Environmental footprints show China and Europe’s evolving resource appropriation for soybean production in Mato Grosso, Brazil

Mato Grosso has become the center of Brazil’s soybean industry, with production located across an agricultural frontier expanding into savanna and rainforest biomes. We present environmental footprints of soybean production in Mato Grosso and resource flows accompanying exports to China and Europe for the 2000s using five indicators: deforestation, land footprint (LF), carbon footprint (CF), water footprint (WF), and nutrient footprints. Soybean production was associated with 65% of the state’s deforestation, and 14–17% of total Brazilian land use change carbon emissions. The decade showed two distinct production systems illustrated by resources used in the first and second half of the decade. Deforestation and carbon footprint declined 70% while land, water, and nutrient footprints increased almost 30% between the two periods. These differences coincided with a shift in Mato Grosso’s export destination. Between 2006 and 2010, China surpassed Europe in soybean imports when production was associated with 97 m2 deforestation yr−1 ton−1 of soybean, a LF of 0.34 ha yr−1 ton−1, a carbon footprint of 4.6 ton CO2-eq yr−1 ton−1, a WF of 1908 m3 yr−1 ton−1, and virtual phosphorous and potassium of 5.0 kg P yr−1 ton−1 and 0.0042 g K yr−1 ton−1. Mato Grosso constructs soil fertility via phosphorous and potassium fertilizer sourced from third party countries and imported into the region. Through the soybean produced, Mato Grosso then exports both water derived from its abundant, seasonal precipitation and nutrients obtained from fertilizer. In 2010, virtual water flows were 10.3 km3 yr−1 to China and 4.1 km3 yr−1 to Europe. The total embedded nutrient flows to China were 2.12 Mtons yr−1 and 2.85 Mtons yr−1 to Europe. As soybean production grows with global demand, the role of Mato Grosso’s resource use and production vulnerabilities highlight the challenges with meeting future international food security needs.


Introduction
Environmental footprints are indicators of resource use and emissions across boundaries of consumption and production. Recently, the ecological footprint, water footprint (WF) and carbon footprint (CF) were combined into a 'footprint family' to collectively understand human appropriation of land, air, and water (Steen-Olsen et al 2012, Galli et al 2011). The ecological footprint, in global hectares (ha), expresses the human appropriated biocapacity related to a given level of consumption (Wackernagel and Rees 1996). The CF quantifies the amount of greenhouse gas emissions (g CO 2 equivalents) related to an activity or product (Hertwitch and Peters 2009), while the WF estimates water consumptive use of a production or consumption activity (m 3 per activity or product; Hoekstra et al 2011).
Global estimates have shown that aggregate human activity and demand for resources and services approach 1.5 planets (Borucke et al 2013), with a global CF of 35 Gtons CO 2 -eq in 2001 (Hertwich and Peters 2009), and a global 1996-2005 average WF of 9087 Gm 3 yr −1 (Hoekstra and Mekonnen 2012). Such estimates bring to light not only our overconsumption of the planet's renewable resources as illustrated by the ecological footprint, but the potential impacts on climate change (CF) and inefficiencies in water resource use (WF). Alongside these footprints are related resource flows that indicate the exchange of natural resource use and emissions across boundaries of production and consumption. Such flows have been employed to quantify water and nutrients involved in the production of traded goods (Hoekstra andMekonnen 2012, Galloway et al 2007). Resource flows do not represent the physically traded resources, but rather provide relative metrics to evaluate resource use and pollution at the site of production and allow for the assessment of resource use efficiency of internationally traded products.
