Main

Food, in particular meat and dairy products, represents a substantial portion of the GHG footprint of a city1,2. Reducing these emissions is difficult because agricultural supply chains are geographically dispersed and have many processes and inputs (for example, fertilizer, irrigation and fuel)3,4. Although the linkages between cities and supply regions for meat, produce and other transboundary flows are well-conceptualized in the literature, empirically connecting them remains rudimentary and underdeveloped5,6,7. This has ramifications in terms of understanding and addressing how consumption in one region (for example, the city) affects environmental and socioeconomic conditions in producing regions (for example, rural communities).

The predominant method when quantifying the GHG footprint of cities (and corporations) is the spend-based approach, which relies on broad, industry or national averages for products and processes. Given that the environmental and socioeconomic impacts associated with food and other land-derived products are geographically diverse, this reliance is highly problematic. Studies have attempted to identify the food supply region (foodshed7) of a city, but these efforts lack sufficient precision: simply considering ‘meat’ as a whole rather than differentiating between beef, pork or chicken each of which has distinct production geographies and impacts8,9,10. The few studies that have tried to nuance this broad categorization use labour-intensive surveys that lack geographic specificity11,12 (Supplementary Note 1 gives a review of relevant literature).

More granular mapping of urban–rural linkages is necessary to enable cities (and citizens) to address the spatially diffuse impacts generated by resource consumption5,13. To date, we have lacked a scalable, high-resolution method to capture subnational flows and impacts of meat and other products consumed in cities.

Two obstacles have hindered development of such an approach. First, data linking production to consumption are scarce or proprietary14,15. Open data, such as freight flow surveys encompass broad industry sectors (for example, animal feeds, meat and milled grain products)8, seldom scale to urban areas16 nor do they cover several supply-chain stages. Second, even if linkages can be identified, spatially explicit data on the environmental intensity of production are usually only available for nations17, regions18,19 or other large geographies.

We overcome these obstacles using the food-system supply-chain sustainability (FoodS3) model20. This mass-balanced supply–demand model reconstructs meat supply chains (including losses) for 93% of the US population by linking 3,143 feed, livestock and poultry producing counties and 3,531 meat-consuming cities (census-designated ‘urban areas’) (Methods). Owing to its high urbanization rate (~85%) and status as the world’s second largest producer21 and per capita consumer22 of meat, the USA makes a compelling case. Using a scalable spatially resolved approach, this research comprehensively tracks commodity flows between cities and their hinterlands across several supply-chain stages. Where existing methods trade geographic specificity (in sourcing or impacts) for scalability, this modelling effort sacrifices neither.

This research provides a vivid picture of the production geographies of meat and animal feed—the meatshed—of US cities. We combine reconstructed supply chains with location-specific GHG estimates of feed, livestock production and primary processing (cradle-to-processing gate) to calculate the GHG intensity of producing beef, chicken and pork for each city. This is then paired with consumption estimates to calculate the annual embodied GHGs from beef, chicken and pork consumption—the carbon hoofprint—for every US city.

Results reveal that cities source meat and animal feed from distinct geographies with varying GHG production intensities. This generates striking differences in hoofprints, which are impossible to capture using traditional carbon accounting methods for cities. Despite comparable meat consumption, per capita hoofprints vary by a factor of three between cities. The spatial granularity of our results also highlights unexpected pathways to reduce GHGs associated with diet (for example, consuming pork rather than chicken in certain cities).

Scenarios show that cities can reduce the hoofprint by 14–51% by implementing strategies to reduce edible food loss and by promoting dietary shifts from beef to poultry. This is a cost-effective decarbonization lever for the estimated 60 million Americans living in cities, where the hoofprint rivals GHGs from household energy consumption. Supply-side strategies, such as silvopastoral systems, can further reduce the hoofprint. Our modelling framework can be expanded to other commodities and geographies, ultimately helping researchers to better conceptualize, map and measure the co-evolution of urban and rural spaces. Reducing the environmental impacts from the consumption of meat (and other products) in cities will ultimately require changing both production and consumption. Our model aids this transition by providing a way to map, measure and mitigate the transboundary impacts of cities on an increasingly urban planet.

Results

Residents in US cities annually consume 11.1 Mt of meat: 4.6 Mt of chicken, 3.7 Mt of beef and 2.7 Mt of pork. The three largest cities—New York City, Los Angeles and Chicago—alone consume 3.2 Mt of meat (Extended Data Fig. 1a and Supplementary Tables 1 and 2).

The aggregate carbon hoofprint of all US cities is 329 MtCO2e, exceeding the territorial CO2 emissions of the United Kingdom (305 MtCO2e) and Italy (313 Mt)23 and on par with total US at-home fossil fuel combustion (334 Mt)24. As with meat consumption, the hoofprint has broad intercity variation (Extended Data Fig. 1b). The total hoofprint of individual cities ranges from 1.68 ktCO2e (Calais, Maine) to 18.1 MtCO2e (New York City). The hoofprint of New York City alone exceeds the GHG emissions of Maine and New Hampshire24.

