The food we eat, the air we breathe: a review of the fine particulate matter-induced air quality health impacts of the global food system

The global food system is essential for the health and wellbeing of society, but is also a major cause of environmental damage. Some impacts, such as on climate change, have been the subject of intense recent inquiry, but others, such as on air quality, are not as well understood. Here, we systematically synthesize the literature to identify the impacts on ambient PM2.5 (particulate matter with diameter ⩽2.5 μm), which is the strongest contributor to premature mortality from exposure to air pollution. Our analysis indicates that the life-cycle of the global food system (pre-production, production, post-production, consumption and waste management) accounts for 58% of anthropogenic, global emissions of primary PM2.5, 72% of ammonia (NH3), 13% of nitrogen oxides (NO x ), 9% of sulfur dioxide (SO2), and 19% of non-methane volatile organic compounds (NMVOC). These emissions result in at least 890 000 ambient PM2.5-related deaths, which is equivalent to 23% of ambient PM2.5-related deaths reported in the Global Burden of Disease Study 2015. Predominant contributors include livestock and crop production, which contribute >50% of food-related NH3 emissions, and land-use change and waste burning, which contribute up to 95% of food-related primary PM2.5 emissions. These findings are largely underestimated given the paucity of data from the post-production and consumption stages, total underestimates in NH3 emissions, lack of sector-scale analysis of PM2.5-related deaths in South America and Africa, and uncertainties in integrated exposure-response functions. In addition, we identify mitigation opportunities—including shifts in food demand, changes in agricultural practices, the adoption of clean and low-energy technologies, and policy actions—that can facilitate meeting food demand with minimal PM2.5 impacts. Further research is required to resolve sectoral-scale, region-specific contributions to PM2.5-related deaths, and assess the efficiency of mitigation strategies. Our review is positioned to inform stakeholders, including scientists, engineers, policymakers, farmers and the public, of the health impacts of reduced air quality resulting from the global food system.


Introduction
Global food demand increased threefold from 1960 to 2010 as a result of increasing population, rising incomes, and shifting dietary choices (Foley et al 2011, Tilman et al 2011. This demand has been met by intensive agricultural practices associated with 'Green Revolution' technologies, changing land management practices, and resource inputs as evidenced by a 700% increase in nitrogen fertilizer use, a 70% increase in irrigated cropland, and a 110% increase in land cultivation that now accounts for nearly 38% of Earth's terrestrial surface (Foley et al 2005, Ramankutty et al 2018. Consequently, agricultural intensification has resulted in widespread environmental damage including surface water eutrophication, groundwater contamination, hypoxia and the formation of dead zones in oceans, increased soil acidity associated with reduced crop productivity, biodiversity loss, climate change, and reduced air quality (Vermeulen et al 2012, Erisman et al 2013, Bauer et al 2016, Springmann et al 2018a. Air pollution is the leading environmental risk factor for mortality, linked to 3.9 million premature deaths in 2017 from exposure to ambient fine particulate matter (PM 2.5 , PM with diameter <2.5 µm) (Landrigan et al 2018, IHME 2020. Atmospheric PM 2.5 can result either through direct emissions as primary PM 2.5 or is formed through chemical reactions as secondary PM 2.5 from precursors that include ammonia (NH 3 ), nitrogen oxide (NO x ), sulfur dioxide (SO 2 ) and non-methane volatile organic compounds (NMVOC). Of all air pollutants, PM 2.5 is the strongest contributor to premature mortality, resulting largely from respiratory disorders, cardiovascular disease and stroke (Burnett et al 2018), and thus is widely regulated with the goal of reducing ambient PM 2.5 concentrations. PM 2.5 is short-lived in the atmosphere with a lifetime of a few days to a week, but it can be transported regionally, resulting in human health damage up to several thousand kilometers downwind from the source itself Zhang 2014, Goodkind et al 2019).
Historically, emissions reductions of PM 2.5 and precursor pollutants have been achieved by regulating major anthropogenic sources, such as power plants, industries and transportation (Bachmann et al 2007). Of emerging concern is agriculture, which has been identified as a significant contributor to global ambient PM 2.5 (Bauer et al 2016, Giannadaki et al 2018 and is linked to nearly 20% of all global ambient PM 2.5 -related deaths (Lelieveld et al 2015). In the United States, emissions from agriculture have been linked to 15%-25% of all PM 2.5 -attributable deaths (Goodkind et al 2019, Thakrar et al 2020). Giannadaki et al (2018) presented an economic case to mitigate agricultural emissions in Europe, finding that a 50% reduction in emissions could reduce PM 2.5 premature deaths by 18%, with a saving of 89 billion USD.
Most research examining air pollution from the global food system focuses on agriculture (e.g. Aneja et al 2015), but the global food system is expansive and encompasses all life cycle stages of food production, use and disposal (Vermeulen et al 2012). Few studies have examined the human health damage that results from air pollution generated by the global food system (Sun et al 2017). Here, we present a systematic review and an order of magnitude estimate of the contribution of emissions from the global food system to ambient PM 2.5 -attributable deaths. We expand beyond the historical focus on agricultural production to account for emissions from sectors associated with the pre-production, post-production, consumption and waste management of food. Specifically, we follow the causal chain of emissions to health impacts to (a) describe emission pathways and determine national-scale emission totals for 15 sectors within the food system that emit five pollutants of interest (primary PM 2.5 , NH 3 , NMVOC, NO x , SO 2 ), (b) summarize studies that estimate impacts of sectorscale emissions on ambient PM 2.5 formation and PM 2.5 -attributable deaths and (c) identify strategies to reduce PM 2.5 pollution burden within and outside farms. By adopting a system-scale approach that expands beyond the historical focus on agricultural production, our review establishes emissions contributions and PM 2.5 -attributable deaths resulting over the life-cycle of the global food system.

