Floods are among the most common and destructive natural hazards globally . Approximately 90 % of disasters worldwide in the last decades were caused by weather-related events. Among them, floods are the most frequent and affected 2.3 billion people between 1995 and 2015 . It is generally acknowledged that, due to projected climate and socio-economic changes, extreme events such as floods may further increase in frequency, magnitude and intensity . In order to minimize the negative effects of floods, disaster risk reduction is increasingly important . The urgency of mitigating flood risks is also recognized by international agreements, such as the Sendai Framework for Disaster Risk Reduction , which underlines the understanding of disaster risk including the hazard characteristics as a first priority. Developing adequate knowledge of past flood events is essential in order to sufficiently address this global problem and to further reduce the consequences of future disastrous events.

Accurate data are key to developing a reliable representation of floods. While hydro-meteorological data are collected and made available in many places, most developing countries still struggle with limited availability due to inconsistent methodologies and datasets . This may for example be because of the lack of rain and discharge gauges due to insufficient resources as a consequence of socio-economic issues . One of the regions where data availability is poor is (Southern) Africa . Not only is there a general lack of data, but the available data and resources are also not evenly distributed across the riparian countries, with most gauges in South Africa . While the country of South Africa is relatively rich in terms of data, technology and knowledge, many of its neighbouring countries are not . This lack of spatially consistent datasets is a particular issue in this region, as many of the larger river basins are transboundary, and extreme events are often linked to phenomena on a wider, regional scale, such as cyclones .

To address the issue of floods in data-poor regions, increasingly available global datasets, such as global reanalysis data, may have significant potential. Reanalysis datasets are the result of a combination of earth observations, as well as various models and datasets containing in situ measurements . Currently there are several reanalysis datasets available at a global scale and applicable to water resources, such as ERA-Interim/Land , GLDAS , Global Water Cycle Reanalysis , GSWP-2 , WATCH and WRR . These datasets provide consistent hydro-meteorological data with a global coverage, spanning several decades. Hence, they have significant potential to fill data gaps in regions such as Southern Africa . Datasets containing different global model outputs have thus far been used to determine climatic extremes as well as their uncertainties at the global or continental scale. For instance, evaluated the influence of different river-routing schemes in the various global hydrological models (GHMs) on peak discharge simulation. compared the 30-year return period level of river discharge calculated using nine different global models regarding their projections of climate change impacts on flood hazards worldwide. assessed the ability of six global models regarding their skill to produce hazard maps for the African continent. However, they note that there has thus far been limited validation of these global flood models against observed floods.

This study assesses the potential of a recently developed state-of-the-art global Water Resources Reanalysis (WRR; ) dataset in identifying damaging flood events for data-poor regions such as the Limpopo River basin. The Limpopo River basin is a transboundary Southern African basin typical of the aforementioned data issues, including a general lack of data as well as an asymmetrical distribution of data availability across the riparian countries. The dataset used in this study is the open-source global Water Resources Reanalysis dataset developed in the eartH2Observe (E2O) research project, a collaborative project funded under the European Union's Seventh Framework Programme (EU-FP7) . The WRR dataset is assessed against two benchmarks. The first benchmark is developed using observed discharges from reliable gauges available in the region. As the upstream catchment area of these gauges varies, this provides insight into (i) the skill of the global dataset in identifying the occurrence and magnitude of flood events in the basin and (ii) how this skill is related to catchment scale. The second benchmark considers reported damaging flood events in the basin. Reported events from three disaster databases, including the Emergency Events Database (EM-DAT), the Global Active Archive of Large Flood Events (GAALFE) and the Natural Catastrophe Service (NatCatSERVICE), were collated to develop a chronology of damaging events. The ability of the global model datasets in identifying such damaging events provides insight into the potential of the global models to be able to predict the occurrence of impacting flood events. There is a critical need for both higher-resolution reanalysis supporting data and flood forecasting systems to properly capture timing, intensity and location of flood impacts. Global models such as those considered in this study are employed by several global and continental flood forecasting systems, such as the Global Flood Awareness System (GloFAS) and the African Flood Forecasting System , and this assessment sheds light on the potential these have in providing information that is useful to managing floods at the regional and sub-basin scale. We also consider of scientific interest that, for both the coarser- and the finer-resolution models, the threshold up to which the models are still able to capture the hydrology is on the order of the cell size. This holds promise for the continuing effort of modelling research groups in developing increased resolution (global models).

The remainder of this paper is structured as follows. Section 2 provides the materials and methods used, a description of the study area, data and verification methods. The results (Sect. 3) reveal the skill of the models in capturing the reported as well as modelled flood events. Section 4 provides a discussion of those results, as well as limitations and suggestions for further research. Conclusions are provided in Sect. 5.

The potential of the global Water Resources Reanalysis dataset was assessed by studying the hydrological performance, identification of hydrological extremes, as well as of damaging flood events, and was evaluated by means of commonly used error statistics and verification skill scores (CSI, POD and FAR). The verification of the models within the WRR dataset was largely dependent on the observed river discharge data. Access to these data proved to be quite challenging, and the quality of the discharge data that were obtained was often insufficient. Only 75 of the 196 river gauging stations for which at least some data were available in the Limpopo for the desired time range were used in this research, with most of these in South Africa. This has implications for the conclusions drawn from the research, especially for the PBIAS as it is highly influenced by the uncertainty in the observed data . Despite these limitations, this research shows that the discharges that were estimated by the different global models are to some extent able to capture the variability of observed discharges, as indicated by the different error statistics. For instance, the NSE demonstrated that for WRR1 as well as WRR2, both the HTESSEL-CaMa and the WaterGAP3 models performed well with roughly similar NSE values, despite the different structure of these models. HTESSEL-CaMa is a LSM and does not include lakes and reservoirs or water usage, whereas WaterGAP3 is a GHM and does include both lakes and reservoirs, as well as water usage (Table ). The differences between the model structures are illustrated by the PBIAS and r values. HTESSEL-CaMa has reasonable PBIAS, while WaterGAP3 has a relatively good r. As noted in Sect. , the basin is highly altered due to human influences, in particular by a large number of storage reservoirs. Models that capture only natural flow conditions, and do not take the reservoirs and water usage into account, may be able to reasonably estimate run-off volumes, though they do tend to largely overestimate the actual magnitude of the discharges. Not including human influences such as regulation, however, results in low correlations. The relative intensity of flood events, on the other hand, can still be well captured by the same model when using the model climatology instead of the observed climatology as a reference. An example of such a model is W3RA, which performs poorly when considering the error statistics, but relatively well for the CSI.

Global models are best suited for the modelling of large-scale processes, but poorly represent the small-scale ones such as the variability associated with convection . These conclusions have been drawn in similar research, such as , and . This study indicates that the small-scale flood events were generally not well captured by the global models that were analysed in this research. The results do show, however, that the performance of these global models improves with model developments in terms of resolution, forcing and model parameterization. The statistics for model performance measures for the higher-resolution WRR2 starts to approach reasonable values for gauges with upstream areas of some 500 km2, while for the lower-resolution WRR1 these same values are attained only at areas of some 2500 km2. The higher-resolution WRR2 also shows for two of the three models better skill in identifying reported flood events, represented in the chronology of reported flood events developed. Whether the improved performance of the higher resolution is due to the improved and higher-resolution MSWEP forcing data or due to improved representation of hydrological processes is unclear. However, as the improvements vary between the models, it is clear that model structure has an influence.

That there is skill in these global models in identifying flood events that have impacts and that this skill improves as the resolution of these large-scale models improves is significant. Global-scale forecasting systems as well as those at continental scale typically employ such large-scale models for developing forecasts, using thresholds based on model climatology to inform the severity of predicted events and subsequent issuing of flood warnings. Such warnings may be issued where there are no (reliable) river gauges, as is the case in much of the Limpopo basin, making calibration of a local model difficult. The ability of these global and or continental models to predict the occurrence of flood events that have impacts bolsters the confidence of using these warnings to initiate a response, though the high false alarm rate found could again diminish confidence.

It is important to note, though, that likely not all small-scale flood events that occurred between 1980 and 2012 will have been included in the chronology of reported flood events that was developed. As has more often been found to be the case in the Global South , the availability of disaster data in the Limpopo River basin is fairly limited. In order to construct a basin-wide timeline of historic damaging floods, events reported in the EM-DAT, GAALFE and NatCatSERVICE databases were collated. Even though the three used here are currently the most comprehensive databases containing reported damaging historic flood events in Africa , several shortcomings are noted. These include inconsistencies between events reported, gaps, and limited reporting in some areas . Additionally, most disaster databases are available at the country scale, whereas flood events occur at the basin or finer scales. It is recommended to enhance the reporting of flood disasters by providing more details on the losses that were incurred as well as a more precise description of the location and extent of the floods. The basin-wide approach to identify past flood events by using empirical disaster databases used in this research has also been applied in other research (; ; ; ; ), noting similar deficiencies.

The flood classification that was used in this research is a discrete classification, taking the number of fatalities and overall losses into account. However, it is expected that a continuous flood severity classification would be better able to reveal the relationship between extreme river discharges and the intensity of reported damaging flood events. However, due to the gaps in the reported damaging flood event data as well as broad area descriptions, this could not be assessed at this point. In order to identify the added value of such a classification, additional research is required in addition to improving disaster loss data.

