Future flooding increases unequal exposure risks to relic industrial pollution

Thomas Marlow; James R Elliott; Scott Frickel

doi:10.1088/1748-9326/ac78f7

1. Introduction

This study investigates whether and where climate change is likely to increase residential exposure to hazardous chemicals through the flooding of former, or 'relic', industrial sites. These are sites where toxic contaminants from prior industrial uses are likely to be buried in the soil or suspended in the groundwater, posing significant health risks to humans and other living organisms (Sengul et al 2012, Muller et al 2018, Wodtke et al 2020). Climate change can spread and magnify exposure to these land-based contaminants by increasing the frequency and intensity of local flooding (Masson-Delmotte et al 2021, ch 11) that mobilizes those contaminants on-site and spreads them to nearby parcels (Ponting et al 2021). These dynamics are especially pertinent in socially marginalized urban communities, where sites of hazardous industrial production have historically concentrated (Salazar et al 2019) and now face increasing flood risks (Buchanan et al 2020, EPA 2021, Tate et al 2021, Wing et al 2022). The result, if confirmed, would be a new kind of climate injustice emerging at the conjunction of past pollution, ongoing social inequities, and future flooding, calling for a new kind of policy thinking and action.

One reason these developments have received little attention to date is that environmental justice research typically measures residential proximity only to known hazards such as waste disposal sites and industrial facilities documented in government databases (Mohai et al 2009). While research has used these databases to document many inequitable risks and exposures (Taylor 2014, Muller et al 2018, Banzhaf et al 2019), their 1980s starting-point renders associated analyses historically incomplete and thus prone to bias (Downey 2003). One study utilizing more comprehensive manufacturing site data back to the 1950s finds that contemporary government databases miss more than 90% of relic and active sites of hazardous industry operating since World War II (Frickel and Elliott 2018; see also Marlow et al 2020).

Meanwhile, cities where these sites disproportionately concentrate remain highly segregated owing to a number of cross-cutting historical forces that include Jim Crow laws, discriminatory redlining practices in real estate markets, and urban renewal. A recent report from the University of California-Berkeley indicates that of the 113 U.S. cities with populations over 200 000, only two are racially integrated (Menendian et al 2021)⁴ . And, most have grown more rather than less racially segregated since 1990⁵ . Similar trends are also occurring along economic lines (Reardon and Bischoff 2011)⁶ . As a result, past social inequities linger on in place, threatening to intertwine with rising urban flood risk that the National Academies of Sciences identifies as one of America's leading climate challenges (National Academies of Sciences, Engineering and Medicine 2019; see also Masson-Delmotte et al 2021; see Mallakpour and Villarini 2015).

To scope the risks presented by the intersection of urban flooding, residential segregation, and past industrial activities, our analyses combine for the first time address-level data on relic sites of industrial production operating in sectors known to release toxic contaminants to on-site lands with address-level data on future flood risks. We integrate these data for six U.S. urban areas selected to maximize variation in local industrial, social, and flood contexts. The aim is to determine if general patterns emerge, not only in the geographic extent of probable exposures but also in their inequitable distribution across local communities of varying racial, economic, and housing status in ways indicative of a new era and type of environmental injustice.

2. Data and methods

The six urban areas we analyze include: the Providence metropolitan area of Rhode Island; Harris County, Texas, which contains most of the city of Houston; Philadelphia County, Pennsylvania; Multnomah County, Oregon, which contains most of the city of Portland; Orleans Parish, Louisiana, which contains most of the City of New Orleans; and Hennepin County, Minnesota, which contains most of the City of Minneapolis. Like most U.S. urban centers, these six areas have experienced significant industrial development and documented site-based legacy contamination (US EPA, O 2015). Located near coasts or along major rivers, they also illustrate a range of local flood scenarios and patterns of residential segregation.

2.1. Relic industrial site data

We collected address-level data on relic industrial sites in sectors that regularly report on-site hazardous waste releases. We used state manufacturing directories⁷ covering every 2–5 years from 1953 to 2010, excluding facilities still in operation after 2010 to spotlight relic sites that are no longer visibly active (Frickel and Elliott 2018, Berenbaum et al 2019). We exclude addresses containing P.O. boxes to avoid non-industrial headquarter sites. One limitation of focusing on relic sites detected from manufacturing directories is that we do not know the true presence or significance of contamination on respective sites. Furthermore, there may be heterogeneity across cities in toxic risk due to differences in the types of hazardous chemicals left behind and how they were deposited into on-site lands. To minimize this uncertainty, we select only relic sites in industrial sectors identified as having a high probability of historical onsite contamination known harmful to humans (Noonan and Vidich 1992). Those sectors include chemicals and allied products, petroleum and coal products, rubber and plastics products, stone/clay/glass and concrete products, primary metal production, fabricated metal production, machinery and computer products, and transportation equipment manufacturing. In Noonan and Viditch's study, the probability of site contamination in these sectors ranges from 80% to 95%. In Providence only, we also include jewelry manufacturing, which dominated local industry during much of the 20th century, utilizing heavy metals as well as polyvinyl chloride and other organics. The result is an address-level database of 15 419 unique sites where hazardous industries operated but are no longer active across the six study areas.

