Impervious cover change as an indicator of environmental equity

Abstract

Pattern-focused environmental equity research has been underpinned by high-resolution remotely sensed data to uncover spatial relationships between environmental amenities (e.g., urban tree cover) and socio-economic status (SES). A constraint imposed by reliance on high-resolution data is the inability to examine temporal patterns, primarily because of the cost of data production and the nascent state of high-resolution land cover mapping. The lack of temporal monitoring is a clear gap in pattern-focused environmental equity research. We examined temporal (2001 – 2019) relationships between a disamenity, impervious cover (IC), and demographic attributes for the entirety of the conterminous United States. Our main finding was 2001 – 2019 increases in IC were more pervasive in minority communities but these communities were not necessarily poor, and only rarely poorly educated or non-English speaking. We supported our use of IC from moderate resolution data by comparing it to high-resolution data for 24 cities within the conterminous United States. Mean Absolute Deviation (MAD) was 4.8% overall, ranging from 2.2% to 11.3% across the 24 locations. Differences in classification objectives contributed to differences in %IC estimates between the two sources.

1. Introduction

There is an extensive literature on spatial relations between environmental amenities and socio-economic status (SES). Within the United States, there is a tendency for this research to rely on land cover data derived from high resolution aerial imagery (Table 1). The reliance on high resolution imagery in the United States is a least in part attributable to the National Agriculture Imagery Program (NAIP) (Popkin 2018), which provides digital imagery at 1- x 1-m and higher spatial resolutions (https://naip-usdaonline.hub.arcgis.com/). High resolution data are preferred because they record extant environmental amenities in greater detail than lower spatial resolution data, such as land cover from the widely utilized Landsat satellites, which resolve earth surface features at a spatial resolution of 30- x 30-m. The attraction of the greater detail realized from high resolution data becomes intuitive when considered in the context of the small geographic extent of the spatial units used to summarize SES. The smallest unit over which SES data are summarized in the United States is a Census block group (CBG) (https://www2.census.gov/geo/pdfs/reference/GARM/Ch11GARM.pdf), and its median area is about 1.4 km². A 1.4 km² CBG has 1.4 million observations (pixels) of land cover at 1- x 1-m spatial resolution but only 1,555 pixels at 30- x 30-m spatial resolution.

Table 1:

Pattern-focused environmental equity research¹

Author	Date	Location	Data	X variable	Scale	Unit	Synopsis
Venter	2023	Oslo, NO NNorway	Sntinel	Grn	100 m2	Dstrct	SES ↑ Grn ↑
Sinha	2022	10 US Cities	NLCD	Grn, IC	900 m2	CBG	Tree ↑ Mor ↓
Roodsari	2022	Tehran, IR	Vector	Grn	NA	Dstrct	SES <> Grn
Coleman	2021	3 MA cities	Survey	Grn	NA	NA	Safety <> GRN
Herreros-C	2021	New York City	NAIP	Grn	1 m2	CB	SES ↑ Grn ↑
Locke	2021	37 US cities	Various	Grn	1 m2	Nghbr	SES ↑ Grn ↑
Zhang	2021	Hong Kong	Vector	Grn	NA	TPU	SES <> Grn
Clement	2020	U.S.	NLCD	IC	900 m2	CT	SES ↓ IC ↑
de Vries	2020	NL	NDVI	Grn	625 m2	Nghbr	SES ↑ Grn ↑
Kolosna	2019	2 US cities	NAIP	Grn	1 m2	CB	SES ↑ Grn ↑
Nesbitt	2019	10 US cities	NAIP	NDVI	1 m2	CBG, CT	SES ↑ Grn ↑
de Sousa S	2018	2 EU cities	Vector	Grn	NA	Nghbr	SES <> Grn
Rigolon	2018	100 US cities	Vector	Grn	NA	City	SES ↑ Grn ↑
Xu	2018	Munich, DE	AP	Grn	0.2 m2	Nghbr	SES <> Grn
Reid	2017	New York, US	NAIP	Grn	1 m2	Nghbr	Tree ↑ SRH↓
Salvati	2017	Athens, GR	Landsat	Grn	900 m2	City	SES ↑ Grn ↑
Rigolon	2017	Denver, US	Vector	Grn	NA	GBG	SES ↑ Grn ↑
Nesbitt	2016	Portland, US	NAIP	NDVI	1 m2	CBG	SES ↑ Grn ↑
Berland	2015	Cincinnati, US	AP	Grn	< 1 m2	CBG	SES <> Grn
Mitchell	2015	3 US Cities	Landsat	LST	900 m2	CT	SES ↑ LST ↓
Li	2015	Hartford, US	AP	Grn	NA	CBG	SES ↑ Grn ↑
Schwarz	2015	7 US Cities	NAIP	Grn	1 m2	CBG	SES <> Grn
Grove	2014	New York, US	LIDAR	Grn	0.5 m2	CBG	SES <> Grn
Jesdale	2013	100+ US cities	NLCD	IC, Grn	900 m2	CBG	SES ↑ Grn ↑ IC ↓
Wen	2013	US cities	Vector	Grn	NA	CT	SES ↑ Grn ↓
Lowry	2012	Salt Lake C, US	NAIP	Grn	1 m2	GBG	SES ↑ Grn ↑
Duncan	2012	Boston, US	Vector	Grn	NA	CT	SES ↑ Grn ↑
McConnachie	2010	9 Cities, ZA	Vector	Grn	NA	Nghbr	SES ↑ Grn ↑
Boone	2009	Baltimore, US	Vector	Park	NA	CBG, CT	SES <> Grn
Landry	2009	Tampa, US	IKONOS	Grn	1 m2	CBG	SES ↑ Grn ↑
Ogneva-H	2009	MA US	AP	IC	1 m2	CBG	SES ↑ IC ↓
Reid	2009	US	NLCD	Grn	900 m2	CT	SES ↑ Grn ↓
Mitchell	2008	GB	LULC	Grn	10 m2	Local	SES ↑ Grn ↑
Lafary	2008	Evansville, US	ASTER	NDVI	15 m2	CBG	SES ↑ Grn ↑
Grove	2006	Baltimore, US	IKONOS	Grn	16 m2	CBG	SES <> Grn
Heynen	2006	Milwaukee, US	Vector	Grn	NA	CT	SES ↑ Grn ↑
Heynen	2003	60 cities, US	AVHRR	Grn	1 km2	CDP	SES <> Grn

