Regionalization of Africa based on Interannual Variability of Precipitation: An Improved Approach and A New Validation Index

Climate regionalization is very important to understand the physical patterns and drivers of spatiotemporal variability in the regional scale as well as ties to global patterns. While many studies of climate variability define regions of interest on the basis of the seasonal precipitation cycle or, for convenience, geographic and political boundaries, regionalization based on interannual variability often yields very different associations that are, arguably, more relevant to studies of climate processes and impacts. Here, we propose a new approach to identify regional patterns of interannual precipitation variability over Africa. Using monthly precipitation data for 1951-1991, a period that was carefully selected based on the number of observational stations and data coverage, we first perform principal component analysis (PCA) on monthly data and then reconstruct the dataset using different numbers of PCs, ranging from 1 to 40. Hierarchical clustering was then applied to the reconstructed dataset to create 2 to 50 regions, defined on the basis of homogeneity in interannual variability of precipitation; the correlation distance was used for dissimilarity measurement. This approach offers an improvement over previous studies, which have used unrotated, orthogonally rotated, and/or obliquely rotated EOFs with a single selected number of principal components (PCs) to filter the data for a PCA-based regionalization, without giving full consideration of the sensitivity of the classification to the choice of PCs. To evaluate the quality of our regionalization and to identify the 'optimal' number of PCs and 'optimal' number of regions for each month, we have developed a spatiotemporal homogeneity and contiguity index (SHCI). The SHCI is computed as the area-weighted average of the difference between inter-region and intra-region average correlation coefficients; inter-region correlations are between the region's mean and all stations in the same region (more than 0.5 and close to 1 for a good regionalization), while intra-region correlations are between the region's mean and all stations in the other regions (close to zero for a good regionalization). This index improves the validation of regionalization results by providing a direct measure for both homogeneity and contiguity: an SHCI value of 0.5 or more is an indication of a good regionalization. The regionalization results were also compared with the techniques used in previous studies. It was found that the generated regions are sensitive to the number of PCs used, and this should be taken into consideration for future PCA-based regionalization studies. It was also found that the spatial patterns of precipitation over Africa are monthly specific, and that regionalization results could be used to define characteristic 'seasons' for each region as different sets of months. The most problematic region for objective regionalization is western equatorial Africa (WEA) due to both limited data availability and the intrinsic complexity of rainfall variability.