Changes in extreme daily precipitation over Africa: Insights from a non-asymptotic statistical approach

Extreme precipitation heavily affects society and economy in Africa because it triggers natural hazards and contributes large amounts of freshwater. Understanding past changes in extreme precipitation could help us improve our projections of extremes, thus reducing the vulnerability of the region to climate change. Here, we combine high-resolution satellite data (1981-2019) with a novel non-asymptotic statistical approach, which explicitly separates intensity and occurrence of the process. We investigate past changes in extreme daily precipitation amounts relevant to engineering and risk management. Significant (?=0.05) positive and negative trends in annual maximum daily precipitation are reported in ~20 % of Africa both at the local scales (0.05°) and mesoscales (1°). Our statistical model is able to explain ~90% of their variance, and performs well (72% explained variance) even when annual maxima are explicitly censored from the parameter estimation. This suggests possible applications in situations in which the observed extremes are not quantitatively trusted. We present results at the continental scale, as well as for six areas characterized by different climatic characteristics and forcing mechanisms underlying the ongoing changes. In general, we can attribute most of the observed trends to changes in the tail heaviness of the intensity distribution (25% of explained variance, 38% at the mesoscale), while changes in the average number of wet days only explain 4% (12%) of the variance. Low-probability extremes always exhibit faster trend rates than annual maxima (~44% faster, in median, for the case of 100-year events), implying that changes in infrastructure design values are likely underestimated by approaches based on trend analyses of annual maxima: flexible change-permitting models are needed. No systematic difference between local and mesoscales is reported, with locally-varying impacts on the areal reduction factors used to transform return levels across scales.