Data and methods
To define the overall scope of this review, we first determined the five key stages that span the lifecycle of the global food system by building on the concept of Vermeulen et al (2012). We then identified emissions sectors within each stage of the food system based on the emissions categories defined by the EMEP/EEA (European Monitoring and Evaluation Programme by the European Environment Agency) inventory guidebook, and used the Nomenclature for Reporting to establish system boundaries for each sector (EEA 2016). We also identified and gap-filled the missing sector of land-use change. Our efforts resulted in 15 emissions sectors aggregated by five stages, as shown in figure 1: (a) pre-production: landuse change, fertilizer production, (b) production: onfarm energy use, manure management, grazing, fertilizer use, agricultural waste burning, and other, (onfarm handling of agricultural products, standing crop emissions), (c) post-production: food industry, retail and distribution, transportation, (d) consumption: commercial cooking, residential cooking (not reported in this review), and (e) waste: open burning, controlled disposal (open disposal, uncontrolled incineration, controlled incineration, landfilling and composting). We then adopted a systematic approach to identify and select relevant scientific literature and analyze relevant findings as defined by Uman (2011). First, the literature was located using scientific databases (Scopus, Google Scholar and Web of Science) by iteratively choosing the preliminary keywords of 'agriculture' , 'emissions' , 'food' , 'air pollution' , 'PM 2.5 ' and 'premature mortality' . This search yielded 4746 peer-reviewed English language publications from the past decade (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020). However, this process did not identify several key studies that examined specific emission sectors. Thus, we systematically expanded the search using additional keywords, by using a combination of each of the 15 emissions sectors, pollutants (PM 2.5 ,NH 3 , NO x , SO 2 , NMVOC), and mitigation strategies (see table 1) to identify an additional 1384 publications. We then applied the following criteria to sub-select relevant studies based on their abstract and introduction sections. Inclusion criteria were: (a) description of mechanisms and magnitudes of primary PM 2.5 and secondary PM 2.5 precursor emissions, (b) air quality studies to quantify the enhancement of secondary ambient PM 2.5 and (c) mitigation strategies to reduce the PM 2.5 pollution burden. Exclusion criteria were: (a) hazardous air pollutant emissions from agriculture, (b) ground-level ozone formation and (c) sub-national scale studies using both modeling and measurement approaches to study contributions of the food system to ambient PM 2.5 concentrations. As a caveat, we do not explicitly show trade and associated emissions flows; instead, emissions are attributed to the geographic domains where sources are located. External to the scope of this review are related topics such as air pollution impacts on agricultural productivity, atmospheric deposition impacts of reactive nitrogen on ecosystems, pathways of global food demand and dietary shifts.
In addition to the archival literature, we also obtained data from highly curated institutional repositories to ensure consistency in data quality across geographic domains. The main data set of interest is the Emissions Database for Global Atmospheric Research (EDGAR4.3.2) that reports annual emissions of primary PM 2.5 and PM 2.5 precursors that are differentiated by activity, use of fuel and technology, pollutant type and end of pipe abatement (Crippa et al 2018). We also scoped the following databases: EMEP/EEA emissions factor database (EEA 2016) for sectoral-scale and pollutant-specific emissions factors, the World Bank for demographic (World Bank 2020) and waste management data (World Bank 2018a), FAOSTAT for data on landuse and land-use change, food production, fertilizer production and livestock management (FAO 2020a), the Global Fire Emissions Database (GFED4) for landscape fire data (van der Werf et al 2017), and environmentally extended input-output models including the World Input-Output Database (WIOD)  (Timmer et al 2012) and EXIOBASE3.3.17 (Merciai and Schmidt 2016 for emissions from fertilizer production and the food industry. Overall, of the 6130 records identified, 322 studies, data sets and reports have been synthesized in this review. Of these, only 19 studies established PM 2.5attributable health damage either as premature deaths or economic damage from sectors relevant but not exclusive in terms of contributions to the global food system. Only two studies by Sun et al (2017) and Malley et al (2021) examine linkages between air quality and the global food system. Sun et al (2017) qualitatively linked the air quality impacts on the production and processing of food, human health in the form of productivity, and the role of markets, trade, and agricultural and energy policies, while Malley et al (2021) employed an air quality model to estimate the impacts of emissions from global food production on PM 2.5related deaths. Here, we explicitly present nationalscale emissions contributions of primary PM 2.5 and secondary PM 2.5 precursors from the global food system, and synthesize studies that link these emissions to increases in ambient PM 2.5 exposure and premature deaths. We organize the rest of our review as follows: section 3 provides a description and estimates of sector-specific, national-scale emissions of primary PM 2.5 and secondary PM 2.5 precursors; section 4 summarizes the resulting impacts on ambient PM 2.5 and premature mortality; section 5 identifies tools to reduce PM 2.5 pollution from the food system, and section 6 presents highlights and conclusions.
3. PM 2.5 pollution burden from the global food system 3.1. Sectoral-scale emissions: description and estimates There are multiple approaches to developing air pollutant emissions inventories. A common approach is the use of emission factors, which represent how much pollutant is emitted by a unit of source activity. The emission-factor approach is readily scalable across regions and thus widely implemented, such as in the National Emissions Inventory for the United States (US EPA 2018) and EDGAR4.3.2 (Crippa et al 2018). Other approaches, particularly for agricultural production, include the use of process models that simulate physical, chemical and biological processes governing pollutant release at the field scale (Brilli et al 2017), and are integrated to develop regional-scale emissions inventories as input to air quality models (AQMs) (Cooter et al 2012, Balasubramanian et al 2015). Inverse modeling approaches have also been used recently to constrain emissions using observations assimilated from ground or satellite platforms, as in the case of improving the seasonality in NH 3 emissions (Zhu et al 2015b, van Damme et al 2018. We derive sector-specific, national-scale emissions of PM 2.5, and PM 2.5 precursors from EDGAR4.3.2 (Crippa et al 2018) that have been widely used as input to AQMs. EDGAR4.3.2 uses the Nomenclature for Reporting to establish system boundaries for sectors that follow initiatives such as the Convention on Long-Range Transboundary Air Pollution to minimize double-counting of emissions (EEA 2016). However, not all the emissions from the food system are accounted for in EDGAR4.3.2, and for many sectors, these emissions are not explicitly reported for the food system. We have thus supplemented data from other databases including GFED4 (van der Werf et al 2017) to estimate landuse change emissions and environmentally extended input-output models, such as WIOD (Timmer et al 2012) and EXIOBASE3.3.17 (Merciai and Schmidt 2016 for emissions from fertilizer production, food industry and waste, using similar system boundary definitions. We also identified the share of production for food versus non-food purposes based on data reported in the National Food Balance Sheets (FAO 2020b) and applied the fractional contribution of food to estimate emissions for relevant sectors. We present an in-depth discussion of each sector in section 3.1.1 and provide a summary in table 2.
3.1.1. Pre-production 3.1.1.1. Land-use change Agriculture is the primary driver of deforestation especially in the tropical regions of South America and Southeast Asia (Fuchs et al 2018, Song et al 2018), and is largely driven by global food demand and international trade (Pendrill et al 2019). As of 2000, 50% of the habitable land has been diverted to grow food for human consumption and feed for livestock production (Ellis et al 2010). Increasing demand for food crops, cattle and timber has been linked to recent increases in forest clearing since 2017 in the Brazilian Amazon (De Oliveira et al 2020) and industrial oil palm productions in equatorial South-East Asia where 30% of the native peatland has been converted since 1990 (Miettinen et al 2016). Land clearing for shifting agriculture or permanent conversion to cropland is typically achieved through fires, while other practices, such as drainage of peatland increase susceptibility of these landscapes to fires (Martin 2019). Fires emit NO x , PM 2.5 , NH 3 and NMVOC, which are influenced by vegetation type and meteorology (Crutzen and Andreae 1990, Andreae and Merlet 2001, Akagi et al 2011, and have been linked to hazardous levels of PM 2.5 over the Amazon (Reddington et al 2015) and in Southeast Asia (Kiely et al 2019). Our review did not identify any studies that estimated primary PM 2.5 and precursor emissions resulting from food-demand driven landuse change. Instead, we designed an approach based on GFED4 that reports emissions that are derived using satellite-derived burned area and vegetationtype specific emissions factors (van der Werf et al 2017) and reported for 14 ecological regions that are aggregated to the following categories: savanna, grassland and shrubland, boreal forests, temperate forests, deforestation, peatland and agricultural waste burning.
We adopted the following method to derive PM 2.5 and precursor emissions totals for land-use change. First, we extracted national-scale emissions from GFED4 for Asia, Africa and South America that experience large-scale deforestation (Carter et al 2018) for the categories of savanna, grassland, shrubland, deforestation and peatland. Second, we extracted the extent of forest loss driven by wildfires, shifting agriculture and conversion for agriculture for the years 2012-2015 (World Resources Institute 2014). By combining forest loss data with GFED4, we identified emissions from shifting agriculture and permanent land-use change for agriculture. Finally, we identified the share of production for food versus non-food purposes based on the National Food Balance Sheets (FAO 2020b), and apply the fractional contribution of food to estimate emissions from land-use change. Similarly, GFED4 emissions for peatland were combined with the national-scale fractions of peatland fires on oil palm plantations (Miettinen et al 2016, Petersen et al 2016 and the fraction of oil palm diverted for food purposes (70%) (Lai et al 2012).

Fertilizer production
Agrochemicals including fertilizers, herbicides and pesticides have been widely used to maximize crop yields and for disease and pest management. The Haber-Bosch process, which was invented in the early 1900s, enabled the conversion of inert nitrogen to NH 3 to produce nitrogen fertilizers, which has dramatically altered agricultural production (Sutton et al 2011). Since 1960, croplands have received 300% more nitrogen from synthetic fertilizers than from natural biological nitrogen fixation, thus supporting nearly 48% of all crop production (Erisman et al 2008b). Global fertilizer production increased by 520% between 1960 and 2014 (FAO 2020a), resulting in large on-site emissions of primary PM 2.5 and NH 3 , as well as emissions of PM 2.5 , SO 2 , NO x and NMVOC from embodied energy. Satellite data have identified 158 hotspots of NH 3 emissions over fertilizer production sites in China, Ukraine, Iran and the United States (Van Damme et al 2018).
EDGAR4.3.2 aggregate emissions from nitrogen fertilizer production into the 'Industrial Processes and Product Use' category (EEA 2016). While the emissions-factor approach can be replicated by combining emissions factors for the production of NH 3 and other fertilizer types (EEA 2016) and scaled using agrochemical production statistics (FAO 2020a), it is challenging estimating emissions from embodied Table 2. System-scale emissions inventory of the global food system. Annual emissions of primary PM2.5 and secondary PM2.

