Quantifying the effect of autonomous adaptation to global river flood projections: application to future flood risk assessments

Flooding is one of the most damaging types of natural disaster, causing serious economic losses and loss of life worldwide (Jonkman et al 2008, Rojas et al 2013, Jongman et al 2015, Winsemius et al 2015). Fluvial floods, in particular, are the most expensive flood hazard of the various flood types, including coastal and flash floods (UNISDR 2011, Jongman et al 2012). The reported annual economic losses and fatalities associated with floods in 2010 reached $39 billion at purchasing power parity (PPP) values (base year 2005) and 8406 persons, respectively, according to the Emergency Events Database of the Center for Research on the Epidemiology of Disasters (EM-DAT; Guha-Sapir et al 2016).

The Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (AR5, Field et al 2014) stated that flood hazards are projected to increase toward the end of this century, albeit based on limited evidence (but with a high degree of agreement). The main driver of flood risk is the increasing exposure of people and assets to flooding (Muis et al 2015). Both climate change and socioeconomic development are projected to increase the future global flood risk (e.g. Hirabayashi et al 2013, O'Neill et al 2014). Therefore, a comprehensive assessment of potential flood risks on a global scale, and an investigation of the impact of factors such as climate change, socioeconomic development and changes in vulnerability, are an urgent issues for policymakers for the facilitation of flood mitigation planning.

Vulnerability is one of the key components of risk assessments (e.g. Kron 2005, UNISDR 2009). Vulnerability could vary in the future through socioeconomic changes or disaster prevention efforts following socioeconomic development (i.e. levees could be raised as the demand for safety increases commensurate with potential increases in the population and affluence). Dang et al (2011) summarized from references that flood vulnerability was the degree of damage caused by a hazard of a given magnitude for a specific element at risk, like a depth-damage function, and they derived a quantitative measure of vulnerability from several social indicators, such as the population, risk perception and income.

Jongman et al (2015) calculated flood vulnerability (in terms of mortality rate and economic loss rates) as the ratio of flood consequences (reported flood fatalities and economic losses to potentially affected flood exposure (modeled flood-exposed population and global domestic product (GDP)), and revealed decreasing trends in historical vulnerability to river floods on a global scale. Tanoue et al (2016) confirmed the findings of Jongman et al (2015) using different disaster data with longer time periods. These studies suggest that it is essential to consider vulnerability changes that are commensurate with socioeconomic development in future projections. For example, previous studies have provided adaptation cost estimations to mitigate flood consequences under climate change (e.g. Jongman et al 2012, Ward et al 2013, Winsemius et al 2013), while these estimates did not distinguish the contribution of autonomously induced disaster risk management from adaptation costs. For the purpose of realistically estimating future adaptation costs to flooding, first the effect of vulnerability reduction by conventional damage risk reduction should be estimated, and then the additional cost to the mitigation of flood consequences due to climate change should be considered. Therefore, we proposed a conceptual definition of autonomous adaptation as an effort to mitigate disasters, strengthen resilience and reduce the vulnerability induced by conventional risk-reduction activities only, excluding other adaptation activities such as climate change.

In this study, we assess how autonomous adaptation can potentially reduce future flood consequences. We further estimated the potential fluvial flood risk for the end of the 21st century, taking different climatic and socioeconomic scenarios into account. Then, we evaluated the contributions of three flood damage drivers (climate, socioeconomic and vulnerability change) to future flood consequences. The evaluation of this contribution elucidates the modelization impact of autonomous adaptation. Note that the modeled vulnerability in this study is simplified to quantify flood vulnerability and that there are many aspects of vulnerability that cannot be quantified and used in the proposed model framework due to scale issues or lack of data. The defined vulnerability in this study and its relation to conventional flood risk drivers will be described in section 2.

