Leakage does not fully offset soy supply-chain efforts to reduce deforestation in Brazil

Modeling framework

We used an open source, fully documented, and publicly available medium run applied general equilibrium (AGE) model³⁵ with explicit treatment of subnational land markets divided in Agroecological Zones (AEZ), nicknamed GTAP-AEZ²⁶. The GTAP-AEZ framework is based on decision nests at which agricultural producers decide on land cover conversions (Supplementary Information Fig. S2), for example, from pastures to cropland, and then on the allocation of individual crops within the cropland. As producers in different AEZs are connected through land, labor, and capital markets, competition among land uses, and supply chains, the GTAP-AEZ model is ideally suited to study within-country changes in land use across AEZs. Moreover, through an explicit treatment of international trade flows, the GTAP-AEZ framework allows for tracking the effects of regional policies on land use patterns in other countries.

We updated the standard GTAP-AEZ model to include a nesting structure that separates the decision to convert forest to agricultural land from the decision to convert pasture to cropland, which is justified by the observation that deforested lands transition first into pastures, and then onto cropland³⁶. This nesting structure applies to all the regions. We also adopted regional elasticities of transformation, from natural covers to agricultural land, and between pastures and cropland, calibrated based on recent historical changes³⁶. We further updated the income elasticities of demand for agricultural and food products to reflect the latest work in this area³⁷. A critical assumption underlying the GTAP-AEZ framework is the productivity of marginal, hitherto, uncultivated lands, as it determines the extensive margin of land expansion. Another key assumption in the GTAP-AEZ model is the response of yields to changes in commodity and input prices³⁸. For both we use the assumptions in the original GTAP-AEZ model²⁶. Given the uncertainty regarding these parameters, we conduct extensive systematic sensitivity analysis of our results to alternative parametric configurations (Supplementary Information S2).

Underlying the model there is a database that consistently represents production, consumption, and trade patterns of 140 regions and 57 sectors in year 2011³⁹. To make solution times and model output manageable, we aggregated the model into 11 regions: Brazil, Bolivia, Argentina, Paraguay, Rest of Latin America, US-Canada (North America), European Union (28 countries), China, Malaysia and Indonesia, sub-Saharan Africa, and the rest of the world. We also collapsed the 57 commodity sectors into 18 sectors (i.e., paddy rice, wheat, coarse grains, oilseeds, raw sugar, grazing livestock, non-grazing livestock, forestry, extractive industries, processed livestock, vegetable oils, processed rice, processed sugar, other processed food, chemicals, manufactures and services). For Brazil, we considered the GTAP aggregate oilseed commodity as soybeans because soybeans account for more than 96% of oilseed production in Brazil⁴⁰. The database is complemented with data on agricultural land rents by land use and natural land covers at the level of Agroecological Zones (AEZ), also representative of 2011⁴¹.

Spatial footprint scenarios (SFS) and market share thresholds of zero-deforestation supply chain policies

The SFS (Fig. 1) are designed to assess how much deforestation would be avoided by implementing different configurations of the company- and importing country-led zero-deforestation supply chain policies. Except for the ASM, most policies–either voluntary or imposed—have not been implemented^1,4,7,21. Therefore, the SFS are counterfactual, non-observed states of the world. We estimate changes in deforestation and other economic outcomes as the difference between the counterfactual SFS and a baseline (as explained below). The baseline includes patterns of land use, land cover, and other economic outcomes obtained by letting the model simulate the changes in equilibrium as the economy responds to a set of drivers of land use during the period 2011-2016 (GTAP Database and AEZ Database, V9^39,41), without any land restriction to land expansion in Brazil.

The economic drivers include macroeconomic indicators (Supplementary Information Table S1), changes in agricultural factor and input productivity (Supplementary Information Table S2), and demand for biofuels (text S1). For each of these indicators, the value in 2011 is the average from 2010–2012 and the value in 2016 is the average from 2015–2017. These three-year periods are intended to smooth annual fluctuations in the different indicators. We use data up to 2017 in our analysis given the significant turbulence in soybean and other agricultural markets brough about by the US-China trade war started in January 2018. Soybeans, central to our analysis, saw a large divergence in the export prices to China charged by the US and Brazil⁴². Although the price gap eventually closed, the direction of trade flows changed significantly, especially for the U.S⁴³. Such turbulence was exacerbated by the global efforts to contain the spread of COVID-19, which triggered policy responses with potential significant worldwide effects on food consumption, production, and distribution⁴⁴. The export market shares of all the companies active in the Brazil’s soybean market from the Trase v2.4 database, are also available up to 2017.