The per capita carbon hoofprint of US cities is 1,093 ± 237 kgCO2e. Despite consistent per capita consumption of meat across cities (μ = 78.6 ± 1.87 kg annually), the hoofprint varies from 500 kgCO2e per capita in Houghton, MI, to 1,731 kgCO2e per capita in Richmond, MO (Fig. 1a). Per capita hoofprints are highest in parts of Texas, Oklahoma and Missouri and lowest in the upper Midwest (especially Michigan and Wisconsin). This is because GHG intensities of most beef and pork supplied to the Midwest are in the lowest two quintiles, whereas the opposite is true for high-hoofprint regions. Because of these differences in intensity, meat consumption is a poor proxy for the hoofprint (Fig. 1b).

Fig. 1: Meat consumption and carbon hoofprint across the USA.
figure 1

a, Per capita meat consumption (beef, chicken and pork) (kg per capita) (top) and hoofprint (kgCO2e per capita) (bottom) for 3,531 US cities. b, Comparison of per capita meat consumption (left) and hoofprint (right) for California (top), the Northeast (middle) and Texas (bottom). Consumption values are before retail and consumer losses (for example, kilograms delivered to retail). Based on uncertainty analysis, there is a mean absolute percentage error of 26% from baseline across all cities (Supplementary Fig. 1).

Full size image

Correlation between per capita consumption and hoofprint is weak (r = −0.04, P = 0.01): 868 cities have above-average consumption, but below-average hoofprint, and 931 cities have the opposite characteristics (Fig. 2a). Despite comparable consumption (beef 34 ± 0.89%; chicken 41 ± 1.1%; pork 25 ± 0.56% of total meat by mass) (Supplementary Fig. 2), the largest hoofprints exceed the smallest by a factor three. This suggests that other factors explain the variation in hoofprints.

Fig. 2: Carbon hoofprints, hoofprint components and GHG intensity of meat production.
figure 2

a, Per capita carbon hoofprint (kgCO2e per capita) versus per capita meat consumption (kilograms per capita) for 3,531 cities. Weighted-average GHG intensity of beef, chicken and pork production for each city (kgCO2e per kg of edible meat delivered to retail). b, Share of hoofprint by beef, chicken and pork for 3,531 cities. c, GHG intensity of beef, chicken and pork production for 3,531 cities. Boxplots should be interpreted as follows: bottom and top of box represent 25th and 75th percentiles, respectively; horizontal line in box represents median; height of box represents interquartile range; bottom and top of vertical line represent minimum and maximum or 1.5 times interquartile range below 25th percentile or above 75th percentile, respectively; dots represent outliers that are either 1.5 times the interquartile range above the 75th percentile or below the 25th percentile.

Full size image

Explaining the hoofprint

The carbon intensity of meat is strongly correlated with the hoofprint (r = 0.99, P < 2 × 10−16). Because of this dependence, using GHG intensities of meat production based on national or regional averages would produce substantial overestimates and underestimates of the hoofprint across US cities. Urban carbon accounting methods must move beyond the common practice of using national average carbon intensities of meat in their calculations.

Breaking down the hoofprint, we find that beef is consistently the largest component of the hoofprint (73 ± 7% of the total), followed by pork and chicken (Fig. 2b), which is expected given the environmental intensity of beef25,26. Considerable variation exists, however, across cities. For US beef the share of the hoofprint ranges from 41% in Pine City, MN, to 87% in Monte Vista, CO. Although pork is usually the second largest hoofprint component (range 5–26%; μ = 16 ± 5%), chicken supersedes pork for 386 cities (μ = 11% ± 3%).

Dissecting emissions by production stage further highlights intercity variation (Supplementary Fig. 3). Grazing generates 63 ± 14% of beef emissions on average but is much lower where feedlot emissions predominate (for example, 17% in San Diego, CA). Pork emissions are split evenly between growing feed (μ = 52 ± 10%) and hog farming (μ = 46 ± 11%), while chicken emissions are mostly from feed (μ = 70 ± 5%). As with beef, these percentages vary considerably. For instance, hog farming can drive 10–80% of pork GHGs. Such variation arises from differences in sourcing locations and related production practices.

Urban–rural connections via meatsheds

Our model links urban food consumption to rural counties that grow animal feed (corn, wheat and soy), raise animals and process animals. Much like the foodshed that grows the food of a city—akin to a watershed supplying a river—we define the meat supply geography as the meatshed.