2.2. Flood risk projections data

To estimate the future flood risk of relic industrial properties, we use site-specific flood risk projection data from the First Street Foundation (FSF). To produce these projections, FSF uses a combination of 21 global climate models under the middle Intergovernmental Panel on Climate Change (IPCC) RCP 4.5 emissions scenario while considering a range of high (75th percentile) and low (25th percentile) scenarios. They also incorporate a baseline climate period of 1980–2010 and incorporate data on historical flood events (First Street Foundation 2020b). FSF flood modeling methodology was independently reviewed and has since been used by the U.S. Environmental Protection Agency (EPA) in their 2021 report on evaluation of social vulnerability during climate change (EPA 2021). Compared to Federal Emergency Management Agency (FEMA) estimates of property flood risk, the FSF identifies many more properties in the United States at risk (First Street Foundation 2020a).

Using this methodology, FSF calculates a Flood Factor (FF) for each property. The FF is an integer between 1 and 10 that captures the likelihood and severity of flooding by 2050, with 1 representing minimal risk and 10 representing extreme risk (First Street Foundation no date). The number increases as the probability increases or as the depth increases, or both. We identify properties with a value of three or higher as being at an elevated risk of future flooding (i.e. what FSF characterizes as 'moderate risk' or greater). At this threshold, a property at 'moderate risk' (FF ⩾ 3) has at least a 6%–12% chance of flooding by 2050, or a good chance of flooding to a depth of 6–9 inches by 2050.

2.3. Residential socio-demographic data

To measure social inequities in exposure, we use demographic data from the U.S. Census Bureau (US Census Bureau no date). Block counts of population and households as well as land area come from the 2010 decennial census. Block-group racial, economic, and housing compositions come from the American Community Survey 2015–2019 five-year estimates. With those data, we follow the Center for Disease Control and Prevention's (CDCs) Agency for Toxic Substances and Disease Registry to construct several multi-item measures of social vulnerability (Center for Disease Control 2021). The CDC developed their measure of social vulnerability to incorporate multiple neighborhood characteristics, grouped into key vulnerability themes (see table 1). The first theme captures the socioeconomic status (SES) of a place. The second focuses on racial composition, and the third, housing conditions. Table 1 displays the census variables included in each multi-item theme, or measure.

Table 1. Social vulnerability themes and associated census variables at the level of block groups.

Social vulnerability	Theme 1: Socioeconomic status (SES)	Theme 2: Minority status	Theme 3: Housing status
Census variables	% below the poverty line % unemployed % with H.S. degree or less Median household income	% non-white % speak English 'Less than well'	% housing units in multi-unit structures % housing units that are mobile homes % housing units that are crowded % households without a car % group quarters

Following the CDC's approach, each variable is first converted to a percentile rank relative to other block groups in the same city. Where necessary, ranks are inverted by subtracting their values from one so that larger values always indicate higher vulnerability (i.e. a higher SES vulnerability score indicates lower SES; and, a higher housing vulnerability score indicates lower housing quality). Next, the percentile ranks of variables associated with each theme are summed, and that total is given its own percentile rank. The effect of this additive-ranking approach is that each variable for a given theme is given the same weight, and the final theme score then ranges from 0 to 1, with values closer to 1 indicating greater social vulnerability relative to other block groups in the same city. Unlike the CDC, we keep the theme scores distinct to evaluate their unique association with elevated flood risk on relic sites. Consistent with Marino and Faas (2020), we caution against interpreting these scores as essential, immutable traits of the communities they describe. Instead, we conceptualize these metrics as indicating where historical inequities have produced contested sites of struggle for contrasting articulations of risk and visions of local futures. In this way, the scores approximate the status and resources of residents to be engaged now in planning for future flooding; they are not a projection of who will be present in those areas when future flooding occurs.