¹ Intended to be representative of trends and characteristics, not an exhaustive literature review.

Abbreviations & symbols: socio-economic status (SES); environmental asset such as street trees, parks, and open space (Grn); Census block group (CBG); Census tract (CT); Census designated place (CDP); mortality (MOR); self-reported heath (SRH) neighborhood (Nghbr); National Agriculture Imagery Program (NAIP); National Land Cover Database (NLCD); Normalized Differenced Vegetation Index (NDVI); Sentinel-2 satellite (Sntinel); aerial photography (AP); not applicable (NA); increase (↑); decrease (↓); no or multifaceted relationship (<>); two-letter country codes (www.iso.org) – Germany (DE), European Union (EU), Great Britain (GB), Greece (GR), Iran (IR), Netherlands (NL), Norway (NO),

The advance in land cover mapping realized from the introduction of a consistent source of high-resolution imagery also imposes constraints. One constraint is an inability to incorporate change over time. None of the studies cited in Table 1 that used high resolution land cover data included a temporal dimension. Lack of high-resolution temporal land cover data is related to cost. The source imagery needed to map environmental amenities at high resolution is expensive to acquire (Popkin 2018) and process (Robinson et al. 2019). Cost-related and other constraints have limited the geographic extent over which high-resolution land cover mapping has been undertaken and moving such efforts into the temporal domain has yet to be realized.

Karanja and Kiage (2021) have advocated for the adoption of a temporal perspective for analysis of vulnerability to urban heat island effects (Oke 1982). Adoption of a temporal framework is intuitive for most if not all environmental equity issues because information on changing conditions is essential to effective natural resource management (NAS 2004; Williams 2011; Rist et al. 2013). Lower spatial resolution sources, such as that provided by Landsat, appear to be a viable alternate to higher resolution data for research focused on temporal change in the spatial distribution of environmental amenities. Some environmental amenities are not resolvable at lower spatial resolutions, but others are. The amount of impervious cover (IC), a disamenity, derived from 30- x 30-m Landsat data has been shown to be inversely related to SES (Clement and Alavarez 2020; Jesdale et al. 2013; Reid et al. 2009).

The main objective of this research is to uncover the spatial relation between 2001 – 2019 IC change and SES for the entirety of the conterminous United States. Urban sprawl (Leyk et al. 2020) will result in increases in IC nationwide, but the predominantly inverse relation between environmental amenities and SES (Table 1) indicates that large increases in IC over time will be realized predominantly in disadvantaged communities. Disadvantaged communities tend to have less environmental amenities (Wolch et al. 2014) and the trend in the research cited herein (Table 1) indicates that disparities in the spatial distribution of environmental amenities have been ongoing over the preceding 18 years.