When applied to agricultural products, environmental footprints estimate the natural resources utilized by both producing and consuming countries, thus bringing to light externalities related to diets (Foley et al 2011). Between 1975and 1995, the Ecological Footprint for land use increased from 10.3 to 12.6 Gha due to increases in consumption (Van Vuuren and Bouwman 2005); in 2008 it reached 5.4 Gha for cropland and grazing land alone (World Wildlife Fund 2012). In 2001, 20% of the world's CF was attributed to food production at close to 7 Gtons CO 2 -eq (Hertwich and Peters 2009), while virtual water flows averaged 2320 Gm 3 yr −1 from the trade of agricultural and industrial products for the 1996-2005 period (Hoekstra and Mekonnen 2012). Resulting externalities have been studied using a life cycle approach (UNEP/SETAC Life Cycle Initiative 2005) to permit consideration of the natural resources used for agricultural products from 'field to fork' (Zaks et al 2009). Studies include the Ecological Footprint assessment of agricultural production in the Canadian Prairies (Kissinger and Rees 2009), the estimate of carbon emissions of tropical deforestation for soybean and pasture expansions (Karstensen et al 2013, Zaks et al 2009, and the WF of meat products (Hoekstra 2012, Galloway et al 2007. Other studies have also included nutrient flows in accounting for soil fertility and pollution from fertilizer application Bennett 2012, Galloway et al 2007).
Brazilian soybean production is particularly of interest given the country's increasingly important role in the international trade of agricultural products in recent decades. Global soybean production rose from 143 to 227 Mtons between 2000 and 2010 among major producers (Argentina, Brazil, China and the US; FAOSTAT 2013). By 2010, Brazil had become the second largest producer of soybean in the world with 68.5 Mtons produced (FAOSTAT 2013), and is anticipated to be the world's leading soybean producer in 2014 (USDA-FAS 2014). Between 2000 and 2010, Brazil's soybean exports increased by 125%, with the share exported to China increasing from 16% in 2000 to 66% of total exports in 2010, and the share going to Europe dropping from 64% to 20% (FAOSTAT 2013, Aliceweb 2013) (figure 1).
Land use change in Brazil is strongly related to the expansion of soybean and pasture in the Cerrado and Amazon regions (Barona et al 2010). Deforestation and land-use intensifications make the Brazilian Amazon one of the largest sources of greenhouse gas emissions related to land cover and land use change (Galford et al 2013). The subsequent production and transport of export crops has made Brazil an important center for greenhouse gas emissions for exported goods (177.0 Gt CO 2 -eq; Hertwich and Peters 2009), but also for flows of virtual water (112 Gm 3 yr −1 ; Hoekstra and Mekonnen 2012). In addition, Brazil is also an important exporter of nutrients for livestock feed used in live animal production systems when considering nitrogen (total of 2.1 Mtons of embedded N exported over 2000-02; Galloway et al 2007) and phosphorous consumed for crops, feed and livestock production (0.8 Mtons of P in 2007; Schipanski and Bennett 2012).
From 1990 to 2010, the contribution of Mato Grosso (figure 2) to Brazilian production increased from 3.1 Mtons yr −1 (15% of Brazilian soybean) to 18.8 Mtons yr −1 (27% of Brazilian production) IBGE 2013a). While the proportion of Mato Grosso soybeans shipped to Europe ranged from 11% to 29% of Brazilian soybean exports between 1997 and 2010, the portion exported to China increased substantially from 0% to 31% (for 2009, 29% in 2010) (figure 1). Increased soybean production occurred through both agricultural extensification (conversion of natural environment into agricultural land), and intensification (increased production on existing agricultural lands) (Galford et al 2010). Agriculture has been expanding in Mato Grosso through conversion of rainforest and Cerrado/Cerradao savanna ecosystems since the 1990s (Barona et al 2010), with several changes in policies and increased dependencies on international markets also influencing land use change (Macedo et al 2012, Richards et al 2012. Local resource use for cropland expansion creates environmental externalities such as losses in biodiversity, ecosystem services or eutrophication, whose costs are often internalized by producers, thus causing a distortion in the real price of commodities (Galloway et al 2007). It is therefore important to quantify resource use and emissions across the supply chain to manage potential environmental impacts from supply and demand perspectives.
This study presents environmental footprints associated with the production of soybean in Mato Grosso, namely land, carbon, water and nutrients, and describes resource flows to China and Europe between 2000 and 2010 related to exports of the whole beans. Macedo et al (2012) reported a sharp reduction in deforestation in 2005-10, despite a continuous increase in soybean production, suggesting a possible change in the production system between the first and second half of the 2000s. This change also coincided with China's rise as a major soybean importer for the region, surpassing Europe. We therefore use environmental footprints and associated resource flows from soybean exports to further understand relationships between the producer and consumers in a decade of apparent change in soybean production system. Natural resources used in Mato Grosso's soybean production have previously been studied with respect to land use change (Macedo et al 2012) and associated carbon emissions (Galford et al 2010(Galford et al , 2011, carbon emissions allocation into soybean production and consumption (Karstensen et al 2013, Zaks et al 2009, as well as water consumptive use from rainfed agriculture (Lathuilliere et al 2012). To our knowledge, no research has yet integrated these footprints of production over a common time period for the purpose of understanding resource use related to soybean exports. This work provides context for Mato Grosso's role in the global food system in which importing nations are increasingly concerned with quantifying externalities related to agriculture (Ruvario et al 2011), particularly Europe (Steen-Olsen et al 2012).

Materials and methods
This study presents findings of resource use and emissions to land, air, water, and nutrient flows related to the production and export of soybean (Glycine max) in Mato Grosso for the 2000-10 decade translated into environmental footprints. Mato Grosso is located in Brazil's Central Western region and is home to three major biomes (figure 2): Amazon rainforest in the north, with a transitional Cerrado/Cerradao (savanna) extending to the Pantanal wetland in the south. We focus exclusively on soybean production and export in this study and do not consider the production of soybean oil and soy meal, which occur in Brazil as well as China and Europe, and would require additional resources (e.g. water, energy). These additional resources for products further down the supply chain would have to be considered in the country of production in order to derive a soy meal or soybean oil product environmental footprint. Also, indirect resources or emissions accounted in labor or off farm in the production of equipment and buildings are not included here. The agricultural production step ahead of such processing is considered the most resource intense stage of the entire life cycle of both products in Brazil (Milazzo et al 2013). Estimates show that an average of 40% of Brazilian soybean was exported as whole beans in 2000-10 (FAOSTAT 2013. Soybean that wasn't exported was processed into either oil for local consumption (18%) or soybean cake (>50%) (FAO-STAT 2013). We consider footprints and resource flows from exports to China and the European Union (hencefourth as Europe). LF, CF, WF and nutrient footprints are included, and described in more detail below. We also track deforested area as an additional indicator to be included with the land resources.
Virtual and embedded resources are differentiated as follows: virtual flows describe resources used which do not comprise a physical part of the product, but were used or consumed as a component of soybean production (e.g. land use conversion, water consumptive use, CO 2 equivalents emitted during deforestation); embedded flows refer to resources which are an integral part of the product. This distinction is apparent in the case of nutrients (Galloway et al 2007) since elemental nitrogen (N), phosphorous (P) and potassium (K) are applied to the field as fertilizer (typically in an N-P 2 O 5 -K 2 O formulation) and are taken up as part of the soybean plant. Moreover, 80% of the N assimilated by soybean is biologically fixed from atmospheric N via rhizobium associated with soybean roots (Smaling et al 2008). Thus embedded N refers to N contained in harvested soybeans, while virtual N refers to N utilized in soy production that doesn't become part of the harvested crop (e.g. N contents of crop residues, and N remaining in soil or leaching from soil). While water is also an integral part of the soybean, the plants' stored water content is small compared to the plant's consumptive use (Galloway et al 2007). We interpret the difference between virtual and embedded in terms of the productionconsumption boundary where embedded resources physically cross this boundary. Embedded nutrients require assimilation by individual regions importing the raw material, while virtual resources represent Mato Grosso's consumption (for water) or load (fertilizer, CO 2 emissions, land) burdened by the region but which we indirectly allocated to importers.
Methodologies from previous studies were used and extended to include the 2000-10 decade (table 1) and described in more detail below. Analyses requiring information from the Brazilian Institute of Geography and Statistics (IBGE 2013a), such as data needed to calculate land, water and nutrient footprints, were carried out at the municipal scale using political units delimited in Lathuilliere et al (2012) because municipalities in Mato Grosso changed in size and number through the 2000s. As such, municipalities were aggregated into 104 municipal units whose number and size were constant within the study time period (Lathuilliere et al 2012). Resource flows were allocated as a percent of total resources used relative to the fraction of soybean exported to each country or region based on annual trade data available from Aliceweb (2013). We assigned full internalization of the footprints to importing countries as in Karstensen et al (2013).