Mapping the beef supply chains for Atlanta (GA), Los Angeles and New York City demonstrates the spatial extent of urban meatsheds (Fig. 3 and Supplementary Figs. 4–6). Although cities typically source processed meat from a handful of counties (Supplementary Fig. 4), the meatshed can eventually encompass hundreds of counties spanning thousands of kilometres. For instance, our model predicts that Los Angeles sources its beef from ten counties with processing facilities, which are supplied with livestock raised in 469 counties. These counties, in turn, rely on feed grown in 828 different counties. New York City and Atlanta show similarly dispersed patterns.

Fig. 3: Beef meatsheds of three large cities.
figure 3

Meatsheds and supply linkages for beef consumed in Atlanta (left), Los Angeles (middle) and New York City (right). Top panels show linkages between animal feed-producing counties and cattle-production counties. Bottom panels show linkages between cattle-production counties and primary processing counties. Colours indicate the kgCO2e per kilogram of animal feed (top) and kgCO2e per kilogram delivered to retail (bottom). Colours determined using ‘Natural Jenks’ classification. Numbers in brackets denote the range of values for each city and indicator represented by a given colour. Only the top 50% of supply-chain linkages are shown to enhance legibility. Full beef, chicken and pork meatsheds, including processing, for the ten largest cities are shown in Supplementary Figs. 4–6.

Full size image

Since beef is produced across the entire USA27,28, the feedlot stage of the beef meatshed differs by city and aligns with regional production centres. Atlanta sources from producers in Texas and Colorado; Los Angeles from California, Texas and the Great Plains; and New York City from the Midwest and Northeast. Meatsheds for chicken and pork vary less because of the spatial concentration of production facilities27,28. Pork processing, for example, occurs mainly in Iowa (31% of US processing), Illinois (10%), Minnesota (10%) and Missouri (8%) (Supplementary Fig. 5), while most cities source processed chickens from the mid-Atlantic and Southeast (Supplementary Fig. 6). Meatsheds for hog and broiler production mirror this spatial concentration. For feed production, the meatshed for all three cities stretches to the sparsely populated corn belt of the Midwest, the wheat belt east of the Rocky Mountains and other agricultural regions of the USA.

These supply chains can span thousands of kilometres. The average distances for beef supply chains are 920 ± 630 km from processing to city, 420 ± 150 km from feedlot to processing and 430 ± 230 km from feed farmer to animal producer for the 20 largest US cities (Supplementary Table 3). Distances from processing to city exceed those from feedlot to processing (P = 0.0006) and from crop farm to feedlot (P = 0.0005). Although these averages are generally representative, there are considerable ranges within cities (for example, 140–2,180 km from processing to New York City). The broader production geography of US livestock influences distances between nodes. For example, distance to processing is below average for cities near beef processing areas such as Denver, CO (130 ± 1 km), Chicago, IL (410 ± 280 km) and Detroit, MI (280 ± 230 km).

Chicken and pork mirror these trends (Supplementary Tables 4 and 5). While distance from processing to city are similar for all three meats (Supplementary Table 6), other portions of the poultry and pork meatsheds are less geographically expansive than beef. Average distance from broiler farms to processing is 190 ± 370 km and is shorter than distances from beef feedlots to processing (P = 0.0264). Average distance from crop farm to broiler farm are also smaller and statistically distinguishable from beef for chicken (μ = 310 ± 200 km, P = 0.0184) and pork (μ = 140 ± 90 km, P < 0.0001).

These findings align with the vertical integration that typifies the US poultry and pork industries29, as well as high travel mortality rates for hogs and chickens30. Finally, most hog and chicken farmers are located within, or near, the US corn belt. Beef feedlots are closer to grazing lands and further from feed producers in the corn belt20. As with beef, the geography of livestock and poultry processing influences the meatshed of a city (for example, distance from processing to city is below average for cities near chicken processing centres in the mid-Atlantic).

The GHG intensity of the different production stages varies across and within meatsheds (Supplementary Fig. 3). This is expected given differences in soil, climate, agriculture policy and farming practice across the USA31. The cumulative effect is large variation in GHG intensity of beef, chicken and pork for different cities (Fig. 4). The GHG intensity of beef production varies by a factor 4.3 across cities. This disparity increases to 4.9 for chicken and 15 for pork.

Fig. 4: Variation in GHG intensities of meat production for US cities.
figure 4

a, GHG intensities of beef (left), chicken (middle) and pork (right) for 3,531 cities. Calculated as kgCO2e per kg of edible meat delivered to the city. Colours determined using ‘Natural Jenks’ classification. Numbers in brackets denote the range of GHG intensities for beef, chicken, and pork represented by each colour. b, A snapshot of the modelled New York City beef meatshed showing select beef-producing counties connected to one of many counties that process beef destined for the city. A cutoff of 20 kgCO2e per kg of beef at processing gate was used to differentiate between high and low GHG-intensity counties.

