Agricultural expansion is the driver of more than 90% of deforestation across the tropics (1), a major source of anthropogenic carbon emissions and biodiversity loss (2, 3). Establishing who is responsible for this expansion – which crops, which companies, and which consumers – is, however, not a straightforward task. The drivers of deforestation can either be identified using agricultural statistics to determine which products are expanding in a region (4) or intersecting remote sensing data to identify which products are produced on recently deforested land (5, 6).
When available, remote sensing data are typically preferred (7), because they are available at high resolution (ranging from 250 m to 1 km for MODIS data down to 30 cm for products such as Maxar) and can include uncertainty estimates for land use classifications. Statistical data are reported at subnational or national-level and may be partly based on estimates by experts with unknown uncertainty. Fueled by this new era of high-resolution data, governance initiatives are increasingly focussed on monitoring land use change at farm-level (8). The European Union’s new regulation to prevent the import of commodities linked to deforestation, for example, requires plot-level monitoring (9). Similarly, corporate emissions accounting guidelines, such as the World Business Council for Sustainable Development’s Greenhouse Gas Protocol and the accompanying Science Based Targets initiative, prioritize ever more fine-scale analyses (10).
Here, we caution, however, that farm-level analysis is not necessarily best when it comes to identifying the drivers of land use change, and that pixel- or farm-level monitoring should be seen as one component within a bundle of analyses and monitoring at multiple scales of resolution (i.e. from plots and farms to subnational regions or at national-scale). We make our case using a recently published remote sensing product for Brazil, Mapbiomas v7.1 (11), which offers a uniquely detailed (i.e. crop-specific), long-term mapping of all land uses in the country, thereby capturing the land use dynamics of multiple commodities in a harmonized methodology across 36 years.
Specifically, we compare the drivers of deforestation which are identified when aggregating and analyzing the original data (30 x 30 meters) at five different observational scales: at the level of individual pixels, properties, municipalities, states, and the country as a whole. The time series includes changes in agriculture (including soybeans, pasture, sugarcane, coffee, cotton, rice), tree plantations, and natural vegetation types (forests, savanna, and grasslands, and other non-forest natural formations). We use the term ‘deforestation’ as a shorthand when referring to the conversion of natural vegetation, though they are not all strictly forested. We note that pasture expansion is not only motivated by its use for cattle ranching, but also land speculation, as pasture is one of the cheapest land uses to establish post-deforestation. The soy class doesn’t distinguish between fields with one or multiple harvests - soy is often double-cropped with maize or cotton. Where the post-deforestation land use (i.e. crops, pasture) could not confidently be determined, pixels are classified as a ‘Land Use Mosaic’.
When the data are aggregated beyond the pixel resolution, we identify the driver of deforestation using the ‘product expansion’ approach, as commonly adopted in other studies and accounting standards for land use change emissions (4, 12, 13). In the supplementary material we compare our results across alternative approaches (SI Appendix Fig. 1). Product expansion attributes the loss of native vegetation to expanding products based on their share of expansion:
$${PAF}_{r,p}= \frac{{A}_{r,p,{y}_{1}}- {A}_{r,p,{y}_{0}}}{{\sum }_{p}\left({A}_{r,p,{y}_{1}}- {A}_{r,p,{y}_{0}}\right)} \left[\text{E}\text{q}\text{u}\text{a}\text{t}\text{i}\text{o}\text{n} 1\right]$$
Where \({PAF}_{r,p}\)is the ‘product allocation factor’, the proportion of agricultural expansion across the study time period (y0, y1) which was due to each product p, in each region r (where \({A}_{r,p,y}\) is the area of product p, in region r, in year y). We ran the analyses for 1985–2021 and 2000–2021 for all regions, and for 2008–2021 when comparing property and pixel-level results in Mato Grosso in the period after the establishment of the Soy Moratorium. To convert \({PAF}_{r,p}\) from a proportion (0–1) of agricultural expansion into an area of deforestation (in hectares), \({PAA}_{r,p}\), we multiplied it by the area of native vegetation (NV) converted in each region r (Eq. 2).