Deforestation and land footprint
Two components of land use were considered in the present study. First, deforested land area (ha) associated with soybean cultivation was assessed after Karstensen et al (2013) for 1990-2010. Karstensen et al (2013) related deforestation area to soybean production based on Landsat satellite imagery from the Brazilian National Institute of Space Research (INPE 2013), in conjunction with the land use transition model from Rammankutty et al (2007). Initial land use following forest conversion was assumed to be 65% cropland and 35% pasture for Mato Grosso (Karstensen et al 2013). Second, we consider the LF without a carbon uptake from the land term as in Steen-Olsen et al (2012), since carbon uptake is accounted for in the CF (discussed below). While Steen-Olsen et al (2012) focus on consumption, we focus on the production side of the soybean supply chain and therefore consider area planted in soybean (IBGE 2013a) as the gross LF.

Carbon footprint
Deforestation and management practices were taken into account as the two main sources of greenhouse gas emissions in soybean production. Greenhouse gas emissions related to deforestation were obtained from Karstensen et al (2013). Biomass clearing through deforestation was assumed to be 70% burnt and 20% slash (Karstensen et al 2013) following agricultural practices and carbon stocks described by Galford et al (2010). The remaining biomass is typically separated into timber product (8%) and elemental C (2%) (Karstensen et al 2013). Carbon emissions are then allocated to soybean using a carbon cycle model when a biosphere-atmosphere flux is registered (Karstensen et al 2013). Such accounting does not include indirect land use change described by Barona et al (2010). Second, we consider emissions from machinery used in soybean production: sowing, fertilizer application, pesticide application and combine harvesting, noting that soybean is directly seeded into the stubble of the previous crop with no tillage employed in the study region. We used the average emissions of 33.4 kg CO 2 -eq ton −1 soybean estimated by Prudencio da Silva et al (2010) for the Brazilian Central Western region using the Ecoinvent® database (Ecoinvent 2013). We then determined the annual emissions of machinery as a function of this emission rate and production numbers from IBGE (2013a).

Water footprint
The total WF was calculated as the sum of green WF and gray WF. The green WF represents soil moisture regenerated by precipitation and utilized for plant growth, while the gray WF is the water volume required to assimilate a pollutant load to ambient water quality standards . In this study, we do not consider the blue WF (e.g. the consumptive water use supplied via irrigation), as Mato Grosso's soybean production is almost exclusively rain-fed at present (Lathuilliere et al 2012).
Green water was calculated from actual evapotranspiration following equation (1) (Allen et al 1998), where ET C (mm day −1 ) is the crop evapotranspiration, K C is the crop coefficient for soybean which depends on the crop development cycle, and ET 0 (mm day −1 ) is the reference evapotranspiration determined through meteorological parameters for the 2000 to 2010 period (INMET 2013, see online supplemental material equation S1 available at stacks.iop.org/ ERL/9/074001/mmedia). Daily ET C was calculated with variations in crop coefficient for the initial growing stage (K C = 0.56), mid-season (K C = 1.50) and harvest (K C = 0.50) for a total development cycle of 126 days (Lathuilliere et al 2012). Final ET C values were obtained from ten day averages of ET 0 (Allen et al 1998) before being compared to effective precipitation to derive soybean green water: if effective precipitation was less than ET C , then green water was equal to precipitation; if effective precipitation was greater than ET C , then green water was equal to ET C as per Lathuilliere et al (2012). Finally, the sum of modeled green water for soybean harvests between 2000 and 2010 were related to harvested area and total soybean produced to derive an average green WF (m 3 ton −1 yr −1 ) for the decade (see where gray WF is calculated in m 3 yr −1 as a volume, and also per ton of soybean for each harvest year (m 3 ton −1 yr −1 ), L is the pollutant load (g ton −1 yr −1 ), c max is the ambient water quality standard (g m −3 ) and c nat is the natural concentration of the pollutant in the water body of interest (g m −3 ). L was determined as a fraction of N and P application rates according to Hoekstra et al (2011). Pollution sources affecting the gray WF were determined from N and P fertilizer application; other agricultural inputs such as pesticides were not considered. We used 10% as the fraction of applied N fertilizer and 0.3% of applied P fertilizer allocated to ground and surface water as non-point source contamination , Riskin et al 2013a. Fertilizer application rates are described in the following section.
