |
81 |
93 |
102 |
117 |
|
3 |
88 |
99 |
104 |
112 |
|
4 |
82 |
94 |
106 |
115 |
Figure 1: Predictions for 2000 and prediction skill for 4 northern African regions. Probabilities, skill and regression and (Standardized units) Forecasts are percentages.
Figure 2: Dynamical Probability Forecasts for July-September 2000 for 3 North African regions expressed as percentages for 5 categories.
