Clusters of community exposure to coastal flooding hazards based on storm and sea level rise scenarios—implications for adaptation networks in the San Francisco Bay region
Sea level is projected to rise over the coming decades, further increasing the extent of flooding hazards in coastal communities. Efforts to address potential impacts from climate-driven coastal hazards have called for collaboration among communities to strengthen the application of best practices. However, communities currently lack practical tools for identifying potential partner communities based on similar hazard exposure characteristics. This study uses statistical cluster analysis to identify similarities in community exposure to flooding hazards for a suite of sea level rise and storm scenarios. We demonstrate this approach using 63 jurisdictions in the San Francisco Bay region of California (USA) and compare 21 distinct exposure variables related to residents, employees, and structures for six hazard scenario combinations of sea level rise and storms. Results indicate that cluster analysis can provide an effective mechanism for identifying community groupings. Cluster compositions changed based on the selected societal variables and sea level rise scenarios, suggesting that a community could participate in multiple networks to target specific issues or policy interventions. The proposed clustering approach can serve as a data-driven foundation to help communities identify other communities with similar adaptation challenges and to enhance regional efforts that aim to facilitate adaptation planning and investment prioritization.
KeywordsClimate change Adaptation Flooding Exposure Cluster analysis
Coastal communities in low-lying areas are continually vulnerable to flooding, and future threats are likely to increase due to the influence of projected sea level rise (SLR) (Deconto and Pollard 2016), potentially displacing millions of people and causing up to $1 trillion in damages (Nicholls 2004; Hinkel et al. 2013). Projected increases in sea level can lead to more frequent and persistent nuisance flooding (Sweet et al. 2014), permanent inundation of low-lying areas, and increases in the extent and depth of storm-related inundation (Cayan et al. 2008; Tebaldi et al. 2012; National Research Council 2012). As the natural sciences continue to improve our understanding of how SLR may affect coastal hazards, there is increasing attention on how communities can adapt to an evolving coastal zone and mitigate potential impacts. Amidst the uncertainty associated with future flooding threats, coastal community planners must continue to make decisions related to maintaining public infrastructure and providing government services to uphold and hopefully improve quality of life and livelihoods in their communities. Many public agencies are becoming cognizant of threats yet are struggling to determine how to best deal with the potential effects of SLR and storm impacts on communities and regional infrastructure (Hanak and Moreno 2012; The Nature Conservancy 2014).
Efforts to address potential impacts from climate change have increasingly called for regional collaboration and have recognized the importance of governance networks (Daniell et al. 2011; Luthe et al. 2012; Dow et al. 2013; McAllister et al. 2014; Leck and Simon 2013). The ability to share lessons learned and build cross-geography approaches provides an opportunity for minimizing duplication of work, leveraging limited resources, and focusing on specific intervention strategies (Adger et al. 2005; Leck and Simon 2013; Thaler et al. 2016). However, the optimal structure of networks remains unclear in the literature (Engle and Lemos 2010; Baird et al. 2016), and there currently is no standardized methodology for quantifying areas with similar climate-change challenges (The Nature Conservancy 2014), making it difficult for communities to identify shared needs and opportunities. Many climate-change initiatives created to support trans-boundary collaboration are designed primarily for sharing information among self-selecting jurisdictions and not for identifying common issues based on comparative vulnerability assessments (Leck and Simon 2013; 100RC 2017; C40 2017; The Nature Conservancy 2014; ARCCA 2017). While these efforts provide useful platforms for sharing best practices, their focus is often broad, thus limiting their capacity to provide specific information on which cities may be facing analogous threats and could consider forming partnerships to enhance collaboration.
Understanding a community’s vulnerability to climate-driven hazards can be a complex undertaking, but a common first step to support more in-depth analyses is an estimation of its hazard exposure and sensitivity based on the integration of hazard scenarios with infrastructure and socioeconomic data (e.g., Frazier et al. 2010; Wu et al. 2002; Strauss et al. 2012; Schweikert et al. 2015; Abdollahian et al. 2016; Paprotny and Terefenko 2017). The relative vulnerability of multiple communities then can be characterized using comparative metrics, such as indices based on principal component analysis (e.g., Cutter et al. 2003; Wood et al. 2010; Nelson et al. 2010). Although useful for discussing vulnerability across a region, single numerical indices have limited utility for partnership development since results may focus attention only on highly ranked jurisdictions, as well as overlook the factors that create vulnerability and their interrelationships (Sharma and Patwardhan 2008). Additionally, single metric rankings may miss similarities that communities share if they are not wholly similar. Other quantitative methods, such as cluster analysis, can be used to identify groupings of communities based on similar vulnerability characteristics (e.g., Wood et al. 2015) and may therefore be better suited for identifying potential partners in climate-change adaptation.
