Assessing the Suitability of Available Global Forest Maps as Reference Tools for EUDR-Compliant Deforestation Monitoring

Abstract

Deforestation monitoring is critical to support compliance with regulatory frameworks such as the EU Deforestation Regulation (EUDR), which requires that products containing or derived from beef, cocoa, coffee, palm oil, rubber, soy, and timber are deforestation-free after 31 December 2020. Earth observation (EO) offers a means to assess deforestation, yet map-based verification remains technically limited and uncertain. This study addresses the lack of a systematic assessment of global Forest/Non-Forest (FNF), Tree Cover/Non-Tree Cover (TC/NTC) and Land Use/Land Cover (LULC) datasets by identifying and evaluating 21 publicly available global forest/tree cover reference maps for their alignment with EUDR criteria. This goes beyond merely treating these datasets as simply “fit” or “not fit” for the purpose of the EUDR, but rather aims to assess how well each dataset meets the needs compared to others, acknowledging strengths, weaknesses, and trade-offs. The 21 datasets are reviewed based on EUDR-related parameters (temporal proximity, spatial resolution, and forest definition) as well as accuracy metrics. From this broader review, eight datasets are shortlisted based on their alignment with key regulatory requirements. However, most datasets fail to fully meet all EUDR requirements, particularly forest definitions, with only two datasets satisfying all indicators. Notably, all datasets are unable to distinguish forests from other non-forest, tree-based systems. Reported accuracy metrics reveal a general overestimation of forest areas, while canopy height-based maps tend to underestimate tree cover, potentially excluding forested regions. Regional comparisons show more consistent estimates in South America, while Europe and North America display greater variability. These findings support informed decision-making by companies and policymakers for selecting suitable datasets, while also highlighting conflicts and challenges associated with the use of global forest/tree cover maps for regulatory compliance.

1. Introduction

Forest ecosystems provide a wide range of ecosystem services that are crucial for maintaining life on earth [1]. Nonetheless, over the last three decades, forest loss is estimated to have reached almost half a billion hectares worldwide with anthropogenic disturbances contributing significantly to the loss of biodiversity, reduction in forest area, and degradation of forest conditions. This severely undermines the ability of forests to provide essential ecosystem services to local populations and people who depend on them [2,3,4].

Deforestation and forest degradation have complex, context-dependent causes, but are often linked to the production and trade of agricultural and forestry commodities [5,6]. During the past years, voluntary guidelines such as zero deforestation commitments (ZDCs) were developed, aiming to foster environmental, social, and economic sustainability throughout the value chain [7,8]. Such governance mechanisms are meant to address related issues beyond national laws, extending responsibilities to all agents within the value chains [9]. Despite their proliferation, however, private governance mechanisms such as ZDCs have faced difficulties in addressing tropical deforestation and the overexploitation of natural resources, as originally anticipated [10,11,12]. Against this backdrop, the European Union (EU) developed a regulation for deforestation-free products (EUDR, Regulation (EU) 2023/1115) in which commodities or products in scope cannot be placed on EU markets unless they are deforestation-free after the cut-off date of 31 December 2020, and were produced in accordance with national legislation. This includes products made from, fed by, or containing derivatives of palm oil, soy, cocoa, coffee, rubber, cattle, and wood (hereafter EUDR-regulated commodities).

The European Union Deforestation Regulation (EUDR), adopted on 29 June 2023 and entering into force on 30 December 2025, forms part of the EU Forest Strategy embedded within the EU Green Deal, the EU Biodiversity 2030 Strategy, and the Farm to Fork Strategy. It seeks to address the EU’s contribution to global deforestation and forest degradation by restricting EU market access for commodities produced through unsustainable means. Under this framework, companies can only place, process or distribute these commodities on the EU market if they provide a compulsory due diligence statement (DDS) confirming that the product is “deforestation-free” and has been produced in accordance with national legislation. “Deforestation-free” means that EUDR-regulated commodities and their products are produced without causing deforestation or, in the case of timber extraction, forest degradation, after 31 December 2020. While the regulation also addresses forest degradation, this study focuses specifically on deforestation, defined by the regulation as the conversion of primary forest and regenerating forest to agricultural use.

In accordance with the EUDR, companies are required to fulfill specific information requirements, which must include, among others, (i) adequately conclusive and verifiable information that the relevant products are deforestation-free and (ii) must indicate the geocoordinates of the plot of land of production (1 point for plots < 4 ha and a polygon for plots > 4 ha). Building on this, the fundamental roles of the key actors in assessing the “deforestation-free” criteria are clearly defined: affected companies need to submit their DDS, and the national competent authorities (NCAs) of the EU member states are in charge of verifying whether the submitted statements comply with the “deforestation-free” production condition. For this case, data derived from earth observation (EO), e.g., satellite time series imagery can be used to prepare a DSS (in case of companies) or to verify (in case of NCAs) the provided deforestation-free statements. Such an approach, however, requires significant efforts, as it demands expert knowledge and in-house capacity, both of which are not always readily available. One less demanding way to verify deforestation-free production is by comparing geotagged production areas with forest reference information such as available from global forest/tree cover maps derived from EO analysis [13]. A multitude of different map products are publicly available, such as [14], which are used for deforestation monitoring, available with the open-data repository Global Forest Watch (https://www.globalforestwatch.org/map/, accessed on 30 November 2024) [15] and Google Earth Engine data catalog (https://developers.google.com/earth-engine/datasets/catalog/UMD_hansen_global_forest_change_2023_v1_11, accessed on 30 November 2024) [16].

In the era of big data, the production of global forest and tree cover maps relies on analyzing and processing large volumes of information, such as remotely sensed satellite imagery and other geospatial datasets. This requires advanced computational methods and machine learning approaches, so that meaningful patterns and trends can be extracted from these data [17,18,19]. Therefore, the verification and proving of deforestation-free production using global map products are data-driven decisions, which can be subject to technical limitations. For example, one of the main technical challenges when using forest reference maps to verify the deforestation-free requirements of the EUDR is the potential false detection of deforestation originating from the mismatches when intersecting commodity production areas with forest reference maps. These potential discrepancies labeled as “false positives” can result from several factors, including the poor quality of the source reference dataset, the reliability of the map in distinguishing forest from non-forest tree-covered areas, mismatches between map features and EUDR requirements, and others.

In general, global tree cover and/or forest datasets are known for having certain limitations [20,21], as they aim to provide a harmonized view of the world. This often leads to compromises in map accuracy, resulting in an inaccurate representation of forest status in certain regions of the world. This is particularly the case for tropical forests (both humid and dry) due to high interference (cloud cover, cloud shadowing, terrain artifacts and atmospheric noises), seasonal limitations and inter-annual climate variability (varying spectral signature of vegetation cover), connectivity of forests (fragmented forest patches), presence of agricultural plantations and agroforestry, forests with low canopy density (dry deciduous forests), and discrepancies between the specific features of particular forest ecosystems and those of globally harmonized ones, among others [19,22,23,24,25,26,27,28,29,30,31].

One frequently debated global tree cover and/or forest dataset is Global Forest Cover of the Year 2020 [32] published by the Joint Research Centre (JRC). This map was generated with the intention of supporting the EUDR regulation; however, the use of the map is not mandatory and non-exclusive, and it does not serve as a binding regulatory decision mechanism. This suggests that other global forest/tree cover maps may be suitable to support the verification of deforestation-free production. For this reason, we review publicly available global maps classified as Forest/Non-Forest (FNF, land use), Tree Cover/Non-Tree Cover (TC/NTC, land cover), and Land Use/Land Cover (LULC, combining both), assessing their suitability for meeting EUDR mapping requirements. Against this background, we firstly conduct a review of major geospatial data portals, relevant reports, and the literature to produce an extended list of potential global reference maps based on remote sensing sensors. Secondly, we analyze the maps by applying criteria that follow EUDR guidelines, and maps’ reliability in accurately mapping “forest cover”.

While previous assessments on global FNF, TC/NTC and LULC datasets have been conducted [29,33,34,35], they have not followed a systematic approach that aligns with the legal framework of the EUDR. Most published information on this subject is scientific news pinpointing particular and localized cases to illustrate a certain viewpoint. One study [36] closely addresses the suitability of global versus local datasets as reference maps within the EUDR framework, focusing exclusively on the case of Côte d’Ivoire.

The overall objective of this study is to identify, collect, describe and evaluate publicly available global FNF, TC/NTC and LULC reference maps on their relative capability to match the EUDR requirements using two groups of relevant indicators, namely EUDR parameters (temporal proximity, spatial detail, forest cover definition standards) and technical parameters (reported accuracy metrics). We examine the identified FNF, TC/NTC and LULC dataset based on the EUDR parameters criteria because these criteria (i) constitute the regulatory framework that must be followed by all, (ii) convey a clear and objective interpretation message, (iii) are independent from secondary factors such as the quality of input/observation data or the effectiveness of the chosen methodological approach, both of relevance for the generation of the forest reference maps. Additionally, we classify some of these datasets according to their bias toward overestimation or underestimation of forest cover, as determined by the interpretation of the reported accuracy results. Our purpose goes beyond merely treating these datasets as simply “fit” or “not fit” for the purpose of the EUDR, but rather assesses how well each dataset meets the needs compared to others, acknowledging strengths, weaknesses, and trade-offs. Finally, we present a condensed shortlist of potential reference maps, conduct a regional bias assessment of observable variations in mapped forest areas, and describe the pros and cons of the individual maps by answering the following research questions:

(1)

Are the existing FNF, TC/NTC and LULC datasets equally suitable as verification tools for EUDR compliance, given the specific requirements outlined in the EUDR regulation?

(2)

What specific traits hinder alignment of these datasets with EUDR requirements?

(3)

Do these datasets show a tendency to over-/underestimate forest/tree cover?

(4)

Are forest/tree cover estimates consistent across world regions, or do notable regional discrepancies exist?

2. Materials and Methods

2.1. Dataset Compilation and Dataset Information Extraction (Step 1)

We conduct a comprehensive search for publicly accessible global maps, which contain information on forest or tree cover derived from EO and remote sensing sensors, across major geospatial data portals, including Google Earth Engine, USGS Earth Explorer, NASA EarthData, ESA Copernicus Open Access Hub, Global Forest Watch, and FAO GeoNetwork, US Open Data Portal, UNEP Environmental Data Explorer, ArcGIS Living Atlas, UNdata, and World bank Open Data. To supplement the data search, we inspect and cross-reference the datasets included in the Forest Resource Assessment report from 2020 [37], the digital public infrastructure for deforestation-related trade regulations [34], and perform a literature review using online search engines (Google Scholar, Scopus and Web of Science) to identify relevant scientific publications on available global FNF, TC/NTC and LULC datasets [38,39,40,41]. We exclude datasets with a spatial resolution greater than 300 m, as well as those of regional or local extent, from the database compilation because the former lack sufficient spatial detail on tree and forest cover, while the latter do not represent global coverage. All datasets included in this analysis were identified and compiled as of November 2024.

After having finalized a comprehensive list of global map datasets, we collect specific information for each dataset by reviewing the respective accompanying literature references, tutorials, websites, reports or any other published documents with relevance to a respective dataset. The retrieved information is used to understand the map features, conditions for its production, validation and adopted definition and parameters. This information includes (a) metadata (reference citations and published documents, (b) links to repositories, short descriptions of datasets, and organizations associated with the datasets), (c) technical specifications (geographical coverage, spatial resolution, minimum mapping unit (MMU) of the map, temporal resolution, covered period, spectral range, indices/metrics, period of EOs), (d) land cover information (number of classes and short description of them), (e) map classification procedure (adopted methodology, description and number of training samples), (f) map validation procedure (validation method, number of validation samples, adopted validation dataset and its reference period, chosen accuracy metrics, reported accuracy values for overall accuracy, producer’s and user’s accuracy, values for measuring errors), (g) reported map limitations, and (h) forest definition characteristics (general description and thresholds for height, MMU and canopy cover of forest/tree cover class). Additionally, we evaluate whether the “forest/tree cover” class encompasses non-forest tree-based systems, as defined under the scope of the EUDR. This issue is particularly important given that the EUDR is concerned with land use classes, whereas many maps primarily depict land cover. When critical information is missing (i.e., parameters for defining forest/tree cover), we contact the referred map representative or corresponding author, in case of a peer-reviewed publication, for clarification.

2.2. Indicator Nomination and Evaluation (Step 2)

To assess the suitability of the collected datasets as EUDR reference maps, we extract information relevant to EUDR requirements and the technical specifications of forest mapping using EO systems. Therefore, we separate the indicators for assessing suitability into two categories: “EUDR Parameters” and “Technical Parameters”. The EUDR parameters comprise three criteria, namely (i) the timeframe posed by the regulation (temporal proximity), (ii) the level of detail in the information provided by the map according to regulatory definitions (spatial detail), and (iii) the map parameters for defining forest. This builds on the rationale that each indicator can serve as criteria for assessing the reliability of a due diligence statement in the case of the EUDR. Two of the three EUDR parameters are applied in the first stage of the selection process, as they directly determine whether a classified map can be considered suitable or unsuitable according to EUDR terminology. Further assessment of map suitability is provided by the technical parameters, which display the tendencies of the classified forest/tree cover class based on the reported accuracy values. We describe each particular indicator in the following sub-sections.

2.2.1. EUDR Parameters

Temporal Proximity

According to Art. 3, EUDR, (a) EUDR-regulated commodities and products need to be “deforestation-free”, which is defined in Art. 2 (13) as not having caused deforestation or forest degradation after 31 December 2020. This so-called “cut-off date” is therefore the desired point in time to which a reference map should ideally refer. While achieving such a perfect image of the forest status at one point in time is not feasible, reference maps should display the forest status as close as possible to the cut-off date. On the contrary, reference maps that would refer to a period too far away from the cut-off date, e.g., ten years, exhibit the risk that the information provided is essentially outdated and therefore does not accurately reflect the 2020 forest cover status. Moreover, remote sensing-based maps may not accurately detect young forest patches. For example, a period of 4.2 to 7.4 years is required before saplings can be detected as forest in EO-based maps [42]. This is in line with the Food and Agriculture Organization of the United Nations (FAO) [37], which describes that “forest areas can be temporarily unstocked due to clear-cutting as part of a forest management practice or natural disasters, and which are expected to regenerate within 5 years”.

For these reasons, we determine a threshold of ±5 years from 2020 to evaluate the ability of a global map to meet the temporal proximity indicator. In the context of EUDR, the ±5 years from 2020 enables us to identify possible land use transitions over time. Equally important, a ±5-year period may limit the inclusion of areas under shifting cultivation, which have not yet been converted to agricultural production, in a forest class.

EUDR-compliant Forest Definition

A second indicator evaluates how closely the forest definition used in each reference map matches the one required by the EUDR, which follows the FAO [37] classification: “land spanning more than 0.5 hectares with trees higher than 5 metres and canopy cover of more than 10%, or trees able to reach those thresholds in situ, excluding land that is predominantly under agricultural or urban land use”. Considering these parameters, we collect details on the adopted forest definition for each of the identified global maps: tree height, MMU of forest/tree cover class and canopy cover density, and general description of forest or tree cover classes. In cases where information on these parameters is not provided, we contact the map developers directly to request the missing details. Unless otherwise specified, we assume that the MMU refers to the pixel size, that is, the squared spatial resolution. For example, a map with a spatial resolution of 30 m corresponds to an MMU of 0.09 hectares. The information on MMU is critical for assessing a map’s conformity with the EUDR forest definition, as the thresholds used to classify tree-covered areas as “forest” or other tree-based vegetation classes can introduce mismatches and biases when verifying deforestation-free production [43,44,45,46,47,48,49]. Additionally, accurately distinguishing land use from land cover classes is treated separately for each evaluated dataset and is particularly important for differentiating between compliant and non-compliant cases.

Spatial Resolution

The third indicator refers to the level of detail that a certain global reference map is able to convey. Such information is commonly referred to by the spatial resolution in remote sensing-based maps as it describes the real-world area covered by a single pixel’s image information. As the EUDR adopts the FAO’s definition of “forest” as land spanning more than 0.5 hectares (see Art. 2(4)), this threshold defines the minimum area that must be distinctly mapped and monitored. Consequently, 0.5 hectares represents the largest acceptable pixel area to ensure that forest areas can be classified as such within a single pixel. When a certain reference layer displays information in a single pixel equivalent to an area larger than the forest class MMU requirement, verification of deforestation violation becomes complicated. For example, in reference maps with a spatial resolution coarser than approximately 70 m, a forest patch or agricultural land of 0.5 hectares would be aggregated onto neighboring landscape in order to form a single pixel value, remaining undetected in the data [50]. Additionally, for a better representation of vegetation changes and forest disturbances, through optical satellite sensor data such as that available from Landsat, a resolution of 30 m or finer is preferred [51], particularly in heterogenous landscapes [51]. Hence, we set a maximum spatial resolution threshold of 30 m.

Table 1. Thresholds for each indicator in the EUDR Metrics.

Indicator		Threshold
Temporal Proximity		2020 ± 5 years
Spatial Resolution		≤30 m
Forest Definition	Tree Height	≥5 m ± 1 m
	MMU of Forest Cover	≥0.5 ha
	Forest Canopy Cover	≥10%

summarizes the thresholds (range or limit) applied in each indicator within the EUDR Metrics group to distinguish potentially suitable from unsuitable maps.

2.2.2. Technical Parameters

Complementary to the EUDR parameters, we review technical descriptions and statistics relevant to the assessment of map uncertainties and their potential implications for detecting true deforestation-free production through global forest/tree cover datasets. Among other characteristics, the reported map accuracy statistics, such as overall accuracy, producer’s and user’s accuracy, and mean bias errors, correspond to technical specifications of the datasets which allow us to infer map quality. Therefore, we use the reported map accuracy statistics as an additional assessment indicator (but not for the purpose of ranking or as exclusion criteria), since they provide a measurable metric that can be directly linked to implications for EUDR compliance. We do not employ technical parameters as exclusion or selection criteria, as reported accuracy values are derived under varying validation conditions and do not necessarily reflect performance consistently across all regions or EUDR-relevant commodities. Moreover, the EUDR regulation does not specify a minimum accuracy requirement for remotely sensed maps. Consequently, datasets with incomplete or partially reported accuracy metrics are still considered. Instead, we incorporate accuracy qualitatively, highlighting dataset tendencies and potential limitations to indicate likely types of classification error. This approach allows us to guide the use of global datasets in the context of EUDR compliance without introducing bias, while maintaining a standardized evaluation framework and avoiding in-depth, inconsistent discussion of individual maps. Moreover, accuracy assessment and its metrics are standard procedures in map classification and are commonly provided by map developers [52,53]. The other elements of the technical description are used as supporting information for a qualitative assessment, when applicable.

Map Accuracy Metrics

Accuracy assessments are essential for evaluating the precision of maps developed through EO tools and remote sensing. Commonly, a confusion matrix [53] is used to compute overall accuracy (OA) and class-specific metrics such as producer’s accuracy (PA) and user’s accuracy (UA), which reflect the truthfulness of mapped vs. actual landscapes. For our assessment, we focus on PA and UA of the forest/tree cover class, as OA does not directly capture its classification accuracy. These two metrics assess the truthfulness of classified samples from different perspectives: the user’s (actual landscape) and the producer’s (mapped landscape). Accordingly, we also compute the PA/UA ratio to identify potential systematic biases and better represent the map’s reliability for EUDR purposes.

Binary and multiple land cover or land use classes maps are typically assessed using OA, PA, and UA. In contrast, forest height maps are commonly evaluated using metrics such as root mean square error (RMSE), mean absolute error (MAE), and mean bias error (MBE), derived from comparisons with reference canopy height datasets such as the relative height at the 95th percentile (RH95) from the Global Ecosystem Dynamics Investigation (GEDI) Level-2A products [54]. RMSE and MAE quantify the magnitude of differences between predicted and observed values, while MBE quantifies the errors but also indicates the direction of systematic errors, highlighting tendencies in the map.

Table 2. Details on the accuracy metrics and their relationship to potential implications on EUDR DDS verification through global forest/tree cover maps.

	Producer’s Accuracy	User’s Accuracy
Calculation (class-specific)	$PA={\displaystyle \frac{Correctly classified samples}{Reference samples}}\times 100$	$UA={\displaystyle \frac{Correctly classified samples}{Classified samples}}\times 100$
Probability	True forest pixel being classified as forest cover	Classified forest pixel actually representing forest cover
Measure	Completeness	Reliability
Associated Error	Omission error: number (%) of true forest pixels incorrectly identified (Omission error = 100 − PA)	Commission error: number (%) of incorrectly classified pixels as forest (Commission error = 100 − UA)
Interpretation	Low producer’s accuracy values may indicate forest underestimation	Low user’s accuracy values may indicate forest overestimation
Ratio (PA/UA)	Ratio > 1 (PA > UA; Oerror < Cerror)	Ratio < 1 (PA < UA; Oerror > Cerror)
	RMSE	MAE
Definition	Squared differences between predicted/modeled values and the true values	Absolute differences between predicted/modeled values and the true values
Measure	Scale and magnitude of errors, where higher weights are given to large differences (spotting outliers)	Scale and magnitude of errors, where all errors are equally treated
	MBE
Definition	Average of differences between predicted/modeled values and the true values
Measure	Direction and tendency of errors
Interpretation	Positive indicates overestimation; negative indicates underestimation

summarizes the accuracy metrics considered for the technical assessment of global map products.

2.3. Exclusion of Unsuitable Datasets (Filtered Datasets I)—Step 2

The first selection criterion aims to identify those global reference datasets that comply with EUDR terminology based on two of the aforementioned metrics (see Figure 1). We choose two out of the three described EUDR Metrics due to the incompleteness of the reported forest definition parameter in some of the reviewed maps. With this approach, we seek to ensure an unbiased exclusion procedure. Subsequently, we filter out datasets that do not meet the thresholds for temporal proximity or spatial resolution as described in Section 2.2.1. These two indicators are the most commonly retrieved information across all datasets, while information on “forest definition” is not always consistently reported.

2.4. Removing Redundancies (Filtered Datasets II = Shortlisted Datasets)—Step 3

Based on the first selection outcomes, we apply a second set of criteria by reviewing additional map characteristics. From the initial list of filtered datasets (filtered datasets I), we exclude those already incorporated into more recent datasets to avoid redundancies. Additional exclusion criteria include datasets that condense multi-year information into a single layer; those for which the newest version is inaccessible, thereby violating the “temporal proximity” indicator; and those that provide only forest gain/loss information rather than a forest status layer. This step results in a concise list of datasets meeting our inclusion criteria, hereafter referred to as “shortlisted datasets”.

2.5. Forest Extent Area Comparison (Regional Map Variability)—Step 3

As a next step, we compare the shortlisted datasets with additional forest information concerning the size of the mapped forest area and conduct assessments on a per-continent basis. This assessment involves comparing shortlisted maps with one another, as well as against an additional and independent dataset. For this purpose, we rely on continental forest cover information reported for 2020 by the forest resource assessment report (FRA) of the FAO, referring to a sample-based approach of forest estimates reported at the country level by national authorities [38]. Hence, we compare forest area estimates (in thousands of hectares) from the shortlisted datasets by counting the number of pixels for forest cover and scaling that to the respective area size.

3. Results

3.1. Complied Datasets

We identify total of 21 global datasets that either refer to FNF (N = 6), TC/NTC (N = 5) or LULC (N = 10) maps, spanning from 1992 to 2024, with spatial resolutions of 300, 100, 30, 25, 10, and 1 m (

Table 3. List of collected datasets categorized by the EUDR metric indicators and data compilation methods. Shortlisted datasets, selected using criteria from steps 2 and 3 (Methods), are shown in bold and underlined letters.

ID	Dataset Short Name/ Abbreviation	Source	Institution	Type of Map	Spatial Resolution	Temporal Coverage	Forest/Tree Cover Class Parameters			Inclusion of Non-Forest Land Use? (Crop Plantation, Agroforestry, Timber Plantation, Woodlands/Savannas)	Classification Approach	Type of Satellite Sensor/Input Dataset	Exclusion Rationale
							Height (m)	MMU (ha)	Canopy (%)
1	ESA-CCI (v207) and ESA-CCI C3S (v2.1.1)	[55,56]	ESA	LULC	300	1992 to 2020	≥5	≥9	≥15	Not explicitly reported	Unsupervised classification and multiple-year strategy	MSS	SP
2	JRC GFC v2	[32]	JRC	FNF	10	2020	≥5	≥0.5	≥10	Yes	Composite map (CM)	Multiple third-party datasets
3	GFM 100	[57]	IIASA	FNF	100	2015	n.a.	n.a.	≥10	No	Random forest (RF) classifier	MSS	SP
4	Globcover	[58]	ESA	LULC	300	2009	≥5	n.a.	≥15	Not explicitly reported	Supervised and unsupervised classifications, and cluster-based classification	MSS	T + SP
5	ESA WC (v100 and v200)	[59,60]	ESA	LULC	10	2020 to 2021	≥5	n.a.	≥10	Yes	Gradient boosting decision tree algorithm (CatBoost)	MSS and SAR	R
6	CGLS-LC100 v3	[61]	Copernicus EU	LULC	100	2015 to 2019	n.a.	n.a.	≥15	Yes	RF classification and Biome-cluster classification	MSS and SAR	SP
7	ESRI-10 m	[62]	IO in cooperation with ESRI and Microsoft AI for Earth	LULC	10	2017 to 2023	≥4.57	n.a.	n.a.	Yes	Deep learning model (convolutional neural network for image segmentation)	MSS
8	Dynamic World	[63]	WRI	LULC	10	2015 to 2024	Forest defined as significant clustering of dense vegetation with a closed or dense canopy that is taller and darker than surrounding vegetation (if surrounded by other vegetation).			Yes	Semi-supervised deep learning (supervised label data to train fully convolutional neural network	MSS
9	GLC-FCS30D	[64]	CBAS	LULC	30	2000 to 2022	n.a.	n.a.	≥15	Yes	Continuous change detection (CCD) algorithm with a local adaptive updating method	MSS with Panchromatic band (PB)
10	GlobeLand30-ATS2010	[65]	NGCC	LULC	30	2010	n.a.	≥5.76	n.a.	Not explicitly reported	Pixel–object–knowledge (POK)	MSS with PB	T
11	FROM-GLC10	[66]	Department of Earth System Science, Tsinghua University	LULC	10	2017	n.a.	n.a.	n.a.	Not explicitly reported	RF classification	MSS
12	Forest Height—GLAD ARD	[67]	UMD GLAD	FNF	30	2000 and 2020	≥3 *	≥0.5	n.a.	Yes	Regression tree model (RTM) ensembles	LiDAR	R
13	Forest Extent—GLAD ARD	[67]	UMD GLAD	FNF	30	2000 and 2020	≥5	≥0.5	n.a.	Yes	Dataset based on forest height GLAD	Forest height GLAD ARD
14	GFW/Hansen Map (Global Forest Change Data v1-11)	[14]	UMD GLAD and GFW	TC/NTC	30	2000 to 2023	≥5	≥0.09	≥10	Yes	Bagged decision tree (bootstrap aggregating)	MSS with PB	R
15	GFW UMD Tree cover	[14]	UMD GLAD and GFW	TC/NTC	30	2010	≥5	≥0.09	≥10	Yes	RTM	MSS with PB	T + R
16	GEDI-Height	[68]	UMD GLAD	FNF	30	2019	≥0 *	≥0.09	n.a.	Yes	RTM	LiDAR	R
17	PALSAR-2 FNF v2.0.0 (3-class)	[69]	JAXA	FNF	25	2017 to 2020	≥5	≥0.5	≥10	Yes	RF classification	SAR
18	GLCLU—GLAD	[70]	UMD GLAD	LULC	30	2019	≥3	≥0.09	≥0	Yes	Global/regional hybrid decision tree (RTM for global and regional classification and local calibration); regional quality assurance models of water and snow/ice; regional deep learning convolution neural networks.	MSS with PB	R
19	CHM-1m	[71]	Meta Sustainability and WRI	TC/NTC	1	2009 to 2020	≥1 *	≥0.0001	n.a.	Yes	Self-supervised learning and simple convolution network	MSS with PB	OC
20	ETH	[72]	EcoVision Lab, Photogrammetry and Remote Sensing, ETH Zürich	TC/NTC	10	2020	≥0 *	≥0.01	n.a.	Yes	Convolutional neural networks (CNNs)	MSS and LiDAR
21	GFCC30TCC-v4	[73]	NASA	TC/NTC	30	2000, 2005, 2010 and 2015	≥5	≥0.09	≥0 *	Yes	Rescaling coarse dataset with finer dataset	MSS with PB	OC

R = redundant, as it is utilized in JRC/GFC 2020; SP = low spatial resolution; T = before 2015; OC = other criteria, such as datasets for which a newer version is unavailable or that lack a forest status layer; n.a. = not available or a measure that does not apply; * tree height or canopy cover dataset is adjustable to user’s preferred value.

). The implemented methods for the creation of these maps include a range of machine learning algorithms (random forest, unsupervised classification, regression tree, continuous change detection, pixel–object-based knowledge, and different decision tree model algorithms), deep learning models (convolutional neural network), and composite mapping techniques. Optical multispectral sensors (MSSs, e.g., Sentinel-2) are the most frequently applied type of data sources for global forest/tree cover and land cover maps, followed by Light detection and ranging (LiDAR, e.g., GEDI) and synthetic aperture radar (SAR, e.g., Sentinel-1) instruments.

3.2. First Selection Criteria (Step 2)

Based on the temporal proximity to the “cut-off date” in 2020 and adequate spatial resolution (both EUDR Metrics), six global tree cover and/or forest datasets are deemed less suitable and therefore excluded from use as reference maps for assessing deforestation-free compliance according to the EUDR in our study. These maps represent information on forest cover more than 5 years prior to the required cut-off date, and also have a spatial resolution coarser than 30 m. From the remaining 15 datasets, only 2 (JRC GFC v2 and PALSAR-2-FNF) fully agree with all the adopted parameters for forest definition according to EUDR, while the other datasets either agree on at least one forest definition parameter or do not fully report how forest cover is defined in their classification scheme (Figure 2). Among the forest definition parameters, tree height most commonly aligns with the standards adopted by the EUDR, as many datasets define a minimum tree height of 5 m for forested areas. Furthermore, all but five datasets explicitly report the inclusion of non-forest tree cover classes, such as agroforestry, crop plantations, woodlands, or savannas, within their forest or tree cover classifications, which is inconsistent with the EUDR’s definition of forest (see also Table 3).

Nearly all datasets apply a randomly stratified validation approach based on selected thematic class groups. Some maps (Dynamic World, GFW/Hansen Map) make further distinctions of strata per biome. Others use additional landscape information such as Köppen climate groups (GLC-FCS30D), geographic sub-regions (CHM-1m), or land ratio proportions (Forest Extent GLAD) for stratification. For most datasets, the reference validation sample collection is independently obtained through visual interpretation of very high-resolution satellite images. These images have a resolution finer than that of the corresponding map and are labeled by experts or trained interpreters. In fewer cases, reference samples are obtained from detailed third-party datasets, especially for the canopy height datasets (GEDI-height, Forest Height-GLAD, CHM-1m, ETH, GFCC30TCC-v4). The temporal coverage of the validation samples aligns with that of the created map. However, particular datasets include data from different years (Globcover, ESA WC, CGLS-LC100-v3, Forest Height-GLAD) or, in some cases, contain an older dataset of high quality (FROM-GLC10, Dynamic World, JRC GFC v2). In some instances, the sampling points not only consist of information of the corresponding class (thematic-label) but also include an attribute representing the robustness of the expert’s judgment of “high confidence/certain” or “low confidence/uncertain” categories and the homogeneous level of a certain point or validation unit.

Regarding the accuracy metrics, we evaluated the filtered 15 datasets (step 2) against their reported accuracy ranges, although not all datasets stated their values. The reported overall accuracies for forest status products range from 73.8% (Dynamic World) to 97.2% (Forest Extent-GLAD), which is typical for global datasets, where the minimum required accuracy is usually 70% (

Table 4. Accuracy metrics for the 15 datasets selected based on the first criterion (Step 2). Datasets that also met the second criterion (Step 3) are bolded and underlined (shortlisted datasets, N = 8). Reported accuracy ranges represent the highest documented values for each dataset.

ID	Name	General Validation Approach + Reference Dataset/Year	OA (%)	PA (%)	UA (%)	PA/UA	RMSE	MAE	MBE
2	JRC GFC v2	IIASA reference dataset. Year(s): 2015	91.5	91.8	82	1.12	x	x	x
5	ESA WC 2020 (v100 and v200)	CGLS-LC validation dataset (Copernicus Global Land Service). Year(s): 2019 to 2021	76.7	91.9	80	1.15	x	x	x
7	ESRI-10m (a)	Very high-resolution imagery visually interpreted. Year(s): n.a.	85	n.a.	n.a.	n.a.	x	x	x
8	Dynamic World	Samples per biome and region from NASA MCD12Q1 land cover. Year(s): 2017	73.8	93.2	70.2	1.33	x	x	x
9	GLC-FCS30D	Visually interpreted global samples and two third-party datasets: Land Use/-Cover Area frame Survey (LUCAS) and the Land Cover Monitoring, Assessment, and Projection (LCMAP) Collection 1.0 annual land-cover product. Year(s): 2020	80.88	92.83	86.35	1.08	x	x	x
11	FROM-GLC10	Multi-seasonal sampling collected from Landsat 8 images. Year(s): 2015	72.76	84.2	83.47	1.01	x	x	x
12	Forest Height-GLAD	Sampling from GEDI Collection 1 and Collection 2. Year(s): 2019 and 2020	x	x	x	x	6.75 m	4.76 m	x
13	Forest Extent-GLAD	Sampling from Landsat GLAD ARD 16-day time series data, annual and bimonthly image composites, and high-resolution image time series from Google Earth. Year(s): n.a.	97.2	94.8	94.6	1.00	x	x	x
14	Global Forest Change Data v1-11—GFW/Hansen Map (b)	Image interpretation of time series Landsat, MODIS and very high-spatial-resolution imagery from Google Earth and LiDAR (light detection and ranging) data from NASA’s GLAS (Geoscience Laser Altimetry System). Year(s): n.a.	(L) 99.6 (G) 99.7	(L) 87.8 (G) 73.9	(L) 87 (G) 76.4	(L) 1.01 (G) 0.97	x	x	x
16	GEDI-Height	Sampling from 10% of the GEDI observations. Year(s): 2019	87.8	66.7	89	0.75	x	x	x
17	PALSAR-2 FNF v2.0.0, 3-class	Visual interpretation of Google Earth imagery within a radius of 40 m. Year(s): n.a.	86	n.a.	n.a.	n.a.	x	x	x
18	GLCLU-GLAD	Google Earth imagery and Moderate Resolution Imaging Spectroradiometer (MODIS) time series data. Year(s): n.a.	78.35	87.36	74	1.12	x	x	x
19	CHM-1m	Sampling from 10% of NEON ALS collection (LiDAR observations) dataset. Year(s): n.a.	74	77	78	1.14	x	x	x
20	ETH	All samples located within 20% of the Sentinel-2 tiles (each 100 × 100 km). Year(s): 2020	x	x	x	x	7.3 m	5.5 m	−1.8 m
21	GFCC30TCC v4	Sampling from 250 m MODIS VCF Tree Cover layer. Year(s): 2000–2005 (5-year dataset)	x	x	x	x	16.83%	13.16%	−6%

^(a) They do not provide PA and UA for the entire globe, only for certain places such as California, Costa Rica, Belgium and Laos. ^(b) They only provide validation based on forest loss (L) and forest gain (G). “x” indicates that the parameter is not applicable, and “n.a.” indicates that the information is not available.

). Two of the FNF datasets, GEDI-Height (87.8%) and Forest Extent-GLAD (97.2%), reached the highest overall accuracies of all datasets (Table 4).

A PA/UA ratio above 1 suggests a tendency to incorrectly classify non-forest areas as forest cover (false positives), which could result in non-compliant production areas according to the EUDR (Figure 3). Conversely, a ratio below 1 indicates that the dataset favors missing true deforestation (false negatives). For 9 out of 11 datasets, the ratio of PA/UA is above one (Figure 4). Three datasets, FROM-GLC10, Forest Extent GLAD and GFW/Hansen Map Forest Loss (L), show PA/UA ratios very close to one, suggesting balanced commission and omission errors.

When considering exclusively tree and/or forest canopy height map products (Forest height-GLAD ARD, ETH and GFCC30TCC-v4), MBE values are negative for ETH and GFCC30TCC-v4, suggesting that the datasets have a tendency to underestimate canopy height by a reported average of 1.8 m and 6%, respectively (Table 4, Figure 3). Although the Forest height-GLAD ARD does not report MBE, it provides information on the magnitude of errors between predicted values and reference validation datasets, indicating that predicted canopy heights are often miscalculated by about 4.76 m.

3.3. Second Selection Criterion (Step 3)

The second selection criterion filters out redundant datasets (marked with an “R” in Table 3) and particularly maps with EUDR-contextual constraints, such as the absence of forest status information (Table 3). Following the second filtering procedure, eight datasets are shortlisted for further analysis (JRC GFC v2, ESRI-10m, Dynamic World, GLC-FCS30D, FROM-GLC10, Forest Extent-GLAD, PALSAR-2 FNF v2.0.0, ETH). Moreover, we compare the forest area, in 1.000 hectares, mapped by each shortlisted dataset with the reported forest area of FAO FRA 2020 [37] for selected world regions. The mapped forest area estimates are shown in Figure 5, categorized by region/continent. Shortlisted datasets consistently overestimate forest areas when compared to their counterparts as reported by FAO FRA. This is particularly visible in Central America and Caribbean regions, North America, and Europe (excluding the Russian federation), deviating on average by +49%, +56% and +86% from the reported FAO FRA data in 2020, respectively. In the case of South America, the shortlisted datasets exhibit greater consistency with FAO FRA, with significant discrepancies observed for JRC GFC v2 and ETH.

A notable case is the lower (−38%) forest area mapped by FROM-GLC10 in Africa compared to the FAO reference (Figure 5). Moreover, Dynamic World consistently exhibits lower forest/tree cover compared to the other datasets, including FAO FRA estimates. The datasets with the highest estimation of forest area on average and for all regions are ETH followed by PALSAR-2 FNF. These datasets belong to the TC/NTC and FNF types of maps, respectively.

Figure 6 presents the summary of the eight shortlisted datasets in comparison to the chosen EUDR criteria and accuracy metrics, with additional information on whether non-forest tree-based systems are included in the considered forest cover information.

4. Discussion

The findings of this study suggest several key implications for the use of global FNF, TC/NTC and LULC datasets in deforestation monitoring within the EUDR context, which are discussed in detail below.

4.1. Spatial Detail and Temporal Proximity

The spatial resolution and the classification year of the map impact the level of spatial variation contained in a single pixel and influence the applicability of a map as a reference for compliance checks under the EUDR. Following this rationale, a pixel that is detailed enough to display small features in a landscape is thus more suitable to represent small-scale farming systems such as coffee or cocoa, for example, or heterogeneous mosaic landscapes common for agroforestry [74]. However, the compiled global datasets range from 10 to 30 m resolution with the exception of CHM, which has a 1 m spatial resolution. Therefore, a cropland plot with an area below the reported pixel size would be aggregated to the neighboring dominant landscape features to form a single pixel value. If the dominant landscape is a forested area, the small agricultural plot may be assigned to the forest class, even though in reality it represents a non-forest class, leading to false conclusions of non-compliance with deforestation-free production. This can increase the error of smallholder farming plot misidentification. For example, in the case of coffee, one of the seven EUDR commodities, about 60% of total global production originates from plots smaller than 5 hectares [31]. Conversely, small-scale tree cover or forest cover losses may go undetected, as demonstrated by a study in the Mato Grosso region of Brazil, which identified more deforestation using a product with a 5 m spatial resolution compared to one with a coarser resolution of 6.5 hectares [50].

Another advantage of finer-scale datasets is the expected better representation of heterogenous landscapes, such as those characterized by mixed farming plots of multi-cropping or agroforestry systems, landscapes with mixed shrub, grass, and tree cover or even the detection of small changes within a forest patch. A global cross-map comparison concluded that locations with more homogeneous characteristics typically exhibit higher overall accuracy, compared to areas where not all pixels within a block belong to the same class [35]. However, higher spatial resolution is not a definitive determinant of precision in capturing smaller and non-uniform features, as the accuracy of mapping classifications depends on quality, representativity and uniformity of the reference (e.g., training) data [35]. Factors such as spectral confusion in agroforestry mosaics, mixed-pixel effects, post-processing smoothing of small features, uncertainties in classification algorithms, and the shape and spatial arrangement of production and forest plots are also significant contributors to mapping errors. Furthermore, the capability of spatial resolution and maps to detect deforestation, particularly forest degradation, involves a trade-off between the wall-to-wall coverage and cost and time efficiency of monitoring over large areas, but reduces the ability to detect smaller changes and fragments [75]. Finally, the EUDR regulation lacks explicit guidance regarding the MMU for detecting forest change processes such as deforestation and forest degradation. This omission adds to the ambiguity surrounding the required spatial resolution for compliance monitoring, particularly in relation to whether the 0.5-hectare MMU defined for forest classification should also be applied to change detection.

The threshold for the temporal proximity metric is set by the requirements of the regulation, even though the dynamics of natural ecosystems may not follow such a fixed “cut-off date”. From the legal point of view of the regulation, a dataset from before 2020 or 2021 may not necessarily accurately represent the status of forests around the cut-off date. Nevertheless, a global map dataset that is 5 years older or newer than the 2020 reference year may still serve as a good indicator for possible land use transitions. On the one hand, the likelihood that non-forest areas mapped in 2015 still exhibit properties of non-forest areas in 2020 is high, as forests do not usually regrow to a mature stage within this short period of time. On the other hand, naturally regenerating (young) forest stands may have been incorrectly classified as forest cover in 2015, or the regeneration process may have started at a later point in time, both resulting in a misclassification of forest cover in 2020 as non-forest. The latter scenario can imply missing true deforestation pixels (false negatives), while the former may prevent false allegations of regulation violation (false positives). A similar logic applies to maps created in 2025, as identifying mature forests in 2025 can provide better insights into the 2020 classification, helping determine whether an area was truly non-forest or a young regenerating forest in 2020.

Another point related to temporal proximity concerns the harvesting of wood as part of mapping forest degradation. Harvesting of wood can cause structural changes to forest cover and is considered forest degradation under the EUDR. This activity occurs within a specific timeframe, typically over hours or days. Except for the “forest disturbances alert” type of datasets [75] and some other exceptions, most static datasets of FNF, TC/NTC and LULC represent the land use or land cover status of a single year based on monthly satellite imagery, which poses threats to precise degradation detection. Consequentially, the time in which degradation can be observed through remote sensing could be reported later than the respective DDS, permitting certain wood products to be mistakenly labeled as “deforestation-free” production. The compiled datasets in this study do not include “forest disturbances alert” datasets, as we primarily focus on the tree cover or forest cover and deforestation.

4.2. Defining Forests: Physical Thresholds and Land Use vs. Land Cover

Of the selected datasets, only JRC GFC v2 and PALSAR-2 FNF v2.0.0 fully meet all three EUDR forest definition parameters, while the others align with at least one parameter, often forest height. In most cases, global datasets classify vegetation shorter than 5 or 3 m as low vegetation such as shrubs and grasslands, excluding it from the “forest/tree cover” class. Vegetation taller than these thresholds is usually classified as “forest/tree cover”, with heights above 10 m associated with tall trees. Among forest datasets, those generated using LiDAR technology offer tree cover maps with heights starting from zero, as LiDAR captures three-dimensional information on the shape of surface features [68]. This allows users to define specific height thresholds for “forest/tree cover” classification.

Regarding canopy cover, some datasets distinguish between dense/close and open canopies, with threshold values above 40% [58] or 70% of the surface [76] referring to dense canopies, and values below the respective thresholds indicating open canopies. However, most global datasets adopt a more general threshold of 10% or 15% when defining forest/tree cover, with tree cover canopy below 10% typically classified as “Other wooded land”. The MMU parameter further defines the minimum area of tree cover required to classify an area as forest. Below a certain threshold, trees may be too sparse to meet the forest classification criteria, which is particularly problematic for dryland forests [77].

Major international environmental and forestry organizations commonly adopt a minimum tree height of 5 m, a canopy cover threshold of 10%, and a minimum forest area of 0.5 hectares [43], similar to the FAO and corresponding to EUDR adopted definitions. However, these parameters do not always align with national standards. A comprehensive list of forest definition parameters available for several tropical countries reveals a range of thresholds, including minimum tree heights from 2 to 5 m, minimum forest areas of 0.05 to 1 hectare, and canopy cover percentages between 15% and 30% [44]. In Germany, the parameter for canopy cover can be as high as 50% [45], and a study in the flatlands of Ukraine identified the optimum threshold as 40% [46], while in Indonesia, it may reach up to 60% for certain regions [47]. These variations illustrate the discrepancies in forest classification standards across different regions and the EUDR-defined parameters. Therefore, not only may global datasets disagree with the adopted parameters of forest definition by the EUDR, but these parameters may also not match certain national standards or regional guidelines. Taking the example of Indonesia, the minimum canopy cover density for a tree-covered area to be classified as forest is 30%. The EUDR regulation defines forests as tree-covered areas with a minimum canopy cover of 10%, which introduces uncertainties for locations with canopy cover between 10% and 30% due to mismatches between the applied thresholds of forest definition. Mismatches in any of the three forest definition parameters increase the uncertainties in verifying deforestation-free production using global generalized datasets. This likely contributes to the significant discrepancies in global forest area estimates derived from these datasets and national/regional forest estimates, as further discussed in the next section.

Moreover, ref. [43] argued that these types of structural measures may be more relevant to ground-based inventories and not necessarily remote sensing surveys as they may (1) prevent the identification of real deforestation because an area, after considerable tree loss, is below an artificially set threshold, and (2) be unable to capture natural forest regeneration or early stages of restored forests that do not yet satisfy certain forest definitions.

Beyond the structural measures, another frequently occurring issue is whether a forest assessment, either remotely sensed or sample-based, actually discriminates natural forest cover from planted forest and tree or agricultural plantations, thus dissociating “land use” from “land cover” [43]. A study [28] estimated that globally, more than 43% of agricultural land is within agroforestry systems with 10% tree cover. Ensuring proper differentiation of tree cover classes from forest land use is thus essential for accurately detecting deforestation. Nonetheless, the vast majority of maps struggle to distinguish tree-covered areas (land cover) from forest cover (land use). While remotely sensed datasets can better identify cover types through properties such as spectral signatures or texture, they cannot reliably infer human-related land use without incorporating auxiliary local information.

4.3. Accuracy Metrics

Although the technical parameters are not applied in the selection criteria (step 2 and step 3), they provide a further assessment of the shortlisted datasets. More precisely, the accuracy metrics give supporting evidence to the dataset quality in mapping forest and tree cover areas. We look at the producer’s and users’ accuracy of tree/forest class as a ratio. This ratio conveys the relationship between a measure of completeness in mapping forest or tree cover areas over the reliability in correctly classifying true forest areas.

With the exception of GEDI-Height and GFW/Hansen Map, all the shortlisted datasets display a PA/UA ratio above 1, which indicates that most maps potentially have a tendency to falsely identify areas of non-forest or non-tree cover as forest or tree cover (false positives, commission errors). In contrast, the two datasets with a ratio lower than 1 suggest that potentially true forest or tree cover area is missing; thus, they are prone to omission errors. A key factor explaining this pattern is that these datasets aim to accurately map (individual) tree cover, rather than forest areas specifically, resulting in potential underestimation of tree cover and not inherently forest cover. The reason is that trees can also be present in non-forest environments, such as crop plantations, woodlands/savannas, and shrublands. Evidence for this is found in the GEDI-height dataset, which features forest areas, with tree height ≥5 m, extending into savannas and open-canopy forests in Africa [67].

Furthermore, FNF and TC/NTC map datasets generally reveal a PA/UA ratio smaller than or very close to 1 and higher OA. Binary maps like FNF and TC/NTC implement a simpler classification scheme, reducing error potential and positively affecting OA. Conversely, multi-class land cover maps introduce more classes, increasing the likelihood of misclassifications due to, among other reasons, the difficulty of distinguishing similar classes. Hence, the PA/UA, for the forest/tree cover class, and OA provide essential insights into systematic biases in the datasets and further reflect the map type under consideration (FNF and TC/NTC or LULC). Despite that, the reported accuracy values may not truly represent the local and regional realities with diverse and complex landscapes, where typically, land use classification experts expect higher commission than omission errors [35,52,78].

Another accuracy metric examined in this study is the MBE, which is applied to assess systematic bias in exclusively forest/tree canopy height datasets. While RMSE and MAE reveal the magnitude of the prediction errors observed in these datasets, they do not inform whether forest height estimations tend to overestimate or underestimate actual values. This distinction is important, as forest height is a key parameter considered in forest definitions. MBE, therefore, suggests an indirect interpretation of these systematic biases and their implications in relation to the EUDR, as MBE relates to the tendency of a dataset to map higher (positive) or lower (negative) height than the validation data. Positive values imply that predicted heights exceed actual values, potentially classifying more tree-covered areas as forest (overestimation). In contrast, negative values (found in two of the three evaluated canopy height datasets) indicate underestimation, possibly excluding certain tree-covered areas due to lower height than the minimum threshold considered in the EUDR forest definition. Additionally, continuous forest canopy height datasets are less suitable to represent discrete multi-class land cover classifications than to function as datasets on vegetation structure and above-ground biomass [67].

It is thus important to acknowledge common aspects that can influence the accuracy results, such as the quality of the employed reference datasets (including spatial resolution and geolocation errors) [79], the homogeneity of the landscape classes in question, the representativity of certain class (frequency of occurrence), the specific sampling strategy used for validation (e.g., stratified or area-weighted approaches), and independence of validation points [26,53].

4.4. Shortlisted Dataset Cross-Comparison on Forest Area

The three regions with the highest percentage average difference in forest estimations compared to the forest area statistics reported by the FAO FRA 2020 [37] are Central America and the Caribbean, Europe, and North America. Common issues in North America and Europe include over- or underestimation of forest cover in boreal and tundra transition zones [78], underrepresentation of forest edges in temperate ecosystems [80], and misclassification of low tree cover and shrubland areas [77], such as those in the Mediterranean region of Europe. The case of Europe is the most significant one, with the highest average differences shown. A plausible explanation for the observed pattern is the abundance of urban parks/forests, green spaces, and small tree-covered patches across Europe [20,46] in addition to differences in forest cover threshold definitions discussed in previous sections, as well as other factors not examined in depth in this review study. The JRC GFC v2 dataset [78] reported persistent challenges in differentiating agricultural plantations and urban vegetation from forest cover already present in its first version. These two land use types do not fall under the classification of forest according to EUDR, even though they contain trees. Urban green areas and urban tree parks are unlikely to conflict with EUDR regulation, as these areas are not expected to be sources of EUDR commodities. However, other tree-based land use systems, such as plantations, managed planted forests and agricultural plantations, may present conflicts with EUDR definitions. These classes are generally categorized as forest or tree cover in most global maps, without distinguishing whether they represent primary forests, secondary forests, plantations, or other sub-classes of tree-based land cover. This lack of differentiation contributes significantly to the discrepancy between FAO estimates and globally mapped areas.

South America is the region exhibiting the highest consistency between the shortlisted datasets and the reported FAO estimate. This region contains major dense forest areas primarily concentrated in the Amazon Basin, which facilitates the differentiation between forest and agricultural lands and certain plantations, often including clear edges and structural patterns such as oil palm plantations [21,77]. A paper [27] evaluated an index designed to measure the similarity of spatial patterns between two numerical raster maps at the pixel level. Their findings revealed that the Amazon Basin and Central Africa exhibited the highest similarity values among the regions analyzed globally. However, due to the systematic mapping of observed dense forest in this region, the risk of misclassification of some agricultural plantations and agroforestry systems is quite high, as most global TC/NTC, LULC and even FNF maps are often unable to capture such fine-scale information and land use-driven classes. Moreover, a large rate of misclassification of mixed landscapes and confusion between agriculture and natural landscapes is expected in intensively cultivated areas in South America [23]. Finally, transition zones between the Cerrado ecosystem and the rainforest in Brazil may contribute to the map disagreements in the region with potential implications for EUDR, as Cerrado (commonly referred to as “other wooded land”) is not yet included in the regulation.

In the African region, a frequent issue is the proper identification of seasonally dry tropical forests due to related phenological seasonal effects [24,25], particularly in transition zones between deserts and tropical rainforest [26]. Other factors are correctly mapping areas of low tree cover density, degraded forest, small patches of forest, and forest edges [4,67,77], and the overlap between woodlands and forest areas based on the adopted forest definition [48]. A recent study in Côte d’Ivoire [36] highlighted inconsistencies between a national forest map and the JRC GFC first version map. Specifically, the JRC GFC first version map shows areas as forested, while the national map classifies the same areas as non-forest, resulting in an error of commission. A particular shortlisted dataset, FROM-GLC10 [66], considerably underestimates forest cover in Africa. The underlying causes for this particular pattern remain unclear and require further investigation. Nevertheless, most global forest/tree cover maps tend to underestimate forest cover in low-tree-cover areas [77], and mapping of dry forest areas is particularly challenging with a tendency for omission errors [78].

Commonly challenging regions for mapping forest cover include complex topographical areas (mountainous and hilly areas) that hinder the proper identification of forest cover due to terrain effects such as shading of slopes from illumination effects [22], and forest edge zones adjacent to non-forest areas [26], which can be problematic for coffee plantations, as in many regions, hilly terrain is preferred due to its favorable climate. Moreover, binary maps such as FNF and TC/NTC maps require a subjective visual interpretation decision to be made on which class a certain area/pixel belongs to, which does not always correspond to the continuous features of landscapes (transition zones) in reality. Furthermore, hard thresholds on forest definition do not always correspond to the local physiographical and phenological conditions of natural areas around the world, varying from region to region [49].

Two datasets consistently and significantly overestimated forest cover in all regions, namely ETH [72] and PALSAR-2 FNF [69]. The dataset ETH reveals a tendency to overestimate canopy height in areas where the canopy is lower than 5 m. As a result, many areas that do not originally belong to the forest class were incorrectly included as forest, possibly due to the overestimation of canopy height. PALSAR-2 FNF shows tendencies for misclassifications of forest areas in highly fragmented landscapes [46]. Although the forest estimates from pixel counts of classified maps shown in Figure 5 derive from existing maps, they do not represent official estimates of forest areas in the respective regions and they do not accurately represent the forest status of 2020. The purpose of comparing forest area estimates from sample-based approaches, such as those used by the FAO, with mapped forest areas is to highlight the differences in the methods used for estimating forest cover. This comparison helps to illustrate the diversity in global forest/tree cover mapping techniques and their varying results. Additionally, FAO-reported forest estimates may also contain data inconsistencies and aggregation of non-forest cover, e.g., rubber and Christmas tree plantations, in forest cover [43], which contributes to discrepancies in the reported forest areas [69].

The observed forest area differences are especially notable when comparing global forest/LULC datasets with national maps, like the case of Côte d’Ivoire [36]. However, national maps are not always publicly available; thus, in some cases, the use of global maps is the most viable solution for verifying evident cases of deforestation, functioning as the first screening. Finally, maps are produced using various methods and data sources, resulting in differing outcomes. Integrating multiple sensors (optical, microwave, etc.) can provide complementary spectral characteristics of different satellite properties [17,77]. Some methodologies may be more robust than others, with training sampling data that are better structured, clustered and thus more representative in certain cases [18].

We acknowledge that our approach of estimating forest area (in thousands of hectares) by counting forest-classified pixels and scaling them by pixel size may introduce quantification bias [81]. Moreover, FRA reports and remote sensing survey data are not directly comparable, as remote sensing data do not adhere to the same formal procedures as FRA reports, and each dataset is derived using distinct methodologies. Additionally, FRA reports may be subject to errors, as they rely on nationally reported data, which can originate from different data sources, such as national statistics, remote sensing, or ground-based assessment. Nevertheless, our objective in this case is not to obtain precise forest area estimates, but rather to understand the variability among estimates and identify the sources of dataset diversity.

4.5. Dataset Selection in Enforcement Contexts: The Use of Shortlisted Datasets

The eight shortlisted datasets streamline verification for NCAs and operators by narrowing the initial range of mapping options, without prescribing the use of a single dataset. Each dataset has its own strengths and limitations, yet all meet the core EUDR criteria, allowing NCAs to use them as a baseline reference for enforcement. Integrating these datasets into a forest agreement layer supports the rapid identification of deforestation-free areas and clear cases of deforestation-linked production, particularly for non-tree commodities. This enables NCAs to prioritize inspections and allocate resources more efficiently. By providing a targeted shortlist, the verification workload is reduced, and disputes arising from conflicting dataset information are minimized, which can be an important advantage given the large number of DDS submissions that NCAs will encounter at the start of regulation implementation. Removing unsuitable products upfront and offering a curated set aligned with regulatory criteria allows NCAs to avoid prolonged case-by-case evaluations, maintain a consistent evidentiary standard across operators, and substantially reduce both uncertainty from conflicting information and the technical burden of comparing numerous datasets.

However, relying solely on these datasets risks missing true cases where deforestation or degradation occurs outside the detectable scope of static global products. Enforcement bodies should therefore require supplementary evidence, such as national or regional mapping, management plans, or other documentation along the chain of custody, to confirm EUDR compliance. Global static annual LULC or FNF and TC/NTC maps fail to capture intra-annual dynamics, which may result in misclassification, particularly when seasonal variations are not adequately considered. In such cases, inter- and intra-annual time series analyses may provide more robust evidence for detecting deforestation or verifying its absence [82]. Moreover, incorporating commodity-specific maps [83,84,85] and near-real-time alert systems into enforcement protocols can reduce classification errors: commodity maps improve the detection of plantations and reduce false positives, while disturbance-monitoring products, such as near-real-time alert systems [86,87], capture subtle or seasonal forest loss often missed by static annual maps.

5. Conclusions

The results of this research provide insights into the systematic assessment of global FNF, TC/NTC and LULC maps’ suitability to the EUDR. We evaluate 21 maps based on two groups of criteria, namely EUDR-relevant parameters and technical parameters. The EUDR parameters describe the ability of a dataset to align with three critical properties of the regulation: (1) temporal proximity, (2) forest definition specifications and (3) level of spatial detail. The technical parameters work as a complementary evaluation to provide detailed information on map error tendencies based on accuracy metrics.

We applied two stages of exclusion criteria: first, based on two of the three EUDR metric indicators; and second, to avoid redundancies and data inconsistencies. After application of the second filtering criteria, we extracted a concise list of eight global FNF, TC/NTC and LULC maps that are largely suitable for checking deforestation-free production within the EUDR framework. Some datasets are better suited than others, but all eight datasets are viable options in assisting the EUDR regulation. Finally, we ran a cross-map comparison of forest area estimates from each dataset against FAO-2020 forest reports to reveal initial regional challenges.

The main results of the study suggest the following:

▪

Very few global FNF, TC/NTC and LULC maps fully match the EUDR forest definition parameters (tree height, MMU of forest cover and forest canopy cover), which could be one of the major sources of uncertainty when assessing EUDR compliance using EO products.

▪

Tree height emerges as the most widely accepted specification among the forest definition parameters.

▪

The majority of global LULC maps show tendencies to display forest overestimation (commission errors), by classifying areas of non-forest cover as forest cover.

▪

Canopy height products display a tendency to underestimate canopy heights (MBE), possibly excluding certain tree-covered areas but not necessarily forest-covered areas.

▪

In addition to the accuracy metrics assessment, most datasets map tree cover rather than forest cover, which can include non-forest areas with trees, thereby contributing to uncertainties.

▪

Discrepancies between forest area estimates from the datasets and the FAO estimate reveal the difficulty of choosing a single mapping approach.

▪

The regions with the highest overestimation of forest areas, compared to FAO estimates, include Central America and the Caribbean, Europe, and North America. The African region shows the greatest underestimation, while South America has more or less consistent estimates.

This study contributes to the understanding of the key considerations for selecting geospatial data to include in DDS or to verify deforestation-free production. A detailed understanding of these map specifications offers valuable insights into how well a map dataset complies with EUDR requirements, as well as the misclassification tendencies within these maps. Additionally, the regional assessment of forest area estimates provides further understanding of regional differences and commodity-specific issues.

The findings support decision-makers, national competent authorities, operators, and producers in selecting forest cover maps for deforestation-free risk assessments and due diligence statements, by evaluating their compliance with the EUDR. Each map displays certain capabilities and limitations, and all maps contain inherent errors. Some of these capabilities and limitations can be higher for some world regions than others and can be aggravated for some EUDR commodities (e.g., agroforestry coffee systems). