Floods are among the most costly natural disasters in Europe [1
]. Their impact has grown steadily in the past decades due to the increase of population and built-up areas. Climate change is likely to affect the hydrological regimes in various world regions, with potential implications on the frequency and intensity of floods, and other weather-related hazards [2
Understanding and quantifying future flood impacts under different climate scenarios is key to developing adequate risk management actions. A large body of research addressing this topic has been produced in recent years. These range from local case studies to national, continental, and some global-scale assessments based on modelling chains of variable complexity e.g., [4
]. Europe is a region that has received considerable attention, thanks to the large availability of hydro-meteorological datasets, reported flood losses, and future climatic projections.
It is recognized that a single climate model cannot give robust predictions for informing adaptation, since uncertainties in regional climate changes are large. A common approach is to use ensembles of multiple climate models to account for a range of possible regional climate responses. Large ensembles of climate model projections now exist through the 5th Coupled Model Intercomparison Project (CMIP5) [12
], which contain tens of realizations of future climates (often referred to as ‘ensemble members’). The outputs of these can be directly used to provide input to models simulating processes relating to impacts, such as river flooding, or to provide boundary conditions for higher-resolution regional climate models (RCMs), which are in turn used to drive impact models. Many global and regional impact studies therefore use a subset of the CMIP5 projections, e.g., the Inter-Sectorial Impacts Model Intercomparison Project (ISIMIP) [13
] and the CORDEX [14
] initiative. Selection of an appropriate subset of members of a large ensemble of climate projections—‘ensemble sub-selection’—is a far from trivial process [15
]. If not chosen carefully with a specific application in mind, a small subset of ensemble members may not be as representative of the uncertainty in projections as the large ensemble, and may give a biased picture of future possibilities.
Few works have investigated the agreement (or disagreement) of flood hazard and risk projections derived from different studies. Comparison studies are crucial, because they allow researchers to identify strengths and limitations of different methodologies, as well as to investigate reasons for disagreement among model results. Further, policy makers demand best estimates of future risk trends, along with confidence intervals deriving from different studies in order to take action.
While multi-model ensembles have been used to investigate climate impacts on variables such as river flows [13
], water availability, and agricultural yields [16
], studies including analyses of the reasons for observed model differences are rare. Comparison studies available in the literature on flood risk projections are mostly qualitative e.g., [17
] due to the complexity of comparing different variables, resolution, and reference periods. For Europe, a recent comparison work by Kundzewicz et al. [19
] identified some regional trends in future flood frequency and magnitude (i.e., British Isles, Scandinavia, Eastern Europe) and pointed out areas where no robust signals of change could be identified (e.g., Southern Europe). However, quantitative comparisons are necessary to investigate the influence of modelling approaches and data on impact projections, and to gain greater confidence in model estimates.
The present work aims at answering two relevant questions: can we identify consistent, model-independent trends in flood risk in Europe under climate change? What are the reasons for the differences (and similarities) among projected model results?
To address these questions, we compare the results of three state-of-the-art research studies that evaluate the socio-economic impact of river floods in Europe under climate change. Specifically, we consider one case study at continental scale by Alfieri et al. [20
], hereafter referred to as JRC-EU, and the European component of two global scale applications [21
], referred to as ISIMIP and JRC-GL respectively, designed to evaluate economic damages and population affected under specific warming levels (SWLs) of 1.5 °C, 2 °C, and 3 °C, relevant to the Paris Agreement [23
]. To the authors’ knowledge, these three assessment studies are the only ones available to date at the European scale which estimated: (1) the socio-economic impacts of river floods; and (2) at SWLs, as prescribed by the Paris Agreement, though some studies focusing on only one of these two key points have been recently produced [7
]. The analysis is complemented with a comparison of model results for a baseline period with country level loss data reported by global datasets on natural disasters including the Emergency Events Database (EM-DAT), from the Centre for Research on the Epidemiology of Disasters (CRED) [27
], the NatCatService by Munich-RE [28
], as well as loss estimates produced for the Global Assessment Report (GAR) on Disaster Risk Reduction [29
We analyze quantitatively the differences in projected changes at the country scale and discuss reasons for the observed outcomes. The three cases cover a wide range of methodologies and datasets for climate forcing, hydrological and flood modeling, and impact assessment. Therefore, the comparison is expected to shed light on the influence of the data applied and methods to assess impact projections.
To compare results from the three cases we use the following approach. First, we compare quantitatively impacts for the baseline period with reference data available from disaster datasets and risk assessment studies. Then, we provide a general overview of the multi-model agreement at the European scale, to highlight possible spatial patterns of change. Finally, we evaluate the agreement of future impact estimations by comparing relative changes in impacts between the baseline and the three SWLs.
and Figure 3
compare the simulated impact of the three ensemble estimates for the baseline period 1976–2005, with the range of available reference datasets including data from the GAR, EM-DAT, and Munich RE for recorded losses and only EM-DAT for reported population affected. Note that for some countries (e.g., Finland, Iceland, Cyprus), no reported data were available for comparison from these three losses databases.
Regarding the reference datasets, there are some important differences to point out. Data from EM-DAT and Munich RE are observations and therefore refer to time variable socio-economic conditions of exposure and vulnerability. On the other hand, GAR estimates are expected values of average annual losses and are based on present day conditions. As shown in Figure 2
and Figure 3
, differences in the average and in the spread of the results are sometimes remarkable. Some general considerations can be drawn as follows:
ISIMIP generally has the largest spread in the ensemble, due to the larger number of ensemble members and the combination of different GHM and GCM;
ISIMIP average impacts are the largest in most countries (31 countries out of 38), which can be attributed to the methodology that considers the whole river network irrespective of the upstream area of catchments. In addition, the coarser resolution of flood maps produces larger flood extents, and in turn, impacts (see Figure 1
). JRC-EU average impacts are the largest in 6 out of 38 countries, including Czech Republic, Croatia, Ireland, Luxembourg, Poland, and Slovenia, while JRC-GL average impacts are the largest only in Latvia, though with a similar value to the other two ensemble means.
JRC-GL baseline impacts are the smallest of the three in most countries, due to the reduced extent of the river network considered (i.e., only rivers with upstream area above 5000 km2). Indeed, results from the JRC-GL and consequent projected changes under global warming could be considered as representative of the flood risk in large rivers only.
In most countries, the confidence bands of the ensembles intersect the range of reported economic losses. However, ISIMIP results for some countries are well above this range, notably for Ukraine and Italy. This is in line with the results of the evaluation exercise performed by Dottori et al. [21
] for ISIMIP, who observed an overestimation of impacts for some European countries. To provide a measure of the accuracy currently attainable with state-of-the-art flood damage models, recent works showed that the expected difference between simulations and observations can be of a factor of two or even more [54
Uncertainties and limitations in the available impact datasets are a known issue [55
], especially for global datasets [56
], though this issue can be partly addressed through the use of simulated impacts [57
]. Main issues include under-reporting of minor flood events and of those further back in time, absence of economic loss data for a large part of reported events, and uneven data coverage across European countries (e.g., fewer data for Eastern European countries before 1990 and in particular for countries that were part of the Soviet Union). For example, a comparison of national disaster loss databases with EM-DAT data showed that total losses can be up to 60% higher when data from high-frequency, low-severity events are accounted for [29
Results in terms of affected population are comparable to those of economic damages, with similar spread in the ensemble results, though with a clear tendency of model results to be higher than reported figures (Figure 3
). For population, it must be noted that observed data come only from the EM-DAT database and that the evaluation of population affected is more complex and prone to errors due to different standards for reporting the number of people hit by floods [58
summarizes the agreement between the three ensemble averages for each country and SWL scenario, considering the sign of projected changes in flood impact. The agreement is evaluated giving the same weight to each ensemble and using +/− signs as follows:
+++ (−−−) : all cases predict an increase (decrease) in impacts;
++ (−−) : two cases predict an increase (decrease) in impacts, results are not available for the third (see Section 2.2.1
+ (−) : this is used for two cases: (1) two cases predict an increase (decrease) in impacts while a third predicts an opposite change; or (2) only one case study is available and predicts an increase (decrease) in impacts;
0: only two ensembles available and predicting opposite changes in impacts.
The spatial distribution of the model agreement shows that the three flood risk assessments agree on an increasing trend in most of Western and Central European countries, and on a decreasing trend in Eastern countries. Model results are more variable in a number of northern countries like Iceland, Finland, Estonia, and Latvia, and in most south-eastern countries, with the exception of Greece. Interestingly, impact trends for the British Isles and for Eastern Europe mostly agree with those identified by Kundzewicz et al. [19
and Figure 6
focus on the future impacts predicted by the model ensembles, showing the relative change for each SWL and country with respect to the baseline. The plots allow comparisons of the magnitude of predicted changes, complementing the information shown in Figure 2
and Figure 3
with a quantitative assessment. Some considerations can be drawn from those figures:
In most countries in Western and Central Europe, all models consistently predict a relevant increase in future flood impacts.
The largest changes are usually predicted by the JRC-GL, which projects a more than 10-fold increase in impacts in the Slovak Republic, Hungary, and Poland. Conversely, the ISIMIP ensemble predicts smaller changes, with JRC-EU generally in between. In particular, ISIMIP predicts a negative change for several south-eastern and eastern countries, while JRC-EU and JRC-GL foresee a decrease only in few countries.
For the vast majority of countries, projected changes in flood risk for each of the three models along the SWLs differ considerably less than the corresponding changes among models, for each specific SWL. Country average range of percent change in flood risk along SWLs is of 180% for expected damages and 170% for population affected. Such values are smaller in comparison to the average range of percent change in flood risk along the three models, which is of 490% for expected damages and 540% for population affected.
The trend of flood risk for increasing warming levels is similar for the three models, for most countries. However, notable exceptions are found in Poland, Germany, Czech Republic, Finland, Sweden, Spain, and Bulgaria, where at least two out of the three models show a monotonic trend of the opposite sign (e.g., in Poland, expected damage estimates from JRC-EU decrease with higher warming levels, while estimates from JRC-GL increase with the SWLs).
In a number of countries, impacts may largely increase even in the case of limiting future warming to 1.5 °C.
Further insight on the spread of relative changes at country level for each ensemble and each SWL is given in Supplement Figures S1–S6
. Those figures show the large variability in impact estimates of specific ensemble members, which stresses the challenge in characterizing the overall uncertainty of the three combined ensemble estimates.
Summary impact projections and relative changes from the baseline for the three cases are shown in Table 2
for expected damage and Table 3
for population affected. The JRC-EU provides the best estimates of flood impacts at the European level for the baseline period, where reported annual figures are between 4.3 and 8 B€ (5 B€ for JRC-EU) of losses and 262,000 (216,000 for JRC-EU) people affected by flood events in Europe [59
]. Average relative changes in flood impacts of the three ensembles (super-ensemble) rise with the SWLs from 113% (expected damage) and 86% (population affected) at 1.5 °C, up to 145% and 123%, respectively, at 3 °C. These are the result of averaging a marked increase in flood risk by the JRC-EU and JRC-GL, with the ISIMIP predictions which point to an initial growth of impacts at 1.5 °C and then a further stabilization for higher SWLs. One should note that JRC-EU and JRC-GL are more likely to identify non-linear trends in flood risk, thanks to the POT approach that enables the detection of non-linear changes in the frequency of future floods. Similarly, the coarser resolution inundation model of ISIMIP, coupled with flooded fraction maps, is prone to underestimating non-linear changes in the flood impacts.
4. Discussion and Conclusions
This work presents, for the first time to our knowledge, a quantitative comparison of socio-economic impact projections of river floods in Europe under climate change, derived by three research works based on state-of-the-art models and datasets. We included only those three case studies in the comparison, due to the very limited availability of large scale assessments of the future impacts of natural hazards under specific warming levels. Nation-wide risk assessments have the advantage of: (1) enabling the comparison of methods set up at different spatial grid resolutions; and (2) enabling the comparison of modeled impact estimates over a set of past years with reported aggregated values available from disaster losses datasets.