Cluster analysis has been used to characterize and differentiate various aspects of climate-change impacts, from regional differences in landscape sensitivity (Mezosi et al. 2013) to basin variations in water resource governance (Engle and Lemos 2010). Cluster-based typologies have been created to identify variations in potential climate-related impacts to household livelihoods related to farming (Nelson et al. 2010; Kok et al. 2016), fishing (Perret and Yuerlita 2014), and cattle production (Marshall et al. 2014). Place-based cluster analyses have characterized differences in social vulnerability and resilience to climate-driven coastal hazards at the census tract level (Stafford and Abramowitz 2017), for single-sector fishing communities (Pollnac et al. 2015), and for US counties (Lam et al. 2016). Each of these studies provides insight into societal vulnerability to climate change, but they have focused only on specific sectors or have not been conducted at the local jurisdictional level. As a result, local practitioners lack the necessary information to create effective inter-governmental climate adaptation networks.
Building on the recognized need to identify regional partnership opportunities based on similar climate-change vulnerabilities, this study uses statistical cluster analysis to classify community similarities of exposure to climate-driven hazards. To demonstrate this approach, we use the coastal communities and counties within the San Francisco Bay region of California (USA). We chose this study area because there has been considerable work in recent years to characterize a suite of coastal flood-hazard zones that reflect a wide range of storm and SLR scenarios (e.g., Barnard et al. 2014) and to estimate variations in community exposure to these hazards based on population, businesses, infrastructure, facilities, and developed land (Jones et al. 2016; U. S. Geological Survey 2017). In addition, the wide range of jurisdictions and current lack of strong inter-community collaboration for adaptation planning suggest that this area would benefit from efforts to promote knowledge sharing and regional planning (California Energy Commission 2012). This study contributes to the growing understanding of societal vulnerability to climate change by focusing on jurisdictionally defined profiles of community vulnerability and by examining how cluster compositions change based on the societal variables or climate-hazard scenarios considered. Results can be used to promote more effective collaboration at a regional level by identifying comparable communities so that they may work together to determine effective adaptation strategies based on local conditions. This analysis not only provides specific results for the San Francisco Bay region but also presents a transferrable method to identify strategic partnerships informed by hazard exposure and demographic variables elsewhere.
This analysis focuses on identifying clusters of communities with similar trends of exposure to coastal flooding hazards in the San Francisco Bay region in terms of current conditions and future conditions that reflect various storm and SLR scenarios. A community-exposure database summarized by Jones et al. (2016) provides the foundation for this statistical cluster analysis.
Geospatial data characterizing flooding extents for various coastal storm and SLR scenarios in the San Francisco Bay region were generated by the US Geological Survey (USGS) Coastal Storm Modeling System (CoSMoS) project. General background on modeling techniques and assumptions are summarized elsewhere (Barnard et al. 2009; Barnard et al. 2014; U.S. Geological Survey 2015), and geospatial data that summarize zones for the San Francisco Bay region are available online (Our Coast Our Future 2017). In short, CoSMoS modeling yields local-scale projections of SLR and storm-induced coastal flooding based on a global wave model, a satellite altimetry-based global tide model, and atmospheric forcing data from global climate models to determine regional wave and water-level boundary conditions, which are dynamically downscaled using a series of nested wave and tide models. Storm scenarios include no storm (e.g., average daily conditions) as well as annual, 20-year, and 100-year return level coastal storms. SLR scenarios include 25 cm increments from 0 to 200 cm as well as 500 cm. The most recent scientific projections for likely SLR by 2100 in San Francisco Bay range from 42 to 166 cm (National Research Council 2012). Scientific consensus on the magnitude of SLR projections over time is constantly evolving; therefore, we characterize changes in sea level by potential increases and not by a specific time period.
Unit of comparison
We focused our comparative analyses on incorporated cities and counties as delineated by 2010 boundaries of the US Census Bureau (U.S. Census Bureau 2010). Coastal counties contain many unincorporated towns and villages that do not have formal charters for self-governance, as an incorporated city would. Emergency services, economic development, and land-use planning for these areas are performed by county offices; therefore, exposure results for unincorporated areas are aggregated and reported at the county level as “remaining land” for a given county. An overlay of community and county boundaries for the San Francisco Bay region with the maximum flood-hazard zone resulted in 63 communities for analysis in this study, including 55 incorporated cities and the remaining unincorporated land in 8 counties (Online Resource 1).
