Abstract
Rainfall in the Sahel region of Africa is prone to large interannual variability, and it has exhibited a recent multidecadal drying trend. The well-documented social impacts of this variability have motivated numerous efforts at seasonal precipitation prediction, many of which employ statistical techniques that forecast Sahelian precipitation as a function of large-scale indices of surface air temperature (SAT) anomalies, sea surface temperature (SST), surface pressure, and other variables. These statistical models have demonstrated some skill, but nearly all have adopted conventional statistical modeling techniques—most commonly generalized linear models—to associate predictor fields with precipitation anomalies. Here, the results of an artificial neural network (ANN) machine-learning algorithm applied to predict summertime (July–September) Sahel rainfall anomalies using indices of springtime (April–June) SST and SAT anomalies for the period 1900–2011 are presented. Principal component analysis was used to remove multicollinearity between predictor variables. Predictive accuracy was assessed using repeated k-fold random holdout and leave-one-out cross-validation methods. It was found that the ANN achieved predictive accuracy superior to that of eight alternative statistical methods tested in this study, and it was also superior to that of previously published predictive models of summertime Sahel precipitation. Analysis of partial dependence plots indicates that ANN skill is derived primarily from the ability to capture nonlinear influences that multiple major modes of large-scale variability have on Sahelian precipitation. These results point to the value of ANN techniques for seasonal precipitation prediction in the Sahel.
Denotes Open Access content.