As expected, the quantitative comparison of results shows significant differences among the three assessments, which may depend on different modelling components and data used in each case study. Part of these differences can be attributed to the use of different climate and hydrological models. While the uncertainty related to the climatological forcing is well known in the literature, results from Dottori et al. [21
] show that the hydrological modeling component may also have a significant impact. The resolution of flood maps and of the driving inundation models also play a prominent role in determining the overall impact estimates. In this regard, we stress the importance of using high resolution inundation modeling to achieve accurate impact estimates. This is presently limited by the scarce availability of high resolution Digital Elevation Models (DEM) over large areas, where small scale features are able to influence considerably the distribution of the floodwaters. Results from the three ensemble projections and their distributions suggest that multi-ensemble averaging can be one way to improve the robustness of impact estimates. However, a rigorous characterization of the multi-model uncertainty appears a more challenging task, due to the strong heterogeneity between the distributions of the three ensemble projections. Moreover, expert knowledge on specific modeling components and their limitations can help in identifying more realistic ranges of uncertainty. Approaches based on Bayesian statistics are a possible way forward to account for such prior information and improve the predictive uncertainty, provided that these are quantified within realistic ranges.
Regarding impact modelling, the three cases here considered use similar approaches, mainly based on the simulated extent and depth of flooding. There is a wide variety of flood damage models in use that can differ substantially in methodological aspects and economic estimates. The datasets and resolution of exposure data may be an additional factor in explaining differences in results, as shown by a comparative quantitative flood damage model assessment by Jongman et al. [61
Overall, we found JRC-EU to produce the best quantitative estimates of past impacts as compared to the other two cases, most likely due to the higher resolution and better quality of the underlying models and datasets. Nevertheless, we found that all the three cases produced results comparable with observed loss data. This is an important result because the performance for present-day conditions does not necessarily imply skillful prediction of future changes, and therefore, the joint analysis of multiple case studies can help identify more robust trends in the future flood risk in Europe.
Results from the three assessments suggest that climate projections are the main driver influencing future trends of flood risk under global warming. Moreover, the uncertainty attributed to the climate projections is likely to be underestimated due to the relatively small, though widely used, model ensembles [62
]. Other factors such as the bias correction of climate projections, the method for assessing the year of exceeding SWLs, and the spatial resolution of the input data, surely do influence results though probably to a smaller degree, and do not affect the direction of the projected changes. Despite some differences in the absolute and relative change in projected flood impacts at SWLs, the three cases showed a generally good agreement in the spatial distribution of the direction of changes. In detail, most of the Central and Western Europe is consistently projected to experience substantial increase in flood risk at all SWLs, with the magnitude of the change increasing for higher levels of warming. Conversely, some persistence in the signal of decrease in flood risk with warmer temperatures is found in some countries in Eastern Europe, though in most occasions the three case studies provided contrasting results, showing that highest uncertainties are located in Eastern Europe and particularly in the Balkan region. Interestingly, in some countries in Southern Europe (Spain, Portugal, Greece), the initial increase in impacts at 1.5 °C turns into more uncertain projections in the case of higher warming levels, due to a consequent substantial reduction in the mean annual precipitation.
Future works should focus on quantifying the influence of specific modelling components or datasets by systematically comparing different versions of the same modelling framework. While similar studies would be demanding, given the amount of data and computational times required to run a full flood impact modelling chain, we believe that more similar case studies should be carried out to improve the robustness and reliability of flood risk estimates. To this end, the evaluation of flood impact models within inter-comparison projects such as ISIMIP is a valid option to progress further.
This work confirms that the impacts of global warming on river flood risk in Europe are widespread and often significant, though they can vary in sign and magnitude from region to region. The Paris Agreement has set critical thresholds of warming that we must aim for, yet it has demanded that the scientific community provide additional evidence on the possible effects of warming on the consequent impacts on the society. Our results show that substantial impacts can be avoided by limiting global warming to lower temperature thresholds. However, a considerable increase in flood risk is predicted in Europe even under the most optimistic scenario of 1.5 °C warming as compared to pre-industrial levels, urging national governments to prepare effective adaptation plans to compensate for the foreseen increasing risks.