Our focus is on IC change because of the well-established urban heat island (UHI) phenomenon (Oke 1982) and adverse health outcomes associated with heat stress (Abi Deivanayagam et al. 2023; Wilson 2020). UHI arises primarily because construction materials (e.g., asphalt, concrete) tend to have lower heat capacities and thus higher blackbody radiation than earth surface materials (e.g., rock, soil) (Memon et al. 2008), contributing to higher surface air temperatures in urban areas than surrounding non-urban areas. The potential for adverse health outcomes related to UHI was presumably a motivation for the development of the concept, heat risk related land cover (Jesdale et a. 2013), and similar research relating the magnitude of urban development to adverse health outcomes (Reid et al. 2009).

We rely on 2001 and 2019 land cover in the NLCD2019 database (Jin et al. 2023) to estimate IC change. We support the analysis of NLCD temporal %IC data by comparing it to high-resolution IC for 24 cities. The comparison is used to gauge the suitability of NLCD as a proxy for IC from higher resolution sources. The comparison was included because temporal land cover data from high-resolution sources are unlikely to be commonplace in the near term due to the costs and complexities associated with its derivation (Robinson et al. 2019; Tong et al. 2020).

2. Overview of Distributional Justice

There is a large body of pattern-focused environmental justice research that explores the relationship between environmental amenities and SES. Informative overviews of this research can be found in Clement (2023), Debats Garrison (2021), and Nesbitt et al. (2018). The intent of this overview is to provide context for our research, not to replicate the informative reviews in the available literature.

In the context of EJ theories, research on environmental amenities and SES (Table 1) falls into the category of distributional justice, which examines how environmental hazards and amenities are distributed among different communities (Rawls 1999, see also Nesbitt et al. 2018, p. 241). The other broad categorical counterpart to distributional justice is procedural justice which focuses on the decision-making processes that can lead to inequitable distributions of environmental amenities (Debats Garrison 2021). A number of potential explanations have been posed for inequitable distributions of environmental amenities. The placement of unwanted land uses in minority neighborhoods can be due to racism, both in present day decisions (Schlosberg 2013, Mohai et al. 2009) and as a result of historic factors (e.g., redlining, see Nardone 2020). Discriminatory zoning that placed undesirable land uses in minority neighborhoods can cause inequalities to persist even in contexts where racist intent is no longer a motivating factor (Mohai et al. 2009). Economic factors can also contribute to environmental inequities – land in disadvantaged communities is often more affordable, leading more development to occur there (Mohai et al. 2009). Rapid urbanization and the privatization of public space have also been identified as potential causes of environmental inequities (Zuniga-Terran 2021). Green space can also lead to gentrification by increasing property values and displacing lower-income people (Jennings et al. 2016, Vilela et al. 2019). Each of these factors, alone or in combination could contribute to higher IC in disadvantage communities.

The majority of the 37 studies summarized in Table 1 reported inequitable distributions of environmental amenities. A few found no relationship or a multifaceted relationship. Wen et al. (2013) was the only study we identified that found a negative relationship between poverty levels and greenspace – they found a negative association between SES and distance to parks, and negative association between SES and greenspace for rural areas. However, even this study found a positive relationship between greenspace and SES in urban and suburban areas. Most of the studies we identified focused on the United States, but the 9 studies from outside of the U.S. showed broadly similar patterns, with around half showing a positive relationship between SES and measures of greenness, and the remaining showing no relationship or a multifaceted relationship. Only three focused on IC, an inverse of environmental amenities such as parks and open space.

Karanja and Kiage (2021) advocated for inclusion of a temporal domain and larger geographic scales for the study of spatial patterns of urban heat vulnerability. In our view, the importance of a temporal domain and a more inclusive geographic scale apply to most if not all aspects of inequitable distributions of environmental amenities. A temporal domain is essential for adapting decision-making and policy to changing environmental conditions (NAS 2004; Williams 2011; Rist et al. 2013), and adoption of a broader geographic scale is needed to uncover potential differences in inequitable distributions of environmental amenities from one location to the next. Nesbitt et al. (2019) reported locational differences in the spatial pattern of inequity across nine U.S. cities.

3. Methods

3.1. Datasets

Four main datasets were used in the analyses: CBG, moderate resolution IC, high resolution IC, and surface temperature. CBG were from the U.S. EPA EJSCREEN project (https://www.epa.gov/ejscreen). EJSCREEN is a mapping and data exploration tool. It includes demographic attributes from the American Community Survey (https://www.census.gov/programs-surveys/acs). EJSCREEN data are updated following the U.S. Census Bureau data releases. At the time of download, EJSCREEN demographic data were from the 2014 – 2018 five-year American Community Survey (ACS) estimate applied to 2010 CBG. We also downloaded and compiled CBG for the year 2000 to support temporal analysis.

EJSCREEN population attributes included (1) minority, (2) low income, (3) lacking a high school diploma, (4) linguistically isolated, (5) over 64 years old, (6) under 5 years old, and (7) housing units built before 1960. Minority included all non-white individuals and was not further subdivided into component ethnicities. Low income was defined as less than twice the federally established poverty level of $26,500 annual income (US EPA 2022). Linguistically isolated identified persons at least 14 years old living in a household in which English was not the native language and all household members had some difficulty with English. The attribute, lacking a high school diploma, was based on all persons at least 25 years old.

Moderate resolution IC data were from the 2019 National Land Cover Database (NLCD2019) (https://www.mrlc.gov/). NLCD is a U.S. land cover monitoring program (Homer et al. 2020; Yang et al. 2018) that provides generalized land cover, IC, and tree cover canopy density. NLCD products are time integrated (e.g., land cover at time t informs land cover at time t-1) and released every 2–3 years. NLCD classifies %IC in 1% increments from 0% to 100% for each pixel. The release of NLCD2019 provides an 18-year (2001 – 2019) record of %IC change.

High-resolution land cover data were from the U.S. EPA EnviroAtlas project (Pickard 2015; https://www.epa.gov/enviroatlas). The 1 m² land cover data were based on NAIP imagery data (https://www.fsa.usda.gov/programs-and-services/aerial-photography/imagery-programs/naip-imagery/) and LIDAR (where available) with image acquisition dates ranging from 2010 to 2016 (Pilant et al. 2019). IC along with water, barren, tree, shrub, grass, agriculture, and wetland comprised the legend. We used the high-resolution land cover data for 24 cities in the conterminous United States (Figure 1), representing about 64 million U.S. residents. Overall classification accuracies for the 1 m² data ranged from 61% to 90% (Pilant et al. 2019).

Figure 1:

Locations of high-resolution land cover data

Land Surface Temperature (LST) data (https://www.usgs.gov/lst) were included to quantify the relationship between urban heat and IC. The data were downloaded from the U.S. Geological Survey (USGS) (https://earthexplorer.usgs.gov/) from the Analysis Ready Data (ARD) (Dwyer et al. 2018) Collection 1 product (https://www.usgs.gov/core-science-systems/nli/landsat/landsat-provisional-surface-temperature?qt-science_support_page_related_con=0#qt-science_support_page_related_con) for the cities for which high-resolution land cover data were available. All available data for 2013 through 2020 for June, July, and August (JJA) were used. The start year (2013) was selected to correspond to the release of the Operational Land Imager – Thermal Infrared Sensor (i.e., Landsat 8) data. Data from Landsat 8 include several improvements not found in its predecessors (e.g., Landsat 7), including 12-bit quantization and improved signal-to-noise ratios (Roy et al. 2014). All 2013 – 2020 summer (JJA) Landsat 8 LST data were organized into an image stack and the image stack was used to calculate an average LST for each pixel. The associated quality assurance bands were used to mask invalid pixel values prior to image stack averaging. The per-pixel average was used to calculate an average for each Census block group. The average represents the mean of all available summertime data across the 8-year period (Supplemental Information).

3.2. Data analysis

3.2.1. Demographic attributes, IC, IC change, Surface Temperature

We used a simple contingent probability model to relate the seven demographic attributes to IC:

$$P\left(EJ|A\right)=\frac{P(EJ\cap A)}{P\left(A\right)}$$ (1),

where EJ is a demographic attribute from EJSCREEN and A is a percentile-based grouping of IC or LST. To implement the model, IC and LST were grouped into 10% increments (deciles) and the demographic attributes were expressed as percentages and categorized into two groups – above and below a threshold. We used a threshold of 25% for the demographic attributes. The 25% threshold was consistent with previous studies (Hess 2021). Probability was calculated as the count of Census block groups that met the threshold (≥ 25%) divided by the number of Census block groups in the IC or LST decile. For example, if a metropolitan area had 10 CBG in the lowest decile (< 10% IC) and 5 of the units exceeded the threshold for a demographic attribute, P(EJ|A) would equal 50% for that IC decile. LST deciles were location-specific because of differing climates (Table 2). For NLCD %IC change we used a custom percentile grouping because of the rarity of change (< 5%, 5% ≤ x < 10%, 10% ≤ x < 20%, 20% ≤ x < 30%, 30% ≤ x < 40%, 40% ≤ x < 50%, ≥ 50%).

