Humanitarians have been investing significant attention and resources in the uptake and use of climate services to inform their work in disaster risk management. For example, disaster managers regularly participate in Regional Climate Outlook forums and climate service partnerships (Hewitt et al., 2012; ICPAC, 2016; Mwangi et al., 2014). While many early warning systems focus on short-term hydrological flood warnings, these climate service initiatives promote the use of forecasts of seasonal total rainfall. The use of such forecasts has yielded mixed results when used to prepare for heightened flood risk in Africa, such as prepositioning flood relief items (Braman et al., 2013) and evacuating vulnerable people (Anon, 2016). In this article we question whether seasonal rainfall forecasts have been overpromoted for their usefulness in flood preparation.

To clarify whether seasonal total rainfall forecasts indeed indicate increased risk of flooding, we identify the dominant indicators of seasonal flooding in different locations of sub-Saharan Africa. In many locations, it is likely that total rainfall is not the dominant driver, and other seasonal descriptors would give a better indication of the risk of flood hazards. Cumulative rainfall is not the dominant flood-generating process for floods in most river basins in the United States (Berghuijs et al., 2016), and monthly total rainfall has not been shown to be a good indicator of regional river “floodiness”, or the percentage of regional rivers with extreme flooding (Stephens et al., 2015). We provide further discussion of “floodiness” in Sect. 2.2.

In the context of sub-Saharan Africa, we quantify the relationship between seasonal total rainfall and floodiness, and explore whether there might be alternative variables with a stronger relationship with floodiness at the seasonal level. In each river basin, the catchment size and the climate regime will affect the influence of hydraulic routing, soil dynamics, and precipitation patterns; we therefore identify which hydrometeorological variables are most related to seasonal flood risk in each location. We investigate the association between seasonal percentage floodiness and seasonal total rainfall, as well as the relationship with 14 other variables and their combinations.

Given the scarcity of hydrological data available for many parts of Africa, we offer an alternative methodology to that used by Berghuijs et al. (2016) for assessing the indicators of flood intensity and frequency in a region. Rainfall estimates from ERA-Interim Land (Balsamo et al., 2015) are used to force the Global Flood Awareness System, a global hydrological model (Alfieri et al., 2013). We calculate anomaly correlations between rainfall input and the predicted flooding, which is defined as the proportion of river cells that has extreme discharge in a region in a given time period (Stephens et al., 2015). We repeat this analysis with the 14 alternative variables, and develop a generalized linear model (glm) to identify which combinations of variables provided the greatest indication of flood hazard in each region.

Our methodology depends on the reanalysis for a climatology of rainfall and focuses on the hydrological model to estimate the consequences of this rainfall for river flows. This approach is not limited by a patchy observational network, and results can be compared across regions to inform regional policies. While the rainfall has been bias-corrected with observations, we would encourage the replication of this methodology using local rainfall observations for more detailed study of the local indicators of floodiness.

Three-month seasonal total rainfall anomalies show significant correlation with floodiness in several regions (Fig. 1). The relationship is weakest in western and central Africa, and also weakens as flood severity increases.

When the rainfall and floodiness are aggregated by FPU and then correlated, the correlations improve in almost all locations, suggesting that seasonal total rainfall forecasts for FPUs (Fig. 1c and d) might be of greater use than grid-box forecasts (Fig. 1a and b) as a predictor of flood hazard. Different regional forecast aggregations could also be explored to determine whether this can be further optimized.

While the correlations are significant in many regions, there is considerable variation in floodiness that remains unexplained by this variable. To demonstrate this, we calculate the probability of flooding (floodiness greater than 0) conditional on seasonal rainfall being in the top tercile of the distribution, which is the focus of many seasonal forecasts. Ultimately, even if a top-tercile rainfall forecast were given with 100 % certainty, it would represent only a small increase in the probability of flooding relative to climatology (Fig. 1e).

In Figs. 2–4 we display results from three different sets of possible predictor variables. In Fig. 2 we plot the anomaly rank correlations with floodiness for five different measures of extreme precipitation events within a season. None of these rainfall variables are a better predictor of floodiness in all locations (Fig. 2, second row); however, the number of rain events above the 99th percentile (1-, 3-, and 5-day events) tend to outperform seasonal total rainfall in the areas of western and central Africa (where seasonal total rainfall had the weakest correlations; see Fig. 1).

Next, we analyzed five different measures of rainfall patterns within a season, including the length of dry spells and wet spells. Apart from in isolated locations, these measures do not have coherently stronger correlations with floodiness than seasonal total rainfall (Fig. 3).

The last set of variables we explored included soil moisture and several measures of seasonal rainfall intensity. Figure 4a shows that in most regions seasonal total rainfall is more strongly correlated with floodiness than soil moisture. In comparison, seasonal rainfall intensity shows a slightly higher correlation with floodiness across the continent (Fig. 4b), defined as the total precipitation divided by the number of rainy days. Similarly, the percent of seasonal rainfall occurring in the top 95th percentile days, here called the “contribution of extremes”, shows higher correlations in the western and central Africa region (Fig. 4c). Both of these variables show less variation across Köppen climate regions, compared to seasonal total rainfall (Fig. 1). Burstiness (Schleiss and Smith, 2016) of a 15-day interamount time (Fig. 4d) does not show better correlations with floodiness than does seasonal total rainfall.