2.4. Methods

For each study area, we perform a series of analyses. We begin by merging the binary, address-level indicator of elevated flood risk (yes/no) to each address-level relic industrial site to reveal the number and proportion of such sites at elevated flood risk over coming years. Next, we use kernel-density maps to show where those sites cluster spatially. Then, we identify the number of individuals and households at different scales of nearby exposure risk. We conclude with a series of regression analyses to explore the relationship between social vulnerability and the probability of a block group having at least one relic industrial site at elevated flood risk. Because diagnostics reveal spatial dependence in that probability (see Anselin and Li 2019), all models are estimated using an intrinsic conditionally autoregressive process (Besag et al 1991) based on a first-order queen contiguity weights matrix. This specification is a common way of modeling spatial dependence via a spatially structured random effect.

Our statistical analysis then proceeds in three steps. First, we fit a pooled spatial multilevel model to evaluate the average relationship between the probability of a block group containing a flood-prone relic site and respective vulnerability indicators (see Dong and Harris 2015). In addition to the spatial random effect, our spatial multi-level model accounts for city-level covariance as an independent and identically distributed random effect. We then use this model to predict the probability of a block group containing a flood-prone relic site over the full range (0–1) of each measure of social vulnerability. In a second analysis, we fit each study area separately (still controlling for spatial dependence) to evaluate the heterogeneity in our effects across cities (see figure 4(b)). Finally, to assess the cumulative effect of vulnerability across multiple themes, we estimate the probability of block groups containing a flood-prone relic site under two scenarios. In the first scenario, all three themes are set to the 90th percentile (i.e. the most vulnerable populations). In the second scenario, all three themes are set to the 10th percentile (i.e. the most privileged populations). All models are estimated using Integrated Nested Laplace Approximation (Rue et al 2009, 2017) via the R programming language (R Core Team 2020).

3. Results

3.1. Relic sites of hazardous industry at elevated flood risk

Analysis starts with an assessment of relic sites with an elevated probability of flooding by 2050 (figure 1). In total, we identify 6636 such sites across our six study cities. While the actual pollution on each site is unknown, the sheer number extends beyond known contamination lists such as the EPAs National Priority List. In Houston's Harris County, for example, there are 15 state and federally recognized superfund sites identified by the Government Accountability Office as being at risk of future flooding (GAO 2019). By comparison, results in figure 1 indicate that there are 1985 relic industrial sites at elevated flood risk. Even Minneapolis's Hennepin County, with the lowest count of flood-prone relic sites in our study, has 436. In addition, figure 1 indicates the proportion of relic sites at risk of projected flooding. Here we see that all of the 733 relic sites in New Orleans are at risk, compared with a low of 19% of relic sites in Minneapolis. Supplemental information figures 1 and 3 describe the presence of flood-prone relics in cities using higher probability and impact FF thresholds. Even with these more extreme cutoffs, numerous sites remain at risk, demonstrating that the results are not an artifact of low probability flooding.

Figure 1. Refer to the following caption and surrounding text. — **Figure 1.** Panel (A) represents the percentage of relic industrial sites with a flood factor (FF) score greater than or equal to 3; panel (B) is the raw count of relic sites at elevated flood risk.
Download figure:
Standard image High-resolution image

3.2. Zones of elevated exposure risks

As analysis extends to include not just well-documented superfund sites but also the multitude of less-known formerly industrialized sites, understanding of exposure risks changes. Figure 2 illustrates this point using kernel-density maps to show where flood-prone relic sites cluster in each study area. Results indicate that, despite different local histories and hydrological contexts, each city has clearly identifiable zones where large numbers of relic sites face elevated future flood risk from climate change. Moreover and regardless of the city, these zones are consistently located near downtowns where industry once thrived and low-income communities historically settled and established lasting attachments to place.

Figure 2. Refer to the following caption and surrounding text. — **Figure 2.** Kernel density plots of flood prone relic industrial sites in six cities. Colors represent the density percentile relative to the city⁸ .
Download figure:
Standard image High-resolution image

3.3. Population and housing exposures

In considering human exposures to relic sites of elevated flood risk depicted in figure 2, there is no correct way to count the number of residents and housing units potentially impacted. Different contaminants respond differently to flooding, and local hydrology influences where respective contaminants will spread in varying concentrations (see Martin et al 2021). In addition, most residents are not confined to their homes; instead, they walk about their neighborhoods and sometimes, when necessary, wade through local flood waters. As a result, there are multiple, nested scales of exposure risk. The most spatially proximate is at the block level, where a resident's home and one or more relic sites both face elevated flood risk. At this scale, flood waters can create a direct, hydrological plain that transports land-based contaminants from a flooded relic site to the flooded home nearby. Scaling outward on the same block, there may also be homes that do not face elevated flood risk. For them, flooding of a relic site on the same block may not bring contaminants directly to their home, but daily activities on the block could still result in direct exposure as well as indirect exposure to underground plumes that shift and expand as local flood waters saturate local soils. Extending outward, we can also count homes in the surrounding block group, a unit of census geography that can include as few as two to three blocks in densely settled urban areas. Those homes can then be subdivided by whether they also face elevated flood risk (yes or no).