Table 2:

Basis for surface temperature (°C) intervals ([Max – Min]/10)

Location	Min	Max	Location	Min	Max
Austin, TX	35.65	48.61	New York, NY	18.28	42.84
Baltimore, MD	26.85	44.79	Philadelphia, PA	10.64	44.25
Birmingham, AL	30.06	45.15	Phoenix, AZ	45.56	58.43
Chicago, IL	25.65	42.69	Pittsburgh, PA	27.17	43.51
Cleveland, OH	26.73	41.51	Portland, ME	18.50	41.31
Des Moines, IA	30.29	40.42	Portland, OR	27.61	43.27
Durham, NC	29.44	41.20	Salt Lake City, UT	26.50	45.42
Green Bay, WI	25.63	37.43	St. Louis, MO	28.05	46.29
Los Angeles, CA	18.65	55.39	Sonoma, CA	24.47	44.27
Memphis, TN	30.58	44.87	Tampa, FL	24.77	44.98
Milwaukee, WI	22.19	39.78	Virginia Beach, VA	27.88	43.62
Minneapolis, MN	20.52	40.60	Washington, DC	27.02	42.74

The demographic attributes were tested for endogenous interaction even though inferential statistics and hence parameter estimates are not characteristics of the model used (Supplemental Information; Eq. S1, Tables S1 and S2). We also tested model sensitivity to threshold choice and spatial unit. Sensitivity to threshold choice was tested by changing the threshold to 20% and 30% (Figure S1). Sensitivity to spatial unit is a well-known issue in geographic research and is often referred to as the Modifiable Area Unit Problem (MAUP) (Dark and Bram 2007). MAUP encapsulates the concept that statistical relationships are often contingent on the spatial units over which the data were collected. We tested sensitivity to MAUP by summarizing the demographic attributes and %IC across Census tracts (Figure S2). Census Tracts are the next largest enumeration unit after CBG. Each Census tract is comprised of 2 to (rarely) 12 CBG units.

For %IC change, we supplemented the contingent probability analysis by calculating the area devoted to new Census blocks (absent in 2000; present in 2010), the area of %IC change within them, and the proportion for which a demographic attribute was ≥ 25%. New Census block groups were identified by converting the 2010 Census block groups to polygon centroids (points) and overlaying them on the 2000 Census block group polygons. Census block groups from 2010 that had two or more points in 2000 Census block group polygons represented the formation of new Census block groups. The centroid-based approach was used to avoid the inefficiency created by overlaying vector-based polygon files.

3.2.2. High-resolution and NLCD IC data

High-resolution and NLCD %IC data were compared for each Census block group using x – y plots, Mean Absolute Deviation (MAD), and Mean Deviation (MD). We departed slightly from the typical application of MAD and MD. MAD and MD were the medians, not averages, of NLCD %IC minus high-resolution %IC. The acquisition dates for the high-resolution imagery and LiDAR data ranged from 2010 to 2016 for the 24 locations included in the study (Pilant et al. 2019). We used the NLCD2019 %IC data from the date most closely matching the high-resolution image acquisition date for each location to compare %IC values (Table 3). A previous, pooled analysis of most locations in Table 3 reported MAD and MD (as averages, not medians) values of 4.99 and 1.12, respectively (Wickham et al. 2020). Here we report MAD and MD for each location individually.

Table 3:

NLCD 2019 land cover dates matched to date of high-resolution data.

Location	Hi-Res	NLCD	Location	Hi-Res	NLCD
Austin, TX	2010	2011	New York, NY	2011	2011
Baltimore, MD	2013	2013	Philadelphia, PA	2010	2011
Birmingham, AL	2011	2011	Phoenix, AZ	2010	2011
Chicago, IL	2013	2013	Pittsburgh, PA	2010	2011
Cleveland, OH	2013	2013	Portland, ME	2010	2011
Des Moines, IA	2010	2011	Portland, OR	2012	2011
Durham, NC	2010	2011	Salt Lake City, UT	2014	2013
Green Bay, WI	2010	2011	St. Louis, MO	2016	2016
Los Angeles, CA	2016	2016	Sonoma, CA	2013	2013
Memphis, TN	2013	2013	Tampa, FL	2010	2011
Milwaukee, WI	2010	2010	Virginia Beach, VA	2014	2013
Minneapolis, MN	2010	2010	Washington, DC	2014	2013

4. Results

4.1. NLCD 2019 %IC

There was an inverse relation between five of the seven demographic attributes and %IC (Figure 2). Cross-sectionally, CBG were more likely to be comprised of poor, minority residents living in older homes, with a substantial fraction of the residents lacking a high school diploma and comfort with the English language as %IC increased. Except for the lowest class (%IC ≤ 10%), minority, low income, and houses built before 1960 had similar responses to %IC. Linguistic isolation and lack of a high school diploma had exponential responses to %IC, changing slowly initially but then rapidly as %IC increased. The over 64 years old demographic attribute declined modestly as %IC increased, and the under 5 years old demographic attribute had no relation to %IC. Changing the threshold choice or spatial unit did not change demographic attribute - %IC relationships (Figures S1 and S2; Supplemental Information).