The consequences of flooding (fatalities and economic losses) were estimated by multiplying exposure (population and assets prone to flooding) by vulnerability, which is one of the components expressed in the following risk equation (e.g. Kron 2005, Jongman et al 2015):

$$\begin{equation} {\rm{risk}} = {\rm{hazard}} \times {\rm{exposure}} \times {\rm{vulnerability}}. \end{equation} \tag{ 1 }$$

In this study, following Jongman et al (2015) and Tanoue et al (2016), flood hazard was numerically calculated as the extent and depth of a flood by the global river and inundation model, the catchment-based macro-scale floodplain (CaMa-flood) model (Yamazaki et al 2011; a more detailed description of the flood simulation is available in the supplementary material available at stacks.iop.org/ERL/13/014006/mmedia), and the term exposure was given as gridded population or an asset map (a more detailed explanation for estimating the gridded population/assets is in section 2.1). The potentially affected exposure was then estimated by superimposing the modeled hazard (flood extent) on the population or asset data, as our river and inundation model did not take any flood protection into account. The equation used to derive flood consequence in this study can be expressed by modifying equation (1):

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} {\rm{consequences}} \,\, &=& \, {\rm{potentially \, af{\kern-1pt}fected \, exposure}}\\ && \, \times \,{\rm{vulnerability}} \end{array} \end{equation} \tag{ 2 }$$

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} {\rm{potentially \, af{\kern-1pt}fected \, exposure}} \, \, &&= \, {\rm{f{\kern-1pt}lood \, extent}}\\ &&\, \times \, {\rm{exposure}}. \end{array} \end{equation} \tag{ 3 }$$

Zoom In Zoom Out Reset image size

The term 'vulnerability' is the main focus of this study. As flood consequences (fatality and damage) come from disaster statistics, we were able to obtain historical vulnerability as a ratio of the observed flood consequences and the potentially affected exposure at a national level in equation (2).

Using the obtained long-term past vulnerability change at the national level, a vulnerability model was developed, assuming that the historical reduction in vulnerability was only induced by autonomous adaptation. The vulnerability model was developed as a function of socioeconomic development and time, based on the finding of Jongman et al (2015) and Tanoue et al (2016). Several models and datasets were used to create vulnerability scenarios linked to socioeconomic development and to evaluate future potential flood consequences (figure 1 and table S1). First, we modeled the variation tendency of historical flood vulnerability. Second, based on the resulting vulnerability function, we developed future flood vulnerability scenarios according to three socioeconomic scenarios. Finally, we estimated future potential flood consequences. This was done by combining future flood inundation simulations based on four Representative Concentration Pathways (RCPs; RCP2.6, RCP4.5, RCP6.0 and RCP8.5) (Moss et al 2007), three Socioeconomic Shared Pathways (SSPs; SSP1, SSP2 and SSP3) (O'Neill et al 2014), and vulnerability scenarios associated with the SSPs. In this study, all combinations of SSPs 1–3, and all four RCPs, were used to assess potential exposure and flood consequences. A time series of future flood consequences was calculated for combinations of RCP (for flood hazard) and SSP (for an exposure map of GDP or population and vulnerability). Combining various SSPs, RCPs and general circulation models (GCMs), we finally obtained 105 scenarios for a potential flood damage assessment. In this study, the use of the term 'flood consequence' indicates a tangible negative impact on economical aspects (economic losses) and humanitarian aspects (fatalities). The effect of autonomous adaptation on reducing flood consequences was obtained from a simulation using fixed vulnerability at the current level.

2.1. Vulnerability data

2.2. Flood vulnerability scenarios

Based on knowledge derived from the analysis of past hazard statistics, we assumed that the loss rate and mortality rate were correlated with time and economic growth. The ability to reduce flood risks will improve on an autonomous basis up to a certain level in all countries according to the level of socioeconomic development. Jongman et al (2015) and Tanoue et al (2016) found that in many countries flood vulnerability decreases as a function of time and economic development. We assume that an increase in GDP leads to a decrease in vulnerability, and time to stock wealth is required for autonomous adaptation. Under the above assumptions, we introduced the vulnerability scenarios expressed by the recurrence equation:

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} V(t_{i+1})& & = V(t_{i}) + Y \cdot \Delta t_{i} + G \cdot \Delta {\rm GDP}_{\rm cap} (t_{i}),\\ && \,\,\,\, (i = 0,1,2, ...), \end{array} \end{equation} \tag{ 4 }$$

where V(t_i) represents a country's vulnerability in a specific year t_i, t₀ is the initial year (2010) and i represents a sequence of t with an annual interval. Y = Y[t_i, GDP_cap] and G = G[t_i, GDP_cap] represent globally common change rates of vulnerability, both of which are functions of time and GDP per capita, Δt_i = t_i+1− t_i is an increment of time, and ΔGDP_cap(t_i) = GDP_cap(t_i+1) − GDP_cap(t_i) is an increment of GDP per capita.

We assumed that the change rates Y and G varied depending on income levels and time periods. Change rates were estimated every 15 years from 1980–2013 on the time axis and each income level as defined by the World Bank (2016) was used on a GDP per capita axis. This partitioning resulted in 4 × 4 = 16 parameters for each change rate, enabling us to estimate change rate variations that reflected economic development stages. Finally, Y and G were estimated by a least squares method in logarithmic space. To prevent any effect of outliers arising from nonlinear vulnerability changes, for example, with respect to landslides in large-scale flooding events, we set constraint values of 0.00 and 0.01 as the upper and lower limits of the change rates Y and G, respectively. These constraint values were determined manually, by trial and error, and were set to be consistent with historical vulnerability trends at country level. We excluded some erroneously large vulnerability values from the vulnerability data estimated by Tanoue et al (2016) by applying thresholds of 0.1 for mortality rate and 1.0 for the loss rate before modeling.

Initial values for each country V(t₀) were estimated from the latest 10 year average (2004–2013) historical vulnerability values for the future vulnerability scenario. For countries where no data was available in the last 10 years, we used all available data from the whole historical period (1960–2013) to create initial values. For countries where no data was available for the whole historical period, we used the mean value of the corresponding income level (low, lower middle, higher middle and high) defined by the World Bank (2014) as the initial value. In the validation, we estimated the initial values from the average of 1960–1980.

The historical vulnerability data estimated by Tanoue et al (2016) indicated that the vulnerability of most high-income countries stabilizes at a certain level. Assuming that the lowest limit in the vulnerability scenario can be approximated by the mean value of vulnerability for the current period (1990–2013) in high-income countries, we set minimum values of 3.92 × 10⁻³ for the loss rate model and 5.04 × 10⁻⁶ for the mortality rate model. Combining the change rates Y and G and country-specific initial values V(t₀) with the minimum value constraints, we solved equation (3) and obtained the future vulnerability scenarios in 170 countries where the IIASA GDP and population data was available (figure S5). Countries where the IIASA data was not available were not included in the potential flood consequence estimation. The method and result of the validation with regard to the vulnerability estimation is in the supplementary material.

3.1. Projected potential flood consequences

There was an obvious spatial variability in the estimated potential flood fatalities (upper two columns in figure 2). In particular, central Africa, China and Canada showed marked increases (over double in the mean of 2081–2100; thereafter the 2090s) compared with the current level (1991–2005). On the other hand, most countries showed a decrease (<100% in south and north African countries, Latin America, and Southeast Asia; i.e. mainly lower-income countries) in SSP1 and SSP2, because of the moderate population increases compared to the explosive increase in SSP3 and high levels of socioeconomic development (which in turn causes more rapid reductions in vulnerability). In contrast, in many high-income countries, such as the UK and Germany, the level of flood protection was modeled as high, hence limiting the effect of autonomous adaptation in the future. Therefore, the change in potential fatalities was strongly dependent on changes in the population and climate change. In terms of climate change, potential economic flood losses were projected to increase in RCP8.5, but not in RCP2.6, in central Africa, Australia, North America and Southeast Asia. Conversely, regions in Europe, north Asia and central Asia had relatively lower levels of flood consequences in RCP8.5 than in RCP2.6. Under RCP8.5, climate change will cause less precipitation or snow melt in these regions and consequently there will be less flooding (Hirabayashi et al 2013).

Zoom In Zoom Out Reset image size

The variation in potential economic losses shown in the lower two columns in figure 2 had a similar spatial pattern to that of the potential fatalities, although most countries showed an increase (>100%) primarily due to economic development. Several countries (e.g. India, Indonesia, Ecuador and central Africa) were projected to experience increases of potential economic loss in excess of 5000% by the 2090s compared to the current level. In contrast, most high-income countries, e.g. the USA, Japan and European countries, had relatively lower increases in their potential economic losses (ranging from −100% to +1000% depending on the scenarios) because relatively lower levels of economic development were projected. On the other hand, the time variation of the global sum of potential economic losses showed the significant impact of incorporating the effect of autonomous adaptation. The increase of potential economic losses from the current climate without autonomous adaptation in SSP1 and RCP2.6 is US$212.6 billion, while about half (US$107.5 billion for SSP1 and RCP2.6 and US$70.5 billion for SSP3 and RCP8.5) is mitigated by autonomous adaptation.

A contribution assessment (see supplementary materials for the method) showed that climate change (with population and vulnerability fixed at current levels) contributed to an increase in potential fatalities, which was +3% in RCP2.6 and +29% in RCP8.5 (figure S7 and table S4). Socioeconomic changes (with the population fixed at 2090, and climate and vulnerability fixed at current levels) made the largest positive contribution to potential fatalities, with an increase of +25% in SSP1 and +125% in SSP3. Unlike other factors, vulnerability changes (with vulnerability fixed at 2090 levels and climate and population fixed at current levels) contributed to a decrease in potential fatalities, i.e. −65% in SSP3 and −72% in SSP1.

Socioeconomic changes were the dominant factor contributing to potential economic losses (figure S7 and table S5). Socioeconomic changes contributed to an increase in potential economic losses of +1051% in SSP3 and +1497% in SSP1. Climate change contributed to increased potential economic losses of +6% in RCP2.6 and +13% in RCP8.5. Similarly to the number of potential fatalities, the reduction of vulnerability due to autonomous adaptation contributed to a decrease in potential economic losses of −28% in SSP3 and −39% in SSP1.

Zoom In Zoom Out Reset image size

3.2. Flood consequence changes caused by a global temperature increase

Our vulnerability model of autonomous adaptation, which is defined as an effort to mitigate disasters, strengthen resilience and reduce vulnerability induced by conventional risk-reduction activities only, (excluding other adaptation activities such as climate change), contains several sources of uncertainty associated with the data and models used for historical vulnerability estimation. The first uncertainty arose from an estimation error with respect to the initial vulnerability. In some countries, there are erroneously high values in historical vulnerability data due to (1) the inclusion of fatalities and economic losses caused by other, simultaneously occurring hazards (such as landslides) in flood damage statistics (e.g. for Bangladesh in 1974 and Venezuela in 1999), and (2) a failure to reproduce extremely severe flooding events in the retrospective inundation simulation, which would lead to overestimations of initial vulnerability (e.g. for Pakistan in 2010 and Brazil in 1984; figure S2). Our vulnerability scenarios only represented long-term linear variations and did not include nonlinear variations depending on the magnitude of the flood hazards. We assumed that the observed reduction in vulnerability was only induced by autonomous adaptation, however, in reality, it reflects all adaptation efforts dealing with potential changes in flood magnitude and frequency, including ongoing climate change or experienced flood events.

The proposed vulnerability model has limitations that tended to be oversimplified and that did not take into account many aspects of vulnerability to be quantified and used in the proposed model framework due to scale issues or lack of data. A single vulnerability variable per country is problematic in larger countries where initial vulnerability and socioeconomic development varies within the country. To better reflect the regional complexity of vulnerability and to make the vulnerability model more sophisticated, more detailed regional classification and significant region-specific data are required for future studies. For example, regional information on flood protection levels would be beneficial for setting the initial value of vulnerability, and finer land use information may improve the spatial distribution of the asset.

The accuracy of retrospective inundation simulations is highly dependent on the reproducibility of climate forcing and runoff data, where the latter are further dependent on the physics of land-surface models, for example with respect to snow accumulation and melting. The parameterization of river routing and anthropogenic effects (e.g. dam operation and irrigation) in inundation models also affects simulation accuracy. It is also necessary to validate the adequacy of our assumptions that the population distribution would remain unchanged after 2005 and the horizontal resolution of population data can be sufficiently represented by a 10 km resolution. Tanoue et al (2016) conducted a sensitivity test of the effect of past changes on population distribution, and showed an increase in the mortality rate up to 48.9% in Asia and Africa. Enhancing the reproducibility of flood exposure data would contribute to the greater precision of vulnerability estimations. In addition, it is not always appropriate to apply our vulnerability scenario to certain countries with a complex topography (e.g. frequent landslides or low-lying delta areas). We were able to manage this problem through the calibration of a country's initial values in the vulnerability scenario.

Another source of uncertainty in vulnerability scenarios is the use of hazard statistics. Our vulnerability model is based on historical vulnerability variations estimated by historical hazard statistics. The EM-DAT database that we used for our historical vulnerability estimation has few records for low-income countries before the 1990s, causing a biased estimation of flood vulnerability in earlier periods (Tanoue et al 2016). The EM-DAT categorizes types of natural disaster using the terms 'hydrological group', 'geophysical group' and 'biological group'. Flood events are categorized as a 'hydrological group', and include not only river floods, but also coastal and flash floods. Landslide events are also included in the 'hydrological group', creating a source of uncertainty.

One of the limitations of our model was that it considered only direct economic losses. In reality, indirect economic losses, such as the halting of supply chains and lost opportunities caused by the suspension of businesses, are too significant to ignore (e.g. Koks et al 2015, Carrera et al 2015). Traditional approaches for flood consequence estimation were based on information regarding the flooded area, asset distribution and relationships between flood damage and flood depth (depth-damage functions) or vulnerability (as proposed in our study), but most previous studies paid little or no attention to the inter-industry spreading effects of economic impact due to flood hazards. Other non-monetary negative impacts of flooding, such as environmental change and emotional trauma, should also be considered for a thorough analysis. It will be necessary to incorporate these flood effects in the future to better understand flood consequences.

Although our flood model framework only used the flood inundation depth and extent for flood consequence estimation, other flood parameters such as flood velocity and inundation duration are also important factors for realistic flood consequence estimation (e.g. Kreibich et al 2009, Merz et al 2010, Dang et al 2011, Ahmadisharaf et al 2015). Our model framework is basically the same as that in Ward et al (2013) or Winsemius et al (2013), who conducted a global flood risk assessment using the annual maximum of inundation depth and extent. They use the knowledge that direct flood damage correlates with the annual maxima of inundation depth and extent, and as a consequence they obtained reasonable estimates for global flood damage estimation. Here the flood velocity is correlated with inundation depth, and thus is implicitly considered in their method. In our model framework, flood consequence calculation also uses the annual maxima of inundation depth and extension due to limitations of computational cost for calculating flood consequences under several future scenarios on a global scale. In addition, our study focuses on the direct damage caused by fluvial flood, not by flash flood. The inundation duration is important for calculating indirect damage because it relates to the factory's shut-down duration, and is expected to be included in a future model framework.

We assumed that our vulnerability estimates were a function of time and GDP per capita, and therefore did not consider nonlinear variations caused by dynamic processes, including adaptation and levee effects and better adaptation planning among others (Tanoue et al 2016). This nonlinear, discontinuous change in vulnerability is a key factor for assessing whether a flood hazard will become catastrophic. Therefore, the nonlinearity of the vulnerability changes must be modeled appropriately in the future.

Our study can facilitate adaptation planning and evaluations by both researchers and policymakers. One possible application of our study is the quantitative evaluation of additional adaptation costs under climate change scenarios, as well as an evaluation of the costs and the benefits of various types of mitigation strategy. For example, our model shows that the annual flood-related potential economic losses for the period of 2061–2100, in RCP8.5 and SSP3, amounted to 1395% of the current level without including the autonomous adaptation effect. Meanwhile, the inclusion of an autonomous adaptation effect decreased by 26% (i.e. the economic loss is projected to increase to 885% from the current level when the autonomous adaptation effect is considered). This 885% corresponds to the adaptation target for keeping flood consequences at the current level under climate change by 2100.

Our model can provide a baseline for evaluating potential flood consequences, although any planned adaptations are not considered. To evaluate the additional adaptation costs under different climate change and socioeconomic development scenarios, information regarding current flood protection levels is necessary. Recently, Scussolini et al (2016) constructed a global database of flood protection standards based on information on flood measures and policies at national and sub-national levels; gaps in their source data were addressed by empirical modeling. The inclusion of information on flood protection levels in our model enabled an evaluation of additional costs for flood protection, in addition to those associated with autonomous adaptation. On the other hand, infrastructural flood measures, such as the construction of dikes along rivers, may cause undesired environmental changes, which may lead to a decline in human well-being and should therefore be considered within flood adaptation planning. Further studies are needed to clarify the effectiveness of planned adaptations, and to evaluate the feasibility of adaptation measures when balancing economic and ecological aspects.

In this study, we quantified the effect of autonomous adaptation for global flood hazard assessment. In our model framework, autonomous adaptation is defined as the effort to strengthen resilience and reduce the vulnerability induced by conventional risk-reduction activity only, without considering other effects such as climate change or a particular experienced flood event. We developed flood vulnerability scenarios, which represent the effect of autonomous adaptation, based on historical vulnerability values derived from flood damage records and a river inundation simulation by Tanoue et al (2016). We estimated the future potential flood impact on a global scale according to a dynamic inundation model and modeled flood vulnerability scenarios. Including the effect of autonomous adaptation, our multimodel ensemble estimates project a 2.0% decrease in potential flood fatalities and an 821% increase in potential economic losses by 2100 under the highest emission scenario, with a large population increase (RCP8.5 and SSP3). The results of the contribution assessment indicate that autonomous adaptation would reduce the potential flood consequences by 64%–72% in terms of potential fatalities and by 28%–42% in potential economic losses by 2100. From the view of adaptation costs for climate change, autonomous adaptation would bear about half of the climate change adaptation required to keep the potential economic losses from floods in 2100 the same as the current level (1991–2005) (the effect of vulnerability change in table S5). Our model-based results also demonstrate that climate change could have a significant effect in terms of increased global potential flood consequences, with either 1.5 °C or 2.0 °C warming (see supplementary material).

To the best of our knowledge, this is the first study of global flood consequence assessment to quantify the effects of autonomous adaptation caused by disaster prevention efforts. Although our vulnerability model has a large degree of uncertainty, which is attributable to uncertainties in the pattern of actual economic growth in the future and the adequateness of our assumptions, the decreasing trend in historical flood mortality rate, without any planned adaptation to climate change, is consistent with the results of Jongman et al (2015) and Tanoue et al (2016). These studies strongly suggest that both time and economic development are associated with greater reductions in vulnerability according to the effects of autonomous adaptation. We demonstrated that quantifying changes in vulnerability can have a significant impact on flood consequence projections and is therefore essential for estimating the cost of adaptation to climate change.

This study was supported financially by the Funding Program for the Global Environmental Research Fund (S-14) of the Ministry of the Environment, Japan; the Integrated Research Program for Advancing Climate Models of the Ministry of Education, Culture, Sports, Science, and Technology (MEXT), Japan; and a grant-in-aid for scientific research (ID: 16H06291) from the Japan Society for the Promotion of Science (JSPS). We acknowledge an editor and the reviewers for their helpful comments.