To count the number of residents and housing units at each of these four, nested scales of exposure risk, we use the same FF cutoff score (⩾3) for residential addresses as for relic industrial sites. Figure 3 shows that even at the smallest geographic scale (flood-prone homes on blocks with flood-prone relic sites), the exposure risks are notable. At the low end, Minneapolis has approximately 2000 residents living in 1000 homes. At the high end, Houston has approximately 35 000 residents living in 12 000 homes. As analysis scales out to include all homes and residents in block groups with one or more flood-prone relic sites, these numbers increase noticeably. In Minneapolis, they rise to approximately 211 000 residents living in 100 000 homes; and in Houston, they rise to approximately 854 000 residents living in 326 000 homes. Even in New Orleans, with the smallest residential exposure counts in our study, approximately 75 000 people (or 22% of city's current population) live in a block group where a relic industrial site faces elevated flood risk from climate change.

Figure 3. Refer to the following caption and surrounding text. — **Figure 3.** Properties and populations residing in proximity to flood prone relic industrial sites at different spatial scales.
Download figure:
Standard image High-resolution image

3.4. Social inequities in exposure risk

To assess social inequities in exposure risks, we use the spatial modeling procedures described in the section 2.4. The unit of analysis is the block group. The dependent variable is the probability of the block group containing one or more relic sites of elevated flood risk (FF ⩾ 3). With this approach we are estimating exposure risk at the most inclusive geographic scale displayed in figure 3. Our predictors of interest are the three measures, or themes, of social vulnerability described in table 1. For all three measures, higher values indicate higher vulnerability. To start, we enter each measure separately, along with controls for the total population, number of households, and land area in the block group. Then, we enter the three measures simultaneously with the same controls.

Panel (A) of figure 4 graphs the results for each vulnerability measure in our model for all six study areas pooled together, net of controls and accounting for local spatial dependence among block groups. These results reveal a common pattern: all else equal and regardless of the indicator, the more socially vulnerable the block group, the higher its probability of having one or more relic sites at elevated risk of future flooding.

To illuminate city-level variation, Panel (B) displays coefficient plots for respective vulnerability indicators for each city modeled separately. Those appear beneath the estimated coefficient (in red) for the pooled results graphed in Panel (A). Here, each dot represents the estimated log-odds associated with a given vulnerability theme in a given city, net of other themes; the horizontal line indicates the 95% credibility interval. Points above zero indicate that higher vulnerability scores correlate positively with higher probabilities of having a flood-prone relic site in the block group. Overall, these results indicate notable socioeconomic and racial variation across cities along with consistently positive coefficients for housing vulnerability. Analyses using cutoffs for middle-range risk (FF ⩾ 3 & ⩽6), high risk (FF > 6), and the presence of five or more flood prone relics indicate substantively similar results to those reported here (presented in supplementary information figure 2).

However, socioeconomic, racial, and housing vulnerabilities rarely exist independently in urban neighborhoods; instead, they tend to overlap. To help visualize this cumulative overlap, figure 5 graphs estimated probabilities when all three indicators of social vulnerability are assigned to their 10th versus 90th percentiles in each city. Here simulated results show particularly high exposure probabilities for the most vulnerable block groups in New Orleans and Providence. They also reveal the range of exposure inequity within each city.

Figure 4. Refer to the following caption and surrounding text. — **Figure 4.** Regression models estimating the probability that a block group will contain will contain a flood prone relic site while varying values of each social vulnerability theme. Panel (A) displays the average effect in the spatial multi-level model as predicted probabilities across the range of possible values for the theme (0–1). All other covariates are held at their median values and the random effects (spatial and city) are set to zero. High values of a theme indicate greater social vulnerability. Panel (B) displays the estimated effect sizes in log odds for each social vulnerability theme. The top line is the spatial multilevel model and others correspond to separate city models. Values greater than zero indicate a positive correlation between increasing social vulnerability and the probability of a block-group containing a flood prone relic site.
Download figure:
Standard image High-resolution image

Figure 5. Refer to the following caption and surrounding text. — **Figure 5.** To better understand the cumulative effects of the social vulnerability themes, we predict the probability of a block group containing a flood prone relic site under different extreme scenarios. Results in the left column predict the probability given that a place is in the 90th percentile in all three dimensions of vulnerability (i.e. the most socially vulnerable places). The right column shows the probability given that a place is in the bottom decile of vulnerability (i.e. the least vulnerable places).
Download figure:
Standard image High-resolution image