Geospatial data summarizing various population, business, and infrastructure indicators were used to estimate community exposure to a given hazard zone in Jones et al. (2016). Residential populations were estimated using block-level population counts compiled for the 2010 US Census (U.S. Census Bureau 2010). Demographic factors, such as age, ethnicity, and tenancy, can amplify an individual’s sensitivity to hazards (Morrow 1999; Fothergill et al. 1999; Burby et al. 2003; Wood et al. 2012); therefore, the 2010 block-level data were used to estimate demographic attributes related to these socioeconomic indicators of sensitivity, including ethnicity (Hispanic or Latino), non-White race (American Indian and Alaska Native, Asian, Black or African-American, Native Hawaiian and other Pacific Islander), age (individuals less than 5 and more than 65 years in age), tenancy (renter-occupied households), and group quarters (institutionalized, such as correctional facilities or nursing homes, and non-institutionalized, such as dormitories or military barracks).
Business populations and regional trends of exposure were estimated in Jones et al. (2016) using employee counts organized by North American Industry Classification System (NAICS) codes (U.S. Census Bureau 2010) at individual businesses using a georeferenced, proprietary employer database (Infogroup 2012). The number of employees associated with each business type is used to identify the primary business sectors in flood-prone areas, an indicator regularly used to evaluate economic health and market trends (Bureau of Labor Statistics 2015). Business types based on NAICS codes are generalized in this analysis into five classes: (1) government and critical facilities, (2) manufacturing, (3) services, (4) natural resources, and (5) trade.
Hazard exposure of critical facilities and infrastructure was estimated in Jones et al. (2016) using the length of rail and road networks (infrastructure) and the number of schools, medical facilities, police stations, and fire stations (facilities). These facilities are considered critical because they provide public safety services or house vulnerable populations. Data sources for critical facilities and infrastructure include a wide array of county and federal sources, which are summarized by Jones et al. (2016). For each variable, geographic information system (GIS) software was used to overlay data representing community boundaries, the community indicator, and a specified hazard zone. Two variables for each asset were estimated at the community level: (1) a total amount (or length, in km, for road and rail networks) of an asset in a hazard zone and (2) a community percentage. For resident and employee populations in hazard zones, the community percentage reflects the exposed amount compared to the total amount within a community (e.g., 500 residents in a hazard zone that represent 20% of a community’s total population). For the business types, percentages reflect the number of businesses of a certain type divided by the total number of that business type in the community. For the demographic attributes, community percentages reflect the percentage of a specific attribute relative to the total number of residents in the hazard zone, not the community total. Spatial analysis of vector data focused on determining if points (businesses and critical facilities), lines (roads and rails), or polygons (census blocks) were inside hazard zones. If census-block polygons overlapped hazard polygons, final population values were adjusted proportionately using the spatial ratio of each sliver within or outside of a hazard zone.
Jones et al. (2016) summarizes estimates of community exposure to 36 flood-hazard zones (based on combinations of 4 storm scenarios and 9 SLR scenarios) for 62 distinct input variables for 63 geographic units (55 incorporated cities plus the remaining unincorporated land in 8 counties). Results also demonstrated wide ranges (e.g., residential exposure in flood-hazard zones assuming 100 cm of SLR ranged from 0 to 34,877 among the 63 units of analysis), making typical descriptive statistics (e.g., means) less useful for communicating results to practitioners. Therefore, a challenge in this study was to determine a method for synthesizing the diverse array of data in a statistically appropriate way that would also yield meaningful results to practitioners with both local and regional perspectives.
Cluster analysis was chosen as a method to ascertain similarities between geographic areas and to identify “clusters” of communities facing similar magnitude, type, or timing of exposure to flooding hazards. Several methods for identifying potential clusters in data exist, and the partitioning around medoids (PAM) method was chosen for this study (Kaufman and Rousseeuw 1990). Medoids are similar in concept to means or centroids but are always members of the data set. PAM uses medoids instead of cluster means to define partitions, which is considered to be a more robust approach for data sets with outliers (Kaufman and Rousseeuw 1990). The PAM algorithm implemented in the cluster package of the R statistical program (R) (c.f., Maechler et al. 2016) was used to perform cluster analysis on the Jones et al. (2016) exposure database. The PAM algorithm requires as inputs a distance matrix between observational units (in this case, a Euclidean distance matrix between cities and counties) and the number of clusters k. Before clustering, each variable was first converted to a z-score by subtracting the mean and dividing by the standard deviation across observational units. This allowed for comparisons between data types that could not be directly compared based on the original scale of the data, such as comparing between the length of highways and number of critical facilities.
An important aspect of cluster analysis is the validation of the obtained clustering by assessing the separation and stability of clusters. Several methods were used in this study to select the number of clusters k for data presented in the “Results” section. Average silhouette widths were calculated for k = 2 through k = 10 clusters to measure within-cluster tightness and between-cluster separation (c.f., Rousseeuw 1987). Cluster stability was determined using bootstrapping and noise addition techniques implemented in the fpc package of R (Hennig 2016). The new clusters produced using bootstrapping and noise addition were compared to the original clusters through calculation of the Jaccard coefficient (c.f., Hennig 2007), which is a measure of the similarity between two data sets.
Two distinct analyses were performed. The first analysis (referred to as a “current planning horizon”) clustered communities for a single coastal-flooding hazard zone that reflects 50 cm of SLR and a 100-year storm scenario based on four data types, which were each considered separately. This serves as a base case for future threats and reflects a 30- to 50-year planning horizon, typical for infrastructure investments and community planning. We did not assign this amount of SLR to a specific time period given the wide range of projections in the region (Our Coast Our Future 2017); however, 50 cm of SLR along the California coast in the next 30 to 50 years is considered plausible (Vermeer and Rahmstorf 2009; National Research Council 2012). The data types considered include the following: population (four variables: number and percent of residents and number and percent of employees), built environment (two variables: number of critical facilities and length of critical infrastructure), residential demographics (five variables: ethnicity, non-White race, age, tenancy, and group quarters), and business types (ten variables: number and percent of employees in each of five classes: government and critical facilities, manufacturing, services, natural resources, and trade). To understand how clusters may change over time as the threat of flooding increases, a second analysis (referred to as “planning for multiple SLR scenarios”) was performed using only the population and built environment variables, which were each clustered separately for six combinations of three SLR scenarios (50, 100, and 150 cm) and two storm scenarios (no storm and 100-year storm), producing 12 distinct clustering outputs.
Current planning horizon
Clustering based on population exposure suggests four groupings (Fig. 1a), although groups 1 and 2 exhibit only small differences. Communities in group 4, including Foster City, Redwood City, and San Mateo, have the highest exposure and collectively represent 60% of the 135,685 residents and 56% of the 98,960 employees in the hazard zone across the study area. Group 3, which includes four communities in Marin and Santa Clara counties, has moderate exposure for both populations, while group 2 has much lower exposure and includes several communities in Marin, Alameda, and San Mateo counties. Group 1 represents the remaining communities, most of which have low exposure of populations.
For built environment data describing the length of critical infrastructure and the number of critical facilities in flood-hazard zones, clustering also suggests four groups (Fig. 1b). San Mateo is once again in the highest-exposure group, with 19% of the exposed highways and rails and 32% of the exposed critical facilities. Group 3 consists of four communities in Marin, Santa Clara, and San Mateo counties that have moderate exposure for infrastructure and facilities. Groups 1 and 2 represent the majority of communities in the study area and have low amounts of critical infrastructure located in the hazard zone.
Clustering based on demographic variables for residents in flood-hazard zones suggests six groups defined by ethnicity, race, age, group quarters, and renter-occupied households (Fig. 1c). South San Francisco is clustered into group 6 by itself due to the relatively high percentage of residents in the hazard zone that are living in group quarters (56%). Six communities in Marin, Santa Clara, and San Mateo counties constitute group 5, primarily due to a relatively high percentage of individuals in hazard zones that identify themselves as Hispanic or Latino (33–75%). Group 4 consists of seven communities that are primarily characterized by a relatively high percentage of households in the hazard zone that are renter-occupied (60–91%). Group 3 represents communities that have a slightly higher percent of residents in hazard zones that are under 5 or over 65 years of age (15–42%). Group 2 is composed of communities that have relatively higher percentages of residents that identify themselves by a non-White race (36–91%). Group 1 represents communities that do not have any residents in hazard zones.
When considering the type of businesses in hazard zones, four groups emerge (Fig. 1d). Foster City and Redwood City comprise group 4 and have high exposure in terms of numbers and percentages of employees in the flood hazard zone in all sectors except natural resources. Group 3 consists of four communities that have moderate numbers and percentages of employees in hazard zones in all sectors. Group 2 has only slightly higher exposure than group 1, notably for the number of employees in the natural resources sector and the percentages of employees in all sectors except natural resources.
Cluster assignments are fairly similar for groupings based on population (Fig. 1a), critical facilities and infrastructure (Fig. 1b), and business types (Fig. 1d), with one large cluster that represents the majority of communities and is estimated to have low to no asset exposure and then smaller clusters comprised of communities with increasing levels of exposure. Most of the communities that fall within the high-exposure clusters are located in San Mateo or Marin counties. For example, Redwood City, Foster City, and San Mateo in San Mateo County and San Rafael in Marin County are consistently in groupings that represent moderate or high exposure for these three analyses. Clusters related to variations in demographic sensitivity cannot be ordered in a similar fashion from least to greatest community exposure (i.e., vulnerability issues are different, but not necessarily better or worse, for having more renters than residents under 5 years old in hazard zones). Instead, results indicate that most communities have moderate or high exposure for at least one demographic attribute (Fig. 1c).
