An assessment of the value of

INTRODUCTION

The Met Office use SST based statistical methods and dynamical models of the global ocean-atmosphere system to make seasonal predictions of tropical rainfall. Forecasts have been made for tropical East Africa October-December rainfall (the 'short rains') since 1994 and appear in previous September issues of this bulletin.

A set of statistical forecasts and a set of dynamical forecasts are presented here for the whole East Africa Region (5N-15S, 30E to Indian Ocean Coast referred to henceforth as the “whole region” ) and for 2.5^olatitude x 3.75^olongitude boxes shown in figures 1 and 2..

FORECAST METHODS

The statistical forecasts are made by using linear regression and discriminant analysis techniques, with three indices of global sea surface temperature (SST) anomaly patterns and 2 indices of mean SST anomaly for 2 box regions (all predictors are defined in the Appendix) known to be linked to East African rainfall. The forecast model is derived from historical rainfall and SST information.

The dynamical forecasts are presented in figure 2 and produced using GloSea, the Met Office GLObal SEAsonal forecast model which is a version of the HadCM3 Coupled Ocean-Atmosphere model modified for seasonal forecasting. GloSea is run as a 40-member ensemble of predictions initialised with atmosphere and oceanic conditions observed in early September. GloSea hindcasts produced as part of the DEMETER project (see www.ecmwf.int/research/demeter for more about DEMETER) were used to correct the GloSea forecast output for model bias and convert the forecast into probability format.

For “whole region” forecasts, the historical rainfall record is divided into 5 equi-probable categories. The category boundaries are based on the 1961-1990 climatology and are given below as percentages of mean rainfall:

Grid box forecasts are expressed as probabilities of terciles which are climatologically equi-probable over 1961-1990. This is in order to make the forecasts compatible with GHACOF (Greater Horn of Africa Climate Outlook Forum) consensus forecasts which are expressed in the same format.

Grid box forecasts are also presented as probabilities of a wetter season than last year (OND 2004) and as probabilities of an extreme season where “extreme” is defined as being wetter or drier than any of the past 10 years (1995-2004).

FORECAST SKILL (FOR THE WHOLE REGION)

PERFORMANCE OF TRIAL FORECASTS FOR 50 PAST YEARS

The statistical and dynamical forecasts were tested using trial forecasts over the period 1948 to 1997 and 1959 to 2000 respectively. The assessment measure used is correlation. Statistical linear regression forecasts were assessed using a method where a trial forecast is made for each year using a regression equation calculated using data for the remaining years. This assessment provides a good measure of forecast skill from minimal data.

Statistical forecast skill correlation=0.53

GloSea model forecast skill correlation=0.38

These correlations are statistically significant at the 5% level.

Note: The GloSea assessments were of trial forecasts initialised in early August whilst the GloSea forecast presented here was initialised in early September. Hence the skill of the dynamical forecast presented here is probably slightly higher than these GloSea assessments indicate as more recent precursor information is used.

PERFORMANCE OF REAL TIME STATISTICAL FORECASTS (USING INFORMATION UP TO LATE SEPTEMBER)

Forecasts have been produced for this region since 1994. The forecast categories for the whole region are compared with the corresponding observed category in the table below. The forecasts for 1994 and 1995 were strongly influenced by above and below average SST in the NW Pacific respectively and the forecasts for 1997-2000 where influenced by the 1997 El Nino and the 1998-1999 La Nina events. The observed rainfalls for 2000 and 2001 were quite close to the dry/average and very dry/dry boundaries respectively so the forecasts for these seasons were more accurate than apparent from the table.

Year

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

Forecast Category

Very

Wet

Dry

Average

Wet

Dry or

Average

Dry

Average

Dry or

Very Dry

Dry

Observed

Category

Average

Dry

Very

Dry

Very

Wet

Very

Dry

Very Dry

Wet

Very Dry

Wet

GRID BOX FORECASTS

STATISTICAL TERCILE PROBABILITY FORECASTS (FIGURE 1a)

Figure 1a consists of 6 statistical forecast probability maps. The upper row of 3 maps show probabilities for the 3 terciles for all grid boxes for which there are data. The second row of 3 maps (labelled with “skill mask”) is a repeat of the first row but only probabilities for grid boxes where there is significant correlation skill according to an independent test are shown. To be included in the skill mask map, linear regression hindcasts for the box must pass at least 1 of these 2 tests:

a. Correlation skill over 1949-1998 is significant at the 5% level

b. Correlation skill calculated over the subset of forecasts predicting the same tercile as predicted for this year is significant at the 5% level

The dry tercile is generally the most probable across East Africa. Notable exceptions are some boxes around the north of Lake Victoria where the middle tercile is favoured and an area of southern Sudan around 10^o north where the wet tercile is favoured.

