Forest definitions require a re-think

Tracking global deforestation is key to the implementation of forest conservation and emissions reduction policies. To make tracking meaningful, the terms forest extent and deforestation must be defined in a way that is simple, transferable, and applicable with existing measurement and monitoring techniques.

Forest monitoring increasingly relies on Earth observations from satellites, yet the definitions of forested and deforested area largely originate from before the satellite era. Though different definitions may be appropriate in different contexts, deforestation mitigation policies that require a monitoring system need definitions designed to facilitate the application of remote sensing data. By contrast, most existing forest definitions are not made with remote sensing in mind, which compromises both conservation monitoring and policy implementation.

For example, in order for a carbon market to function, carbon credits must be fungible, or mutually interchangeable, such that an emission reduction credit can be used to offset an emission elsewhere. However, when both forest extent and the rate of deforestation vary by location and depend on the definitions used to quantify them, it becomes impossible to evaluate consistently across jurisdictions or against global datasets whether deforestation has declined.

Three important issues limit the current definitions of forest and deforestation: (1) variations across contexts and geographies lead to inconsistent estimates of forest and deforestation; (2) the focus on forest extent over forest loss can result in significant areas of unaccounted deforestation; and (3) current definitions are too often incompatible with the satellite observations used to monitor forests and deforestation.

We argue that we must re-think the way ‘forest’ and ‘deforestation’ are defined to make global deforestation monitoring efforts more accurate, consistent, and efficient.

Limitation 1: Inconsistent forest definitions

In 2000, The Food and Agriculture Organization (FAO) adopted a new consensus forest definition, which included minimum biophysical thresholds of 10% canopy cover, 5 m tree height, and 0.5 ha¹. Though countries were encouraged to adopt this definition, the FAO pushed for ‘harmonization, not standardization’ of forest definitions². As a result, definitions continue to vary significantly across countries and contexts, with minimum area thresholds typically ranging from 0.1ha to 1ha and minimum canopy cover thresholds ranging from 10% to 60%³.

Forest and deforestation definitions not only vary between countries; they can also vary within the same country. For example, in its Forest Reference Emission Level (FREL) submission to the United Nations, Brazil defines forests using a 0.5 ha minimum area threshold but its official monitoring systems employ 1ha, 3ha, and 4ha minimum area thresholds⁴.

Land use adds an additional consideration for forest definitions. For example, in addition to the biophysical thresholds of area, canopy cover and height, FAO’s definition also includes “trees able to reach these thresholds in situ” and “areas that are temporarily unstocked due to clear-cutting as part of a forest management practice”⁵. Considering these criteria, timber and rubber plantations are considered forests under FAO’s definition, whereas agricultural tree crops such as oil palm and fruit tree plantations are not. Under FAO’s definition, countries can clear natural forest to establish a timber or rubber plantation and claim that no deforestation occurred. Alternatively, they can exclude forest plantations from their forest definition (e.g., Colombia⁶), or they can include plantations in the forest definition even beyond those included in FAO’s definition (e.g., cacao and coffee in Republic of Congo⁷).

Limitation 2: Prioritizing measurement of forest over deforestation

Determining a single set of biophysical thresholds for minimum area, tree cover, and height involves certain tradeoffs. For example, applying forest definitions that employ 0.5 ha minimum area and 10% tree cover thresholds bring important open forest ecosystems into monitoring frameworks. However, using such thresholds could also result in the omission of deforestation in dense tropical forests: a forest area of 0.5 ha could lose up to 90% of its tree cover and still be defined as ‘forest.’ This unaccounted deforestation can add up, particularly in regions with high forest fragmentation or that are dominated by small-scale clearings. For example, the Democratic Republic of Congo has had the second highest annual rate of primary forest loss in the world over the past two decades⁸, driven largely by a large local population that clears small forest patches for subsistence agriculture⁹. In 2023, over 20% of that loss occurred in patches smaller than 0.5 ha and over 40% in patches smaller than 1ha, representing important areas of deforestation that risk being omitted from any official accounting system employing these minimum area thresholds (Fig. 1).

Fig. 1: Share of 2023 primary forest loss in patches smaller than commonly employed minimum area thresholds (0.5ha and 1ha).

Data obtained from overlaying tree cover loss data for 2023¹² over remaining 2022 primary forest extent¹³, and filtering for loss patches ≤5 Landsat pixels (0.45 ha) and ≤12 Landsat pixels (1.08 ha) using 8-way connectivity. Countries shown correspond to those with the highest primary forest loss globally in 2023. DRC is the Democratic Republic of Congo; PNG is Papua New Guinea.

Limitation 3: Non-alignment with satellite data

Though the number of Earth-observing satellites in orbit has increased rapidly in recent years, only the Landsat series and the Sentinel-2 series provide optical data that meet the requirements needed to enable the operational monitoring of deforestation. Namely: an unbiased global acquisition strategy, a consistent and adequate revisit time, spatial and spectral resolutions that are appropriate for detecting the land dynamic of interest, and robust georeferencing and pre-processing to high-quality, consistent surface reflectance data¹⁰ (Fig. 2). Additionally, the satellite data archive must be available for the desired monitoring dates and the data must be free and openly accessible in order to enable transparency, reproducibility, and consistent measurements through the years.

Fig. 2: Data requirements for mapping and quantifying deforestation.

Example time-series imagery from Landsat (L9), Sentinel 2 (S2B), and Google Earth (Maxar Technologies) for a mining site in Haut-Uele province, Democratic Republic of the Congo. Subsets (a–c) depict composites of shorter (red, green, blue) to longer (near infrared, short wave infrared 1, short wave infrared 2) wavelengths for Landsat 9 and Sentinel 2B imagery, illustrating that wavelengths beyond the visible capture deforestation dynamics in higher contrast and detail. Subset (d) is the only available Google Earth image available for this site (the deforestation event is completely missed). Google Earth cannot be employed for forest monitoring due to its lack of systematic acquisitions. Subset (e) shows the spatial mismatch of 0.5ha grids (white squares) on 10–30 m image subsets, illustrating that the 0.5 minimum area threshold is incompatible with Landsat and Sentinel 2 data. Subset (f) illustrates per pixel forest loss information in a graphical time-series of Sentinel 2 NDVI data for a pixel at the center of e). Such time series are required for sample interpretation and are available at the pixel scale, not at the scale of 0.5ha squares. Subsets (a–c) and (e, f) illustrate the requirements needed for operational deforestation monitoring, including appropriate spectral and spatial resolutions, a systematic global acquisition strategy, robust georeferencing, and surface reflectance correction.

However, in practice, it is not possible to precisely determine tree cover within the commonly used 0.5 ha square using Landsat data, because the sensor’s spatial resolution (30 × 30 m pixels) does not perfectly nest within 0.5 ha (70.71 × 70.71 m) (Fig. 2e). Though the mismatch between the 70.71 × 70.71 m grid and the Sentinel-2 10 × 10 m grid is smaller over a single 0.5 ha square, this discrepancy still quickly adds up to significant shifts over larger areas.

The 0.5 ha minimum area threshold also results in an important loss of efficiency and precision in the measurement of forests and the changes within them. The decrease in measurement precision is evident in Fig. 1 where, depending on geography, 2% to >40% of 2023’s total primary forest loss could remain unquantified based on the use of 0.5 ha or 1 ha minimum forest area thresholds. Though deforestation could be measured with the precision of 30 × 30 m or 10 × 10 m pixels, instead it is measured with the equivalent of 70.71 × 70.71 m pixels. This represents a >5× loss of precision compared to the Landsat resolution and a 50× loss of precision compared to the Sentinel-2 resolution.

Further, policy compliant deforestation area estimates come from probability-based samples of reference data, not pixel counts from maps (Box 1). The loss in efficiency due to the larger-than-needed minimum area thresholds is pronounced in sample-based area estimation. When using a 0.5 ha definition, human interpreters must create reference labels (e.g., ‘forest’ or ‘deforestation’ labels) for a sample of 0.5 ha squares, a task that is technically challenging and time-consuming. Pixel-based interpretations of medium spatial resolution data offer a more efficient and accurate alternative. Using 0.5 ha forest definitions therefore represents a lose-lose situation: a massive increase in the effort required for sample interpretation with an accompanying loss of precision in the final area estimate. If remote sensing data are to be employed for monitoring deforestation, then forest definitions should align with the specific characteristics of Landsat and Sentinel-2 data sources.

Box 1 Policy implications

Existing forest definitions can negatively impact policy implementation when remote sensing data are used directly for decision-making related to legality, payments for ecosystem services, and other policies and programs designed to reduce deforestation. For example, the convention of allowing significant variation in forest definitions across countries becomes problematic when it is carried over into the context of the Voluntary Carbon Market (VCM), as is the current status quo. A high-quality credit from avoided deforestation should represent the same emission reduction regardless of whether it was generated in Brazil, Ghana, or Vietnam. But when both forest extent and the rate of deforestation differ depending on the definitions used to quantify them, it becomes difficult to evaluate consistently across jurisdictions or against global datasets the degree to which deforestation has declined. As such, the quantity of emission reductions credits generated may vary depending on where they originate, potentially eroding buyers’ trust in the system.