Livestock management
Human demand for animal-based food has quadrupled since 1961, with meat production increasing by 200% in Europe and North America, and significantly larger increases in Asia (1500%) and South America (530%) (FAO 2020a). Subsequently, manure-nitrogen production has increased by 520%, with regional contributions dominated by Asia (34%), Africa (17%) and South America (15%) (Zhang et al 2017a). Livestock systems are highly nitrogen inefficient, as a large fraction (45%-95%) of nitrogen from the feed is excreted as manure and urine (McQuilling and Adams 2015), which decomposes and is subsequently emitted as NH 3 through volatilization of nitrogen (Behera et al 2013). Livestock operations are also associated with primary PM 2.5 emissions from the movement of livestock within facilities (Ni et al 2009, Yang et al 2011, and trace emissions of NMVOC (Hobbs et al 2004) and SO 2 (Lim et al 2003). Globally, NH 3 emissions from livestock management are attributed to the production of cattle (43%), goats and sheep (33%), swine (11%) and poultry (10%) (Zhang et al 2017a), and can occur at multiple stages in the livestock management system: from accumulated manure in housing, yard and storage facilities (31%-55%), land application for crop cultivation (23%-38%) and from livestock grazing (17%-37%) (Beusen et al 2008, Dämmgen andHutchings 2008). The most important factors determining NH 3 emissions are the type of livestock, its age and the nitrogen content in the feed (Beusen et al 2008). Emissions from manure storage and handling depend on the surface area and bedding material. As a result, larger losses are observed in open housing with solid or slatted floors compared to cubicle houses, deep litter and closed manure storage systems (Dämmgen and Hutchings 2008). Emissions from manure application to crops are highly dependent on environmental conditions and application mode, with increased emissions positively correlated with higher temperatures, wind speeds and lower moisture content (Webb et al 2010).
NH 3 emissions from livestock rearing have received extensive attention in the development of emissions inventories and through multiple, targeted measurement campaigns in Europe and the United States (Slattery 2005). These efforts have resulted in detailed emissions factors that are differentiated by livestock type and manure management operation (housing, storage and handling, grazing and manure application to soils) (Battye et al 1994, EEA 2016, that are implemented in EDGAR4.3.2 (Crippa et al 2018). Additional approaches have been developed to better capture spatial and temporal heterogeneity in emissions. Semi-empirical models, such as the Farm Emissions Model fine-tune existing emissions factors by estimating NH 3 losses based on mass balances and mass transfer processes that are influenced by meteorology (McQuilling and Adams 2015). Process models have also been implemented to develop emissions inputs from manure management to AQMs (Deng et al 2015, Giltrap et al 2017. However, given the large data requirements to capture manure management systems and environmental conditions, and the need for calibrated models to capture regionspecific variability, these approaches are yet to be scaled globally. Here, we obtain national-scale emissions of primary PM 2.5 , NH 3 and NMVOC from EDGAR4.3.2 that are differentiated by livestock-type for the categories of manure handling and storage, manure application and grazing.

Fertilizer use
The application of synthetic fertilizers for crop cultivation is one of the most important land management practices to increase soil fertility and crop yields. Global nitrogen inputs to crops increased by 850% between 1960 and 2013 (Lu and Tian 2017) to meet the demand for food, animal feed and biofuels. Large regional variations exist in nitrogen use, ranging from 0.15-6 kg N ha −1 in sub-Saharan Africa to 100-200 kg N ha −1 in cropland in Asia (Lu and Tian 2017). Global NH 3 emissions increased from 1.9 to 16.7 Tg N between 1961 and 2010 (Behera et al 2013), 67% resulting from the cultivation of rice, corn, wheat and soybeans (Xu et al 2019). Depending on the fertilizer type, amount and mode of application, and weather and soil conditions, 1%-64% of the applied nitrogen can volatilize as NH 3 (Sommer et al 2004, Balasubramanian et al 2017, thus representing a major financial loss to farmers (Pan et al 2016). Urea, which is the most commonly used fertilizer globally (Behera et al 2013), has a volatilization potential 22%-55% higher than other nitrogen forms (Goebes et al 2003, EEA 2016, Pan et al 2016. Similar to livestock rearing, NH 3 emissions from fertilizer use are estimated using the emission-factor approach as in EDGAR based on fertilizer-type specific emission factors (Crippa et al 2018). However, this approach introduces large uncertainties as it does not capture the impact of crop management and the resulting spatial and temporal heterogeneity that has been identified at the farm scale ( Similar to the livestock sector, we thus obtained national-scale NH 3 emissions from EDGAR4.3.2 that were proportionally adjusted for contributions for food versus non-food purposes using data from National Food Balance Sheets (FAO 2020b).

Agricultural waste burning
Open burning of agricultural waste is a low-cost way to dispose of crop residues left over after harvesting, land clearing and pest control (Crutzen and Andreae 1990, Akagi et al 2011). Annual agricultural waste burning increased by 150% between 1960 and 2015 (FAO 2020a), releasing large amounts of primary PM 2.5 (1.76 Tg), NH 3 (0.6 Tg), SO 2 (0.11 Tg), NO x (0.08 Tg) and NMVOC (0.11 Tg) (van der Werf et al 2017). Several studies have examined the impacts of agricultural waste burning at regional scales. In the United States, the burning of corn, cotton, bluegrass, rice, soybeans, sugarcane and wheat residues was linked to local increases in ambient PM 2.5 (Pouliot et al 2017). Similarly, the burning of rice, corn and wheat straw residue in China was linked to PM 2.5 emissions (Ni et al 2015), which may have been underestimated (Li et al 2017a). Burnt agricultural residue in India from managing rice (43%), wheat (26%), sugarcane (10%) and cereal residues (11%) (Ravindra et al 2019) has been linked to a 600% increase in ambient PM 2.5 during the harvest season (Jethva et al 2018). In Southeast Asia, rice straw burning dominated PM 2.5 emissions (95%-98%), largely driven by crop production in Indonesia (25%-39%), Vietnam (17%-30%), Myanmar (8%-19%) and Thailand (7%-16%). Emissions of primary PM 2.5 , NH 3 , NO 2 , SO 2 and NMVOC from agricultural waste burning have been reported using the emissions-factor approach in both the GFED4 (van der Werf et al 2010) and EDGAR4.3.2 (Crippa et al 2018). Here, we obtain national-scale emissions from EDGAR4.3.2, which are proportionally adjusted for food versus non-food contributions by using data from the National Food Balance Sheets (FAO 2020b).

Other emissions
On-farm operations including plowing, tilling and harvesting, and on-farm handling and storage of agricultural products are typically associated with emissions of coarse PM that result from the attrition of dry plant particles, silica, biological species including molds, pollen, spores, bacteria, fungi and agrochemical residues. On-farm operations also emit primary PM 2.5 (Aneja et al 2009, van Grinsven et al 2013, with contributions ranging from 2%-5% of the total anthropogenic, primary PM 2.5 emissions in Europe (Erisman et al 2008a, Oenema et al 2012 and Canada (Pattey and Qiu 2012) to 15% in the United States (Penfold et al 2005). Crops also naturally emit NMVOC, including isoprene, monoterpenes and sesquiterpenes, among 50 other identified species as a part of normal growth (Lamb et al 1993, König et al 1995, Laothawornkitkul et al 2009 or as a defense mechanism that can be triggered during harvesting (Guenther et al 2000). Miscellaneous sources include emissions from pesticides and NH 3 emissions from treated straw that is used as ruminant feed. Here, we follow the EDGAR4.3.2 methodology to estimate on-farm primary PM 2.5 emissions and NMVOC emissions from standing crops by combining national-scale crop production data (FAO 2020a) with emissions factors (EEA 2016). We exclude emissions from pesticide application and treated straw as they are assumed to be negligible. These estimates are adjusted for food demand using data from the National Food Balance Sheets (FAO 2020b).

Post-production 3.1.3.1. Food industry
Food and beverage manufacturing (here, the 'food industry') includes industrial manufacturing of food ingredients and products that are processed and packaged typically for retail. The global food industry annually consumes 200 EJ energy (Ladha-Sabur et al 2019), accounting for 4% of the industrial energy consumption in OECD (Organisation for Economic Co-operation and Development) countries and 2% in non-OECD countries (EIA 2016). The reporting of sub-national-scale energy embodied in the food industry is fragmented and only for select commodities (Ladha-Sabur et al 2019). The industrial processing of food products emits primary PM 2.5 , NH 3 , NO x , SO 2 and NMVOC (US EPA 1995) as a result of embodied energy use and on-site operations. While the emissions-factor approach can be implemented to estimate these emissions, the lack of harmonized data on national-scale fuel and technology used to power the food industry, and how food commodities are produced, limit these efforts. Here, we obtain data for emissions of primary PM 2.5 , NH 3 , NO x , SO 2 and NMVOC from EXIBOASE3.3.17 for the production and processing of meat from cattle, poultry and pigs, vegetable oils and fats, dairy products, processed rice, sugar refining, beverages, seafood products and miscellaneous food commodities (Merciai and Schmidt 2016). These emissions are reported for 43 countries and for five regions for all other countries, which we distributed proportionally to the national population to gap-fill data.

Retail and distribution
Energy use in food retail is driven by business size, nature of products sold and use of equipment for onsite food preparation and preservation (Vermeulen et al 2012, Ladha-Sabur et al 2019. Commercial refrigeration is highly energy-intensive, accounting for 15% of global electricity consumption (James and James 2010). We identified only one study (hereafter DEFRA report) that reported primary PM 2.5 and secondary PM 2.5 precursor emissions from food retail and distribution. The report provided relative emissions contributions for the food industry, retail and distribution, and food transportation, but was limited to the United Kingdom (Smith et al 2005). Given the lack of such data at the global scale, we combined the relative contributions from the DEFRA report with the national-scale food industry emissions derived from EXIOBASE3.3.17 to estimate national-scale food retail and distribution emissions. As a caveat, the United Kingdom is a highincome country. Thus, our approach will result in higher magnitudes of emissions than expected at the global-scale and be reflective of supply chain management trends that low-income countries may adopt in the future.

Transportation
The transportation of food or 'food miles,' is a popular albeit often misapplied, indicator to assess the sustainability of food commodities (Schnell 2013). While the impact of food miles on greenhouse gas (GHG) emissions (Pirog et al 2001, Weber andMatthews 2008) and along supply chains of specific commodities (Brodt et al 2013, Brunori et al 2016, Schmitt et al 2016 have been studied, the focus on air pollutant emissions is rather limited. We identified only one study reporting PM 2.5 emissions from food miles, which was limited to the United Kingdom (Smith et al 2005). Transportation modes have a significant impact on emissions, with lower reported emissions per km-tonne for food moved by ship and rail in comparison to cars and trucks. While food commodity flows by transportation modes are reported for Europe (Eurostat 2019) and the United States (Federal Highway Administration 2014), limited data coverage on transportation choice and fuel use at the global scale limits the estimation of primary PM 2.5 and secondary PM 2.5 precursor emissions. Freight transport of goods including food commodities has been linked to PM 2.5 -related health impacts resulting from emissions of PM 2.5 and NO x (Liu et al 2019). It is imperative to establish the global-scale air quality impacts of transportation occurring as a result of food trade (Dalin and Rodríguez-Iturbe 2016), given that 25% of the food produced globally is traded (Odorico et al 2014). Similar to the retail and distribution sectors, we combine relative contributions of food transportation from the DEFRA report with national-scale food industry emissions estimates, without accounting for miles from retail to home.

Food preparation and consumption sectors 3.1.4.1. Commercial cooking
Several studies have examined the contribution of commercial cooking to ambient PM 2.5 pollution in urban settings (Robinson et al 2006, Gysel et al 2018, through emissions of ultrafine particles (PM with diameter <0.1 µm) that are retained longer in the lungs and cause more pulmonary infections than PM 2.5 (Schraufnagel 2020), and NMVOC in the form of n-alkanes, furans, lactones, polycyclic aromatic hydrocarbons and cholesterol (Rogge et al 1991). Commercial cooking often elevates PM 2.5 , especially ultrafine fractions (PM 2.5 ⩽ 0.1 µm) several orders of magnitude higher compared to the urban background and to larger extents than congested roadways (Robinson et al 2018) and smoking (Nasir and Colbeck 2013). These emissions are influenced by practices including cooking style, the temperature, duration of cooking and type of cooking oil (Abdullahi et al 2013, Torkmahalleh et al 2017. Commercial cooking impacts not only in-house workers, but elevates ambient PM 2.5 concentrations (50%-300%) and drives spatial patterns in PM 2.5 exposure in neighboring urban areas (Robinson et al 2018, Saha et al 2019, with disparate socio-economic impacts given the demographics of the population living in proximity to restaurants (Shah et al 2020).
Only the United States reports commercial cooking emissions that are classified by the equipment type and amount of food cooked (Roe et al 2004, US EPA 2018. Commercial cooking accounts for 1% of national PM 2.5 emissions resulting from underfired-char broilers (78%), conveyorized charbroilers (10%) and flat griddle frying (12%) (Roe et al 2004). Commercial food establishments account for a large fraction of the energy consumption (28%-34%) in the United States (Todd 2017), and this fraction is increasing globally (Fryar et al 2018). The lack of similar emissions reporting for other countries limits efforts to develop a global emissions inventory. Here, we do not quantify commercial cooking emissions, given data constraints and endemic challenges in delineating indoor-outdoor emissions contributions. However, given that this sector accounts for 1% of the PM 2.5 national emissions and an increasing shift towards consumption of food from commercial cooking, this source may be of increasing importance for urban air pollution, and should be revisited.

Household cooking
Much of the focus on cooking and PM 2.5 pollution has been on household air pollution resulting from solid fuel use, which is a major health risk in developing countries (Smith andPillarisetti 2017, Goldemberg et al 2018). In 2017, 3.6 billion people, primarily in South Asia, East Asia and sub-Saharan Africa, were exposed to elevated household PM 2.5 concentrations resulting from the use of solid fuels, such as wood, charcoal, coal and other biomass (Health Effects Institute 2019). Similar to commercial cooking, household cooking emits primary PM 2.5 , NMVOC and trace levels of NO x and SO 2 , that are dependent on fuel type (Sidhu et al 2017) and cooking practices, such as food and oil type, cooking temperature and duration, type and efficiency of cooking appliance, and indoor ventilation (Rehfuess et al 2011, Hu et al 2012. A large body of the literature has examined emissions from solid fuel use in various settings. Example studies include laboratory measurements (Roden et  The reported PM 2.5 emissions factors (g MJ −1 ) are highly variable (0.01-1.5), with lower emissions rates observed for electric and gas stoves, and nearly an order of magnitude higher for natural-draft and traditional cookstoves fueled by charcoal, wood and residue (0.06-1.8) (Arora and Jain 2016). Average emissions factors (g kg −1 ) for primary organic aerosols, SO 2 , NMVOC and NO x have been compiled for mud stoves (5.7, 0.3, 2.7 and 1.0 respectively), conventional wood stoves (3.9, 0.2, 23.6 and 2.8), wood boilers (1.5, 0.3, 14 and 1.2) and coal-burning stoves (0.8, 0.2, 0.5 and 2.2) (Bond et al 2013). Average emission rates for outdoor cooking to model personal exposure were found to be substantially higher than for indoor cooking (Edwards et al 2017). Hu et al (2012) compiled a PM 2.5 emissions database for residential environments in the United States and identified lower emissions rates for microwave and oven use (0.64-0.7 mg h −1 ) and 200%-300% higher for frying irrespective of oil type.
EDGAR4.3.2 does not account for ambient PM 2.5 and precursor emissions from household cooking. These contributions, which are specific to ambient air pollution, are instead reported by the GAINS emissions model based on the methodology by Chafe et al (2014). Household fuel use for cooking and heating is a significant contributor to anthropogenic PM 2.5 emissions, ranging from 20%-55% globally (Tao et al 2016, Pervez et al 2019. Here, we do not further compile a global emissions inventory for cooking. There are multiple opportunities to develop further research on the impacts of household cooking on PM 2.5 -attributable premature deaths. Topics of interest to the broader conversation of the sustainability of food systems include (a) cookstove technologies (Arora and Jain 2016) and the impacts on PM 2.5 -attributable health damage (Grieshop et al 2011), (b) socio-economic and air quality impacts of carbon-financing schemes and national-scale fuel

End-of-life disposal practices
Food loss and waste occur at all stages of the food supply chain (Parfitt et al 2010). Food losses of >40% are common in developing countries during the production and post-harvest stages, typically through agricultural waste burning due to inefficient technologies and poor infrastructure. Food waste of >40% at the retail and consumer stage is typical in developed countries and nearly equals the net food production in sub-Saharan Africa (Lipinski et al 2013). Household loss is the most important source of food waste with large per-capita variation, ranging from 6-11 kg yr −1 in sub-Saharan Africa to 95-115 kg yr  (Wiedinmyer et al 2014). However, this analysis is not exclusive to food waste. Here, we derive emissions of primary PM 2.5 , NH 3 , NO x , SO 2 and NMVOC for 43 countries from EXIOBASE3.3.17, and gap-fill data for other countries by combining national-scale solid waste data that are classified by waste management method (World Bank 2018a) and technologyspecific emissions factors (EEA 2016).

Global emissions inventories
We present national-scale emissions inventories of primary PM 2.5 and secondary PM 2.5 precursors from the global food system, reported for the year 2015 or the most recent year of available data, following the methods we describe at the sector-scale in section 3.1. Global emissions totals of primary PM 2.5 and secondary PM 2.5 precursors are shown in table 2 with fractional sector contributions shown in figure 2. Figures 3 and 4 show national emissions totals and regional-scale fractional sector contributions, respectively. Overall, we find that the global food system is a major contributor to the anthropogenic emissions of primary PM 2.5 (58%), NH 3 (72%), SO 2 (9%), NO x (13%) and NMVOC (19%) in comparison to total anthropogenic emissions reported in EDGAR4.3.2 (Crippa et al 2018). We estimate that the global food system emits 24 Tg primary PM 2.5 , driven by fires for land-use change (60%), agricultural waste burning (28%) and open burning of food waste (6%). The dominant emission sources of primary PM 2.5 vary regionally. Land-use change was identified as the predominant source in South America, Africa and Asia, while crop management and on-farm energy use dominate primary PM 2.5 emissions in North America and Eastern Europe, and China and Russia, respectively.
Global NH 3 emissions (42 Tg NH 3 ) largely result from livestock manure management (40%), grazing (20%) and synthetic fertilizer use (33%), with large variations in relative regional contributions. Fertilizer use is also a dominant contributor to NH 3 emissions in Asia and North America (40%-45%) in contrast to smaller contributions in Africa (<10%), where the slower adoption of nitrogen fertilizers and inefficient manure handling practices (Ndambi et al 2019) result in more than 50% contributions from livestock management. Of the 32 Tg NMVOC emitted from the food system, the dominant contributors included manure management (58%) and agricultural waste burning (12%). Smaller emissions totals were estimated for SO 2 (9 Tg) and NO x (16 Tg), which are typically a result of combustion. SO 2 was linked to onfarm energy use (35%), post processing of food (30%) and open burning (15%), with similar trends for NO x (35%, 30% and 6%, respectively).

PM 2.5 exposure and PM 2.5 -attributable deaths from the food system
We describe the causal pathway of emission impacts on ambient PM 2.5 concentrations and PM 2.5attributable premature deaths in section 4.1, summarize studies that report PM 2.5 -attributable premature deaths from sectors within the global food system to develop an overall estimate of PM 2.5 -attributable premature deaths in section 4.2, and discuss uncertainties in section 4.3.

4.1.
Connecting the emissions-PM 2.5 exposure-premature mortality pathways Ambient PM 2.5 concentrations are a result of precursor emissions, and are impacted by transport, chemistry and removal processes in the atmosphere. Of key importance to the discussion here are the emissions of NH 3 , 72% of which is emitted from the food system (section 3.2). As the most dominant alkaline component in the atmosphere, NH 3 neutralizes acids formed from atmospheric To date, the analysis of emissions contributions to PM 2.5 -attributable deaths has been limited to economic sectors, such as energy and transport (Lelieveld et  The typical approach is to sequentially: (a) generate emissions inputs to AQMs, (b) develop spatially resolved AQM predictions of ambient PM 2.5 concentrations, (c) estimate population-weighted PM 2.5 exposure and (d) finally scale PM 2.5 exposure using IER functions to estimate PM 2.5 -attributable premature deaths. The AQM framework has been implemented using two approaches (Conibear et al 2018).
In the 'zeroed out' approach, emissions from a sector of interest are zeroed or reduced and the resulting PM 2.5 deaths are attributed as source contributions. Alternatively, in the 'attribution' approach, sectorspecific mortality is estimated in proportion to the fraction of the sectoral contribution to PM 2.5 concentrations, either by examining emissions contributions or in models that 'tag' PM 2.5 concentrations as marginal changes in emissions. Given that emissions totals in the two approaches differ and due to the nonlinear emissions-PM 2.5 exposure responses, estimates of premature mortality can vary, especially in populated regions (Conibear et al 2018).
The analysis of health damage beyond broad economic sectors has been limited due to the large computational, data and resource requirements when using AQMs. Advances in high-performance computing, the use of alternative statistical approaches and the development of other models, such as reduced complexity models (RCMs) have enabled AQM assessments at high spatial resolution and for multiple scenarios. RCMs use simplified representations of atmospheric processes with variable grid sizes and leverage outputs from an existing AQM simulation to predict marginal changes in ambient PM 2.5 concentrations at high spatial resolution in response to marginal changes in precursor PM 2.5 emissions, with reduced computational times (Tessum et al 2017). RCMs have been widely implemented to study contributions of emissions to PM 2.5 -attributable premature deaths from various economic sectors at high spatial scales ( 4.2. Global food system linked to significant PM 2.5 -attributable premature deaths 4.2.1. Summary of studies discussing the impact on ambient PM 2.5 -attributable premature deaths Given the large contribution of the global food system to primary PM 2.5 and NH 3 emissions, and the central role of NH 3 in the formation of secondary PM 2.5 , we identify the lack of a system-scale analysis on the contribution of the global food system to PM 2.5 -attributable premature deaths as a key literature gap. Here, we briefly discuss AQM studies that link emissions from different stages and emissions sectors within, but not exclusive to the food system to ambient PM 2.5 -attributable premature deaths. Key findings are summarized in table 3, which highlights differences in the approaches used by the listed studies in terms of spatial extent of analysis, choice of emission inventories and AQM configurations, and the reporting of health damage. Much of the focus on the impacts of the global food system on air quality has been on agricultural production. Emissions from agricultural production contribute to about 20% of PM 2.5 deaths worldwide (Lelieveld et al 2015), with larger contributions in China, the United States and Europe (45%-55%) and smaller contributions in India and Africa (5%-15%) ( (2021) linked global agricultural production to 537,000 PM 2.5 -related deaths. Notably, a 100% reduction in these emissions would reduce 800 000 (95% confidence interval (95% CI): 420 000-980 000) global, annual PM 2.5 -attributable premature deaths (Pozzer et al (2017). Achievable health benefits were identified to be the largest for Europe and North America (70%-80%) where significant reductions in NO x and SO 2 emissions have already been achieved and PM 2.5 formation is highly sensitive to NH 3 emissions (Pozzer et  The morbidity and mortality costs of 1 kg NH 3 emitted into the atmosphere showed large spatial variability (0.1-73 USD) and were valued to be much larger than the marginal damage that results from emissions of SO x (0.2-12 USD) and NO 2 (0.02-2 USD) that have been the historic focus for PM 2.5 regulation (Muller and Mendelsohn 2010, Gilmore et al 2019). Overall, these findings suggest that air pollution regulations should consider regional-scale impacts of NH 3 emission reductions that are expected to provide the largest gains in Europe and North America, consistent with Pinder et al (2007) and Megaritis et al (2013).
Landscape fires (wildfires, prescribed burning and biomass burning but not limited to the global food system) have been linked to 330 000 (interquartile range: 260 000-600 000) excess deaths (Johnston et al 2012). Open biomass burning is a significant contributor to PM 2.5 -excess deaths in China (1 million (95% CI: 840 000-1.3 million)), India (990 000 (95% CI: 660 000-1.35 million)) (Reddington et al 2019) and Africa (780 000 (95% CI: 760 000-800 000)) (Bauer et al 2019). However, these estimates are not delineated for contributions specific to the pre-production (land-use change), Table 3. List of studies reporting PM2.5-premature mortality estimates for sectors relevant but not exclusive to the global food system. Of the 320 studies, data sets and reports synthesized in this review, only 19 were identified that followed the causal chain of emissions-PM2.5 exposure-premature mortality for sectors relevant to the global food system. These are listed in chronological order with a focus on the geographic scale of analysis, emissions sector and pollutants of interest, air quality model configuration including the name of the air quality model, emissions database, and horizontal grid resolution, and findings from the studies in terms of PM2.5-excess deaths and economic damage. An important caveat is that the excess death estimates are based on different IER functions and are not adjusted for contributions ofor the global food system, given that the scope of all but three of the reported studies (marked with a ) are not focused on the food system. Reported estimates could thus also include contributions resulting from land-use change for urbanization, mining, timber production and the production of non-food agricultural commodities, such as fiber and biofuels. Air quality and emissions model acronyms are defined alphabetically in the footnote.  In addition to these sectors, food export in the United States was linked to an average increase in PM 2.5 exposure by 0.36 µg m −3 , mostly attributed to NH 3 emissions, and resulted in 36 billion USD damage in 2006, which was equivalent to 50% of the food export value (Paulot and Jacob 2014). Hill et al (2019) estimated that maize cultivation, which accounts for 95% of all feed grain production in the United States, was linked to 4300 PM 2.5 -attributed premature deaths. The resulting economic damage valued at 39 billion USD in 2017 exceeded the monetized damage as a result of GHG emissions, and in 40% of the maize growing states the combined PM 2.5 and GHG economic damage exceeded the market value, indicating large negative externality costs.