Full size image

The wide range in meat-related emissions is due to several factors. Chief among them is variability in emissions from feed, driven by land-use change, yield differences, differing nitrogen fertilizer application rates and related N2O emissions20. These factors combine across supply chains to influence the GHG intensity of feed consumed in different regions. This variability strongly influences chicken and pork hoofprints as feed is a larger proportion of total GHGs for them (Supplementary Fig. 3). Pork production is further compounded by differences in manure management, with some locations having negligible or negative manure emissions due to methane capture and energy use.

The GHG intensity of beef is lower for cities that source higher proportions of beef from culled beef and dairy cattle as opposed to feedlots31,32. Cities near major dairy-producing areas such as Wisconsin or southern Pennsylvania, are more likely to consume this beef. Conversely, intensity is highest for beef cattle that are fed GHG-intensive feed at facilities with open manure lagoons. Cities consuming this type of beef leads to disproportionately GHG-intense hoofprints, such as the McAllen and other cities in Southern Texas (top-left of Fig. 2a). Variation in enteric fermentation emissions across regions also plays a role.

The high resolution of FoodS3 also reveals differences in production intensity within the meatshed of a city. New York City, for example, sources meat from the Northeast and Midwest via processors in Pennsylvania (Fig. 3b). Clear differences are evident between counties within the modelled New York City beef meatshed with low-GHG production intensities that produce significant volumes of culled dairy cattle and those that produce more feedlot beef using GHG-intensive feed.

Notably, and in contrast to most studies25,26, the emissions intensity of chicken exceeds that of pork for some cities (for example, Monte Vista, CO). This exception, however, only occurs for ten cities. Similarly, we find ten instances where the emissions intensity of pork exceeds that of beef (for example, Calexico, CA). These cities consume a high portion of beef from culled cows, and a high portion of pork supplied from areas using low-tech manure management systems (for example, uncovered lagoons) as well as significant land-use change emissions from feed.

Discussion

To reach their climate goals, cities need to reduce the size of their carbon hoofprints. Nationwide, the hoofprint equals 43% of the GHGs generated from at-home fossil fuel combustion and electricity consumption across all US cities (Fig. 5a). This percentage increases to more than 50% for 60 million Americans living in high-hoofprint cities. This ratio is highest in cities with mild climates and low-carbon energy grids, such as in San Diego, CA (77%) and Los Angeles, CA (81%). Conversely, it is smaller in colder cities, such as Boston, MA (24%) and Chicago, IL (19%), and drops further when combined with carbon-intensive grids and low hoofprints (for example, 5% in Springs, NY).

Fig. 5: Comparing the hoofprint to residential energy and pathways towards lowering the hoofprint.
figure 5

a, Carbon hoofprint as percentage of carbon footprint of household energy use in US cities. Colours determined using ‘Natural Jenks’ classification. Numbers in brackets denote range of values represented by a given colour. b, Scenarios to curb the carbon hoofprint. Scenarios are as follows (from left to right): retailers and consumers halve wastage of edible meat to reduce total production to satisfy total meat consumed; consumers substitute 50% of beef with pork and chicken (even split); consumers substitute 50% of beef for chicken; and consumers cut meat from diet one day a week. Left side shows the effects of individual scenarios. Right panel shows the cumulative effect of the scenarios only considering chicken substitution for beef (Extended Data Fig. 2a shows results for pork and chicken substitution).

Full size image

We model four policy scenarios that cities can take to reduce their hoofprints (Fig. 5b). Scenario 1 halves the amount of meat lost (as edible food waste) by retailers and consumers (resulting in less meat produced to satisfy total meat consumed). Scenario 2 substitutes half of beef consumption evenly between pork and chicken. Scenario 3 substitutes half of beef consumption with chicken. Lastly, scenario 4 assumes not eating meat once a week (‘meatless Monday’).

The largest hoofprint reductions come from substituting beef (28–33%) and halving food waste (16%). Combined, these actions would reduce the total hoofprint by 123–142 MtCO2e (37–43%) from baseline. Reductions are largest for cities with GHG-intensive beef (Extended Data Fig. 2b). Notably, these reductions preclude consuming less meat, a persistent challenge to curbing the hoofprint, allow moderate beef consumption and align with ongoing dietary shifts22. Since substituting chicken for beef and cutting in-home food loss lowers grocery expenditures, mitigating the hoofprint is less expensive and can be implemented more swiftly than some city decarbonization strategies used in other sectors (for example, home energy retrofits). Combining the above actions with marginal reductions in meat consumption (scenario 4) produces a 51% reduction from baseline.

