the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Aug 2025
GFC2020: A Global Map of Forest Land Use for year 2020 to Support the EU Deforestation Regulation
Abstract. Earth observation (EO) data are used to map tree cover extent, estimate canopy height, detect disturbances, and classify land cover and land use. However, comprehensive global information on forest cover, capturing both physical characteristics and land use components as defined by the United Nations Food and Agriculture Organization (FAO), remains limited. Here, we present a harmonized and globally consistent map of forest presence or absence at 10 meter spatial resolution for the year 2020, hereafter referred to as GFC2020. Our approach combines multiple spatial datasets, primarily derived from EO, to harness their complementary strengths within a transparent, flexible, and open science framework. GFC2020 maps 4,562 million hectares (Mha) of forests globally, which is 12 % more than the estimate from latest FAO Global Forest Resources Assessment (FRA). Approximately 11 % (~578 Mha) of tree cover is excluded from forest area in GFC2020, primarily because it does not meet the height threshold or occurs on agricultural or urban land. Conversely, around 0.6 % (~25 Mha) of the area classified as forest in GFC2020 is unstocked, due to forest management practices or natural disturbances such as fire. Based on the reinterpretation of an existing reference set of 21,752 sample units, GFC2020 achieves an overall accuracy of 91 %, with 18 % probability of overestimating the forest area and 8 % for underestimation. Future improvements in EO products, such as better detection of trees in dry and open landscapes, distinguishing natural from human drivers of forest disturbance, mapping tree crops at high spatial resolution or identifying agroforestry systems, will contribute to enhancing future versions of GFC2020. The shift from tree cover to forest cover mapping is not only essential for ecological and climate related applications but also provides new opportunities to support policy needs. GFC2020 (https://forobs.jrc.ec.europa.eu/GFC) is one of many tools to inform the deforestation risk assessments under supply chain oriented regulations such as the European Union Deforestation Regulation (EUDR). This map is not mandatory, not exclusive and not legally binding.
- Preprint
(4122 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
CC1: 'Comment on essd-2025-351', Meine van Noordwijk, 15 Aug 2025
-
AC5: 'Reply on CC1', Clement Bourgoin, 05 Nov 2025
- In line 81-90 authors correctly clarify the challenge that ‘tree cover’ as observable characteristic of land differs from ‘forest’ as a land use category in policy designs and instruments. Yet, in subsequent text it seems that ‘tree cover’ is equated with ‘forest’.
In the paper we make a clear distinction between tree cover and forest. However, we use tree cover data in the first step of the mapping methodology to identify areas where forests could potentially exist and then exclude treed areas that do not correspond to the forest land use definition. This is broadly explained in the last paragraph of the introduction, in more detail in section “2.2 Approach” and in various parts of the results and discussion sections including a reflection on challenges and limitations (section 4.1.2).
- Line 103-109 claims that the GFC2020 maps ‘align with’ EUDR and FAO forest definitions. The text acknowledges the challenges in this claim especially where tree crops are involved that are the primary concern of EUDR regulations, but don’t follow up on these concerns.
We do follow up on these concerns and in the case of commission errors of GFC2020 due to confusion with tree crops, we detail this in the section 4.1.2 “Challenges and main limitations”. In this section, we compare against several commodity layers and assess potential maximum reductions in commission errors. Furthermore, our accuracy assessment reflects and quantifies the challenges in the separation between forest and land uses that may appear similar from a land cover perspective; section “3.2.2 Assessment of forest and land use types” and especially Figure 6 assess the commission error with other land uses.
- Specifically, a recent publication (van Noordwijk et al. 2025) suggested multiple types of evidence for an ‘agroforestry’ (and thus non-forest) status of land in an institutional interpretation of the EUDR and FAO forest definitions, regardless of tree cover. It would be appropriate if authors comment on these evidence categories and the way they Cn (or should) be taken into account if the target is to create an EUDR-policy relevant map of 2020 global forest cover.
Thank you for highlighting this relevant citation. The categories presented in Table 3 of van Noordwijk et al. (2025) illustrate additional types of information that can be provided by operators or traders during the deforestation risk assessment as part of the due diligence process. It is important to note that the legislative text of the EUDR does not mention any mandatory use of maps for its implementation, meaning there will not be any official global "EUDR-policy map." However, maps—whether at global, national, or regional scales—can serve operators on a voluntary basis as supporting information in the risk assessment phase. Their use should be complemented by other sources of evidence, such as local maps from very high-resolution imagery, or relevant non-spatial data regarding year 2020, given the uncertainties at local level inherent in any spatial dataset that cover large areas such as the globe or an entire country.
We added a reference to (van Noordwijk et al. 2025) in section 4.3 as follows: “The GFC2020 dataset is intended to support operators and traders as one of several tools for deforestation risk assessment during the due diligence process. Specifically, it can help in the preliminary identification of plots where more detailed or locally relevant data should be gathered for a robust risk evaluation. Given this intended use, the overestimation of forest area in GFC2020 may result in additional assessments by operators. False positives, i.e. areas that are wrongly mapped as forest in GFC2020, will likely be identified as non-forest during subsequent assessments using more detailed or locally relevant data. In contrast, false negatives, i.e. forest areas that are omitted in GFC2020, could be a major concern for operators as deforestation risk areas may be missed. It has to be noted that the overestimation of forest by GFC2020 varies by regions and commodities, in particular with significant overestimation in regions with agroforestry systems like coffee and cocoa. Therefore, we strongly encourage operators to complement GFC2020 with national or regional forest cover datasets that align with the relevant definitions set out in Article 2 of the EUDR, particularly datasets that offer high spatial resolution and known accuracy. Alternatively, ground samples, geotagged photographs or non-spatial data could be used by operators to support or enrich the risk assessment (van Noordwijk et al. 2025).”
- The procedure described in lines 200-203 can deal with part of the ‘agroforestry’ area, but certainly not all.
We agree, and this is partly due to the high omission error for the agroforestry class of the Lesiv et al. (2022) Global Forest Management Map. Future versions of this product from IIASA, along with other advancements in remote sensing for detecting understory crops, are expected to address this limitation.