Another example of how forest definitions can negatively impact policy implementation comes from European Union’s Deforestation Regulation (EUDR). This regulation establishes that products imported into the EU must not result from recent deforestation or forest degradation¹⁴. To enforce this regulation, the EU will require importing companies to provide evidence, including remote sensing data, proving that certain commodities were not produced on land deforested or degraded after 2020. For the EUDR, forest and deforestation are defined using 0.5 ha minimum area, 5 m height, and 10% tree cover thresholds, regardless of the exporter country’s forest definition. Companies trading in products impacted by the EUDR may therefore need to comply with the EU’s laws following one forest definition and with national environmental laws following a different definition. Political tensions and monitoring complications may arise from disagreements regarding differing definitions. Further, the EUDR’s forest definition is not well suited to enable monitoring with remote sensing, because the minimum area threshold does not align with satellite data spatial resolutions. To try to adhere to the EUDR definition, practitioners will have to take additional processing steps that take additional time, lower the quality of map products, and still ultimately fail to fully adhere to the 0.5 ha threshold. For example, the 2020 Forest Cover map developed by the EU’s Joint Research Centre (JRC) was subject to filtering and smoothing algorithms (removing treed areas smaller than 0.5 ha and ‘filling in’ non-treed areas smaller than 0.5 ha surrounded by trees)¹⁵, resulting in a loss of precision and accuracy in mapping the real extent of the biophysical asset of interest.

Box 2 Impact of forest definitions on sample-based area estimation

Rather than calculating deforestation area using “pixel counts” from maps, the good-practice approach for estimating areas of deforestation involves using an unbiased estimator from a probability sample of reference data, resulting in estimates with known uncertainty^16,17,18. To employ this methodology using a 0.5 ha definition such as that of the FAO, human interpreters create reference labels (e.g., ‘forest’ or ‘deforestation’ labels) for a sample of 0.5 ha squares based on the visual interpretation of any available satellite data and other data sources. Interpreters determine, for each sampled 0.5 ha square, whether forest cover meets or exceeds the canopy cover threshold as set in the forest definition (e.g., 10%). This labeling can at times be performed using very high-resolution data such as that available on Google Earth, which can enable a precise determination of percent tree cover for any sample unit size. However, we cannot rely on this source of imagery for operational global deforestation monitoring, primarily due to its incomplete temporal and geographic coverage (Fig. 2d). Interpreters must instead rely on Landsat and Sentinel-2 data to apply the 10% tree cover and 5 m height thresholds for each 0.5 ha square. Instead of interrogating a single pixel, interpreters have to examine upwards of four Landsat pixels and/or upwards of 49 Sentinel-2 pixels to determine the reference label for each 0.5 ha sample unit. And since Landsat and Sentinel-2 data do not nest within 0.5 ha grids, a 0.5 ha sample unit will contain portions of Landsat/Sentinel-2 pixels, for which it is impossible to assign any label. As a result, confidently labeling ‘forest’ or ‘deforestation’ using 0.5 ha sample units and Landsat and Sentinel-2 data is not only grossly inefficient, it is also technically infeasible.

Standardize minimum area thresholds and exclude tree plantations

Remote sensing is critical to forest monitoring in support of policy frameworks (Box 2). To support conservation and climate change policy frameworks, we argue that definitions of forest extent and change must be standardized at the scales of public earth observation imagery at 10–30 m and distinguish between natural forests and treed land uses, thereby excluding plantations of wood or other agricultural products. Following an assessment of tree cover change, assigning a designation of forest type or land use to a pixel does not require the use of a larger minimum area threshold. In sample interpretation, this can be done using the landscape context where the pixel is found (also known as the spatial support unit¹¹).

We see several benefits. First, exposing all deforestation dynamics consistently would allow for more internationally coordinated conservation action and more effective policy implementation. Second, prioritizing the monitoring of deforestation, rather than standing forest extent, would lead to a more precise quantification of the associated carbon emissions and loss of valuable ecosystem services. Third, removing the issue of non-nesting grids associated with applying the 0.5ha forest definition on Landsat and Sentinel-2 data would eliminate technical barriers to applying forest definitions with remote sensing data and significantly increase the precision and efficiency of mapping and estimating areas of deforestation.

We need a pragmatic approach that balances the ecological significance of forests with the technical requirements of monitoring and conservation initiatives. We should adopt a standardized minimum area threshold that aligns with satellite data capabilities and remove the possibility for forest definitions to include plantation land uses. Only then can we ensure a more consistent and reliable quantification of forests and deforestation that supports international initiatives and carbon markets and contributes to global conservation efforts.