Cities are increasingly promoting low-carbon diets. New York City has an initiative to expand availability of plant-based foods at city events and facilities and enlists universities, companies and other institutions as allies33. Cities should also adopt an expanded notion of urban food policy to tackle the economic and retail barriers that limit access to affordable, sustainable food34,35. For example, educational campaigns in public spaces and schools tailored to the social and cultural diversity within cities36, could emphasize the fiscal and health benefits of sustainable diets37,38. More directly, cities can influence menus at hospitals and schools, promote plant-based products with local retailers and support initiatives, such as urban farming39, linked to sustainable dietary shifts. Cities could also encourage state governments to model policies after those in Europe that discourage retailers from discarding unspoilt meat40 and limit meat advertisements in public spaces41.

Future work

Reducing GHG-associated meat production will also be necessary to curb the hoofprint. While most of the work to do this will happen in rural areas, cities can play a supporting role. Large US cities, such as Los Angeles, have adopted measures to address food-system GHGs through production and consumption. By identifying emissions hotspots in meat supply chains, our research provides the requisite knowledge for cities to deploy targeted production-based solutions (Fig. 4b and Supplementary Fig. 3). For example, beef silvopastoral systems could reduce the total hoofprint of US cities by 6% or considerably more for some cities (Extended Data Fig. 3). Future research could model additional climate-friendly practices (for example, alternative feeds, agrivoltaics and bioenergy from manure) to identify which strategies would be most effective for different cities. This could inform urban–rural collaborations to accelerate urban food-system decarbonization42.

Meat production is also a significant source of air pollution, nutrient runoff and water stress25. A more holistic hoofprint assessment could incorporate these indicators using the national emissions inventory43 and leverage the spatial capabilities of FoodS3 to link urban consumption with impacts in specific farming communities. Combining this with fate and effect models could clarify the associated public health impacts44. Another advance would be a combined uncertainty analysis including farming inputs (for example, irrigation), instead of modeling uncertainties in meatshed and consumption independently (Supplementary Figs. 1 and 7) we performed.

Uncertainty exists with respect to the estimated GHG reductions in our scenarios. Reduced domestic demand might spur increased exports of US meat and GHGs from distribution and refrigeration. However, exports could substitute meat produced less efficiently elsewhere and curb land-use change associated with pastures and feed farming. Consequential models incorporating international trade could quantify net GHG impacts (or other stressors) as well as economic outcomes for farmers in the USA and importing countries. Conversely, future iterations could consider the subnational origins and impacts of meat imported by the USA. Existing supply–demand models could support this45 or new ones for countries with comparable agricultural statistics.

The FoodS3 modelling architecture can also be deployed for other agricultural or non-agricultural commodities (for example, steel) where supply and demand data exist at subnational scales. For instance, forest products could be modelled using US Forest Inventory and Analysis statistics on timber harvests and mill outputs at the county level or metals using mining and mineral processing data from commercial46 and government sources47. Incorporating consumption data48, pollution inventories43 and remote-sensing data could then link urban processes to environmental or even social change elsewhere.

This study concretely demonstrates how to make such opaque urban–rural linkages more transparent. By doing so, how rural and urban livelihoods are intertwined becomes clearer, as does the collective responsibility for cities and rural communities to collaborate in the shared goal of reducing environmental impacts and improving livelihoods. Researchers can build on these findings to advance a multiscalar conceptualization of urban sustainability5,6,49. This will ultimately advance efforts by policy-makers, citizens and communities to identify and implement strategies to decarbonize the city beyond the spatial confines of the urban boundary.

Methods

The study deploys and expands the FoodS3 model to estimate the carbon hoofprint of 3,531 US cities. Our expansion develops unique estimates of beef, chicken and pork consumption for all counties across the contiguous USA that are then allocated to census-defined ‘urban areas’. These estimates add a further link to the supply chains modelled in the FoodS3 model20. This expansion models tens of thousands of supply chains that link animal feed production, animal husbandry and animal processing to final meat consumption. This allows us to quantify the carbon hoofprints and meatsheds for each city.

Meat consumption estimates

We estimate beef, chicken and pork consumption at the county level using the National Health and Nutrition Examination Survey (NHANES)50, Center for Disease Control. NHANES is a complex, stratified bi-annual survey of the dietary habits and health of approximately 10,000 individuals. It is meticulously designed to be representative of the US population and uses weighting schemes to avoid oversampling and undersampling of demographic groups. We combine five iterations of NHANES, resulting in a sample size of 51,623 individuals. We estimate the mass of per capita beef, chicken and pork consumed using the Center for Disease Control prescribed protocols for manipulating and analysing raw NHANES data and the survey package in R Studio.

NHANES records the mass of meals consumed at home and away from home (for example, restaurants, canteens and so on). Meals are classified as standardized US Department of Agriculture (USDA) food codes. We link NHANES responses to the USDA food intake converted to retail commodity database (FICRCD)51 to disaggregate meals to individual meats. The FICRCD gives the grams of beef, chicken and pork per 100 g for each of the ~10,000 USDA food codes, which we use to estimate per capita consumption of each meat in grams per capita per day.