The results over a five-year term horizon are representative of an economic medium run⁴⁵, which is long enough to allow economic agents to adjust their production and consumption patterns to the changes in prices brought about by land use restrictions. By focusing on a medium run we avoid the rigidities of economic short-run assumptions (i.e., lack of supply and demand response) as well as the significant uncertainties of economic analysis in the long-run (i.e., uncertain, or unpredictable future economic growth trajectories, technologies, and changes in international trade patterns).

The ASM was in place during our period of study, and it should be considered a baseline relative to further hypothetical policy developments analyzed here. The drawback of including the ASM in the baseline is that we would not be able to report the ASM outcomes, which, by virtue of its pioneering status, is a natural benchmark of future zero-deforestation policies in Brazil’s soybean supply chain. For this reason, we exclude the ASM from the baseline by not imposing land restrictions in the Amazon.

The SFS we evaluate are as follows:

1. 1.

   Amazon Soy Moratorium. This scenario uses the spatial footprint in the Brazilian Amazon of the companies that implemented the moratorium in 2006. These companies are: Abc Industria, ADM, Amaggi, Bunge, Cargill, Louis Dreyfus, Seara, Fiagril, Nidera, Noble, Cofco, Baldo, Imcopa, Agrex, CHS, Coamo, Engelhart CTP, Gavilon, Glencore, Invivo, Marubeni, Multigrain, Nova Agri, Olam, Perdue, Sodrugestvo, Timbro, and Selecta. Other companies that are part of the ASM do not export soybeans from the Amazon are: Binatural, JBS, Oleos menu, Agribrasil, and Culturale. The duration of the ASM has been extended indefinitely¹⁷.

2. 2.

   Global voluntary Zero-deforestation Commitments + Soy Moratorium. This scenario includes the ASM companies above and adds all global voluntary zero-deforestation commitment as if they were implemented in 2011. The global commitments have been pledged by a subset of the companies that agreed on the ASM. These are (pledge year and in parentheses): ADM (2015 company pledges); Amaggi (2017 company pledges), Bunge (2015 company pledges) Cargill (2014 New York Declaration on Forests); Louis Dreyfus (2018 company pledges); Cofco (2019 Soft Commodities Forum); Glencore (2019 Soft Commodities Forum); and Denofa do Brazil (2014 New York Deforestation of Forests).

3. 3.

   Import restrictions imposed by the European Union (EU). Agriculture-driven deforestation has become an increasingly polarizing issue between the EU and Brazil, and is a central issue in a potential trade agreement between the EU and the MERCOSUR, a trade bloc agreement among Argentina, Brazil, Paraguay, and Uruguay⁴⁶. We therefore also explore the effects of the adoption by the EU of mandatory rules currently considered by the European Parliament that would de facto require the implementation of zero-deforestation supply chain policies by the companies sourcing soybeans from Brazil⁴. We consider 155 traders exporting to EU plus Switzerland and the United Kingdom that would only procure their soybeans from areas already converted to agriculture prior to 2011. The EU countries appearing as importers consist of Belgium, Bulgaria, Croatia, Cyprus, Denmark, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Netherlands, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, and Sweden.

4. 4.

   Hypothetical import restrictions imposed by China, assumed to be similar to those being currently evaluated by the EU²¹. We consider 140 traders that export to China including Hong Kong (in addition to the ASM) that would only procure their soybeans from areas already converted to agriculture prior to 2011.

5. 5.

   Hypothetical import restrictions imposed by both China and the EU. These simulations provide an upper bound estimate of the ZCDPs.

In the SFS 3-5 (scenarios 3–5 in the main text) we assume that if a company supplying either EU or Chinese markets decides to produce deforestation-free soybean to preserve market share in one destination, the company will apply those restrictions to their entire supply chain. In other words, we do not allow for different supply chains from the same trader when exporting to different destinations. This is a realistic assumption as companies consider supply chain differentiation very costly due to the unwieldy procedures that would be needed to monitor, trace, and certify production^25,47.

Market share thresholds

Uncertainties exist regarding the critical market share (i.e., the percentage of total regional market share held by corporations with zero-deforestation supply chain policies) needed to discourage farmers from selling soybeans produced in recently cleared land to non-committed traders¹. If no soybean buyers within a region have a zero-deforestation supply chain policy, farmers have no incentives to avoid forest for soybean clearing. Alternatively, if only committed traders buy soybeans within a region, producers should be forced to comply with the zero-deforestation land use restrictions to sell their soybeans. In many Brazilian regions, traders with and without zero-deforestation supply chain policies purchase soybeans, thus producers with soybeans that are not zero-deforestation can typically sell their products. We posit that with increasing regional zero-deforestation supply chain policies market share, the difficulty of selling non-compliant soybeans increases. At some critical market share threshold, farmers may be completely disincentivized from producing soybeans on non-compliant lands that were recently deforested due to the difficulties in selling their product^48,49.