Figure 2:

Proportion of CGB for which demographic attributes were ≥ 25% by NLCD2019 %IC decile class¹.

¹ lower limit=inclusive

The disparity between minority and low income when %IC was < 10% effectively maps the ethnic dichotomy between urban and rural poverty (Hess 2021) (Figure S3). Rural poverty tended to be dominated by persons identifying as white. Census block groups for which percentages of minority and low income both exceeded 25% tended to be urban. Impervious cover was < 10% for about 25% of all CBG and the median percentages for low income and minority for these units were 30% and 9%, respectively.

4.2. NLCD 2001 – 2019 IC change

The longitudinal relationships for 2001 – 2019 %IC change and demographic attributes were quite different than the cross-sectional relationships reported in Figure 2. Percentage minority was the only demographic attribute that had a positive association with 2001 – 2019 %IC change (Figure 3). The demographic attributes low income, no high school diploma, and homes built before 1960 were inversely related to 2001 – 2019 %IC change and there was no association between linguistic isolation and 2001 – 2019 %IC. The associations between 2001 – 2019 %IC change and demographic attributes indicate that large changes in %IC occurred in CBG that had higher proportions of more affluent minorities –minorities with high school diplomas, facility with the English language, and incomes greater the 2x the federal poverty level.

Figure 3:

Proportion of CGB for which demographic attributes were ≥ 25% by 2001 – 2019 %IC change class¹.

¹ p50M and P50LI are %IC change per-class Medians for minority (M) and Low Income (LI). The demographic attributes under 5 and over 64 years old had low proportions (≤ 5%) across all %IC change classes.

There was a distinct geography to the relationship between minority and %IC change. About 88% of the 847 CBG that experienced ≥ 30% increase in %IC occurred in Texas and the 11 western states, and 41% were in Las Vegas, Nevada and Phoenix, Arizona (Figure 4). The median percentage minority across these CBG was 49%, which increased to 53% for the subset of CBG for which percentage minority was ≥ 25%. Medians for low income were much less for these CBG. The median percentage for low income was 16% for the entirety of the 847 CBG, and 18% for the subset for which percentage minority was ≥ 25%. The western predominance of the minority – %IC change geographic pattern emerged once the 2001 – 2019 %IC change threshold was ≥ 10% (Table S3).

Figure 4:

Location of CBG for which %IC change was ≥ 30% and percentage minority was < 25% (●) or ≥ 25% (●). The dashed line marks the eastern boundary of the 11 western states.

As expected, new CBG (since 2000) had a much larger unit increase in IC area than those for which enumeration boundaries remained unchanged (Table 5). In new CBG, IC unit area increase between 2001 and 2019 was about 0.80 ha/km², whereas unit area increase for static CBG was about 0.25 ha/km². Differences in unit area IC increase were more disparate when percent minority was added to the binary classification of CBG. Unit IC change in new CBG with a minority population of at least 25% was about 2.2x greater than for new CBG with a minority population less than 25%. The ratio for static CBG was about 1.6x greater. The ratios for other demographic attributes were consistent with the trends in Figure 3.

Table 5:

Frequency (Frq), Area, area of 2001 – 2019 IC increase (IC Δ), and unit increase (↑) by cross-classification of minority and change in CBG enumeration boundary.

Class	Frq	Area (km2)	IC Δ (ha)	Unit ↑ (ha/km2)
Minority < 25%, no CBG Δ	84,482	5,243,047	1,123,245	0.209
Minority ≥ 25%, no CBG Δ	87,615	1,889,935	641,988	0.340
Minority < 25%, CBG Δ	18,644	615,478	357,912	0.587
Minority ≥ 25%, CBG Δ	25,589	333,442	418,741	1.225
Total	216,330	8,081,903	2,541,887

4.3. Surface temperatures

As expected (e.g., Maskooni et al. 2021), demographic attribute response to surface temperature was very similar to its response to extant %IC (Figure 5). The proportion of CBG exceeding the 25% threshold for minority, low income, and houses built before 1960 increased as the surface temperature decile increased, adding quantification to the “heat risk-related land cover” concept introduced by Jesdale et al. (2013). Like its response to extant %IC, lack of a high school diploma exhibited an exponential response to increasing surface temperature deciles, but the response of the linguistic isolation was more linear than exponential. Respectively, the proportion of CBG meeting the 25% threshold for over 64 years old and under 5 years old declined and remained unchanged as the surface temperature decile increased.

Figure 5:

Proportion of CBG for which demographic attributes were ≥ 25% by location-specific surface temperature deciles.