We develop beef, chicken and pork consumption profiles based on race/ethnicity and income level as these demographic attributes influence meat consumption. For race/ethnicity, we include Black, Latino, White and Other. Other includes all races covered in NHANES with sample sizes too limited to develop reliable consumption patterns.

For income, we use a categorical variable to estimate consumption, because of the nonlinear relationship between affluence and meat consumption52. We categorize individuals as high income if the income of their household was 200% the poverty level for a given iteration of NHANES, and low income otherwise. We choose 200% as we found statistically significant differences in meat consumption patterns above and below this threshold (Supplementary Note 2).

Combining income and race/ethnicity traits produces eight unique consumption profiles for each type of meat (for example, beef consumption by White, high-income individuals). These estimates represent national averages for each demographic group and do not necessarily capture differences in regional cuisines nor other determinants of dietary habits (for example, non-Western diets consumed by immigrants). Future work should incorporate uncertainty analysis of NHANES estimates or utilize additional data (for example, supermarket scanner data) to account for these factors (Supplementary Note 6).

Being a self-reported survey, NHANES is prone to under-reporting consumption of certain foods, including meat53. However, this issue is systematic across the survey and does not reduce its ability to capture differences in consumption patterns between demographic groups. Moreover, we adjust for under-reporting by scaling NHANES consumption to national domestic meat supplies as part of the supply–demand model (see below). Given the paucity of mass-based, city-level dietary surveys, it is not possible to compare our consumption estimates to published results. Nonetheless, NHANES is generally accepted by public health experts as a reliable estimate of US diets and its method has been replicated in other countries54. We estimate that standard errors in our consumption estimates range from 1% to 10% (Supplementary Note 2).

We combine our consumption profiles of beef, chicken and pork with demographic data from the 2017 US census to estimate aggregate demand for each meat at the county level55. We choose this year to align with the 2017 USDA Census of Agriculture, a critical data input into the current version of the FoodS3 model31. We start with demographic data at the census tract level to capture the substantial variation in income and race that often occurs within counties. Census tract estimates of household income and population by race are used to determine the ratio of average household income by race to federal poverty levels. A small number of tracts lack income data by race due to small sample sizes. For these tracts we use state-level data to calculate the poverty–income ratio. After calculating this ratio, we combine population estimates by race with appropriate consumption profiles to estimate total beef, chicken and pork consumption for 73,057 census tracts. Estimates at the tract level are then aggregated to the county.

Enhancing the food-system supply-chain sustainability (FoodS3) model

The FoodS3 model is a US-based transport model that simulates the movement of crops and livestock, connecting places of production to places of consumption20. It models the distribution of crop outputs (corn, wheat, soy, alfalfa and corn silage) from the counties of production to different nodes of downstream demand (for example, livestock and dairy production, ethanol processors, soybean crushers and flour millers), and includes the entire supply and demand of each crop. It also models the subsequent stages of demand, distributing ethanol, soymeal and flour milling byproducts to livestock producers and moving dairy milk, beef cattle, hogs and broilers to primary processors. This approach includes a data accounting component, a spatially explicit environmental impact life-cycle assessment (LCA) and a transportation optimization component. The result is a link between the downstream companies and the upstream environmental impacts of the crops/products supplied.

We expand the FoodS3 model, which previously ended at the primary processing stage of meat production20,31,56,57, to estimate the downstream supply flows of processed meat to counties where meat is consumed. We do so by first estimating the equivalent balance of total supplies and total demands of each meat type across the USA and across counties. We then deploy linear optimization that minimizes total impedance between processing facilities and final consumption locations for each type of meat produced, resulting in an estimated meatshed for each consumption county. Details for estimating the equivalent balance of total supplies and total demands are described further below.

County supply of meat

We estimate the total 2017 domestic supply of each meat type across counties based on the total annual livestock and poultry population processed at primary processing facilities (that is, broilers, finished hogs, finished beef and culled cows), state-specific typical animal mass at slaughter and the associated average dressing weights for each meat type, at approximately 61% for beef and about 75% for both pork and chicken58. We then scale the total estimated chilled-carcass weight produced in each meat-processing county to match the total US chilled-carcass weight for beef, pork and chicken58. In addition to the domestic supplies, we also include imported meat and supplies from leftover stocks from previous years (‘beginning stocks’) minus ending stocks (that is, total supplies remaining for subsequent years). Altogether, imports represent approximately 10% of beef supplied, 4% of pork and 0.3% of chicken (Supplementary Table 7).