STATISTICAL FORECASTS OF EXTREMES AND CHANGE FROM LAST YEAR (Figure 1b)

Figure 1b is similar to figure 1a but probabilities of a wetter season than last year and of 10-year extremes are presented. The probability of 2005 being wetter than 2004 is generally very low except in some more western regions including southern Sudan. Probabilities for an extreme dry or an extreme wet season are mostly close to or lower than expected from climatology (<10%). The main exception is for Indian Ocean coastal regions where there are elevated probabilities for an extreme dry season.

GLOSEA TERCILE FORECAST PROBABILITIES (Fig 2a)

The GloSea forecast (fig. 2a) probabilities favour the dry or average category for most parts and hence show a less dry signal than the statistical forecasts. The average category is favoured for some boxes around Lake Victoria, just to the south of the boxes where the statistical forecasts favour the average category. In common with the statistical forecasts, the Wet category is favoured for an area of southern Sudan around 10^oN.

GLOSEA FORECASTS OF EXTREMES AND CHANGE FROM LAST YEAR (Figure 2b)

A drier year than 2004 is favoured for most boxes. Probabilities for extreme seasons are generally close to or lower than expected from climatology (<10%). The most notable exception is coastal parts of Kenya where probabilities for an extreme dry season are slightly elevated. There are slightly elevated probabilities of an extreme wet season in southern Tanzania but is a consequence of a lack of wet years in the last 10 years in this region, The locations of these elevated extreme probabilities correspond with the statistical forecasts (fig.1b) but the signal for an extreme dry season in Indian Ocean coastal regions is stronger in the statistical forecasts .

OVERALL BEST ESTIMATE:

(a) Whole Region:

This year, the dynamical and the statistical forecasts are both indicating that below-average precipitation is likely with the signal for dry being stronger from the statistical forecast. The DRY quint category has the highest probability according to both the statistical and dynamical forecasts and is located near the middle of the probability distribution in each case, hence the DRY quint category is our “best estimate” for the region as a whole.

(b) Grid boxes:

Consistent with the large area forecast, the dry tercile is generally most probable, particularly in the statistical forecasts. The highest probabilities for the DRY tercile are over southern Kenya and northern Tanzania. There is an elevated risk of an extreme dry season near the east coasts of Kenya and Tanzania.

REFERENCE:

Mutai, C.C., Ward, M.N and Colman, A.W. Prediction of East Africa seasonal "short rainfall" rooted in evidence for widespread SST-forced variability during October-December. I.J.Climatol.18 975-997 (1998).

APPENDIX:

The figures show the rotated global SST EOF patterns used as predictors for the statistical forecast. For all 3 patterns, positive SST anomalies in regions with positive weights and negative SST anomalies in regions with negative weights favour above average rainfall and vice-versa.

The 2 box regions are

1) “NE Indian ocean” region 0-10N 50-60E + 0-10S,40-60E

2) “Pacific” region 0-20N, 150E-150W

The discriminant and regression forecasts are the average of 3 predictions from the following 3 sets of predictors:

1) Rotated EOF2, EOF 4 and EOF 5

2) Rotated EOF2, EOF 4 and EOF 5 + NE Indian Ocean region

3) Rotated EOF2, EOF 4 and EOF 5 + NE Indian Ocean region + Pacific region

The predictors are weighted as follows

Rotated EOF2, EOF4 and EOF 5 all 25% , NE Indian Ocean region 16.7%, Pacific region 8.3%

The pattern shown in figure A3 (rotated EOF 5) is the most important predictor contributing to over 50% of the forecast variance.

FIGURE 1a: PROBABILITY FORECASTS BY STATISTICAL METHOD FOR TERCILES

FIGURE 1b: PROBABILITY FORECASTS BY STATISTICAL METHOD FOR A WETTER SEASON THAN LAST YEAR AND FOR 10 YEAR EXTREME SEASONS

FIGURE 2a: PROBABILITY FORECASTS FROM GLOSEA FORECAST FOR TERCILES

FIGURE 2b: PROBABILITY FORECASTS BY GLOSEA FOR A WETTER SEASON THAN LAST YEAR AND FOR 10 YEAR EXTREME SEASONS

Figure A1:

Figure A2:

Figure A3:

Very Dry/Dry	Dry/Average	Average/Wet	Wet /Very Wet
74%	86%	102%	124%