4.4. Comparison of high-resolution and NLCD IC data

Average MAD and MD values were 4.78% and 3.43%, respectively, indicating that NLCD %IC tended to overpredict %IC from the high-resolution data (Table 6). The tendency for NLCD %IC to overpredict %IC from the high-resolution data was consistent across the locations examined since MD values were rarely negative and MAD-MD differences were typically small. Durham, NC was an exception. It was the only location where the MAD and MD were nearly equidistant from zero (0). MAD values for individual locations ranged from 2.17 for Des Moines, IA to 11.27 for Chicago, IL (Figures S4 and S5). Notwithstanding the aberrant results for Chicago IL, the magnitude of overprediction was reasonably consistent across locations, with about ⅔ of the locations having MAD values less than 5%. Contingent probabilities were reflective of the MAD and MD results (Figure 6). Both sources exhibited rapidly increasing likelihoods of CBG ≥ 25% for minority and low income as %IC deciles increased.

Table 6:

Comparison of high-resolution and NLCD %IC

Location1	# Obs	MAD	MD	Location	# Obs	MAD	MD
Austin, TX	725	4.41	4.22	New York, NY	6,073	7.58	7.42
Baltimore, MD	1,579	3.09	-1.49	Philadelphia, PA	3,796	2.42	-0.29
Birmingham, AL	556	4.20	3.81	Phoenix, AZ	2,403	5.77	4.72
Chicago, IL	6,181	11.27	11.27	Pittsburgh, PA	1,035	7.49	7.49
Cleveland, OH	1,378	4.65	4.61	Portland, ME	139	4.20	4.20
Des Moines, IA	302	2.17	0.32	Portland, OR	1,152	5.88	5.88
Durham, NC	179	4.28	-3.97	Salt Lake City, UT	601	2.67	1.80
Green Bay, WI	148	5.19	4.46	St. Louis, MO	1,483	5.21	4.45
Los Angeles, CA	6,278	3.63	-0.27	Sonoma, CA	375	3.89	3.88
Memphis, TN	673	4.09	3.89	Tampa, FL	1,787	2.90	1.63
Milwaukee, WI	1,140	7.52	7.52	Washington, DC	2,885	3.93	-0.40
Minneapolis, MN	1,711	5.76	5.72	Virginia Beach, VA	970	2.46	1.44
				All locations2	43,547	5.13	3.84

¹Total population = 64,391,808

²Census block group total and pooled medians

Figure 6:

Proportion CBG for which minority or low income was ≥ 25% by resolution of IC data

Our comparisons of high- and moderate-resolution IC estimates does not address accuracy of moderate-resolution IC change data. Lack of temporal high-resolution data render such comparisons unattainable. To address this gap, we provided ca. 2001 and 2019 Google Earth^™ images for three CBG with low values of %IC from the NLCD2019 database (Figures S6, S7, and S8). This anecdotal evidence suggests that temporal NLCD2019 %IC data detects change when and where it occurs.

5. Discussion

A principal objective of this work was demonstration of the value of adding a temporal domain to research on the spatial pattern of environmental equity (Karanja and Kiage 2021). Our cross-sectional and longitudinal analyses produced contrasting results, highlighting the importance of the temporal domain. Our cross-sectional (NLCD2019 %IC versus demographic attributes) results were consistent with most previous studies (e.g., Table 1), showing that CBG with higher extant (NLCD2019) %IC tended to be comprised of poor, minority residents who were often lacking a high school diploma and comfort with the English language. In contrast, our longitudinal analysis indicated that large 2001 – 2019 increases in %IC tended to occur in CBG that had more affluent minority populations.

Our results are broadly consistent with two prior studies (Clement and Alvarez 2020; Hess 2021). Based on the demographic attributes from the 2014–2018 ACS used in this study, the association between minority presence and IC change reflects how the suburbs in major metropolitan areas (e.g., Charlotte, NC, Houston, TX, Phoenix, AZ, Sacramento, CA) are majority minority (Frey 2022) (Table S3). In other words, minorities disproportionately live in suburban areas where the pace of IC change is higher (Clement and Alvarez 2020). Meanwhile, the attenuated association between poverty and IC change is the result of the fact that, even though the rate of poverty change is higher in the suburbs compared to urban centers, the level of poverty is still higher and far more concentrated in urban centers (Hess 2021), where the pace of new land development is lower (Clement and Alvarez 2020).