In this study, the Gumbel distribution is used to determine the annual exceedance probability thresholds of both the modelled and observed discharge data. Different extreme value distributions, however, can significantly influence the probability of the extreme discharges . The Gumbel distribution is a two-parameter distribution and was applied due to its simplicity and robustness, though some authors (e.g. ) argue that a three-parameter distribution such as the generalized extreme value (GEV) or the log-Pearson type III should be used for flood frequency analysis. However, the goodness of fit of the distributions found was tested using the Kolmogorov–Smirnov test, with stations that did not meet the 5 % significance threshold not considered. Further inspection of these stations revealed that these were often directly downstream of a dam, or otherwise strongly influenced by human activities. Additional research could additionally explore the influence of using more complex extreme value distributions. This could also consider the influence of the length of the moving window that was used to identify the maximum discharge in the observed and modelled time series. This moving window was chosen to allow for the travel time from the upstream parts of the sub-catchment. In reality, however, the catchment upstream of each gauging station has its own time of concentration, and the window used could be made specifically for each station accordingly.

Of the global models considered in this study, the higher-resolution WaterGAP3 in WRR2 demonstrated the best performance, both for capturing the hydrological behaviour across the Limpopo basin, as indicated by good values for the error statistics, and for identifying the occurrence and severity of hydrological extremes, which was indicated by the skill scores. It was also observed that WaterGAP3 in WRR2 is reasonably good at estimating low annual exceedance probabilities for the damaging flood events for the stations with a large upstream catchment area. One reason for this improved performance may be the inclusion of lakes and reservoirs, as well as water abstractions in the model. However, results for other models, such as W3RA, which has the worst model performance error statistics, may rank higher than other models when used to identify the occurrence of flood events, where these are identified using the model's own climatology. It is also important to note that, if similar research were to be applied elsewhere, the ranking of model performance may be quite different. The ranking of the models also clearly depends on the aim of the research. WaterGAP3 for instance performed poorly with respect to other global hydrological models in research focussing on a snowmelt-driven catchment . Furthermore, when the key interest is the relative performance of the model for the Limpopo River basin, taking only the model climatology into account, the W3RA model would be the preferred model, as it has a high CSI. However, when the main goal is the absolute magnitude of discharges, the W3RA model would not be considered, as it is found to severely overestimate the discharges in the Limpopo River basin. The seven-model ensemble mean, on the other hand, proved to be quite consistent in its performance. For the CSI values particularly it scores remarkably high, but it also scores relatively well for Pearson's correlation coefficient r. Although which model would be the best applicable for each instance should carefully be assessed, the model ensemble mean would be the safest bet in an area where no model clearly stands out.

The study explores the use of a global reanalysis dataset developed within the EU-FP7 EarH2Obverve (E2O) project, which is constructed using a set of global hydrological and land surface models, to support flood risk analysis in data-sparse regions, such as the Limpopo River basin. There is a necessity for such reanalysis data, since measured river discharge data in this basin and others like it are currently insufficient, poorly spatially distributed, have an insufficient period of record or are partly inaccessible. The E2O reanalysis dataset provides hydro-meteorological data of sufficient length and coverage required for statistical analysis. When the variability of the discharge results of the ensemble of models included in the reanalysis dataset is evaluated, the error statistics found show that the models all have reasonable skill in capturing the variability of the observed discharges, though there may be significant bias in magnitude. This was indicated by strong correlations, low Nash–Sutcliffe efficiency and high percent bias values. Furthermore, the error statistics revealed that the variability is better captured by the models at hydrological gauging stations that have larger upstream catchment areas compared to those in smaller catchments. The upstream catchment areas of the river gauging stations at which WRR1 and WRR2 are able to provide representation of the hydrological behaviour that is better than the average of the observed are found for catchment areas of some 2500 and 520 km2 and above respectively, with significantly poorer performance for smaller catchment sizes. This shows that the continued improvements in the global models with a higher resolution, either due to improved higher-resolution forcing or due to improved model structures, can be expected to lead in most cases to better capabilities of capturing the variability of the observed discharge as well as the magnitude of observed discharges.

A novel aspect of this study is the exploration of the skill of the global models in identifying the occurrence and severity of flood events in two benchmark chronologies of flood events. The first was developed through flood frequency analysis, with flood events identified to occur at selected probabilities, while the second was developed through collating reported flood events in three disaster impact databases. This shows that the global models do have skill in capturing the observed as well as reported damaging floods. This is, however, only the case when the thresholds of the discharges corresponding to the flood events are determined using the model climatology, and not the observed climatology. The simulated discharges of these global models are thus found to better represent the variability of the observed discharges than the magnitude; though this is less an issue for the better-performing higher-resolution models of WRR2.

Despite the absence of high-quality data in the Limpopo River basin and the coarse resolution of the models in the global reanalysis dataset, this research shows that, regardless these limitations, the global reanalysis dataset can provide valuable information for flood risk assessment in data-sparse regions. The skill of the models to predict flood events in the basin that have led to flood damage, as recorded in the chronology of reported floods, is an important finding, as global models such as those assessed here are often used in global and continental forecasting systems to generate flood forecasts and issue warnings in basins with little or no gauged data, but where floods and consequent impacts do occur. This indicates (i) that openly available global-scale hydro-meteorological data can provide valuable information regarding extreme events in data-sparse regions and may therefore be of use to local decision makers in mitigating the negative consequences of future flood events and (ii) that this may improve as the resolution of these global models improves.