Planning for multiple sea level rise scenarios
SLR is not a static prediction for a specific location or time period; therefore, communities benefit from understanding trends in exposure across a range of SLR scenarios. To illustrate progressions in hazard exposure over time, population and built environment variables were also clustered individually for three SLR scenarios (i.e., 50, 100, and 150 cm) based on projections for the region. Changes in community clusters across a range of SLR scenarios provide insight on how community adaptation networks and collaborative opportunities could change over time.
Results also suggest that although the community groupings are relatively stable with increasing SLR, the range in z-scores between the four groupings decreases, which indicates that deviations from the mean become less pronounced between the clusters. For example, z-scores for communities in group 4 in the no storm scenario are on the order of 6.0 for 50 cm of SLR but decrease to values between 2.0 and 4.0 for 150 cm of SLR. This decrease in the range of z-scores suggests that only a few communities have high exposure for lower SLR scenarios but that more communities throughout the region will begin to have similar exposures as SLR intensifies.
The statistical cluster analysis presented in this study complements existing collaborative initiatives by providing a data-driven method for quantifying community similarities of exposure to current and projected coastal hazards. It varies from previous cluster analyses of societal vulnerability to climate change by focusing on local jurisdictions and by including multiple hazard scenarios to recognize the uncertainty in future climate-driven hazards. Results could be used by individual community leaders to identify collaborative opportunities on their own and by regional practitioners and nonprofit representatives to connect similar communities unaware of each other’s similar challenges. Results may also help policymakers work across scales to connect national policies to local efforts, align planning processes, and reduce gaps in governance, which have all been identified as challenges in integrated resource management (Rouillard and Spray 2016).
Implications for regional adaptation networks and collaborative planning
Based on our case study, cluster analysis provides a regional perspective of shared community vulnerability that may not be readily apparent if one were to rely on indices that emphasize increasing exposure (Wood et al. 2012) or relative rankings (Cutter et al. 2003). Our results demonstrate that the neighboring communities of Redwood City, Foster City, and San Mateo in San Mateo County consistently have some of the highest numbers of residents and employees in hazard zones across all scenarios (Figs. 1a and 2). Therefore, collaborations may already exist among certain planners in these communities or may be relatively easy to encourage given the shared jurisdictional boundaries, infrastructure, and populations that move between adjacent communities for work or recreation. However, if one looks at community exposure among the other 60 jurisdictions, clusters involving communities that do not share jurisdictional boundaries emerge. For example, the communities of Belvedere, Corte Madera, and San Rafael in Marin County, located in northern San Francisco Bay, also exhibit similar exposure to each other (group 3 in Figs. 1a and 2b). A network among these three communities to share best practices in adaptation planning may be beneficial and could be strengthened by including communities elsewhere, such as Union City, Palo Alto, and San Francisco (Figs. 1a and 2b).
Another advantage of cluster analyses over single-score indices is the ability to differentiate and communicate how all communities are specifically vulnerable to a given hazard (Sharma and Patwardhan 2008). This information can be used to create networks organized around specific issues of concern and shared adaptation opportunities. In our case study, this is most apparent with networking opportunities associated with the varying demographics of affected populations. For example, Spanish language outreach materials could be developed and shared among communities with higher representation of individuals in hazard zones that identify as Hispanic or Latino, such as San Rafael, Sunnyvale, Mountain View, East Palo Alto, and Menlo Park (group 5 in Fig. 1c). Common materials could also be shared among communities with higher representation of renters in hazard zones (e.g., Mill Valley, Petaluma, Martinez, Oakland, San Jose, Pacifica, and San Francisco), since renters are considered to be less likely than homeowners to be prepared for extreme events because of high turnover rates, fewer resources, less access to information, and lack of incentive to invest in mitigation (Burby et al. 2003). Similar targeted interventions related to group quarters, age, or other demographic attributes could be developed to help those considered to be more vulnerable. Another benefit of using clusters instead of single-score indices to characterize demographic sensitivity is that it reduces issues of inappropriate compensatory logic that have been noted with additive indices (Jones and Andrey 2007; Holand and Lujala 2013; Fekete 2012; Tate 2013).
