2001 East African March to May Rainfall: Experimental Statistical Forecasts

Based on February Values of Global Regional Predicators

Contributed by Nathalie Philippon and Pierre Camberlin

Centre de Recherche de Climatologie 21000 Dijon - FRANCE

A forecast of the March-May 2001 rainy season in East Africa is presented, using complementary statistical techniques: the Multiple Linear Regression (MLR) and the Linear Discriminant Analysis (LDA) as in Folland et al. (1991).

The predicted (March-May cumulative rainfall amounts) refers to a part of the East African region which has been shown to exhibit spatially consistent rainfall variations on the 1951-1997 period, in the months of March to May (Camberlin and Philippon, 2001). It encompasses Uganda, northern Tanzania, and most of Kenya, except for the coastal area (30-39E, 4S-4N). Station data originate from the global GHCN dataset, with additional records from national meteorological agencies. The period of March to May corresponds to the main rainy season in Equatorial Eastern Africa (so-called "long rains").

Although this rainy season exhibits lesser interannual variability than that of October-December ("short rains"), it is the main agricultural period in East Africa, and droughts severely affect the entire economy (e.g., in 2000 in Kenya). Yet, these interannual variations are much less known than those of the "short rains". In particular, the teleconnections with sea surface temperature anomalies (both regional and global) are much weaker (Ogallo et al., 1988; Rowell et al., 1994 ; Mutai and Ward, 2000), and this precludes the definition of acute seasonal rainfall prediction models based on that sole variable.

The current prediction scheme therefore uses both SST indexes and atmospheric predictors. Given the short-term memory of the atmosphere, compared to the ocean, we could not valuably use predictors beyond the month of February, thus the lead time for the forecast is very small, but it may still be useful since the peak of the rainy season is in April.

The SST indices used in the first pool are computed from the GISST database (UKMO, Bottomley et al, 1990). We focused on the tropical basins, dividing them along the meridional plane in tropical (30N-10N & 10S-30S) and equatorial (10N-10S) belts, the Indian and Pacific basins being additionally separated along the zonal plane in eastern and western parts. The second pool describes atmospheric dynamics, as from NCEP/NCAR reanalysis data (hereafter NCEP; (Kalnay et al., 1996)). The zonal (U) and meridional (V) components of the wind, the geopotential height (Z), the air temperature (T) and the specific humidity (Q) were considered at the following three levels :1000, 500 and 200 hPa. In addition to these five parameters, we computed Moist Static Energy (MSE), which enables to give a synthetic picture of (esp. horizontal) energy gradients. In order to capture regional as well as global signals, we constructed indexes, for each parameter and level, according to three grids : a 5x 10 grid over East Africa, a 10 x 15 grid covering tropical Africa and lastly, a 20 x 40 grid extending from West Africa to East Asia longitudes. All these indexes form the second pool. Although NCEP reanalyses now extend back to 1948, data prior to 1968 are not fully reliable for the African continent, as demonstrated in Camberlin et al. (2000). Reciprocally, rainfall data for validation purposes were insufficient after 1997. Prediction schemes were therefore defined on the 30-yr period 1968-1997.

The predicted value R for the MLR model is given in the regression equation:

where bn are the regression coefficients and Xn the predictors. Using a stepwise procedure maximizing the explained variance, we limited the number of predictors entering the MLR model to four.

Besides, the discriminant model was fitted using the same four MLR selected predictors to assess the robustness of the forecasts. The LDA model is based on three rainfall categories Qi (Dry, Normal and Wet) each of which has a prior probability pi of 0.33. Using Bayes' probability theorem, the LDA computes the posterior probability of observing a category Qi when a predictor vector x (here composed of 4 variables x) is given. This probability is noted :

where is the probability density of x when is known. The denominator being the same for all the i groups, the model has to find maximal.

The two models have been trained using the cross-validation procedure which is particularly useful for small data sets and allows the assessment of the real skill of a forecast model before it is actually applied. We opted for Linear Error in Probability Space skill-scores measures, which are less biased than the RMSE or anomaly correlation (Potts et al., 1996). The LDA and RLM model's ability were respectively tested using LEPSCAT and LEPSCONT, both expressed in percentages for comparison purposes (Ward and Folland, 1991).

The best MLR equation for March-May rainfall in Kenya-Uganda (R) is as follows :

R = 0.3 U1000 + 0.55 Z500 + 0.87 MSE - 0.55 Niño1.2

Among the four predictors, one documents SST (Niño 1.2, with a negative regression coefficient indicative of below-normal rainfall for warm events in the eastern equatorial Pacific). The other three predictors correspond to atmospheric variables : "U1000" depicts the near-surface zonal wind over the Congo basin, "Z500" is the geopotential height at 500 hPa over the Near East, "MSE" corresponds to a near-surface East-West gradient of Moist Static Energy between the East African highlands and the eastern Sahel. Since all predictor series are standardized, the associated coefficients give the part of each predictor in the determination of the rainfall season quality (the hierarchy is the same in the LDA model). MSE introduces the role of the East African land surface conditions, particularly the differential between the highlands and the lowlands. An enhanced gradient accounts for a westerly circulation anomaly between the two regions and an eastward shift of the meridional arm of the ITCZ, which are conducive to more convection over the highlands. Niño1.2 and Z500 have different effects over the ITCZ, the former bringing about a northward shift of the rainbelt over the Indian Ocean whereas the latter stimulates the ITCZ activity by generating a ridge-trough system. U1000, which has a smaller contribution, signs moisture advection from the Congo basin which is conducive to higher rainfall amounts.

Over the period 1968-1997, the correlation coefficient between observed and predicted rainfall from the MLR equation reaches 0.66 in cross-validation mode (LEPSCONT : 44%). The associated scatter-plot (fig.1), with the observations on the x-axis and the hindcasts on the y-axis, in standardized values, shows the rather good linearity of the model. However notice that it fails to forecast the very dry year 1984 (residual value of -3) and, more generally, it tends to over-estimate the rainfall amounts. With the discriminant analysis method, these same four predictors give, in cross-validation way, a classification success rate of 70% (table 1). The LEPSCAT is fairly good with a value of 51%. The single big error again concerns the dry year 1984 categorized as wet. This particular drought year was shown to be associated with an unusual high and late frequency of tropical cyclones over the Southwest Indian Ocean (Macodras et al, 1989). Tropical cyclones divert moisture from the North Indian Ocean, resulting into an anomalous divergent north-south flow over East Africa. A significant negative correlation was also found between tropical cyclones occurrence and rainfall further north in Ethiopia (Shanko and Camberlin 1998).

Acknowledgments. The authors thank the United Kingdom Meteorological Office and the National Center for Atmospheric Research (USA) for providing respectively the global analysed monthly SST database and the NCEP/NCAR outputs. They also thank Isabelle Poccard, CRC, for updating of the reanalysis data.

Table 1 : contingency table of observed (rows) vs predicted (columns) categories of March-May rainfall in Kenya/ Uganda (accurate hindcasts are in bold, large errors are underlined)

	Dry	Normal	Wet
Dry	8	1	1
Normal	2	6	2
Wet	0	3	7

In view of the weak inter-monthly coherency of the March-May season, we considered correlation between predictors and bimonthly rainfall amounts (table 2). Interestingly, at the end of the March-May season the MSE predictor contributes the most whereas at the beginning of the season the oceanic (Niño1.2) and mid-latitude (Z500) predictors are more important.

Table 2. Partial correlation between seasonal and bimonthly rainfall amount in Kenya-Uganda and February predictors (1968-1997). Significant correlations: underlined: 10%, bold: 5%.

	March- May	March- April	April- May
Niño1.2	-0.61	-0.61	-0.42
Z500	0.58	0.59	0.35
U1000	0.31	0.33	-0.28
MSE	-0.66	-0.61	-0.53

For 2001, February standardized values of the four predictors are as follows :

Niño1.2 = -0.03 / Z500 = 0.70 / U1000 = -0.74 / MSE = -0.56

The MLR model therefore yields a standardized anomaly of +0.69, thus suggesting above-normal rainfall for the March-May season in East Africa. The LDA also indicates a fairly high probability of above-normal rainfall (table 3)

Table 3 2001 March-May rainfall forecast for East Africa (Kenya-Uganda area), from LDA model

LDA
	Dry	Normal	Wet
Probability in each Category	0.01	0.29	0.70

References:

Bottomley M, CK Folland, J. Hsiung, D.E Parker, 1990: Global ocean surface temperature atlas. Meteorological Office and Massachusetts Institute of Technology, 313 plates.

Camberlin P, N. Philippon: The east African March-May rainy season, its teleconnections and predictibility over the 1968-1997 period. Journal of Climate, January 2001.

Downing, T.E, K.W Gitu, C.M. Kamau, Eds.,

Lynne Rienner "Coping with drought in Kenya." Boulder, Colorado.

Folland CK, J. Owen, N. Ward, A., Colman, 1991: Prediction of seasonal rainfall in the Sahel region using empirical and dynamical methods. Journal of forecasting, 10, p.21-56.

Hastenrath, S., A. Nicklis, L. Greischar, 1993: Atmospheric-hydrospheric mechanisms of climate anomalies in the Western equatorial Indian ocean. Journal of Geophysical Research, 98, 20,219-20,235.

Kalnay, E., M. Kanamitsu, R. Kistler, W. Collins, D. Deaven, L. Gandin, Saha M. Iredell, S. White, G. Woollen, J.Y. Zhu, M. Chelliah, W. Ebisuzaki W.Higgins, J. Janowiak, K.C. Mo, C. Ropelewski, J. Wang, A. Leetmaa, R. Reynolds, R. Jenne, D. Joseph, 1996 : The NCEP/NCAR 40-year reanalysis project, Bull. of the American Met. Society, 77,p 437-471.

Macodras M.W., P.N. Nthusi, J. Mwikya, 1989: Synoptic features associated with the failure of 1984 long rains in Kenya.

Mutai, C, C. and M. N. Ward , 2000: East African rainfall and tropical circulation / convectionon intraseasonal to interannual timescales. Journal of Climate, 13, 3915-3939.

Ogallo L.J., J.E. Janowiak, Halpert: Teleconnection between seasonal rainfall over East Africa and global sea-surface temperature anomalies. J. Met. Society Vol. 66-6, Ser. II, 807-822.

Potts, J.M., C.K. Folland, Jolliffe and D. Sexton, 1996: Revised "LEPS" scores for assessing climate model simulations and long-range forecasts. Journal of Climate, 9, 34-53

Rowell D.P., J.M. Ininda, M.N.Ward, 1994: The impact of global sea surface temperature patterns on seasonal rainfall in East Africa. Proc. Int. Conf. Monsoon Variability and Prediction, Trieste, Italie, 9-13 Mai 1994. WMO/TD n619. 666-672.

Shanko, D., P. Camberlin, 1998: The effects of the southwest Indian Ocean tropical cyclones on Ethiopian drought. International Journal of Climatology, 18, 1373-1388.

Ward, N.M., C.K. Folland,1991: prediction of seasonal rainfall in north Nordeste Brazil using eigenvectors of sea-surface temperature.International Journal of Climatology, 11, 711-743.

Figure 1 : scatter-plot of observations (x-axis) versus hindcasts (y-axis) of March-May Kenya / Uganda rainfall (period 68-97).