Estimate of ambient PM 2.5 -attributable premature deaths resulting from the global food system
The studies summarized in section 4.2.1 collectively highlight the large PM 2.5 -attributable premature deaths from sectors related to but not exclusive to the global food system. Here, we develop the first estimate, to our knowledge, of annual PM 2.5 -attributable premature deaths from the global food system, as summarized in table 4. For the agricultural production stage, we adjust the median PM 2.5 -attributable deaths from agricultural production reported by Pozzer et al (2017) with the fraction of global crop area devoted to food production to estimate 750 000 excess PM 2.5 deaths from food production. Similarly, for the waste stage, we adjust estimates by Kodros et al (2016) with the fraction of domestic waste that is composed of food (40%) to conservatively estimate 76 000 median excess PM 2.5 deaths. We derive estimates for food-demand-driven land-use change using findings for landscape fires by Johnston et al (2012), by first deducting PM 2.5 -death contributions from fires resulting from open waste burning (Kodros et al 2016) and then further deducting contributions resulting from natural wildfires (23%) and non-agricultural commodity land-use change (30% of prescribed burning) (World Resources Institute 2014), resulting in an average estimate of 70 000 excess PM 2.5 deaths. We ensure no double counting of deaths occurred by conforming to the system boundaries that were used to describe stages in the food system and by excluding open waste burning, wildfires and non-food commodity land-use change from landscape fires.
Overall, by adding these estimates, we identify that 890 000 median excess deaths can be attributed to the global food system, 84% being a result of emissions from agricultural production. This order of magnitude estimate, developed based on studies with different approaches and IER functions (see table 4), is equivalent to 23% of the overall 3.9 million PM 2.5 -attributable deaths in the Global Burden of Disease Study 2015 (IHME 2020), and is similar or higher in comparison to global contributions from natural sources (18%), power generation (14%) and transportation (5%) (Lelieveld et al 2015, Crippa et al 2019. Our estimates are higher than the PM 2.5related deaths reported by Malley et al (2021), as ours accounts for life cycle emissions over the entire food system. Overall, we identify that our estimate of PM 2.5 deaths from the global food system is underestimated given the limited accounting of contributions from sectors including agrochemical production, post-processing, consumption and inherent uncertainties in the causal pathways of emissions to exposure estimates as identified in section 4.3.
Our analysis has also identified key research gaps: (a) to date, the focus has been on agricultural production, with few studies examining sectors from other stages in the food system, and at national or sub-national scales. There is a dearth of studies examining the impacts of food demand and agricultural production activities in highly populated regions in Africa, South America and Asia, where countries also have a high share of GDP (15%-58% in Africa and Asia) attributed to agriculture (World Bank 2018b). Given that a 10% increase in global NH 3 emissions could result in 22 000 additional excess deaths (Lee et al 2015), it is important to focus on these regions that are also expected to see increases in NH 3 emissions in the future. (b) It is important to identify the regional-scale efficacy of NH 3 emissions controls in regulating ambient PM 2.5 (Pinder et al 2007). Notably, Bauer et al (2016) demonstrated that emissions from increased food production could be managed without deteriorating future air quality, assuming emission controls on combustion sources of NO x . Given the substantial uncertainties in the emission inventories from agriculture (Crippa et al 2019) (see section 4.3.1), the extent of the impacts of NH 3 , NO x and NMVOC emissions on ambient PM 2.5 at regional scales needs further investigation. Table 4. Estimate of PM2.5-attributable premature deaths from the global food system as developed in this review. Estimates are derived by using the median estimates reported by the listed studies and applying an allocation factor that identifies food versus non-food contributions from the emissions sector. These findings are largely underestimated given the paucity of emissions from the post-production and consumption stages, reported global underestimates in the magnitude of NH3 emissions, the lack of sector-scale analysis of PM2.5-attributable deaths in several regions including South America and Africa, and different methods and uncertainty in IER functions.  (Punger andWest 2013, Paolella et al 2018). A wide range of approaches have been adopted to reduce uncertainties in NH 3 emission inventories, including the use of inverse models that use observation data to constrain seasonality in NH 3 emissions , Zhu et al 2015a, process models that capture interactions between crop, soil and weather to predict NH 3 emissions at site and regional scales (Cooter et al 2012, Balasubramanian et al 2017, Xu et al 2019, and meta-analysis of field measurements (Pan et al 2016). In addition, continued advances in capturing emissions from sources, such as small fires, domestic burning and peatland fires through products like GFED4, further research delineating emissions contributions from agriculture-driven land-use change, and estimating emissions from food waste will help improve our understanding of the PM 2.5 pollution burden from the food system.

Resolving uncertainties in AQMs and choice of model parametrization
Uncertainties in air quality modeling that result from model formulation and model parametrization can introduce uncertainties in estimates of PM 2.5 premature mortality. However, these concerns are not specific to the analysis of the global food system. It is infeasible to examine the entire extent of formulations and parametrizations to quantity embedded uncertainties (Solazzo et al 2017). However, marginal PM 2.5 responses to additional emissions have smaller biases than PM 2.5 predictions in response to the absolute magnitude of emissions (Hogrefe et al 2008). An important aspect of CTMs (Chemical Transport Models) is the choice of spatial resolution (Reddington et al 2014). Kushta et al (2018) found that premature mortality estimates varied by less than 3% when using a coarser CTM resolution (>100 km) in comparison to a finer population-scale spatial resolution (<20 km). Similarly, a fine spatial scale analysis (4-36 km grid dimensions) over the United States constrained PM 2.5 -attributable mortality to ±10% (Thompson et al 2014). In contrast, Punger and West (2013) found higher differences (∼30%) when scaling PM 2.5 exposure from a coarser scale of global models (>250 km) to 12 km × 12 km, with similar differences reported (27%) when switching from coarsest (∼69 km) to finest (∼5.9 km) grids for the United States using an RCM (Paolella et al 2018). Despite similar methodologies, (Kodros et al 2016) estimates of total annual, global PM 2.5deaths were 13% lower in comparison to (Lelieveld et al 2015) as a result of coarse AQM configuration. Thus, rigorous PM 2.5 evaluation on a case-bycase basis is recommended in comparison to standard model performance benchmarks (Emery et al 2017) before further evaluation for health assessment. Further model improvements should also focus on reducing uncertainties in capturing PM 2.5 formation that is non-linear in response to NH 3 emissions, as well as representations of secondary organic aerosol formation (Fuzzi et al 2015).

Exploring IER functions to link PM 2.5 exposure to PM 2.5 -attributable deaths
Many studies use log-linear IER functions, wherein a given reduction in PM 2.5 concentrations would yield the same gains in health benefits (Marshall et al 2015). Supralinear IER functions, however, better represent premature mortality outcomes as a function of PM 2.5 exposure (Goodkind et al 2014), thereby resulting in greater benefits at lower PM 2.5 concentrations (Marshall et al 2015). The IER responses at relatively high levels of PM 2.5 represent a source of uncertainty as they are derived based on studies for North America and Europe where the annual average PM 2.5 exposure is less than 30 µg m −3 and have different baseline health conditions compared to several parts of the world. Recent studies now account for impacts from regions with high PM 2.5 exposure, such as in China (Shiraiwa et al 2017, Yin et al 2017. Burnett et al (2018) estimated that global PM 2.5 excess deaths could be as high as 8.9 million if the IER functions were derived using cohort studies covering the entire range of global PM 2.5 exposure. Goodkind et al (2019) estimated that varying the IER functions resulted in a 21% difference in mortality estimates for the United States. In addition, PM 2.5 -attributable damage should consider both chronic and sporadic exposure for episodic emissions sectors, such as fires (Johnston et al 2012), and account for toxicity resulting from PM 2.5 components (Shaffer et al 2019). Lelieveld et al (2015) identified that when carbonaceous PM 2.5 was assumed to be more toxic than inorganic PM 2.5 , the resulting mortality attributed to agricultural emissions reduced from 20% to 7%. However, similar analysis for fires from land-use change and waste combustion could result in large estimates of PM 2.5attributable deaths. The responses of human health to PM 2.5 toxicity, especially to components that are carcinogens or allergens, and the synergistic interactions resulting from organic fractions remain active areas of research (West et

Opportunities for PM 2.5 mitigation and policy implications
If the current shifts in diets, affluence and population growth trends continue, agricultural production will need to increase by 60%-100% by 2050 to meet future food demand (Tilman et al 2011, Tilman and Clark 2014, FAO 2018. This demand is expected to increase the environmental burden through increases in GHG emissions by 87%, cropland demand by 67%, water withdrawals by 65% and nitrogen fertilizer inputs by 860% (Springmann et al 2018a), but the potential increase in PM 2.5 -health damage is less well understood. Likewise, few studies have evaluated the emissions reduction potential of farm-scale interventions (Kupper et  . In addition, there is concern with regard to the inequity in air pollution exposure impacts that occur from demographic differences in emissions attributed to the consumption of goods and spatially distant impacts of emissions on PM 2.5 exposure (Tessum et al 2019). These environmental justice implications are of particular interest in the global food system, dependent on where and how food is cultivated, and further exacerbated by socioeconomic differences in access to adequate and nutrient-rich foods. Reducing these environmental and health impacts will require a 'third Green Revolution' that focuses on the adoption of sustainable diets, improved agricultural practices and the implementation of regulatory instruments (FAO 2018). Here, we briefly identify instruments that have been proposed for the global food system to meet climate targets (Bryngelsson et al 2016, Wollenberg et al 2016, Niles et al 2018 that have potential co-benefits in minimizing ambient PM 2.5 health burden, both within and beyond the farm gate (Kanter et al 2020).