4. Discussion

Climate change is raising the probability that urban flooding will remobilize and spread hazardous wastes contaminating the soils of potentially thousands of former industrial sites, increasing exposure risks for urban residents generally—a finding suggestive of a new kind of climate injustice for more socially vulnerable communities. Researchers and policymakers alike are beginning to register concern over these developments as witnessed by recent discussions over the Biden Administration's American Jobs Plan, which proposed $5B to remediate and redevelop 'idle' industrial sites and attend to legacy pollution confronted by low-income communities of color (The White House 2021: 10). Yet, despite growing attention, the scale and scope of the problem and thus potential solutions remain poorly specified, constrained by government databases that overlook the vast majority of relic industrial sites, many of them likely to contain legacy contaminants.

To overcome these limitations, we paired historically comprehensive industrial site data with address-level projections of future flood risk. Results reveal entire zones of flood-prone relic industrial sites, with thousands to hundreds of thousands of residents locally exposed (depending on the scale of analysis). Our analyses further reveal that these at-risk areas tend to be disproportionately inhabited by more socially vulnerable groups (i.e. racial minorities, those with lower-incomes, and/or those living in lower quality housing). The theoretical contributions of these findings are to explicate and document an emergent type of environmental inequality created when past industrialization, ongoing social marginalization, and future flood risk intersect. Specifically, such inequality emerges when legacy contaminants on flood-prone relic industrial sites are remobilized into surrounding residential and public spaces.

A limitation of the present study, however, is that we do not know the extent of contaminants on each relic site, nor do we know how those contaminants will remobilize and spread during future flood events. Yet, the sheer number of flood-prone relic sites does suggest that the cumulative burden will be high, regardless, even if not every site is equally or heavily polluted. It also calls for us to rethink conventional approaches to environmental assessment and intervention.

The broader implication is that current theory, thinking, and action on environmental risk and injustice must catch up to the new era in which we are entering—one where the hidden legacies of industrial pollution and social inequities collide with rising flood risks. Future flooding threatens to inundate already marginalized communities with toxin-laced waters through no fault of their own (GAO 2022). This is not to say that the same people living in these communities today will be the ones ultimately exposed. Rather, it is to acknowledge that efforts to address these compounding risks must include and engage these communities now as we plan for future eventualities (Chan et al 2021). In doing so, we must also recognize that current socio-demographics do not make some urban dwellers inherently more vulnerable than others. Instead, we should understand vulnerable neighborhoods as contested sites of struggle, places that have been historically marginalized and where residents' understandings and experiences of risk and mitigation strategies often continue to be ignored. Such community-centered planning, as well as a parallel focus on the broader processes of industrial pollution, social marginalization, and rising urban flood risks, are relevant outside of the U.S. context, too. Thus, our findings could prove useful for climate justice scholars whose research focuses on urban areas worldwide because of the historical location of industry near coasts and waterways (Harlan et al 2015, Shi et al 2016, Chu and Michael 2019).

5. Conclusion

The goal of this study was to illuminate the scale and scope of risk that U.S. cities face from future flooding of relic industrial sites as well as the socially inequitable distribution of that risk across different types of communities. In our six sample cities, we identified more than 6000 relic sites at an elevated risk of flooding by 2050, exposing tens of thousands of people and homes to this under-studied but ascendant type of environmental risk. Using spatial regression models, we also found that block groups with higher measures of social vulnerability have higher predicted probabilities of exposure to flooded relic sites. These results indicate a new kind of climate injustice emerging at the conjunction of past pollution, ongoing social inequities, and future flooding, calling for a new kind of policy thinking, theory, and action.

Acknowledgments

The authors would like to thank the First Street Foundation for generous access to their data, as well as the providing useful feedback through their community workshop series.

Data availability statement

The data generated and/or analysed during the current study are not publicly available for legal/ethical reasons but are available from the corresponding author on reasonable request.

Funding

The authors have no conflicts of interest to declare. Thomas Marlow was supported by the NYUAD Center for Interacting Urban Networks (CITIES), funded by Tamkeen under the NYUAD Research Institute Award CG001. We also acknowledge support from the Superfund Research Program of the NIEHS, Grants P42ES013660 and 3P42ES013660-14S1 and National Science Foundation collaborative research Grants #0849823 and #0849826.

Author e-mails

Author affiliations

Author notes

ORCID iDs

Dates

Peer review information