County domestic meat demand

As mentioned above, under-reporting biases produce underestimates of actual meat consumption in NHANES than actual consumption. Additionally, NHANES-based estimates represent final consumption quantities of meat, which has several layers of loss embedded throughout the supply chain and thus is less than the total chilled-carcass weight available from processing. As a result, we use the estimated distribution across counties from the NHANES-based regression analysis (that is, proportion of each meat type consumed in each county relative to the total amount of each meat consumed in the USA) and combine with the national total chilled-carcass weight used domestically58. In doing so, we avoid the issues with NHANES systematic under-reporting biases and embedded losses, and ensure that the total amount of chilled-carcass weight demanded across counties matches the total national domestic carcass weight, with around 12 Mt, 9.5 Mt and 15.5 Mt of beef, pork and chicken carcass weight, respectively59 (Supplementary Table 7).

Total loss-adjusted meat consumption

To estimate the final total quantities of meat purchased for final consumption in each county associated with the above estimated total chilled-carcass weight demanded, we must account for the losses that occur in processing and retail. In processing, losses occur from converting chilled-carcass weights to boneless edible meat, where bones and trim are removed. Additional losses occur at retail, where a portion of the meat delivered is disposed of because of expiration or other reasons it was deemed unfit for sale, and at consumption from cooking and uneaten food waste. Supplementary Table 8 shows the losses assumed at each stage for each meat type, resulting in 1 kg of chilled-carcass weight producing 0.63, 0.59 and 0.61 kg of edible beef, pork and chicken meat purchased for final consumption, respectively, and 0.50 kg, 0.42 kg and 0.52 kg of edible beef, pork and chicken meat consumed60.

Spatial carbon hoofprint

We use the FoodS3 estimated spatially explicit emissions of meat production at primary processing facilities and track these impacts through the supply chain to downstream counties of final consumption. The spatially explicit emissions of meat production are based on the Intergovernmental Panel on Climate Change (IPCC) AR5 Global Warming Potential (GWP) characterization factors with climate carbon feedbacks, and include differentiation of practices and emission profiles of upstream crop/feed production, livestock production, processing and transport across stages, producing unique cradle-to-processing gate emission intensity (kgCO2e per kg of edible boneless meat) at each processing facility (Supplementary Figs. 8 and 9). Upstream practices in crop/feed production consider regional differences in nitrogen fertilizer application rates and types; nitrous oxide rates based on soil crop type and management practices; cover crops on irrigated and non-irrigated lands; tillage regimes for no till, reduced till and conventional intensive till and implications for fuel use; irrigation quantities, water source, irrigation system and energy source; land-use change expansion rates and emission rates, with livestock practices differentiated across manure management systems, ambient temperatures, manure excretion rates, energy use and enteric fermentation rates, and processing differentiated by electricity grid profiles31,57. These spatially explicit inventories are based on a combination of USDA, Environmental Protection Agency, US Geological Survey, national laboratory, intergovernmental and non-governmental organization data and LCA databases, with extensive methodological details published previously31,56,57,61.

Emissions are allocated to fresh meat based on economic allocation approaches, which attribute 71%, 65% and 91% to beef, pork and chicken, respectively. In addition to domestic emissions, we also include emission for imports based on the distribution of countries that each meat type is imported from62 and the respective emission intensity63.

The cradle-to-processing gate emissions (kgCO2e per kg of boneless meat) thus represent the emissions associated with producing meat, delivered to the retailer. While emissions from transport are included across feed, livestock production and processing stages, we omit transportation impacts between processing facilities and final consumption to avoid double-counting with conventional urban GHG inventories.

We quantify the total emissions from meat purchased and consumed in each urban area by first estimating the total emissions in each consuming county using the estimated loss-scaled quantity of each meat type sourced from each processing facility and the respective cradle-to-processing gate emission intensity. Because cities often are composed of or straddle multiple counties, we estimate the fraction of each county population within each urban area and use this to allocate a portion of the total county emissions for each meat type to each city. Summing the allocated emissions across counties in each city and dividing by the total boneless edible meat purchased results in the unique hoofprint of each city. See Supplementary Note 3 for additional details about the spatial carbon footprint methodology.

Limitations and uncertainty analysis

While spatially explicit models can explore potential impact ranges and identify geographic and stakeholder-specific hotspots and mitigation opportunities, we acknowledge several sources of uncertainty. Supply-chain estimates are inherently uncertain as a result of the opaque nature of US agriculture commodity flows. We assume that supplier–buyer relationships are solely dictated by cost and that impedance between counties (that is, difficulty or ease of shipping between counties using existing rail, road and waterways) is a reasonable proxy for this. However, supplier–buyer relationships are also influenced by contractual agreements and brand preferences.