'Eating enough' and 'eating right'
Individual dietary demands play a key role in determining the impacts of the global food system (Kearney 2010). Since 1961, global food consumption has increased by 400 kcal d −1 , with the largest increases in South Asia (>50%) and Latin America (>30%) (FAO 2020a), and is projected to further increase by 25% by 2030 (FAO 2017). Despite improvements in food supply equity in the past century, a triple burden of malnutrition exists in the form of undernutrition (690 million), obesity (1.9 billion adults and 42 million children) and micronutrient deficiencies (2 billion) (WHO 2018, FAO 2020c. Agricultural production will need to increase to meet global food demand, while also accounting for shifts towards animal-based foods that are expected to increase by nearly 30% for meat and 20%-58% for eggs and dairy by 2050 (Clark  et al 2018). These increases will likely be accompanied by increases in PM 2.5 and precursor emissions, especially in Asia and Africa, which face the largest increases in food demand (Godfray et al 2010).
The paradigm of 'eating right' and 'eating enough' could be the key to mitigating environmental damage, including air quality impacts. Consuming only the required calories that meet individual metabolic and nutritional demands could improve health and climate outcomes (Niles et al 2018). Producing crops only for human consumption (i.e. plant-based foods) can increase caloric availability by 70% (Cassidy et al 2013), thus meeting not just current, but future global food demand (Berners-Lee et al 2018). Plant-based foods have been identified to have lower environmental impacts per serving in comparison to animal-based foods, especially ruminant meats from cattle, sheep and goats that have larger contributions compared to pork, poultry, eggs and dairy (Clark et al 2018, Poore andNemecek 2018, Willett et al 2019). Springmann et al (2016) estimated that a complete shift to vegetarian diets and increasing vegetables, fruits, lentils and grain consumption by >50% would reduce GHG emissions (3-11 Gt yr −1 ) by 2050. An examination of emissions factors for animal type-specific manure management suggests that similar trends could be expected for PM 2.5 pollution burden. However, the extent of the impact of shifts in diets needs further investigation. Reducing dependencies on animal-based foods could maximize both health and environmental benefits (Clark and Tilman 2017, Godfray et al 2018. It is thus imperative to establish spatially explicit impacts of the global food system on PM 2.5 health impacts, with a focus on NH 3 emissions, as well as emissions resulting from land-use change.

Managing food waste
Globally, food waste tripled between 1960 and 2011 (FAO 2011, Porter et al 2016 and is a contributor to emissions of primary PM 2.5 , NMVOC and SO 2 as a result of disposal practices. Reducing consumer food waste by 50%, either by individual choice or through supply chain interventions, could result in a 10% reduction in fertilizer and land use while improving food security through 1300 trillion kcal savings yr −1 by 2050 (Clark et al 2018). Developing policies and infrastructure to shift the open burning of waste to controlled disposal, possibly coupled with energy recovery, could provide benefits in reducing PM 2.5 pollution (Coventry et al 2016). However, tradeoffs in the form of increases in NH 3 emissions that result from organic waste decomposition should be carefully evaluated (Wang and Zeng 2018).

Farm-scale interventions
The demand for food is expected to increase substantially, along with subsequent emissions, especially in Asia (by 40%) and Africa (by 47%) by 2050. While the agricultural contributions to ambient PM 2.5 in these regions are small (3%-9%), in comparison to residential (27%-45%) and power generation (17%-26%) (Crippa et al 2019), even a 50% reduction in agricultural production emissions could reduce up to 130 000 PM 2.5 -attributable premature deaths (Pozzer et al 2017). It is thus imperative to balance the need for food security with resulting health impacts, to reduce the expected large externalities and economic losses, through improvements in agricultural productivity as well as farm-scale mitigation strategies.
An increase in crop yields and reductions in farm-scale inefficiencies that are prevalent in lowerincome countries in Sub-Saharan Africa, Mexico, India and Southeast Asia could reduce nitrogen and energy inputs to meet future food demand (Mueller et al 2012). Of high priority is the reduction of yield gaps (Lobell et al 2009, van Ittersum et al 2013 that are prevalent in 43 countries where crop yields are less than a third of their potential (Clark et al 2018). Suggested strategies include improving access to agricultural inputs, such as fertilizers, seeds and pesticides, especially in sub-Saharan Africa (Pradhan et al 2015). However, tradeoffs in increased yields and economic gains must be carefully weighed against increases in GHG, NH 3 and PM 2.5 emissions, and other environmental concerns. Global NH 3 emissions are expected to increase as a result of livestock production by 2050(Bouwman et al 2013. Reducing NH 3 emissions in current food systems will not only benefit air quality, but reduce economic losses for farmers that result from nitrogen volatilization (Good and Beatty 2011). Guthrie et al (2018) compiled a comprehensive list of mitigation interventions for Europe that include improvements in livestock management by modifying animal feed (NH 3 reductions of 30%-45%) and increasing grazing time (<50%), structural interventions, such as redesigning animal housing and manure storage (>80%), adding control technologies, such as wet scrubbers (25%-65%), and modifying crop cultivation practices, including changes in nitrogen fertilizer type from urea to other forms of nitrogen, the use of fertilizer inhibitors and changing fertilizer application timing, loading rate and application mode (20%-70%). Similar assessments for emissions reduction potential and costs have also been reported by other studies in Europe (Kilmont and Winiwarter 2015) and the United States (Pinder et al 2007). The current suite of engineering solutions and best management practices could result in a 30% reduction in livestock-NH 3 emissions and 20% in fertilizer-NH 3 emissions (total 0.7 Tg yr −1 ) for the United States alone (US EPA 2011). However, further regionspecific studies are required.

Technological interventions
Technological solutions can reduce PM 2.5 pollution and have climate co-benefits within and beyond the post-production stage. Proposed interventions include improvements in energy efficiency by 20%-50% in food processing, distribution and retail, through correct specification and equipment use, cold chain-based food supplies, modal shifts in food transportation (Wakeland et al 2012, Pelletier 2015, Niles et al 2018, and new packaging technologies (Heller et al 2019. Reducing household and ambient PM 2.5 exposure in regions that are reliant on solid fuel use for cooking have been identified as an important area of research. Ongoing efforts have focused on reducing disparity through the widespread adoption of cleaner fuels and cleaner technologies by the World Health Organization (WHO 2016) and by governments in India, China and across Africa (Aung et al 2016, Anenberg et al 2017. Recommended guidelines include switching from dirty household fuels including kerosene and coal to cleaner fuels higher on the energy ladder, such as LPG, ethanol, biogas and electricity, and introducing cheaper and cleaner cookstoves as promoted by the Global Alliance for Clean Cookstoves (Lewis and Pattanayak 2012, Anenberg et al 2013, Pachauri et al 2013 to reduce emissions of primary PM 2.5 , NO x and SO 2 , and the resulting health burden.