We bound the uncertainty about the competition structure needed to ensure compliance through three market share thresholds built using the export market shares of all the companies active in the Brazil’s soybean market from the Trase v2.4 database⁵⁰:

1. 1.

   The most restrictive market share thresholds requires that at least 75% of the soybeans exported from a given municipality are bought by companies with voluntary zero-deforestation commitments. In this scenario, 10% of the area under soybeans in Brazil is subject to the ASM (Fig. 1f). By adding pledged global voluntary zero-deforestation commitments in other biomes to the ASM, 27% of Brazil’s soybean area would be under agreements to halt forest conversion for soy production (Fig. 1g).

2. 2.

   A less conservative market share thresholds requires an export threshold of 50%. Under this scenario, the area under soybeans that is affected by global voluntary zero-deforestation commitments under the current pledges amount to 48% of Brazil’s total soybean area (Fig. 1g).

3. 3.

   The least restrictive scenario requires at least one committed company to be present in the municipality (>0% of market share covered by voluntary zero-deforestation commitment). Under global voluntary zero-deforestation commitments, 75% of Brazil’s soybean area would be subject to forest conversion restrictions (Fig. 1g).

Land cover definitions

We use two different definitions of forests in Brazil to accommodate different biome characteristics and land restriction targets⁵¹. Definition A was exclusively based on mapped forest cover. Definition B included natural grasslands outside the Amazon Biome, which may have high conservation value and are included in some traders’ zero-deforestation voluntary commitments [e.g., “Transforms our supply chain to be zero-deforestation while protecting native vegetation beyond forests.”⁵²]:

• Forest definition “A”: Forest is defined as forest only, as mapped by PRODES for the Amazon⁵³ and by Mapbiomas for other biomes [Mapbiomas v4⁵⁴, classes 1, 2, 3, 4, and 5]. Forest area was derived excluding forest regrowth, with forest base year of 2006⁵⁴. We used PRODES for the Amazon biome, because PRODES deforestation maps define the baseline for ASM monitoring, implementation, and enforcement. We used Mapbiomas outside the Amazon Biome. To our knowledge, Mapbiomas provides the most accurate and consistent large scale land use and land cover classification for Brazil.

• Forest definition “B”: Forest is defined as forest only in the Amazon biome, as mapped by PRODES, and forest and grasslands in all other Brazilian biomes, as mapped by Mapbiomas.

Land use and land cover databases

In addition to the data on soybean export market share and forests, we gathered municipality-level data on agricultural land cover from Mapbiomas v4⁵⁴: total cropland [classes 18–20] and pasture area [class 15], soybean area (ha), soybean production (tonnes), areas with both maize and soybean (ha), maize second harvest area (ha), calculated as the area of second harvest maize that is greater or equal to the area of soy harvested⁵⁵, and cattle headcount (heads)⁵⁶. These data were used to build the different versions of the GTAP-AEZ database, as explained below.

The supply-side spatial footprint scenarios, market share thresholds, and forest definitions give rise to thirty different databases with land use and land cover in each Brazilian municipality (five SFS * three market share thresholds* two Forest Definitions = thirty databases.) We use these land use/land cover databases to build biome-specific distributions of land cover and soybean production with and without zero-deforestation supply chain policies that can be used to calibrate the counterfactual experiments using the GTAP-AEZ model.

Model calibration

We use the databases discussed in Supplementary Information S1 to recalibrate the GTAP-AEZ model so that Brazil is split into biomes instead of the standard AEZs. This requires rebuilding the original GTAP and GTAP-AEZ databases. The algorithm to split Brazil’s agricultural output values into biomes proceeds as follows. For each spatial configuration of zero-deforestation policies, market share threshold, and forest definition, we use the following algorithm:

1. 1.

   Overlay the AEZ map used in the GTAP-AEZ database⁴¹ on a municipality-level map of Brazil⁵⁷. In case that a municipality is split across more than one AEZ, assign the municipality to the AEZ with the largest intersection.

2. 2.

   Overlay a biome map over the AEZ and municipality maps for Brazil. The biomes generally encompass several AEZs and the same AEZ can occur in different biomes. Biomes other than the Amazon and Cerrado are in an “Other” category. In case that a municipality is split across more than one biome, we assign the municipality to the biome with the higher policy implementation stringency, prioritizing the Amazon, second the Cerrado, and all other.