$${PAA}_{r,p}={PAF}_{r,p}*\left({NV}_{r, {y}_{1}}- {NV}_{r, {y}_{0}}\right) \left[\text{E}\text{q}\text{u}\text{a}\text{t}\text{i}\text{o}\text{n} 2\right]$$
We calculated \({PAA}_{r,p}\) for several scales of region (property, municipality, state, and country) before summing and comparing results at the national level (\({PAN}_{p}\), as a percentage of the total deforestation (Eq. 3)).
$${PAN}_{p}= \frac{\sum _{r}{PAA}_{r,p}}{\sum {PAA}_{r,p}}*100 \left[\text{E}\text{q}\text{u}\text{a}\text{t}\text{i}\text{o}\text{n} 3\right]$$
The state-level results, for example, were calculated separately for each of Brazil’s 27 states to capture the land use dynamic within each state, before being summed to build a picture at the national level. For the calculation at property-scale, we intersected Mapbiomas data with polygons of different land tenure (14) to calculate \({PAF}_{r,p}\)in all polygons and land tenure classifications (capturing land use change in individual private properties, legal reserves, and indigenous lands). All analyses were done in R 4.1.3 (15) and Google Earth Engine (16).
How scale affects the attribution of land use change to different drivers
Though pasture is typically considered the main driver of deforestation in Brazil (17, 18), across both time periods (1985–2021 and 2000–2021), its relative importance decreases as the scale of analysis increases (Fig. 1B and C). Mirroring this, the proportion of deforestation attributed to soy is 3–4 times greater when calculated at national- rather than pixel-level (36% vs 14% for 1985–2021 and 55% vs 14% for 2000–2021). Sugarcane and forest plantations also emerge as major drivers of deforestation at national-level (Fig. 1B and 1C). These patterns arise because analysis at larger scales captures indirect land use change – where the expansion of one land use (the ‘indirect’ driver) displaces another, which is the ‘proximate’ or ‘direct’ driver of deforestation (5, 18, 19). In this example, though the net area of pasture in Brazil has been stable or declining since the mid-2000s (Fig. 1A), cattle pasture across the south and central Brazil has been displaced by the expansion of soy, sugarcane, and tree plantations; pasture has shifted notably northward, expanding, in particular, at the expense of forests in the Amazon biome (17, 19–22).
How scale affects the solutions to land use change
In recent decades, a series of governance efforts have emerged to tackle deforestation. Prior to the 1990s, governance of land use was the domain of domestic governments (23), who set constraints on land use, including the establishment of protected areas, recognition of indigenous territories, and rules for conservation on private land – such as Brazil’s Forest Code, established in 1965 (24). As globalization accelerated, production and consumption became increasingly distanced; with domestic policy seemingly failing to deliver products with the standards demanded by international agendas (e.g. on deforestation, farmer incomes, pollution, forced labor), governance shifted towards non-state actors and from being place-based to flow-based (25), with a variety of voluntary governance initiatives organized around global value chains. Examples include the rise of certification schemes, company standards, and zero deforestation commitments (26). In the 2000s, various subnational approaches to governing land use emerged – REDD + projects to incentivize subnational action on climate change and biodiversity loss, and then jurisdictional sourcing initiatives, which seek to expand the horizons of companies, to look beyond their supply chains and account for impacts in their sourcing regions (27–30). Most recently, flow-based legislative efforts have also emerged, with the European Commission’s recently adopted due-diligence legislation seeking to ensure that products imported into the EU do not come from recently deforested land (9).
