The predictands (June-September cumulated rainfall) refer to a large West African region (17.5N-5N; 17.5W-17.5E) using the 2.5░ latitude., 3.75░ longitude grid box rainfall database developed by Hulme and Jones (1993). They consist on one hand, of a regional Sahel index computed over the domain bounded by the latitudes 17.5N-10N and longitudes 15W-15E, and, on the other hand, in 41 local indexes covering all the West African region mentioned above. The Sahelian region suffering from a significant reduction in total rainfall since the end of the 1960's, the statistical models have only been trained over the years 1968-1998. For convenience we have expressed our forecasts as percentages of the 1961-1990 mean.

In order to be able to propose a >useful= forecast (and not the best one), the predictors document only the near-surface information available on the web by the end of April, 2 months before the beginning of the season to predict. April data have been arranged into 2 pools : the first set includes 27 potential tropical (30N-30S) SST indexes using the GISST6 observed database analysed on a 5x5 degrees grid (Bottomley et al, 1990); the second group is composed of 50 potential regional (15W-15E; 25N-5S) atmospheric indexes (near-surface humidity, moist static energy, and geopotential values) for better documenting the land/sea surface conditions in the West African/eastern Atlantic sector just at the beginning of the northward excursion of the rainbelt into the continent. The atmospheric data has been extracted from the NCEP/NCAR reanalysis (Kalnay et al, 1996) and the selected indexes describe the atmospheric field in terms of local values and of zonal and meridional gradients. This approach was retained as recent works of the authors (Fontaine et al, 1999; Fontaine et Philippon, 2000) underline the existence of significant signals in the NCEP/NCAR air mass energy content in northern spring : the hindcast skills are significantly improved when selected predictors of regional scale describing the monsoon air mass energy are included.

where b_n are the regression coefficients and X_n the predictors. Thanks to a stepwise procedure maximizing the explained variance, we limited the number of predictors entering the MLR model to four. Besides, the discriminant model and the neural networks were fitted using the same four MLR selected predictors to assess the robustness of the forecasts. The LDA model is based on three rainfall categories Q_i (Dry, Normal and Wet) each of which has a prior probability p_i of 0.33. Using Bayes= probability theorem, the LDA computes the posterior probability of observing a category Q_i when a predictor vector x (here composed of 4 variables x) is given. These probabilities are noted:

where fi(x) is the probability density of x when Q_i is known. The denominator being the same for all the i groups, the model has to find p_if_i(x) maximal. As regards the neural method, we use a single hidden layer feed-forward backpropagation network. The activation functions for the two hidden and the single output units are respectively log-sigmod w/ï and linear. The weights initial value was randomly chosen whereas the biases were set to one. The three models have been trained using the jack-knife procedure which is particularly useful for small data sets and allows assess to the real skill of a forecast model before it is actually applied.

Over the period 1968-1998, the correlation coefficient obtained between the observed and predicted rainfall rises 0.84 (table 1) with the MLR model (RMSE of 0.53). Among the four retained predictors, 3 indexes document the continental surface state: AHUM@ designates the specific humidity content at 17.5N, over the longitudes 5E-15E; AMSE1" and AMSE2" refer to Moist Static Energy gradients along the 10W meridian, between the Guinean shore (1) or the Sahel band (2) and the Sahara margins. They reflect the preseason meridional arrangement of energy over West Africa which is of prime importance for the monsoon excursion onto the continent. The last selected index, AATL@, documents a surface temperature gradient over the tropical Atlantic southeastern basin. This conforms to the fact that April-June SST anomalies in the south tropical Atlantic are associated with July-September west African rainfall (see, among others, Zheng et al., 1999).

The forty-one local (2.5 lat., 3.75 long.) multiple regression forecasts are presented in figure 1, using 4 April predictors selected among the second set (table 3). On the whole, the JAS 2000 season is predicted as normal (which is in agreement with the global Sahel index forecast); we see however the northeastern parts would know drier conditions.

Figure 1: 2000 JAS rainfall forecasts for each 2.5*3.75 West African box using the four best predictors in Set 2 (forecasts are expressed in % of 1961-1990 mean). VW for Very Wet, W for Wet, N for Normal, D for Dry and VD for Very Dry, we have adopted the same scale as the Agrhymet Center. The crosshatched areas indicate missing rainfall information; non underlined forecasts must be considered carefully because of models explaining less than 56% of the rainfall variance or of forecasts not consistent with the surrounding regions.