3. 3.

   Each municipality receives a unique id for each biome-AEZ combination, for example: AEZ5 becomes AEZ5-Amazon, AEZ5-Cerrado, and AEZ5-Other.

4. 4.

   Compute compliant market share thresholds for each municipality (using pledges as of 2017), and then categorize the biome-AEZ ids into compliant and non-compliant based on the SFS. For example: AEZ5 becomes AEZ5-Amazon-ZDC, AEZ5-Cerrado-ZDC, and AEZ5-Other-ZDC, and AEZ5-Amazon-Non-ZDC, AEZ5-Cerrado- Non-ZDC, and AEZ5-Other-Non-ZDC.

5. 5.

   Use aggregate municipality-level land cover (cropland, pasture, and forest area) and land use (soybean area, soy production, areas with both maize and soy, cattle headcount) to assign land cover and land use areas to each biome-AEZ-market share thresholds level.

6. 6.

   For all regions other than Brazil, build a conventional GTAP-AEZ database representative of 2011. This step uses a database of land use and land cover areas at the level of AEZs^41,58 to split the country-level output value of relevant products (crops, grazing livestock, and forestry) in the standard GTAP database³⁹ into AEZs.

7. 7.

   For Brazil, we use the AEZ-Biome area and production shares created in steps 1–4 to split the aggregate output values of oilseeds, coarse grains, grazing livestock, and forestry into biomes. Each new database represents a counterfactual year 2011 in which some of the area in each AEZ-BIOME was under a voluntary zero-deforestation commitment pledged before 2020. The simulations answer the question: how different area, production, and consumption would have been in 2016 if the pledged commitment had been in place since 2011.

8. 8.

   The area of the crops other than oilseeds and coarse grains (paddy rice, etc.) are shared out in each biome-AEZ in proportion to the cropland.

Implementation of zero-deforestation supply chain policies and deforestation leakage channels in the GTAP model

In each experiment we halt land conversion between forest and agriculture in the areas assumed under ZDCs by way of a subsidy that compensates producers for the economic losses of not transforming forests on to agriculture. Halting forest conversion in Brazil induces a shortage of agricultural land which drives up land rents in agriculture relative to other land uses. The effects of heightened land scarcity in Brazil may be transmitted to other regions of the world through changes in commodity prices. The strength with which these price changes affect other countries depends on Brazil’s global market shares of the commodity in question, and on the extent of competition in destination markets. In turn, changes in commodity prices alter the relative profitability of alternative land uses in these regions. Either in Brazil or abroad, leakage occurs as the higher returns to agricultural land incentivize land expansion into forests without zero-deforestation policies. Changes in land use are accompanied by a relocation of factors of production (land, labor, and capital) and other inputs (e.g., fertilizer) toward the production of the most profitable commodities. In addition to the reallocation of inputs, the model allows for substitution of non-land inputs for land in response to higher commodity prices.

Deforestation leakage rates

Following the literature on carbon leakage⁵⁹, we define the deforestation leakage rate as the increase in deforestation in regions without restrictions (No ZDSP, where ZDSP stands for zero-deforestation supply chain policy) induced by the measures taken in regions with zero-deforestation policies (ZDSP) as a percentage share of the absolute value of deforestation in regions with ZDSP. Formally:

Where (Delta) denotes the difference between deforestation outcomes in the baseline and counterfactual scenarios, and

Greenhouse gas emissions

Greenhouse gas emissions (GHGs) from the changes in land cover associated with the different experiments are calculated using the open-source AEZ Emission Factor (AEZ-EF) Model⁶⁰. The AEZ-EF model closely follows IPCC GHG inventory methods and relies on its default values. The model includes cover-specific (cropland, pastures, and forests) subnational carbon estimates for biomass (above and below-ground), dead organic matter, and soil carbon⁶¹. It also includes data on carbon remaining on harvested wood products, non-CO₂ emissions, and foregone sequestration. The carbon stock data is combined with assumptions about carbon sequestration from forest growth (foregone if converted), mode of conversion, and CO₂ emissions from land clearing using fire, and the fraction of carbon that remains sequestered in wood products during a 30-year time horizon. The AEZ-EF model is designed to estimate land use emissions from land use transitions predicted by comparative static economic models, whereby one starts with a baseline and estimates the resulting final equilibrium. The AEZ-EF model underlies the emission estimates in several analysis of the indirect land use effects of biofuels emissions and land conservation measures^15,62.