It is possible that a combination of these variables would outperform any of them in isolation, so we also test the combination of three different types of variables that each have strong correlations with floodiness: (1) 3 days above 99th, (2) soil moisture, and (3) contribution of extremes. To test whether a combination of these variables is better able to predict 50-year return period floodiness, we fit a logistic regression model for each grid point using these three variables. Because these variables are correlated with each other in several regions, we select the generalized linear model (glm) fit with the fewest variables that is still within 1 standard error of the optimal fitted model.

Results of the glm generally confirm the spatial patterns reflected in the correlation figures above, and indicate that a combination of these variables could be a useful indicator of floodiness in many regions. Figure 5 shows that the number of 3-day events above the 99th percentile was a meaningful contributor when added as a predictor independently, or in conjunction with another variable, in most of sub-Saharan Africa. Soil moisture is included as an additional predictor primarily in southern Africa, while the contribution of extremes was included primarily in central Africa. A combination of all three variables was recommended in eastern Africa and parts of southern Africa, while none of the predictors was selected as a meaningful contributor for much of western and central Africa.

In the analysis above, we have demonstrated that indicators of floodiness differ widely across the African continent, using a methodology that can be replicated for other data-scarce regions to assess the key indicators of flooding. Improvements to both the climatology of reanalysis rainfall and the skill of global hydrological models could further improve the understanding of predictability of these processes, and we encourage replication of this methodology using observations to further describe and validate the flood-generating processes in specific locations.

It is clear that seasonal total rainfall is not a reasonable proxy for floodiness in most of western Africa, central Africa, and Madagascar. Large portions of these regions fall into the “equatorial” Köppen classification, which includes tropical savannahs. Floodiness in these regions demonstrated a stronger relationship with measures of the intensity of rainfall during a season than in the rest of the continent. In these regions, the climate services community should reconsider their association of seasonal total rainfall with flood risk and flood preparation measures (Braman et al., 2013). When using forecasts in an operational context, imperfect forecast skill of the rainfall proxy itself further reduces the usefulness of this information for flood preparedness.

On the other hand, much of eastern Africa, southern Africa, and the Sahel tends to show similar patterns in the dominant indicators of flooding. Seasonal total rainfall had some of the highest correlations in these regions, as well as the number of extreme events within a season. There are large “arid” areas in each of these regions, and these findings are consistent with studies done in other arid areas. Berghuijs et al. (2016) found that daily and multi-day rainfall events were the dominant flood-generating processes for river basins in arid regions of the United States, similar to the results in Fig. 2d.

To maximize usefulness in these regions, forecasters could consider simple formatting alternatives to current forecasts that would provide a better indication of floodiness, such as replacing tercile forecasts with forecasts of the top percentiles of the distribution (Grieser, 2014), and offering aggregate forecasts for river basins or FPUs. The latter could also lend itself to greater forecast skill than for rainfall itself, and encourage regional-scale disaster preparedness.

Researchers developing new forecast products should consider several of the predictor variables discussed here. Forecasts of the frequency of extreme rainfall events would likely provide a better indication of floodiness, compared to seasonal total rainfall forecasts, for much of Sub-Saharan Africa. Studies have shown the potential predictability of this variable in several locations (Anderson et al., 2015; Higgins et al., 2000; Verbist et al., 2010). Seasonal forecasts of soil moisture could give a useful indication of flood risk in dry regions of Africa (Fig. 4), and these forecasts are also likely to have seasonal predictability in areas where they can be well initialized, notably due to the persistence of soil moisture (Kanamitsu et al., 2002; Koster et al., 2010; Poveda et al., 2001). This also takes evaporation into account.

Forecasts of rainfall intensity could give a better indication of flood risk in western and central Africa (Fig. 5). However, intensity is the least spatially coherent and therefore least likely to be predictable (Moron et al., 2007). Further research into the area is merited, as there are a few examples showing some potential predictability of rainfall intensity (Pineda and Willems, 2016).

Seasonal skill in forecasting total 3-month rainfall anomalies is varied around the world; the highest skill has been achieved during ENSO events in areas that have ENSO teleconnections (Barnston et al., 2010a; Weisheimer and Palmer, 2014). Given the low correlations we have found here between floodiness and either seasonal total rainfall or other rainfall indicators, forecasts of any of these proxies are unlikely to provide strong signals of increased risk. However, there have been several studies using large-scale climate patterns and sea surface temperatures (SSTs) as predictors of flood risk, most focusing on the role of ENSO in changing global flood risk (Emerton et al., 2017; Ward et al., 2014, 2016). Further research on using SSTs and other climate patterns to directly forecast changes to flooding is merited, to explore whether such forecasts would give stronger indications of change in flood hazard than seasonal climate models of rainfall.

Ultimately, the most informative forecasts of flood hazard at the seasonal scale could be seasonal streamflow forecasts using hydrological models calibrated for individual river basins (Sahu et al., 2016). While this is more computationally and resource intensive, investments in better forecasts of seasonal flood risk could be of immense use to the disaster preparedness community.

In their work, disaster managers can support these forecasting efforts by better defining the meteorological and hydrological variables that relate to disaster. Sharing this information with forecasters can inform the development of forecast products that provide specific information about these “danger levels”, thus better enabling stakeholders to take appropriate preparatory actions. Forecast-based finance initiatives are underway globally, with the aim of taking action and releasing financing proportional to the risk information in a forecast, before the potential disaster (Coughlan de Perez et al., 2016). Changes to forecast products to provide clearer and more targeted risk information can support this process, and enable humanitarians to better anticipate and prepare for disasters before they strike.