Gray water was then estimated as the volume of water needed to dilute the larger pollution load (N or P) considering the range of fertilizer application rates (Section 2.4). Therefore, the pollutant with the smallest volume is considered coassimilated along with the contaminant with the largest pollutant load. The values of c max (10 mg NO 3 -N L −1 and 0.1 mg total P L −1 ) were selected from water quality standards established under the Brazilian National Environmental Council (CONAMA) which specifies limits for class 2 water for human consumption, environmental conservation, recreation, irrigation and aquaculture (SEMA 2010a(SEMA , 2010b(SEMA , 2010c. The value for c nat was assumed to be zero, which is known to underestimate the gray WF  and is preferred in order to allow gray WF estimates to be conservative. Virtual water flows from total green and gray water consumed by soybean production then described the amount of local water resources appropriated for cultivation and assimilation of pollutant loads from fertilizer application. As described above, embedded water is small compared to virtual water (Galloway et al 2007) and therefore only virtual flows, representing indirect water use for the commodity, are discussed in this study.

Nutrient footprints
Virtual NPK and embedded NPK were determined as per Schipanski and Bennett (2012) and Galloway et al (2007) in equation (3): where F v is the virtual fertilizer (N, P or K) remaining on the field (kg), F a is the applied N, P or K fertilizer (kg) based on planted area and application rate (see below), and F e is the embedded N, P or K (kg) based on production and elemental concentrations in Mato Grosso soybean (see below).
Embedded NPK represents the nutrients physically traveling with the soybean which requires full assimilation by importers. Virtual NPK are the resources needed to grow the crop but that do not become embedded in the soybean. Just as in the virtual water case, the virtual NPK is related to local resources or the amount of NPK remaining in Mato Grosso's soil after soybean harvest but allocated to importers indirectly. This differentiation prevents double counting of NPK remaining in the soil and the amount physically crossing the production-consumption boundary through export. This separation also allows for an interpretation of potential environmental impacts experienced by the producing region (via virtual NPK) and consuming regions (via embedded NPK). We constrain virtual nutrients to be greater than or equal to zero, assuming that the plant has assimilated all applied fertilizer not lost to pollution. Eighty percent of N assimilated by the soybean occurs through biological N fixation (Smaling et al 2008) in which case a negative virtual N would imply a spurious impoverishment of the soil N (a detailed N soil balance is available in Smaling et al 2008).
Similarly, lower K application rates (see below) could result in negative virtual K values. All fertilizer applied to soybean fields was assumed to be industrially derived as per Smaling et al (2008). Values obtained for P 2 O 5 and K 2 O application rates were divided by 2.3 and 1.2, respectively to determine elemental P and K application rates. Total applied NPK was determined by top-down and bottom-up approaches following similar steps by Riskin et al (2013b). The top-down estimate focused on information from ANDA (2010) that reports combined fertilizer sales aggregated for all crops. Mato Grosso's share of total fertilizer consumption in Brazil's Central Western Region (as the states of Mato Grosso, Mato Grosso do Sul, Goias, and Federal District) was constant over 2000-10 at 56% (±1.5%) (ANDA 2010). We assumed that fertilizer sold to Mato Grosso had the same elemental NPK composition as that sold in the Central Western region for all years, except 2008 and 2010 for which Mato Grosso fertilizer sales were specifically reported by ANDA (2010). We then assumed a constant ratio of fertilizer used for soybean to total NPK consumed for 2010 (ANDA 2010), considering that 60% of fertilizer used for soybean was of the 2-20-18 formulation (N-P 2 O 5 -K 2 O), and 40% was the 0-20-20 formulation (ANDA 2010). In the bottom-up estimate, average NPK application rates were obtained for the 2000-10 decade following Brazilian fertilizer consumption data (Cunha et al 2010, Heffer 2009, Fertistat 2007, FAO 2004) and information obtained from 13 farms combined with agricultural production information IBGE 2013a). These two approaches, when combined to planted soybean area from IBGE (2013a) gave a range of applications rates of 0-5 kg N ha −1 , 28-34 kg P ha −1 and 39-62 kg K ha −1 . These ranges of fertilizer application rates were used in our analyses for all municipalities of Mato Grosso and were assumed to be constant within the study period.