Results of this case study also reveal how cluster composition is not static and instead shifts based on the variables being considered. This observation demonstrates the importance of developing independent cluster models for various community characteristics (e.g., population counts, demographic attributes, business types, and infrastructure in this study) instead of the more common approach of developing a single cluster typology for a given set of variables and locations (e.g., Kok et al. 2016; Stafford and Abramowitz 2017). The implication for network planning is that a community may want or need to participate in multiple issue-specific collaborative networks for sharing best practices instead of partnering just with one other community. For example, San Rafael and Corte Madera in Marin County are clustered together in group 3 for residential and employee exposure (Fig. 1a) and could collaborate on strategies to protect lives and livelihoods. However, if adaptation planning shifted to the type of residents in hazard zones to address potential equity issues, then these two cities have less in common, since San Rafael has a relatively high number of renter-occupied households and residents that identify as Hispanic or Latino in hazards zones, whereas Corte Madera has a relatively high percentage of people over 65 years in age. Instead, San Rafael may have more in common with Mountain View and East Palo Alto, whereas Corte Madera may want to work with Palo Alto and Belvedere (Fig. 1c). If discussions then shifted to business-sector resilience, San Rafael and Corte Madera once again have similar workforce exposure profiles (Fig. 1d) and may wish to include Palo Alto and San Mateo in their planning efforts. This example demonstrates the importance of taking a more nuanced approach to evaluating similarities in exposure between communities based on specific variables or domains of interest, rather than simply clustering or ranking at an aggregate level using all socioeconomic variables.
In addition to shifts in cluster compositions based on considered variables, results also demonstrate that cluster compositions shift across hazard scenarios. For example, at 50 cm of SLR (Fig. 2a), most communities are in group 1 for population exposure (noting low exposure) with only a few individual cities at higher levels of exposure, such as Foster City (group 4), neighboring Redwood City (group 3), and San Mateo, Palo Alto, and San Rafael (group 2). As such, relatively few communities would be identified for developing a collaborative network. However, if 150 cm of SLR is assumed, then population exposure increases in more communities (Fig. 2b), and potential networks could be expanded to share best practices (e.g., land use ordinances, zoning regulations, capital improvement planning).
Critical infrastructure and facility results provide another example of how networks expand with increasing SLR. Assuming 50 cm of SLR and no storm in flood-hazard zones, unincorporated areas of Sonoma County and San Mateo are in their own groups (Fig. 3a) due to relatively high amounts of roads and rails (Sonoma County) or critical facilities (San Mateo) in hazard zones. If SLR assumptions are increased to 150 cm, Sonoma County is joined by San Mateo and Redwood City in group 4, with the highest exposure, and 11 other communities join group 3 due to high amounts of roads in hazard zones (Fig. 3b). The expansion of group size with increasing SLR scenarios for infrastructure suggests that some collaborations may focus on short-term mitigation efforts to protect existing infrastructure or facilities (e.g., groups assuming 50 cm of SLR) whereas others may focus on long-term infrastructure investment planning (e.g., groups assuming 150 cm of SLR). At the same time, recognition of the expansion of clusters may provide an incentive for communities that join high-exposure groups under higher SLR scenarios to learn from communities that face similar risks in the short-term and to participate in the development of adaptation or risk-management strategies now, with the knowledge that early investment in these collaborations may enhance future planning and mitigation efforts.
An advantage of providing cluster compositions that reflect various hazard scenarios is that it allows for variations in risk tolerance by practitioners and policymakers. As discussed earlier, we did not assign specific SLR magnitudes to specific time periods (e.g., 50 cm of SLR by 2050) given the wide range of projections in the region (Our Coast Our Future 2017). For example, the scientific literature indicates that the region may experience 50 cm of SLR by 2050 (National Research Council 2012), 2060 (Vermeer and Rahmstorf 2009), 2070 (Cayan et al. 2008), or 2090 (Intergovernmental Panel on Climate Change 2007). Although research will continue to better constrain the rate of SLR in the study region and elsewhere, practitioners and policymakers currently engaged in adaptation planning will need to make assumptions of potential SLR now for planning purposes. These decisions could be based on their general perceptions of or risk tolerance to climate change or could vary based on specific adaptation issues. For example, the potential consequences and adaptation options to protect lives, livelihoods, or infrastructure from climate-driven hazards may vary in a community, resulting in varying levels of acceptable risk and subsequently different SLR assumptions in adaptation plans for each of these topics. Multiple cluster compositions based on different hazard scenarios and varying domains (infrastructure and populations) provide adaptation planners with the ability to develop and participate in collaborative networks that better reflect their risk tolerances and vulnerability issues of interest.