These different governance structures seek to address land use change at different levels and so target different land use change processes (Fig. 2). Supply chain commitments are focused on pixels (e.g. the Amazon Soy Moratorium (31, 32)) or properties (e.g. zero deforestation commitments in the cattle sector (33)), and so address ‘direct’ or ‘proximate’ drivers of deforestation. The difference between pixel- and property-level assessments and policy focus is well illustrated by the implementation of the Soy Moratorium. The Moratorium prohibits the planting of soy on land deforested after 2008 in the Amazon biome, though it is monitored at the pixel-level; farmers may thus plant soy on pasture and then clear land elsewhere within their property (34). When comparing soy expansion in the state of Mato Grosso (the main soy producer in the Amazon) since 2008, the proportion of deforestation attributed to soy is almost double when analyzed at the property- rather than the pixel-level (20.3% vs 11.4%). Jurisdictional efforts help capture regional and cross-commodity land-use dynamics, though they currently receive little funding relative to their prominence in corporate social responsibility materials (30, 35). Legislative approaches can govern land use change at multiple levels – they may target areas at the pixel level (e.g. regulations prohibiting clearing in riparian areas), constrain land use within properties or protected areas, and enact regional land use planning. For example, expansion of sugarcane in the Amazon has been constrained by the Sugarcane Agroecological Zoning, launched in 2009 and lifted in 2019 (36). This policy prevented sugarcane from becoming a direct driver of deforestation, but did not address its indirect role through the displacement of other agricultural land uses (36). Finally, national policies can also address the ‘underlying’ drivers of deforestation, which ultimately fuel the direct and indirect drivers of deforestation. Underlying drivers include demand for land-intensive products, such as livestock products and sugarcane ethanol biofuels, or land tenure uncertainty. Governments can modulate demand for land-intensive products through subsidies and taxes and address land tenure issues by land titling programs and assigning undesignated land to conservation (37).
Action at different scales is complementary - and necessary
Because of their different characteristics, pixel, property, jurisdictional, and national efforts can be complements in governing land use change. Indeed, this complementarity is seen as a strength of polycentric governance more generally (38). We are concerned, therefore, that the sustainability agenda is stampeding toward a narrow focus on high-resolution mapping (i.e. farm-level traceability and monitoring) and pixel- or property-level action as the key to addressing commodity-driven deforestation. The European Union’s regulation on deforestation has launched a race for companies to trace and measure compliance at the farm-scale. The race is also on for companies to set climate mitigation targets – where, again, property-level action is being prioritized. The greenhouse gas protocol distinguishes between ‘direct land use change’, which is measured at farm or land management unit level, and ‘statistical land use change’ (such as the product expansion methodology shown here) which is calculated across sourcing regions. Though the protocol acknowledges the complementarities of these two metrics, it currently requires companies to select a single metric for scope 3 (i.e. supply chain emissions) reporting, with a clear tendency for companies to switch to reporting direct land use change emissions wherever data allow (10).
Moreover, not only is such property-level action limited to tackling the direct drivers of deforestation, but it is far from guaranteed to provide robust and accurate indicators. The traceability required for farm-level monitoring is extremely challenging, particularly in commodity-producing landscapes where diffuse smallholder production, informal trading relationships, multi-tier-supply chains, corruption, and illegality are common (35).
The European Union’s commodity deforestation legislation and the greenhouse gas protocol are both key initiatives because they are agenda-setting, defining what ‘success’ looks like in reducing deforestation and carbon emissions. It is essential, therefore, that they encourage action at multiple scales. The window of opportunity has not closed: the EU deforestation legislation, approved in May 2023, includes periodic reviews 1–5 years after implementation and the greenhouse protocol is currently under revision before final publication in May 2024. The EU deforestation legislation should support jurisdictional sourcing initiatives by counting them toward its subnational risk benchmarking and by allowing companies to include sourcing from jurisdictions with time-bound, transparently-monitored land use planning and demonstrated deforestation reductions as an approved criteria for due-diligence reporting. In the revisions of the greenhouse gas protocol, statistical land use change should be elevated to an obligatory metric given equal weight with direct land use change emissions. Only by measuring progress and defining success at multiple scales can initiatives for sustainable commodity sourcing create the right mix of incentives for addressing deforestation.