Embedded NPK data (i.e. elemental concentrations in soybean) were obtained from Cunha et al (2010), where concentrations were reported to be 59.2 g N kg −1 , 5.5 g P kg −1 and 18.8 g K kg −1 of soybean leaving the field. The elemental concentrations were assumed to be constant within the decade for all soybean cultivated in Mato Grosso.

Sensitivity analysis
We performed a sensitivity analysis to provide uncertainty estimates for results when reporting environmental footprints and resource flows with the exception of LF obtained from IBGE (2013a). For the deforestation CF, we consider legacy emissions as described in Karstensen et al (2013). Most greenhouse gases emitted from deforestation would have occurred from biomass burning following forest clearing, with releases occurring in the year of deforestation. Legacy emissions include the biomass decay from slash, which is allocated to subsequent years following land clearing (Karstensen et al 2013). We used two planting dates for our annual green WF estimate (October 1st and November 1st) to account for differences in agricultural practices as per Lathuilliere et al (2012). We also included both extremes of the range of application rates derived from the top-down and bottom-up estimates to assess the differences in gray WF and virtual nutrient flows to China and Europe. Discussion of assumptions and comparisons to literature values can be found in the supplemental material. Ecosystem services related to climate and water cycling may also be impacted as a result of a loss of forest cover in the region (Lathuilliere et al 2012, Nepstad et al 2008, which may in turn impact the productivity of soybean agriculture. Simulations show that climate feedbacks from expanding agriculture further into the Amazon may reduce soybean productivity by 16-26% in 2041-50 (Oliveira et al 2013) such that the benefits of agricultural production are far outweighed by the costs to biodiversity and ecosystem services. While there are noted differences in deforestation and carbon footprint between 2001-05 and 2006-10, the above mentioned impacts on biodiversity and ecosystem services which would have been greater from soybean production based on extensification (2001-05) would likely carry over into the second half of the decade and into the future.  The increasing virtual P illustrates the increasing amount of nutrients remaining in the state of Mato Grosso with possible impacts to water bodies. Loss of applied P is known to cause eutrophication of water bodies over time in areas where P is not strongly bound to the soil. Local soil conditions in the study area, however, have a very high P fixing capacity thus largely preventing it from reaching aquatic environments (Riskin et al 2013a, 2013b, Schipanski and Bennett 2012. Monthly water quality campaigns were performed in 2007-09 in the Amazon River, Paraguay River and Tocantins-Araguaia River basins (74 sites in total) (SEMA 2010b). Total P measurements exceeded the CON-AMA limit of 0.1 mg total P L −1 on 40 out of 705 occasions, averaging 0.30 mg P L −1 (n = 40; sd = 0.25). Finally, there is evidence of ongoing accumulation of P in soils (Riskin et al 2013b, Schipanski andBennett 2012), which eventually could lead to saturation of P in the long term with consequential eutrophication, although results to date have not shown increases in P loading from on farm experiments (Brando et al 2013).

Results and discussion
3.2.2. Non-renewable fertilizer inputs. In addition to concerns over contamination of water bodies, the argument that agricultural trade results in exports of soil fertility has been made (Galloway et al 2007). However, this claim may not apply to Mato Grosso's Cerrado soils since its fertility is mostly derived from fertilizer inputs. Placing soybean production within the context of the fertilizer supply chain therefore provides additional insight into resource use for soybean exports, especially since Brazil is the fourth largest buyer of fertilizer in the world, with 10.1 Mtons of NPK fertilizer purchased in 2010 (ANDA 2010). Devaluation of the Brazilian Real through the early 2000s led to an increase in the costs of production, especially fertilizer, which increased in price from US$ 55.3 ha −1 in 2000-01 to US$ 180 ha −1 in 2009-10 (adjusted for inflation) for typical macronutrients (Agroserra 2001, IMEA 2010). Brazil's demand for fertilizer raw materials between 2000 and 2010 increased 6.6% per year, with 60-80% of materials imported for domestic fertilizer production (Amaral and Peduto 2010). In 2008, Brazil imported 4% of the world's N fertilizer, 7% of P and 14% of K (Amaral and Peduto 2010). For 2005-10, P fertilizer was mainly sourced from Israel (35%), Tunisia (19%), the E.U. (17%) Morocco (10%), and other international partners (19%) (UN Comtrade 2012). K fertilizer during this period was mainly sourced from Canada (27%), Belarus (22%), the E.U. (20%), Russian Federation (15%), Israel (12%), and other countries (4%) (UN Comtrade 2012). Brazil's dependence on K imports is particularly critical, with only 8% of K fertilizer use supplied by national production (IBGE 2012).
Mato Grosso fertilizer use ranks among the highest in Brazil for P and K fertilizer sales. In 2010, fertilizer sales in Mato Grosso were equivalent to application rates of 58.9-82.3 kg P ha −1 of cultivated area, and 72.9-94.3 kg K ha −1 (IBGE 2012). High P application rates and high K use efficiency (defined as the ratio of embedded to applied K fertilizer, see Section 3.3.2. below) suggest that Mato Grosso's soybean production works as a node for nutrient transfer from soil to soybean. Most applied K is Table 2. Total resources used for the production of soybean in Mato Grosso for the 2001-2005 and 2006-2010 periods with amounts exported to China and Europe: deforestation, land, carbon (nonleg-without legacy emissions; leg-including legacy emissions), water (average of planting assumptions and fertilizer application); average virtual phosphorous (P v ), potassium (K v ); embedded nitrogen (N e ), phosphorous (P e ) and potassium (K e  exported out of Mato Grosso through the commodity trade since embedded K was larger than virtual K. Within this context of international trade, soybean production in Mato Grosso and exports remain strongly tied to non-renewable resource extraction in other countries, noting that global reserves of phosphate rock for P and potash for K are believed to be declining due to long regeneration time scale (Obersteiner et al 2013, Schipanski andBennett 2012). This strong dependency also creates an important node for spikes in regional soybean price and costs of production due to fertilizer price fluctuations that strongly depend on oil supply and local fertilizer production capabilities (Amaral and Peduto 2010 (table 3) as soybean production intensified. Deforestation per ton of soybean harvested was 455 m 2 ton −1 for 2001-05 and 97 m 2 ton −1 for 2006-10. At the same time, the CF decreased from 19.9 tons CO 2 -eq ton −1 (on average resulting from biomass burning and subsequent emissions from decaying slash) to 4.6 tons CO 2 -eq ton −1 between the two 5 year periods. In other words, Chinese soybean importers indirectly had smaller externalities on global climate change, Mato Grosso biodiversity and ecosystem services compared to European importers earlier in the decade when the soybean production system was more focused on agricultural extensification. The rise in China's imports of soybeans coincided with more effective deforestation policies and enforcement, yet China's virtual resources imports, on the basis of planted area in Mato Grosso and exported volumes, remained larger than Europe's in 2010.
3.3.2. Efficiency in resource use through soybean trade. The average green WF during the study period was 1590 m 3 ton −1 , which is lower than the world soybean average (2037 m 3 ton −1 ) and soybean produced in China (2428 m 3 ton −1 ), but within the range of some European countries (Italy: 1177 m 3 ton −1 ; Germany: 1948 m 3 ton −1 ) . Nutrient exports embedded in soybean exports were found to increase over the 2000-10 period with the largest NPK values in 2010 (table 2). We assessed the phosphorous use efficiency for phosphate in fertilizer (PUE) and potassium use efficiency (KUE), defined as the ratio of embedded nutrients to applied P and K fertilizer. Averages over the study period were 0.48-0.58 for PUE and 0.90-1.43 for KUE based on fertilizer application rates. Average soybean PUE values in this study were similar to values previously reported for Brazil in 2002 (0.57) (Schipanski and Bennett 2012), but larger than the PUE for a farm in Northern Mato Grosso (0.39) (Riskin et al 2013a). The average KUE >1 suggests that essentially all of the applied K is taken up by the plant and exported. Additional K sources not quantified here may come from decomposing crop biomass left on the field during no-till agricultural practices (Bertol et al 2007).