- Where the paper employs standard concepts of ‘producer’ and ‘user’ accuracy, van Noordwijk et al. (2025) used (and pleaded for wider use of) a more specific ‘user accuracy’ in the context of EUDR. The most relevant use of the maps in EUDR context is to evaluate whether or not tree crops (incl coffee, cocoa, rubber) marketed after 2025 were derived from land deforested before or after 2020. A ‘real user’ accuracy would assess the likelihood that plots known to already produce any of these commodities before 2020 is correctly classified as ‘non-forest’. A number of studies, incl van Noordwijk et al. (2025), but also studies for Cameroon, Ivory coats and Peru that are on their wway to publication, have found erroneous classification of such points to be around 60%. Before recommending the use of the current GFC2020 map for EUDR implementation, this issue may need to be addressed (or at least acknowledged).
Thank you for this insightful comment. A map needs to be assessed for the spatial domain for which it was produced. Given that this study produces a global map, we report the accuracy assessment at both global and continental scales, using established accuracy metrics and adhering to best statistical practices as outlined by Stehman et al., 2013. It would be interesting to assess the map’s performance against plots that are already producing EUDR-targeted commodities, as this could provide some evaluation of how well GFC2020 captures relevant land use. However, such plot-level data are typically not publicly available from such studies and there is definitively not any available global dataset of plot level data on EUDR commodities. As an alternative, we assessed the potential overestimation of GFC2020 by comparing it against independent regional or global spatial datasets that were not used in the development of GFC2020 v2 (see Figure 10).
It is important to emphasize that we only recommend to use GFC2020 as one potential initial source of information in the risk assessment phase that needs to be complemented by other sources to correctly assess the risk of deforestation. GFC2020 is non-exhaustive, non-mandatory, and has no legal status. It is an open-access, freely available representation of forest land use, developed using a forest definition that aligns with that of the EUDR. We are transparent about the map’s limitations, including known omission and commission errors—that are common to all geospatial products. Future improvements in global spatial datasets will be integrated into subsequent versions of the map. Ultimately, it is the responsibility of the operator or trader to decide whether to use this global map in combination with other spatial and non-spatial sources to support their deforestation risk assessment.
Citation: https://doi.org/10.5194/essd-2025-351-AC5
-
AC5: 'Reply on CC1', Clement Bourgoin, 05 Nov 2025
-
RC1: 'Comment on essd-2025-351', Anonymous Referee #1, 09 Sep 2025
This paper describes a dataset which is hugely valuable to a wide range of users in light of the forthcoming EUDR implementation. The paper is well written and the methods and approach are robust and fit for purpose. The authors have identified the most appropriate datasets for the purpose of building a map which describes forest land use in the year 2020 – with some known limitations. The use of only datasets which are global in scope is an understandable approach, but the map could probably be improved if local/regional/national datasets were utilized, and this highlights the value of any global datasets which themselves incorporate local data bring to this mapping challenge.
Here are a few minor comments and reflections:
Line 23-24 - consider rephrasing to say "GFC does not include 578Mha of tree cover (11% of the global tree cover area) because it does not meet the height threshold or occurs on agricultural or urban land"
Would it be worth explaining the benefit of a conservative approach (higher commission vs omission error) in light of a more cautious approach for ensuring that products conform to EUDR? While acknowledging that this makes it more challenging for those who are demonstrating that their products are deforestation free.
Line 40 - would a broader term indigenous peoples and local communities make sense here?
Line 140 - can we add 'continued collaborations' because they are already ongoing
Is it possible to provide some more clarity on the forest land use interpretation - for example forest/logging roads, fire breaks, skid trails and other forest related activities. Are they marked as forest in your interpretation and in the validation?
It seems clear to me that the majority of the commission errors are in areas where potential production for EUDR commodities will occur. I wonder whether this needs to be more prominently mentioned.
Very high commission error in Africa – can be done about this in future?
Perhaps in the discussion you could mention the value of specific information on tree crops’ year of planting, to mitigate related limitations in the map.
Line 389 - thematic ambiguity by the interpreter - not sure what this completely means – can you elaborate or rephrase?
Perhaps worth a mention that EUDR also has a legality component which cannnot be addressed by this map?
Citation: https://doi.org/10.5194/essd-2025-351-RC1 -
AC1: 'Reply on RC1', Clement Bourgoin, 05 Nov 2025
- This paper describes a dataset which is hugely valuable to a wide range of users in light of the forthcoming EUDR implementation. The paper is well written and the methods and approach are robust and fit for purpose. The authors have identified the most appropriate datasets for the purpose of building a map which describes forest land use in the year 2020 – with some known limitations. The use of only datasets which are global in scope is an understandable approach, but the map could probably be improved if local/regional/national datasets were utilized, and this highlights the value of any global datasets which themselves incorporate local data bring to this mapping challenge.
We thank the referee for their appreciation of our work. As explained in the manuscript, the GFC2020 map is a globally consistent and harmonized representation of forest cover for year 2020. It is based on global wall-to-wall products or datasets that are harmonized globally or focus on specific ecological zones at a global scale. For instance, the Spatial Database on Planted Trees from the World Resource Institute (WRI SDPT, Harris et al., 2019, Richter et al., 2024), harmonizes local, national and regional tree crop and planted forest information at national level, and provides a globally consistent collection following FAO/EUDR definition of planted trees. Another example is the use of the JRC Tropical Moist Forest dataset, which offers comprehensive, global wall-to-wall mapping of tropical moist forest cover and its disturbances over recent decades. The approach adopted for GFC2020 ensures consistency across all regions of the world, rather than relying on a patchwork of local, regional, or national datasets that may not meet the EUDR forest definition. That said, we recognize the value of local, regional, or national datasets as complementary sources to be used in a convergence of evidence framework for EUDR risk assessments, provided that these datasets use a forest definition consistent with that outlined in the regulation. We addressed this issue of combined use of a national dataset and the GFC 2020 map in more detail in a JRC policy report for the Ivory Coast as a case study (Verhegghen et al. 2024). This study is already multiple times in the manuscript.
Verhegghen, A., Orlowski, K., Dontenville, A., Reboud, V., Riano, C., Njeugeut Mbiafeu, A. C., Kouamé Koffi, G.-B., Tillie, P., Rembold, F., and Achard, F.: Use of National versus Global Land Use Maps to Assess Deforestation Risk in the Context of the EU Regulation on Deforestation-Free Products – Case Study from Côte d’Ivoire. Publications Office of the European Union. https://doi.org/10.2760/7042220. 2024. a
We would like to inform the reviewer that, with the release of the FAO FRA 2025 on 21 October 2025, we took the opportunity to update the relevant tables, figures, and accompanying text (Table 4; Figures 7, 8, A2, and A3) to incorporate the latest comparisons of global, continental, and country-level forest area statistics with GFC2020.
- Line 23-24 - consider rephrasing to say "GFC does not include 578Mha of tree cover (11% of the global tree cover area) because it does not meet the height threshold or occurs on agricultural or urban land"
We agree with this suggestion and have revised the sentence accordingly.
- Would it be worth explaining the benefit of a conservative approach (higher commission vs omission error) in light of a more cautious approach for ensuring that products conform to EUDR? While acknowledging that this makes it more challenging for those who are demonstrating that their products are deforestation free.
Thank you for this suggestion, we have added a few sentences on this in section 4.3 dedicated to the potential use of GFC2020 for supporting the assessment of deforestation risk under the EUDR (starting at line 655). The revised paragraph reads as such:
“The GFC2020 dataset is intended to support operators and traders as one of several tools for deforestation risk assessment during the due diligence process. Specifically, it can help in the preliminary identification of plots where more detailed or locally relevant data should be gathered for a robust risk evaluation. Given this intended use, the overestimation of forest area in GFC2020 may result in additional assessments by operators. False positives, i.e. areas that are wrongly mapped as forest in GFC2020, will likely be identified as non-forest during subsequent assessments using more detailed or locally relevant data. In contrast, false negatives, i.e. forest areas that are omitted in GFC2020, could be a major concern for operators as deforestation risk areas may be missed. It has to be noted that the overestimation of forest by GFC2020 varies by regions and commodities, in particular with significant overestimation in regions with agroforestry systems like coffee and cocoa. Therefore, we strongly encourage operators to complement GFC2020 with national or regional forest cover datasets that align with the relevant definitions set out in Article 2 of the EUDR, particularly datasets that offer high spatial resolution and known accuracy. Alternatively, ground samples, geotagged photographs or non-spatial data could be used by operators to support or enrich the risk assessment (van Noordwijk et al. 2025). Several studies have demonstrated the value of multi-criteria or "convergence of evidence" approaches in this context (Verhegghen et al. 2024; D’Annunzio et al. 2024).”
- Line 40 - would a broader term indigenous peoples and local communities make sense here?
We modified the sentence as suggested.
- Line 140 - can we add 'continued collaborations' because they are already ongoing
Very true, we modified the sentence as suggested.
- Is it possible to provide some more clarity on the forest land use interpretation - for example forest/logging roads, fire breaks, skid trails and other forest related activities. Are they marked as forest in your interpretation and in the validation?
In line 265 we added: “According to the definition, forest also includes land that is unstocked or where trees for forest land use are temporarily below the 5m threshold, for instance following forest harvesting operations or fire”. I confirm that they are marked as forest in the interpretation.
- It seems clear to me that the majority of the commission errors are in areas where potential production for EUDR commodities will occur. I wonder whether this needs to be more prominently mentioned.
We already stress in section 4.1.2 the challenges and main limitation of the approach and in general the remote sensing capacity to exclude agricultural tree and rubber plantations from a forest extent. In section 4.2.1, we provide examples of future updates that will potentially lower commission errors linked to current misclassification of cocoa, coffee, oil palm or rubber. At global scale, commission errors are not restricted to EUDR commodities but also concern shifting agriculture and agroforestry systems or confusions of forest with shrublands or ‘other wooded land’. This is further detailed in section 3.2.2.
- Very high commission error in Africa – can be done about this in future?
The commission error of 24.6% ± 3.1% in Africa is primarily driven by confusion between forest and agricultural tree plantations or agroforestry systems, as discussed in Section 4.2.1. Additionally, there is an overestimation of forest cover in dry and open landscapes, where distinguishing between forests with very low tree cover, other wooded land, and savannahs remains a well-known challenge in remote sensing—this is also highlighted in Section 4.2.1 under research directions.
- Perhaps in the discussion you could mention the value of specific information on tree crops’ year of planting, to mitigate related limitations in the map.
The year of planting for tree crops is indeed a crucial piece of information to ensure that the land status aligns with the reference year 2020. However, this data is not consistently available in the SDPT database and is also absent from remotely sensed products, such as the tree crop probability layers developed by the Forest Data Partnership (Clinton et al., 2024). These products rely on direct land use mapping for a specific year (primarily 2020) rather than a retrospective assessment of planting dates. Furthermore, there is currently no indication that planting year information will be integrated into such spatial products in the near future.
- Line 389 - thematic ambiguity by the interpreter - not sure what this completely means – can you elaborate or rephrase?
We rephrase the sentence as such (line 405): “The confusion regarding other wooded land stems from uncertainties in the interpretation by experts and the potential inaccuracies in the mapping algorithm in deciding whether the physical criteria for forest classification are satisfied”.
- Perhaps worth a mention that EUDR also has a legality component which cannnot be addressed by this map?
Indeed, the legality aspect is already briefly mentioned line 53. We added at line 653, section 4.3: “The colocation of geolocation data (points or polygons) with forest areas identified in GFC2020, or any other map, does not automatically indicate non-compliance with the EUDR. Such plots must undergo further assessment to determine the actual risk of deforestation. In addition being deforestation-free, commodities and relevant products also need to meet the legal criteria of the EUDR.”
Citation: https://doi.org/10.5194/essd-2025-351-AC1
-
AC1: 'Reply on RC1', Clement Bourgoin, 05 Nov 2025
-
RC2: 'Comment on essd-2025-351', Anonymous Referee #2, 10 Sep 2025
General comments
This manuscript presents the GFC2020 map of forest land use for 2020 and provides a globally consistent 10 m resolution presence / absence product. The manuscript is very well written with methods and results thoroughly and clearly described. The description of the map and associated methodology is accompanied by a detailed and rigorous accuracy assessment. The accuracy assessment is based on a stratified sampling design and the authors report overall accuracy, user’s and producer’s accuracies, omission and commission error rates, and standard errors globally and by region. The effort and quality of the assessment are top tier.
I have a few minor technical suggestions or requests for clarification that I hope will aid the authors to further improve an already excellent manuscript. I will list the comments in order of appearance in the manuscript by line number rather than by importance.
Line26: This is very minor but the word “existing” is unnecessary because if the reference set does not exist, then it obviously cannot be used. Perhaps replace “exist” with “previously collected”?
L27: The phrasing “… with 18% probability of overestimating the forest area and 8% for underestimating” does not seem correct. First, it is unclear what the setting is for this probability. It would be plausible to frame this as selecting a unit (pixel) at random and asking, “What is the conditional probability that a randomly selected pixel that has reference class of forest is an omission error of forest?” I would claim that that conditional probability is 0.082 from the Table 2 error matrix (omission error rate of Forest). But the statement in the text is specifically about the probability of overestimating forest AREA. I don’t think there is a probability associated with overestimating forest area. The map shows 33.6% forest area and the reference data show 30.0% forest area, so the map overestimates forest area by 3.6% and there is no probability statement associated with that outcome.
L230-L235: Some additional details are needed to clarify the stratification, perhaps to allow the reader to know what was done without having to access Tsendbazar et al. (2020 and 2021). First, was the stratification applied to the 100m x 100m PSUs? Second, do the strata also include a feature related to the distribution of forest and non-forest? For example, at L232 it is stated that “Additional sample units were drawn for rare land cover classes.” Were these added from strata defined based on prevalence of these rare classes? A key feature of the stratification is that the strata are not simply forest and non-forest and this is what motivates use of Stehman (2014) for the estimation formulas rather than, say, Olofsson et al. (2014).
L292-293: This might not merit adding text to explain but I am confused why a sample unit cannot be associated with a strata class.
L294: replace “of” with “by”
L296-297: consider revising to “… and to account for unequal inclusion probabilities of sampled units.”
L306: It might be relevant to provide a formula for these weights? Also, if the pixels differ in area according to a latitudinal gradient, would that not also impact the sample-based estimates of accuracy and area because differences in area of the sampled pixels would need to be accounted for?
L361: Specifically, the “moderate overestimation” would be 3.6% according to Table 2.
L405: Figure 6a reports results in terms of number of sample units. Such an analysis ignores the sampling design and the necessary “weighting” of sample observations to produce estimates due to different sampling intensity in different strata. I would not advocate for re-analysis, but I do believe it is necessary to flag the analysis and alert readers that this particular set of results is not using the sample design information.
L455 uses “R-squared” and L442 uses “r^2”. It would be good to be consistent with notation.
L535: “potentials” should be “potential”
Citation: https://doi.org/10.5194/essd-2025-351-RC2 -
AC2: 'Reply on RC2', Clement Bourgoin, 05 Nov 2025
- This manuscript presents the GFC2020 map of forest land use for 2020 and provides a globally consistent 10 m resolution presence / absence product. The manuscript is very well written with methods and results thoroughly and clearly described. The description of the map and associated methodology is accompanied by a detailed and rigorous accuracy assessment. The accuracy assessment is based on a stratified sampling design and the authors report overall accuracy, user’s and producer’s accuracies, omission and commission error rates, and standard errors globally and by region. The effort and quality of the assessment are top tier.
We thank the referee for their appreciation of our work. We would like to inform the reviewer that, with the release of the FAO FRA 2025 on 21 October 2025, we took the opportunity to update the relevant tables, figures, and accompanying text (Table 4; Figures 7, 8, A2, and A3) to incorporate the latest comparisons of global, continental, and country-level forest area statistics with GFC2020.
- Line26: This is very minor but the word “existing” is unnecessary because if the reference set does not exist, then it obviously cannot be used. Perhaps replace “exist” with “previously collected”?
We agree with the suggestion and modified “exist” with “previously collected”.
- L27: The phrasing “… with 18% probability of overestimating the forest area and 8% for underestimating” does not seem correct. First, it is unclear what the setting is for this probability. It would be plausible to frame this as selecting a unit (pixel) at random and asking, “What is the conditional probability that a randomly selected pixel that has reference class of forest is an omission error of forest?” I would claim that that conditional probability is 0.082 from the Table 2 error matrix (omission error rate of Forest). But the statement in the text is specifically about the probability of overestimating forest AREA. I don’t think there is a probability associated with overestimating forest area. The map shows 33.6% forest area and the reference data show 30.0% forest area, so the map overestimates forest area by 3.6% and there is no probability statement associated with that outcome.
We modified the sentence as such (line 26): “Based on the reinterpretation of a previously collected reference set of 21,752 sample units, GFC2020 achieves an overall accuracy of 91%, with a commission error of 18% and an omission error of 8% for forest”.
- L230-L235: Some additional details are needed to clarify the stratification, perhaps to allow the reader to know what was done without having to access Tsendbazar et al. (2020 and 2021). First, was the stratification applied to the 100m x 100m PSUs? Second, do the strata also include a feature related to the distribution of forest and non-forest? For example, at L232 it is stated that “Additional sample units were drawn for rare land cover classes.” Were these added from strata defined based on prevalence of these rare classes? A key feature of the stratification is that the strata are not simply forest and non-forest and this is what motivates use of Stehman (2014) for the estimation formulas rather than, say, Olofsson et al. (2014).
Thank you for your comment. We believe the section 2.3.1 describing the sampling design already states that stratification was applied at the level of the primary sampling units (PSUs), which are 100×100 m in size and subdivided into 100 secondary sampling units (SSUs) of 10×10 m each. Regarding your other points, we have added additional information to clarify that no forest-related features were included in the stratification process, and the revised section now reads as follows (line 236):
“The statistical approach builds on a global set of 149 continental strata (Tsendbazar et al. 2018, 2021). Koeppen climate zones and human population density served as basic parameters for spatial sample unit distribution per continent. Tsendbazar et al. (2021) introduced additional strata to increase sampling intensity in rare land cover types, specifically wetlands, urban areas, water bodies, and shrublands, as identified in the discrete land cover map from the Copernicus Global Land Service (Buchhorn et al., 2020). Figure 2A shows the geographical distribution of all 21,752 sample unit locations. The sample units consist of Primary Sample Units (PSU) of 100x100m (blue frame in Figure 2B) divided into 100 10x10m Secondary Sampling Units (SSU, yellow mesh in Figure 2B). We selected the top-left SSU in the centre of each PSU (red cell in Figure 2) for validation of the 10m GFC2020 map.”
Added reference: Buchhorn, M., Lesiv, M., Tsendbazar, N.-E., Herold, M., Bertels, L., Smets, B.: Copernicus global Land Cover layers—collection 2. Remote Sens. 12, 1044. https://doi.org/10.3390/rs12061044. 2020. a
The rationale for using the formulas proposed by Stehman (2014) instead of those by Olofsson et al. (2014) is already explained in Section 2.3.2.
- L292-293: This might not merit adding text to explain but I am confused why a sample unit cannot be associated with a strata class.
The stratification used for the CGLS-LC100 dataset is based on the Köppen climate zones and human population density (Olofsson et al., 2012) which was later modified by overlaying the CGLS-LC100 land cover map to increase sample sizes for rare land cover classes such as wetland, urban (Tsendbazar et al 2021). Due to the extent difference between these two layers, 62 points previously selected using Koppen climate zones (located beyond 80 degrees latitude) fell outside the modified stratification. Hence these were removed from the analysis.
- L294: replace “of” with “by”
Done.
- L296-297: consider revising to “… and to account for unequal inclusion probabilities of sampled units.”
We modified the sentence as suggested.
- L306: It might be relevant to provide a formula for these weights? Also, if the pixels differ in area according to a latitudinal gradient, would that not also impact the sample-based estimates of accuracy and area because differences in area of the sampled pixels would need to be accounted for?
Forest area was computed by summing the surface area of all pixels mapped as forest in a WGS84-referenced geographic coordinate system; the geodesic area calculation inherently accounts for latitude-dependent pixel size. This computation was performed using the ee.Image.pixelArea() function from Earth Engine (Gorelick et al., 2017) which internally accounts for area distortions inherent of the WGS84 system.
Secondary sample units for accuracy assessment have a size of 10x10 m (100x100m for primary sample units) irrespective of their latitude.
We see that the term ‘weights’ may be misleading so we rephrased this sentence as such (line 316): “Forest areas derived from GFC2020 are calculated by summing the surface area of all pixels mapped as forest in a WGS84-referenced geographic coordinate system; the geodesic area calculation inherently accounts for latitude-dependent pixel size”.
- L361: Specifically, the “moderate overestimation” would be 3.6% according to Table 2.
As stated in this section (3.2.1) and in Table 2, the commission error for the forest class is 18%. In section 3.3.1 and in Table 4 we state that GFC2020 overestimates the forest area by 9.5% compared to FAO-FRA-2025 for year 2020.
- L405: Figure 6a reports results in terms of number of sample units. Such an analysis ignores the sampling design and the necessary “weighting” of sample observations to produce estimates due to different sampling intensity in different strata. I would not advocate for re-analysis, but I do believe it is necessary to flag the analysis and alert readers that this particular set of results is not using the sample design information.
The figure caption and the text already state: “Number of …” and thus notes that these are sample counts. The expression in % by category helps in the visualization and the introduced biases are taken care of by plotting the number of samples in each category. We added a specific note stating, unlike elsewhere in this study, data shown in Fig 6a show samples counts and not probabilistic estimates of accuracy. The revised sentence (at the start of section 3.2.2, line 400) reads as such:
“This assessment focuses on the labels assigned at the level of the SSU (second-level assessment in Table A2) to each sample unit. Figure 6A illustrates the number of correctly and incorrectly classified sample units in GFC2020 for each land use type; however, it does not present probabilistic estimates of accuracy or area for these categories.”
- L455 uses “R-squared” and L442 uses “r^2”. It would be good to be consistent with notation.
Thank you for noting this issue, we harmonized and only refer to “R-Squared” now.
- L535: “potentials” should be “potential”
Done.
Citation: https://doi.org/10.5194/essd-2025-351-AC2
-
AC2: 'Reply on RC2', Clement Bourgoin, 05 Nov 2025
-
RC3: 'Comment on essd-2025-351', Anonymous Referee #3, 16 Sep 2025
The study presented a harmonized and globally consistent map of forest cover at 10 meter spatial resolution. The proposed method is innovative and dataset is of high accuracy. However, there are still some problems that deserve to solve before publications.
(1) In the introduction part, the author should mention more details about data and methods used for forest mapping, as well as the deficiencies of the existing research.
(2) In the method part and figure 1, the authors emphasized more on the extraction and excluding rules for the forest mapping. I wonder how to conduct these complex rules.
(3) Multisource data may have different acquisition time and the tree cover may have some phenological features. Did the authors consider these?
(4) The consistency check and normal distribution test should be conducted for the sample construction.
(5) Some of the figures and tables are in poor quality and the authors should improve them. Some comparisons with existing products and deeper analysis of these difference are also necessary.
(6) There are still some grammatical and lingual problems, and authors should make a thorough revision.
Citation: https://doi.org/10.5194/essd-2025-351-RC3 -
AC3: 'Reply on RC3', Clement Bourgoin, 05 Nov 2025
- The study presented a harmonized and globally consistent map of forest cover at 10 meter spatial resolution. The proposed method is innovative and dataset is of high accuracy. However, there are still some problems that deserve to solve before publications.
We thank the referee for their appreciation of our work.
- (1) In the introduction part, the author should mention more details about data and methods used for forest mapping, as well as the deficiencies of the existing research.
The introduction already provides a detailed description on the most relevant and recent developments in remote sensing related to land cover and land use mapping, monitoring tree cover dynamics, mapping tree structure and characterizing of tree cover loss drivers. In lines 66 to 79, we reference these developments by citing key datasets that are subsequently used in our forest product workflow. Providing detailed and technical methodological descriptions of each global dataset including input data and classification methods is beyond the scope of this paper, which focuses on harmonizing existing products rather than developing new methods from raw satellite imagery. Furthermore, in lines 81 to 89, we highlight gaps in current research related to mapping forests as a land use. While an in-depth critique of each dataset falls outside the scope of this work, we recognize the need to better position our approach in relation to similar efforts in global forest land use mapping. To address this, we have added the following sentences before the last paragraph of the introduction (line 91):
“To date, only three global-scale forest maps exist that align with FAO definitions: (i) a hybrid forest map calibrated with FAO FRA data at 1 km resolution for the year 2000 (Schepaschenko et al., 2015), (ii) a forest management map at 100 m resolution for the year 2015, which categorizes forest use according to FAO classification (Lesiv et al., 2022), and (iii)a natural forest map for the year 2020 at 10 m resolution produced from Sentinel-2 imagery and deep learning methods (Neumann et al., 2025). The natural forest map excludes planted and plantation forests from its forest cover extent. Currently, there is no global map available at 10 m resolution for the year 2020 that encompasses all components of forest as defined by the FAO”.
Added references:
Schepaschenko, D., See, L., Lesiv, M., McCallum, I., Fritz, S., Salk, C., Moltchanova, E., Perger, C., Shchepashchenko, M., Shvidenko, A., Kovalevskyi, S., Gilitukha, D., Albrecht, F., Kraxner, F., Bun, A., Maksyutov, S., Sokolov, A., Dürauer, M., Obersteiner, M., Karminov, V., Ontikov, P.: Development of a global hybrid forest mask through the synergy of remote sensing, crowdsourcing and FAO statistics, Remote Sensing of Environment, Volume 162, Pages 208-220, ISSN 0034-4257, https://doi.org/10.1016/j.rse.2015.02.011. 2015. a
Neumann, M., Raichuk, A., Stanimirova, R., Sims, M., Carter, S., Goldman, E., Rey, M., Jiang, Y., Anderson, K., Poklukar, P., Tarrio, K., Lesiv, M., Fritz, S., Clinton, N., Stanton, C., Morris, D., Purves, D.:https://eartharxiv.org/repository/view/9085/ Natural forests of the world - a 2020 baseline for deforestation and degradation monitoring. https://doi.org/10.31223/X5ZX6P. 2025. a
- (2) In the method part and figure 1, the authors emphasized more on the extraction and excluding rules for the forest mapping. I wonder how to conduct these complex rules.
The expert-based boolean decision rules behind the two step methodology have been built following the forest definition as stated in the EUDR. Multiple rounds of adjustments have been carried out with the help of qualitative assessment from European experts. In the first step, the construction of the maximum extent of potential forest cover integrates with equal weights multiple tree cover, mangrove and planted forests. In the second step, the exclusion of covered pixels that do not qualify as forest require more complex rules that are often tuned by ecological zones or drivers of tree cover loss. These rules are also sensitive to thresholds that are adjusted at global scale or ecological zone depending on expert knowledge and reported accuracy metrics (Clinton et al., 2024).
- (3) Multisource data may have different acquisition time and the tree cover may have some phenological features. Did the authors consider these?
We carefully considered data acquisition time when selecting input datasets for GFC2020. Table A1 provides details on the acquisition periods or validity ranges for each dataset used. Our goal was to select products as close as possible to the year 2020, which serves as the cutoff date for the EUDR.
In addition to near-2020 data, we also used historical datasets and time series extending up to 2020 to analyze forest disturbance trajectories, land-use change dynamics, and historical tree cover. This approach helps mitigate potential underestimation of forests in unstocked areas, as well as overestimation in regions practicing shifting cultivation.
Regarding phenological characteristics, we used the ESA WorldCover 2020 and 2021 datasets in Step 1 of the GFC2020 workflow to define the maximum potential forest extent. These are annual products that capture tree cover for the respective years, using data collected within each year’s time window. Phenological features are incorporated during the preprocessing phase of the global land cover classification. These include descriptive statistics—specifically, the 10th, 50th, and 90th percentiles, as well as the interquartile range—calculated from the time series of Sentinel-1 and Sentinel-2 imagery for 2020 (and 2021). Further details can be found in the ESA WorldCover product user manual: https://worldcover2021.esa.int/data/docs/WorldCover_PUM_V2.0.pdf
- (4) The consistency check and normal distribution test should be conducted for the sample construction.
This validation uses an existing set of sample unit locations from the Copernicus Global Land Cover Product Validation for the year 2015 (Tsendbazar et al. 2018, Tsendbazar et al. 2021). The manuscript already provides details about the stratification and refers to the original literature. In this study we carried out a reinterpretation for all sample units, which is described in detail in the section “2.3.1 Sampling and response design”. For 12% of the samples we also carried out consistency checks between interpreters (described in the last paragraph of section 2.3.1). These results are presented in a separate JRC technical report (Colditz et al., 2025) already cited in the manuscript.
We added the following sentence in Section 3.2.1 to reference this previously published report where agreements between the two rounds of interpretation are presented (line 375):
“Information on the agreement between the first and second interpretations, showing 92.8% overall global agreement and a balanced pattern of under- and overestimation without significant regional differences, is available in Colditz et al. (2025).”
- (5) Some of the figures and tables are in poor quality and the authors should improve them. Some comparisons with existing products and deeper analysis of these difference are also necessary.
We believe the issue raised by the reviewer may be related to the quality of the PDF they accessed. All tables will be formatted according to the journal's guidelines during typesetting, and all figures are already provided in high resolution (>300 dpi).
Regarding global comparisons, the only datasets we can reliably compare with are the FAO FRA (Global Forest Resources Assessment) and its Remote Sensing Survey (FAO FRA RSS), as they are aligned in terms of forest definitions. These comparisons are already presented in Table 4, Figure 7, and Figure 8 of the manuscript. With the release of the FAO FRA 2025 on 21 October 2025, we took the opportunity to update the tables, figures, and accompanying text (Table 4; Figures 7, 8, A2, and A3) to reflect the latest comparisons of global, continental, and country-level forest area statistics with GFC2020.
Additional comparisons with national and regional datasets—such as those for Europe, Côte d’Ivoire, Brazil, and North America—have been carried out in a separate JRC policy report already cited in the manuscript (Bourgoin et al 2025a) and mentioned in last sentence appearing before figure 1 (lines 141 to 145 in the original manuscript)https://data.europa.eu/doi/10.2760/1975879. To the extent possible, national and regional datasets were adjusted to align with the forest definition used in GFC2020. However, significant limitations remain, primarily because most of these products depict land cover, reflecting the physical characteristics of the Earth's surface, while GFC2020 is designed to map forest areas based on a land use definition.
Finally, in Figure 10 of the manuscript we analyzed overlaps between GFC2020 and global datasets on tree crop commodities and drivers of forest loss to identify potential under- or overestimation in GFC2020.
We added the following sentence in Section 3.3.1 to reference a previously published report comparing GFC2020 with existing products at national and regional scales (line 454):
“Additional information on comparisons between GFC2020 and regional or national land cover products across Europe, Côte d’Ivoire, Brazil, and North America, which show spatial agreement in forest cover ranging from 66% to 87%, is provided in Bourgoin et al. (2025).”
- (6) There are still some grammatical and lingual problems, and authors should make a thorough revision.
The full author list (including native speakers) have revised the whole text and checked for grammatical and lingual issues. Remaining minor issues will be tackled by the proofread team of the journal before publication.
Citation: https://doi.org/10.5194/essd-2025-351-AC3
-
AC3: 'Reply on RC3', Clement Bourgoin, 05 Nov 2025
-
RC4: 'Comment on essd-2025-351', Anonymous Referee #4, 23 Sep 2025
Comment on "GFC2020: A Global Map of Forest Land Use for year 2020 to Support the EU Deforestation Regulation" by Bourgoin et al., ESSD
General comments:
This manuscript follows the definition of forest from EU Deforestation Regulation and constructed an updated forest cover dataset at 10-m resolution. It provides a good example of how to choose appropriate datasets, merge data products from multiple sources for dataset production and perform accuracy assessment. This paper is already well written and I can barely find any obvious errors. I have the following doubts and suggestions, mostly related to writing.
Specific comments:
Some of my major concerns are already conveyed through other reviewers so I here only have one major comment. Since authors emphasized the data processing, production and evaluation to be a general framework for producing other categorical maps in high resolution, I expect to see some further discussions about the potential usage of this framework in mapping other categorical variables. Maybe one example in 2-3 sentences will add more merits to the discussion section of this article.
Technical correction:
Line 16: EO refers to both remote sensing based and local experimental based approaches for measurement at different scales, which is quite a big concept. Here I believe the major datasets you used are RS based. I suggest changing it to RS based observations, or EO on a global scale.
Line 33: It will be good to emphasize in your abstract that your definition of forest follows European Union Deforestation Regulation (EUDR), from the very beginning of your article.
Line 118: Why do you need weight? Isn't the maximum extent to be the Union of all input products?
Line 120: Just want to confirm with the authors how you analyze historical tree cover loss?
Line 127: Should move references about expert judgement from Line 139 (see Bourgoin et al. 2024b; Colditz et al. 2024).) to here.
Line 140: "WRI, IIASA" Need full names.
Line 191: "we create a global cropland extent by combining" here combining means the union of the area extent from all data products or another approach?
Line 270: "labelled 'other wooded land' " Is the label information available together with the data on GEE? I did not find this information.
Line 302: "GAUL" shall add the full name "Global Administrative Unit Layers".
Figure 5: It might be good to tell readers the central coordinates of each subplot.
Table 3: Missing (CI) after overall accuracy.
Line 429: "generally align will with" shall be "generally align well with"
Citation: https://doi.org/10.5194/essd-2025-351-RC4 -
AC4: 'Reply on RC4', Clement Bourgoin, 05 Nov 2025
- This manuscript follows the definition of forest from EU Deforestation Regulation and constructed an updated forest cover dataset at 10-m resolution. It provides a good example of how to choose appropriate datasets, merge data products from multiple sources for dataset production and perform accuracy assessment. This paper is already well written and I can barely find any obvious errors. I have the following doubts and suggestions, mostly related to writing.
We thank the referee for their appreciation of our work. We would like to inform the reviewer that, with the release of the FAO FRA 2025 on 21 October 2025, we took the opportunity to update the relevant tables, figures, and accompanying text (Table 4; Figures 7, 8, A2, and A3) to incorporate the latest comparisons of global, continental, and country-level forest area statistics with GFC2020.
- Some of my major concerns are already conveyed through other reviewers so I here only have one major comment. Since authors emphasized the data processing, production and evaluation to be a general framework for producing other categorical maps in high resolution, I expect to see some further discussions about the potential usage of this framework in mapping other categorical variables. Maybe one example in 2-3 sentences will add more merits to the discussion section of this article.
We present a general framework for forest land use mapping that will remain consistent in future versions of the Global Forest Cover Map despite the fact that new inputs may be incorporated and/or decision rules may be refined or improved for specific categories of the framework. However, this framework is specific to forest land use mapping, as it captures the multifaceted nature of forest definition—including both land use and physical characteristics. Also, the design of such a framework depends on the availability of thematic input data (maps, stratifications, continuous fields layers, etc.) for the date or period of interest. Therefore we do not consider that a generalization of this approach is feasible for other categorical land cover or land use variables.
This being said, further discussion on the use of this framework for forest type mapping can already be found in Section 4.2.2 of the manuscript, “Forest types mapping to address EUDR definition of forest degradation.” There, we explain how the workflow is already being applied to map sub-categories of forest types—specifically primary forest, naturally regenerated forest, and planted or plantation forest—to support the forest degradation component of the EUDR. A preliminary map of Global Forest Types (JRC GFT2020 v0) was published in December 2024 and a revised version will be produced towards the end of 2025.
- Line 16: EO refers to both remote sensing based and local experimental based approaches for measurement at different scales, which is quite a big concept. Here I believe the major datasets you used are RS based. I suggest changing it to RS based observations, or EO on a global scale.
We agree with this suggestion and modified as such: “Remote sensing-based observations are used to…”.
- Line 33: It will be good to emphasize in your abstract that your definition of forest follows European Union Deforestation Regulation (EUDR), from the very beginning of your article.
Thank you for this suggestion, we modified the last sentence of the abstract as such (line 33): “Although this map follows the EUDR’s definition of forest, it is a non-exclusive, non-mandatory, and not-legally binding source.”
- Line 118: Why do you need weight? Isn't the maximum extent to be the Union of all input products?
You are right and we deleted ‘equal weight’ from the sentence.
- Line 120: Just want to confirm with the authors how you analyze historical tree cover loss?
This section addresses the challenge of mapping potentially unstocked forest areas resulting from temporary natural or anthropogenic disturbances—such as landslides, fires, or clear-cut harvesting—that do not result in a land use change. Depending on the intensity and timing of these disturbances, land cover products from around 2020 may fail to detect existing or recovering tree cover. To map such areas, it is essential to integrate historical data on tree cover and tree cover loss (e.g., UMD Global Tree Cover Loss from 2001 to 2020), particularly in regions flagged by complementary datasets—such as the IIASA Forest Management Map or driver-specific global forest loss maps (e.g., Curtis et al., 2018, Sims et al., 2025)—as being affected by natural or forestry-related disturbances.
- Line 127: Should move references about expert judgement from Line 139 (see Bourgoin et al. 2024b; Colditz et al. 2024).) to here.
We moved them as recommended.
- Line 140: "WRI, IIASA" Need full names.
We applied the change as suggested.
- Line 191: "we create a global cropland extent by combining" here combining means the union of the area extent from all data products or another approach?
Yes correct, we mean the union of the area extent of all data products mentioned in this sentence.
- Line 270: "labelled 'other wooded land' " Is the label information available together with the data on GEE? I did not find this information.
The validation dataset for the global map of forest cover 2020 that is being made available at this link (https://data.jrc.ec.europa.eu/dataset/8fbace34-a2fe-47b9-ad82-3e9226b7a9a6) only provides the first level interpretation of the 10x10m sample units, i.e. binary forest/non-forest classification. The second-level was used internally to understand the types of issues behind commission and omission errors. The GFC2020 map in Google Earth Engine maps only two classes: forest and non-forest.
- Line 302: "GAUL" shall add the full name "Global Administrative Unit Layers".
Thank you, we rephrased this sentence as such (310): “The approach extrapolates the proportion of sample units labelled “forest” over the total land area, here the area of the 149 strata inside the FAO Global Administrative Unit Layers (GAUL) country limits.”.
- Figure 5: It might be good to tell readers the central coordinates of each subplot.
Good idea, we added them in the caption of Figure 5. It now reads as such (line 359):
“Figure 5: GFC2020 mapping in the context of various natural and human-made landscapes: full-sun cocoa plantation (6.7°W, 5.6°N) (A), soybean and pasture structured landscape (47.3°W, 3.3°S) (B), full-sun coffee plantation (108.1°E, 11.5°N) (C), industrial oil palm plantation (98.6°E, 1.8°N) (D), rubber plantation (100.8°E, 21.9°N) (E), planted forest (0.8°W, 44.6°N)(F), urban trees (73.9°W, 45.2°E) (G), dry and open tropical forest (32.5°E, 4.2°S) (H), other wooded land (42.7°W, 10.7°S) (I), clear-cut harvesting on the left side and very young regrowth on the right side (173.4°E, 41.2°S) (J), agroforestry system and mixed urban (113.9°E, 8.3°S) (K), shaded coffee plantation (85.8°W, 13.3°E) (L). Background data: Google, © 2024 Maxar Technologies. Locations of each zoom is shown on Figure 4.”
- Table 3: Missing (CI) after overall accuracy.
Thank you, issue fixed.
- Line 429: "generally align will with" shall be "generally align well with"
Thank you for spotting this typo. Error fixed.
Citation: https://doi.org/10.5194/essd-2025-351-AC4
-
AC4: 'Reply on RC4', Clement Bourgoin, 05 Nov 2025
Data sets
Global Forest Cover 2020 Clement Bourgoin et al. https://forobs.jrc.ec.europa.eu/GFC
Validation dataset for the global map of forest cover 2020 – version 2 René Colditz et al. https://data.jrc.ec.europa.eu/dataset/8fbace34-a2fe-47b9-ad82-3e9226b7a9a6
Model code and software
Joint Research Centre – Global Forest Cover for year 2020, version 2. Code source Clément Bourgoin https://doi.org/10.6084/m9.figshare.29315528.v1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,779 | 116 | 39 | 1,934 | 39 | 43 |
- HTML: 1,779
- PDF: 116
- XML: 39
- Total: 1,934
- BibTeX: 39
- EndNote: 43
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
In line 81-90 authors correctly clarify the challenge that ‘tree cover’ as observable characteristic of land differs from ‘forest’ as a land use category in policy designs and instruments. Yet, in subsequent text it seems that ‘tree cover’ is equated with ‘forest’.
Line 103-109 claims that the GFC2020 maps ‘align with’ EUDR and FAO forest definitions. The text acknowledges the challenges in this claim especially where tree crops are involved that are the primary concern of EUDR regulations, but don’t follow up on these concerns.
Specifically, a recent publication (van Noordwijk et al. 2025) suggested multiple types of evidence for an ‘agroforestry’ (and thus non-forest) status of land in an institutional interpretation of the EUDR and FAO forest definitions, regardless of tree cover. It would be appropriate if authors comment on these evidence categories and the way they Cn (or should) be taken into account if the target is to create an EUDR-policy relevant map of 2020 global forest cover.
The procedure described in lines 200-203 can deal with part of the ‘agroforestry’ area, but certainly not all.
Where the paper employs standard concepts of ‘producer’ and ‘user’ accuracy, van Noordwijk et al. (2025) used (and pleaded for wider use of) a more specific ‘user accuracy’ in the context of EUDR. The most relevant use of the maps in EUDR context is to evaluate whether or not tree crops (incl coffee, cocoa, rubber) marketed after 2025 were derived from land deforested before or after 2020. A ‘real user’ accuracy would assess the likelihood that plots known to already produce any of these commodities before 2020 is correctly classified as ‘non-forest’. A number of studies, incl van Noordwijk et al. (2025), but also studies for Cameroon, Ivory coats and Peru that are on their wway to publication, have found erroneous classification of such points to be around 60%.
Before recommending the use of the current GFC2020 map for EUDR implementation, this issue may need to be addressed (or at least acknowledged).
van Noordwijk, M., Dewi, S., Minang, P.A., Harrison, R.D., Leimona, B., Ekadinata, A., Burgers, P., Slingerland, M., Sassen, M., Watson, C. and Sayer, J., 2025. Beyond imperfect maps: Evidence for EUDR‐compliant agroforestry. People and Nature 7:1713–1723. https://besjournals.onlinelibrary.wiley.com/doi/pdf/10.1002/pan3.70088