Regulatory instruments
Two regulatory instruments are of interest to minimize the impacts of food system emissions on PM 2.5 deaths. First, unlike economic sectors such as electricity generation and transportation, not all emissions sources within agriculture have been considered for emissions regulation in most parts of the world. While agriculture is not explicitly excluded from regulations in the United States, emissions regulation on primary PM 2.5 or secondary PM 2.5 precursors from farms is required only in non-attainment areas (US EPA 2017). For example, state regulations are imposed in California on crop growers, poultry, dairy and cattle farms, and agri-businesses (CARB 2019). However, on-farm emissions typically do not exceed the specified threshold and are thus exempt from most regulatory programs in the United States (US EPA 2017). Second, in the United States, the Clean Air Act regulations consider six criteria, air pollutants including NO x , SO 2 and primary PM 2.5 ; NH 3 is currently not regulated. Given the large body of evidence identifying the key role NH 3 plays in regulating atmospheric chemistry, the US EPA Science Advisory Board has recommended regulatory approaches to treat NH 3 as a harmful PM 2.5 precursor (US EPA 2011).
Such a regulatory approach should be considered worldwide.
Programs to study nitrogen management strategies and impacts on the environment and agricultural productivity have been adopted in Europe. The Convention on Long-Range Transboundary Air Pollution and the Gothenburg Protocol that have set targets to reduce SO 2 , NO x , NMVOC and NH 3 by 63%, 41%, 40% and 17%, respectively, by 2010 compared to 1990 to reduce acidification and surface water eutrophication, as well as preventing 48 000 excess deaths from PM 2.5 and ozone exposure (UNECE 2017). NH 3 emissions have already been reduced by 24% between 1990 and 2008, facilitated through multiple programs ranging from the adoption of alternative fertilizer types in Germany to providing financial incentives for improved nitrogen use in the Netherlands (EEA 2015). The United Kingdom also recently announced a plan to reduce NH 3 emissions by 15% by 2030 (Plautz 2018), demonstrating increased attention to cost-effective PM 2.5 abatement through the regulation of NH 3 . A multifaceted regulation policy should be considered at the national scale to optimize the economic and environmental costs of farm-scale practices and alternative approaches.

Legislation, environmental and health protections
Legislation and environmental protections are important drivers of reducing the demands of agriculture on land-use change (Nolte et al 2017, Seymour andHarris 2019). However, these strategies rarely account for the nature of agricultural commodities and consumption patterns (Henders et al 2018). Deforestation rates decreased between 2004 and 2014 in Brazil following the establishment of conservation zones (Anderson et al 2016). However, recent increases since 2017 (Amigo 2020) are a result of non-compliance with conservation agreements to meet the increased demand for soy, cattle and timber (Carvalho et al 2019) and the impact of export-driven trade demands (Tester 2020). In Indonesia, national moratoriums as well as pledges from corporations to increase sustainable products in their supply chains have helped reduce conversions of primary forests and peatland for industrial palm production (Carlson et al 2018, Gaveau et al 2019. However, the impacts of such policies on air pollution and health exposure are relatively unexplored and limited to a few studies (Marlier et al 2015(Marlier et al , 2019. In addition, health protection policies to promote healthier diets through dietary guidelines and legislation could offer co-benefits to both health and the environment (Clark et al 2018). The implications of these policies on GHG emissions have been the subject of recent inquiry, with examples Table 5. Key findings from the system-scale analysis of excess deaths occurring from exposure to PM2.5 from the global food system.
1. Important sources include land-use change, livestock and crop production, and agricultural waste burning. 2. Sectoral emission contributions vary by region. Agricultural production emissions dominate contributions in North America and Europe, while land-use change, manure management and agricultural waste burning emissions dominate contributions in Asia, Africa and South America. Emissions from the global food system are linked to at least 890 000 annual PM2.5-attributable deaths, which is equivalent to 23% of PM2.5-deaths reported in the Global Burden of Disease Study 2015.
1. 84% of estimated PM2.5-attributable deaths from the global food system are a result of agricultural production. 2. Estimates of excess deaths are likely underestimated given the paucity in emissions data from the post-production and consumption stages, underestimated global NH3 emissions, and lack of sector-scale analysis of PM2.5-attributable deaths in South America and Africa. There are uncertainties in establishing the impact of emissions from the global food system on PM2.5-attributable deaths.
1. It is important to fill critical information gaps in emissions inventories and reduce existing uncertainties (30%-50%) in primary PM2.5, NH3 and NOx emissions from the food system. 2. Developing emissions inventories for land-use change and food waste disposal are recommended as high priority research focus, with spatially explicit analysis for South America, Asia and Africa. 3. Continued air quality model development and evaluation is required to reduce current biases of ∼30% in ambient PM2.5 predictions. 4. Using a homogenous set of IER functions across studies that also account for toxicity of PM2.5 components will reduce existing biases in estimates of excess deaths. There are likely many cost-effective opportunities to mitigate PM2.5 pollution by reducing emissions from the global food system.
1. Air pollution regulations should include NH3, given the key role in formation of ambient PM2.5. 2. Shifts in dietary patterns and regulatory instruments can mitigate PM2.5 emissions from land-use change. 3. Reduced dependence on fertilizers, improvements in crop yields and better technological practices can help reduce NH3 emissions from livestock and crop production. 4. Emissions of NOx and SO2 from the food industry including retail, distribution and transportation could be mitigated through clean and energy-efficient technologies. 5. Reduction in food waste and shifts to controlled waste disposal is key to mitigating emissions from food waste.
including the evaluation of national dietary recommendations (Behrens et al 2017), national-scale strategies to reduce dependencies on animal-based foods (Springmann et al 2018b), and expanding sustainability metrics to also account for macro and micronutrient delivery from food production (DeFries et al 2015, de Ruiter et al 2018. In addition, managing food pricing has been recommended as a tool to reduce GHG emissions, namely through GHG taxes (Springmann et al 2017), taxes on less healthy foods, such as refined sugar (Briggs et al 2016), and through subsidies and tax revenues (Hadjikakou 2017) albeit with concerns about disproportionate effects on those of lower socioeconomic status. PM 2.5related pollutant emissions could also be achieved through programs that target reductions in waste (Porter et al 2016), food portioning (Story et al 2008) and food labeling akin to calorie labeling (Upham et al 2011, Leach et al 2016. Expanding environmental impact assessment to include impacts on PM 2.5 pollution burden deserves further study, given the downstream impacts of such policies on shifts in diet, food production, processing and waste disposal. Finally, it is essential to consider the environmental justice implications of food and agricultural systems for the world at large.

Highlights and research needs
The recent growth in understanding the global food system and its complex interplay with energy, material, water and land use has expanded our understanding of the large burden it places on the environment (Springmann et al 2018a). Indeed, a comprehensive accounting of food sustainability requires further consideration of major diets and food commodities, the processes that drive the food system and expanding the suite of environmental impacts (Halpern et al 2019). Our review adds to the conversation about global food system sustainability by identifying the large health burden resulting from exposure to ambient PM 2.5 . Here, we show that PM 2.5 -related emissions from the global food system are linked to 890 000 PM 2.5 -attributable premature deaths annually, which is equivalent to 23% of the 3.9 million ambient PM 2.5 -attributable deaths reported in the Global Burden of Disease Study 2015. These findings are, however, underestimated, given the paucity of emissions from food post-production and consumption stages, the overall global underestimate in emissions of NH 3 and the lack of PM 2.5 exposure impact studies for several emissions sectors and in regions including South America and Africa. A summary of our key findings is listed in table 5.
Additional empirical research is needed to reduce uncertainties in the characterization of emissions across multiple spatial and temporal scales to support air quality forecasting, and with a focus on expected future trends in production, consumption and food losses in low-and middle-income countries. Research opportunities abound in identifying improvements in energy and resource use in the food industry, retail and distribution, and transportation. Furthermore, systematic and region-scale efforts, especially in Asia, Africa and South America, are required to establish how the identified emissions mitigation strategies could deliver costeffective reductions in ambient PM 2.5 concentrations and PM 2.5 -attributable premature deaths. With diets shifting towards animal-based and more processed foods, and increases in global caloric consumption, additional environmental and health burdens resulting from degrading air quality are expected. However, by considering variability in regional shifts in future food demand and production, strategies that encompass a wide range of regulatory, technological and educational tools that encourage health and environmentally conscious diets can be implemented to sustainably manage these increases with minimal impacts on air pollution. Given the recent interest in food system research in the context of climate and other environmental impacts, we argue that the se studies should further account for damage from PM 2.5 pollution as a key indicator of both human and environmental health.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files).

Acknowledgments
This publication was developed as part of the Center for Air, Climate, and Energy Solutions (CACES), which was supported under Assistance Agreement no. R835873 awarded by the U.S. Environmental Protection Agency (EPA). It has not been formally reviewed by the EPA. The views expressed in this document are solely those of the authors and do not necessarily reflect those of the EPA. The EPA does not endorse any products or commercial services mentioned in this publication. This research was also supported by the U.S. Department of Agriculture (MIN-12-083).