Quantifying the standard error in the FoodS3 estimates remains challenging, as publicly available ground-truth sourcing data for validation is limited. However, consultations with stakeholders in the beef, pork and poultry industries, suggest that FoodS3 model is generally consistent with proprietary sourcing data and stakeholder expectations. Additionally, FoodS3 aligns with product-specific sourcing estimates and broader commodity-group estimates (for example, ‘animal feed’) from the Oak Ridge National Laboratory freight analysis framework (FAF5). Uncertainty is expected to be greater for locations that source small volumes from a narrow set of regions, where sourcing patterns may be more variable. Conversely, locations sourcing large volumes are more likely to draw from a broader supply region, increasing the likelihood that modelled estimates reflect average sourcing behaviour.

To test this, we use Monte Carlo scenarios to estimate variation in results from uncertainty in buyer–supplier relationships. We modelled 1,000,000 unique, mass-balanced sourcing scenarios for each of the 3,351 cities. This equates to 3,531,000,000 unique simulations across all cities. Supplementary Note 4 provides a complete description of these simulations. Monte Carlo analysis estimates a mean absolute percentage error of 26.5% for our model across all cities (Supplementary Fig. 1). This drops quickly with city size (for example, <20% for cities with populations above 125,000). Full results of the uncertainty analysis are in Supplementary Fig. 2 and Supplementary Data 1.

We also estimate the effect of uncertainty in consumption on the results. We run 100 Monte Carlo simulations for each city (353,100 total simulations) varying consumption of each meat by each demographic group based on the statistical ranges estimated using NHANES (Supplementary Note 2). Results show that the mean absolute percentage error from baseline ranges from 0.02% to 5% for individual cities and is below 1% for all cities (Supplementary Fig. 7). Supplementary Note 5 provides a complete description of this uncertainty analysis.

Additional uncertainty from GHG inventories is also expected. Future work should overcome the computationally challenging task of quantifying this uncertainty at the county level and propagating it through downstream stages. LCA model uncertainty arises from several sources, including variability in input and output parameters (for example, material and energy types and quantities, manure management practices and environmental conditions), model selection (for example, IPCC tier 1, 2 and 3 methods; GWP characterization factors) and scenario assumptions and allocation decisions. Despite these uncertainties, our approach improves representativeness by integrating detailed, subnational production data and using tier 2 and 3 methodologies to reflect local environmental conditions. This enhances the spatial accuracy of emission estimates compared with using national averages, which may significantly overestimate or underestimate impacts. See Supplementary Note 6 for a more qualitative discussion of other sources of modelling uncertainty.

Comparisons with residential energy use and scenario

Residential energy use

We use the University of California, Berkeley’s Cool Climate Network Data (www.coolclimate.berkeley.edu) to compare the carbon footprint of residential energy use to the carbon hoofprint. The Cool Climate Network Data include estimates of emissions from in-home electricity, fuel oil and natural gas use per household for each zip code in the USA for the year 2013. The estimates were derived from multilinear regression models of fuel and electricity consumption combined with GHG intensity factors for fuels and electrical grid intensity factors from the US Department of Environment eGrid model. Details of the models underlying the Cool Climate Network Data can be found in ref. 64.

We compute total GHGs from household energy in each zip code by multiplying the number of households in each zip code by the per-household emissions. We then sum up these values for all the zip codes in each urban area and divide them by the total population to estimate the per capita GHGs from household energy in each urban area. We then compute the ratio of the per capita carbon hoofprint to per capita energy emissions for 3,531 urban areas. Spot checks with published urban GHG inventories for New York City65 and Los Angeles66 suggest that our method provides a conservative estimate of the scale of the hoofprint relative to emissions from residential energy use.

Scenarios

We explore four scenarios for hoofprint decarbonization. Scenario 1 looks at reducing edible food waste at retailers and households. We assume that baseline wastage is cut in half across both locations, leading to a decrease in total meat production as total meat demand remains constant. This means that edible beef losses are 15.1%, edible chicken losses are 13.8% and edible pork losses are 18.9%, effectively reducing emission intensity per kg of meat delivered by 18.5%, 23.4% and 16.0% respectively. Scenario 2 assumes that 50% of beef consumption is replaced by pork and chicken consumption by mass (loss-adjusted). For instance, the average annual US beef delivered to retail for final consumption would drop from 26.9 kg to 13.4 kg and average annual pork at retail would increase from 19.5 kg to 25.9 kg (chicken increases from 32.2 kg to 38.2 kg). Note that chicken required to supplant beef would decrease if substitution was based on protein and slightly increase if based on calories (pork would be largely unaffected because of similarity with beef). Scenario 3 mirrors scenario 2 except that chicken replaces pork. Lastly, scenario 4 models abstaining from meat once weekly, whereby consumption of beef, chicken and pork is multiplied by 6/7 for each city before calculating the hoofprint. As a point of comparison, we run an additional scenario considering production-side hoofprint reductions using beef silvopastures (that is, integrating trees and livestock grazing). See Supplementary Fig. 9 for results and earlier publications for methods31.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.