The distinct geographic pattern in the 2001 – 2019 %IC change-minority relationship highlighted the importance of broad geographic scale undertaken for this research. In addition to the lack of a temporal domain, most high resolution studies cited herein (Table 1) focused on one to a few urban centers and therefore were not set up to address differences over broad geographic scales. The CBG in which large increases in %IC occurred in counties that had increasing African American and Hispanic populations and tended to have declining white populations (Table S4). The increase in Hispanic ethnicities was about twice as large as the increase in African American ethnicities, and not surprisingly, most of the locations highlighted in Figure 4 occur in larger metropolitan areas in states that comprise about ½ of the total US population increase between 2010 and 2020 (Table S4). Schwarz et al. (2015), Nesbitt et al. (2019) and Locke et al. (2021) conducted cross-sectional, environmental equity studies for 7 to 37 US cities, finding locational differences locational differences in equitable distribution of environmental assets. We add these and other studies by extending the geographic scope to the entire US and adding a temporal domain. The broadened geographic scope and temporal domain were used to identify hotspots (e.g., Figures 4 & 5) of potentially emergent (since 2001) inequitable distributions of environmental assets. The “hottest” hotspots were in Houston, TX, Dallas-Ft. Worth, TX, Phoenix AZ, and Las Vegas, NV. The identified hotspots add geographic precision to spatiotemporal changes in ethnic patterns (Martin et al. 2017).

Our focus on the temporal domain relied on IC data from an established operational program as a proxy for high resolution temporal IC data. Cross-sectional comparison indicated NLCD IC was a suitable proxy for high resolution IC. Hansen and Loveland (2012) astutely differentiate land cover mapping into operational and research programs. Operational programs are characterized by long-term institutional support, well-defined objectives, stakeholders, deliverables with articulated timelines, and programmatic efforts dedicated to product improvement. In contrast, research-oriented programs tend to be “one-off” efforts that conclude when the map has been produced and the questions posed have been addressed. Many of the characteristics of operational programs are missing or less stringent in research-oriented efforts. Now ongoing for 20 years, NLCD is an operational program that includes all of the aforementioned characteristics (Homer et al. 2004; Yang et al. 2018). To our knowledge, none of the high-resolution mapping efforts cited herein have the characteristics of an operational program. NLCD will certainly not meet many of the needs for environmental equity study supplied by high-resolution data, but it appears to be a suitable proxy for high-resolution data for temporal analysis of IC and demographic attributes. The results presented here add to those from Reid et al. (2009), Jesdale et al. (2013), Salvati et al. (2017), Clement and Alvarez (2020), Sinha et al. (2022), and Venter et al. (2023), who used Landsat or similar data to assess spatial patterns of environmental equity.

Table 4:

Proportion of vulnerable Census block groups (DI ≥ 25%) in IC^§ class 1 (p|EJ₁), IC classes >= 8 (p|EJ₈)^●, and locality-wide (p).

Location	p|EJ1	p|EJ8	p	Location	p|EJ1	p|EJ8	p
Los Angeles, CA	0.57	0.94	0.87	Cleveland, OHⴋ	0.06	0.90	0.54
New York, NYⴋ	0.80	0.86	0.86	Milwaukee, WIⴋ	0.02	0.61	0.52
Memphis, TNⴋ,ⴌ	0.61	0.82	0.79	Philadelphia, PA	0.06	0.87	0.52
Durham, NCⴋ, ⴌ	0.39	1.00	0.50	Sonoma, CAⴌ	0.23	0.56	0.49
Birmingham, ALⴌ	0.48	0.86	0.69	St. Louis, MO	0.12	0.78	0.51
Washington, DC	0.36	0.68	0.68	Green Bay, WIⴋ,ⴌ	0.44	0.75	0.48
Virginia Beach, VAⴌ	0.26	0.27	0.66	Salt Lake City, UTⴋ,ⴌ	0.15	0.84	0.48
Austin, TXⴋ,ⴌ	0.56	0.92	0.65	Portland, ORⴋ	0.20	0.77	0.47
Chicago, IL	0.15	0.82	0.63	Des Moines, IAⴋ,ⴌ	0.14	0.43	0.43
Tampa, FL	0.60	0.63	0.63	Minneapolis, MNⴋ	0.05	0.88	0.42
Baltimore, MD	0.11	0.83	0.62	Pittsburgh, PAⴋ	0.08	0.73	0.40
Phoenix, AZ	0.41	0.84	0.61	Portland, MEⴋ,ⴌ	0.05	0.62	0.11

^§Based on %IC from NLCD 2019;

^●IC classes 8 – 10 combined to increase sample size for smaller cities;

^ⴋLess than 50 observations in IC class 1;

^ⴌLess than 50 observations in IC classes 8 and higher.

Highlights.

• High impervious cover (IC) was likelier in poor, minority Census block groups (CBG)

• IC 2001 – 2019 increase was positively associated with percent minority

• IC 2001 – 2019 increase was not associated with poverty

Data Availability

The data analyzed for this study will be available on https://data.gov.