Abstract
Expanding on earlier research, empirically based methods are developed for predicting March-September rainfall in northern Northeast Brazil. From a network of 27 raingage stations in Brazil's Nordeste, new rainfall index series are constructed for March-September (MS) and October-January (OJ). Data input to stepwise multiple regression models further include January values of the Tahiti minus Darwin pressure index, an index of sea surface temperature (SST) in the equatorial Pacific (PWT), and indices of the fields of meridional (v) and zonal (u) wind components and of SST in the tropical Atlantic between 30°N and 30°S (AFV, ARJ, and AFT, respectively). Empirical orthogonal function (EOF) analysis of the v, u, and SST fields was used in constructing the latter three indices. Throughout the study, a sharp distinction is kept between a “dependent” dataset (1921–42 and 1948–57) used as a training period and an “independent” portion of the record (1958–87) reserved for prediction experiments.
Preseason rainfall (OJ) is by far the most powerful predictor, allowing 52% of the interannual MS variance in an independent dataset to be forecast. From experiments with unrestricted data input, a model with the predictors OJ, AFV, and PWT shows the best performance, capturing 71% of the MS variance in the independent dataset. Experiments on the optimal length of the training period suggest that one of 20–30 years is adequate. Updating offers no advantage over a fixed training period. For the purposes of operational application, the maintenance and timely processing of raingage measurements is of paramount importance. The next important task, demanding considerably greater resources, would be the real-lime monitoring of the surface wind field over the tropical Atlantic.