There is little or no averted water contamination through soybean trade due to the high P fixing capacity of soils, as discussed by Schipanski and Bennett (2012). However,  (FAOSTAT 2013), much of which is used for the production of animal feed within importing countries (Nepstad et al 2008). The use of water and nutrients abroad for Chinese soybean production remains an important question in the country's agricultural policy as depicted in the supply side of soybean production in the context of Mato Grosso's land use dynamics described here. Rather than exporting native soil fertility, Mato Grosso is virtually exporting its water resources, photosynthetic capacity and the N fixing capabilities of soybean to China and Europe through the product. As soybean production in Mato Grosso is currently rain-fed (e.g. production relies almost exclusively on soil moisture derived directly from local precipitation without supplemental irrigation), the interannual variation in rainfall can affect yields. We calculated the average green WF for soybean as 1450 m 3 ton −1 for a planting date of November 1st, versus 1750 m 3 ton −1 for soybean with a planting date October 1st. This difference in green WF reflects the possible water stress that the plant may undergo which could greatly impact yields, especially during the development and flowering stages. Macedo et al (2012) identified that 22% of increased soybean production in the state was due to yield increases related to land and water management practices. Moreover, simulations on climate feedbacks of deforestation show a potential connection with precipitation such that soybean yield may become affected (Oliveira et al 2013).

Conclusions
This study presented a synthesis of environmental footprints for the production of soybean in Mato Grosso, Brazil, an . Ne, Pe, Ke are embedded nitrogen, phosphorous and potassium respectively; P v , K v are virtual phosphorous and potassium. Exports to China, Europe and the Rest of the World only concern soybean as whole beans. important global food security node upstream of the animal protein supply chain (figure 4). The estimate of resource use and related flows is a key step in assessing externalities which are often not included in commodity prices, but whose related impacts must be considered within the 'planetary safe operating space' (sensu Rockstrom et al 2009). Soybean production represented 65% of deforestation in Mato Grosso during the 2000-10 study period, and 14-16% of total Brazilian land use related carbon emissions despite a close to 70% decline between the 2001-05 and 2006-10 periods. The decline in deforestation suggested a shift in soybean production system from one based on extensification (agricultural expansion into natural ecosystems) to one based on intensification (increase in yield on available land) with differences in resource use and allocation to trade partners with China replacing Europe as the main soybean importer.
Over the study decade, soybean production required ever larger amounts of water and fertilizer inputs related to increases in planted areas and average yields, both of which are key to ensuring adequate total crop production. On the one hand, current practices of soybean production in Mato Grosso depend on rainfall, a free resource whose magnitude and timing can affect yields. On the other hand, soybean production requires inputs that strongly depend on international suppliers to construct soil fertility in Mato Grosso before exporting nutrients abroad. Our estimated environmental footprints suggest potential regional impacts on climate, biodiversity, ecosystem services, and a possible incremental soil P saturation that could increase the risk of eutrophication in the long term.
Supply side footprint assessment and resource flows can help reconcile policy actions that seem to have opposing goals, such as the inclusion of the environment in agricultural policy but its exclusion in trade policy (Wurtenberger et al 2006), or actions led by countries supporting deforestation reduction initiatives and their dampening effects from increased demand of soybean and meat products (Karstensen et al 2013). Such pressures will only increase with affluence, which has shown to contribute to further land displacement (Weinzettel et al 2013) as illustrated here by China's continuing growth in demand for Mato Grosso's soybean. Options for future soybean production will have to consider the trade-offs between agricultural extensification and intensification illustrated by the combined footprint analysis and an associated impact assessment to address costs and benefits of future food production strategies.