Areas for future research
Results from this cluster analysis of community exposure to flood hazards are not a final statement on potential losses. Results are meant to help initiate and inform discussions between researchers, community planners, elected officials, and members of affected communities to assist in decision making and policy formulation and to identify additional research avenues. One area for future research is to take into account potential changes in the magnitude and distribution of populations, businesses, and infrastructure over time. For example, one study suggests that over three times more people in the USA might be affected by SLR-driven inundation by 2100 when considering population growth trends, including a five-fold increase in California (Hauer et al. 2016). Results presented here based on current distributions provide insight on potential community vulnerabilities but likely have greater relevance for shorter-term decision-making, such as planning for the next several decades as opposed to potential impacts 50 years from now. Future work could integrate advances in land-change modeling to develop projections of future development, populations, and infrastructure in hazards zones (e.g., Sleeter et al. 2017). Downscaling of shared socioeconomic pathways, which describe potential future trends in human development for integration with emissions projections, could also be useful if high enough resolutions are attained (O’Neill et al. 2014).
Another area for future research is more in-depth characterization of the business community and potential impacts to local economies and quality of life. Business categories used in the current analysis are broad (e.g., two-digit NAICS codes) and do not provide a complete picture of potential impacts. For example, a community may overall have a low percentage of its service sector in the hazard zone; however, the community impact may still be large and adaptation planning may be more challenging if the few service businesses that are in the hazard zone are the only grocery stores or gas stations in a community, as opposed to other more general services. However, because this analysis is concerned primarily with regional patterns of business vulnerability, the discussion focuses more on general implications for local economies (e.g., potential unemployment) and not impacts to community quality of life. Future work could expand the analysis of business exposure to include aspects of community quality of life and livelihoods along with more traditional aspects of how community-scale economies would be impacted.
Cluster analysis provided a data-driven mechanism for identifying community groupings with similar hazard exposure, given multiple hazard scenario combinations and societal assets of interest.
Cluster compositions changed based on the selected societal variables, demonstrating the importance of developing independent cluster models for various community elements instead of a single typology based on all data.
Cluster compositions changed based on selected SLR scenarios, which could allow for managers to develop or seek out adaptation networks that match their risk tolerance or planning horizons.
Shifts in cluster compositions due to selected societal variables and SLR scenarios suggest that a community could participate in multiple networks to share best practices and leverage resources to target specific issues or policy interventions.
These conclusions demonstrate that cluster analysis provides an effective tool for understanding community exposure that transcends composite rankings and considers instead the specific variables that characterize vulnerability. Clustering is also useful for understanding how flood exposure changes over time as forcing (in this case, due to SLR) intensifies. While this study focused on a regional level analysis of flood hazards in the San Francisco Bay area, a similar approach could be applied over a range of scales for a variety of natural hazards. The information obtained from clustering is crucial for regional efforts that aim to facilitate adaptation planning and prioritize investments to mitigate the impacts of natural hazards in the face of uncertain future conditions.
The authors would like to thank Sandrine Dudoit for early discussions about clustering stability. Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the US Government.
This study is part of the Resilient Infrastructure as Seas Rise (RISeR) project, supported by the National Science Foundation Critical Resilient Interdependent Infrastructure Systems and Processes (CRISP) Award 1541181.
US Geological Survey (USGS)-affiliated authors are supported by the USGS Land Change Science Program and the USGS Coastal and Marine Geology Program.
- 100 Resilient Cities Program (2017) The Rockefeller Foundation. http://www.100resilientcities.org/#/-_/. Accessed 24 January 2017
- Abdollahian N, Jones J, Wood N (2016) Community exposure to potential climate-driven changes to coastal-inundation hazards for six communities in Essex County, Massachusetts. U.S. Geological Survey Open-File Report 2016–1187, 87 p. https://doi.org/10.3133/ofr20161187
- Alliance of Regional Collaboratives for Climate Adaptation (ARCCA) (2017) http://www.arccacalifornia.org/about/. Accessed 25 January 2017
- Barnard PL, O'Reilly B, van Ormondt M, Elias E, Ruggiero P, Erikson LH, Hapke C, Collins BD, Guza RT, Adams PN, Thomas JT (2009) The framework of a coastal hazards model; a tool for predicting the impact of severe storms. U.S. Geological Survey Open-File Report 2009–1073, 21 p. http://pubs.usgs.gov/of/2009/1073/
- Barnard PL, van Ormondt M, Erikson LH, Eshleman J, Hapke C, Ruggiero P, Adams PN, Foxgrover AC (2014) Development of the Coastal Storm Modeling System (CoSMoS) for predicting the impact of storms on high-energy, active-margin coasts. Nat Hazards 74(2):1095–1125. https://doi.org/10.1007/s11069-014-1236-y CrossRefGoogle Scholar
- Bureau of Labor Statistics (2015) State and metro area employment, hours, and earnings. U.S. Department of Labor website. https://www.bls.gov/sae/790over.htm. Accessed 27 January 2017
- C40 Cities (2017) http://www.c40.org/. Accessed 24 January 2017
- California Energy Commission (2012) Identifying and overcoming barriers to climate change adaptation in San Francisco Bay: Results from case studies. Report CEC-500-2012-034. http://www.energy.ca.gov/2012publications/CEC-500-2012-034/CEC-500-2012-034.pdf
- Hennig C (2016) Package: fpc. CRAN. R Repository. https://cran.r-project.org/web/packages/fpc/fpc.pdf. Accessed 8 October 2016
- Hinkel J, Nicholls RJ, Tol RSJ, Wang ZB, Hamilton JM, Boot G, Vafeidis AT, McFadden L, Ganopolski A, Klein RJT (2013) A global analysis of erosion of sandy beaches and sea-level rise: an application of DIVA. Glob Planet Chang 111:150–158. https://doi.org/10.1016/j.gloplacha.2013.09.002 CrossRefGoogle Scholar
- Infogroup (2012) Employer database: Infogroup online database. http://www.infogroup.com/ Accessed 8 October 2016
- Intergovernmental Panel on Climate Change (2007) Synthesis report—contribution of Working Groups I, II and III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change [Core Writing Team, Pachauri, R.K and Reisinger, A. (eds.)]. IPCC, Geneva, Switzerland, 104 p. https://www.ipcc.ch/pdf/assessment-report/ar4/syr/ar4_syr_full_report.pdf
- Jones JM, Wood N, Ng P, Henry K, Jones JL, Peters J, Jamieson M (2016) Community exposure in California to coastal flooding hazards enhanced by climate change, reference year 2010: U.S. Geological Survey data release. doi: https://doi.org/10.5066/F7PZ56ZD
- Lam N, Reams M, Li K, Li C, Mata L (2016) Measuring community resilience to coastal hazards along the Northern Gulf of Mexico. Nat Hazard Rev 17(1): 12 p. doi: https://doi.org/10.1061/(ASCE)NH.1527-6996.0000193, 04015013CrossRefGoogle Scholar
- Maechler M, Rousseeuw P, Struyf A, Hubert M, Hornik K, Studer M, Roudier P, Gonzalez J (2016) Package: cluster. CRAN. R Repository. https://cran.r-project.org/web/packages/cluster/cluster.pdf. Accessed 8 October 2016
- National Research Council (2012) Sea-level rise for the coasts of California, Oregon, and Washington: past, present, and future. Committee on Sea Level Rise in California, Oregon, and Washington. The National Academies Press, Washington. doi: 10.17226/13389
- Our Coast Our Future (2017) http://data.pointblue.org/apps/ocof/cms/. Accessed 27 January 2017
- Strauss BH, Ziemlinski R, Weiss JL, Overpeck JT (2012) Tidally adjusted estimates of topographic vulnerability to sea level rise and flooding for the contiguous United States. Environ Res Lett 7(1):12p. https://doi.org/10.1088/1748-9326/7/1/014033
- Sweet W, Park J, Marra J, Zervas C, Gill S (2014) Sea level rise and nuisance flood frequency changes around the United States NOAA Technical Report NOS CO-OPS 073, 58 p. https://tidesandcurrents.noaa.gov/publications/NOAA_Technical_Report_NOS_COOPS_073.pdf
- Tebaldi C, Strauss BH, Zervas CE (2012) Modelling sea level rise impacts on storm surges along U.S. coasts. Environ Res Lett 7: 11p. doi: https://doi.org/10.1088/1748-9326/7/1/014032
- The Nature Conservancy (2014) California coastal resilience network technical report. Conservation Gateway. https://www.conservationgateway.org/ConservationPractices/Marine/crr/library/Documents/2014%20TNC%20California%20Coastal%20Resilience%20Network%20Technical%20Report.pdf
- United States Census Bureau (2010) TIGER products. Geography, Maps & Data. http://www.census.gov/geo/maps-data/data/tiger.html. Accessed 27 January 2017
- United States Geological Survey (2015) CoSMoS 2.1: San Francisco Bay, Pacific Coastal and Marine Science Center. https://walrus.wr.usgs.gov/coastal_processes/cosmos/sfbay/index.html. Accessed 10 November 2016
- United States Geological Survey (2017) Hazard Exposure Reporting and Analytics (HERA) www.usgs.gov/apps/hera. Accessed 26 January 2017
- Wood N, Ratliff J, and Peters J (2012) Community exposure to tsunami hazards in California. U.S. Geological Survey Scientific Investigations Report 2012–5222, 49 p. http://pubs.usgs.gov/sir/2012/5222/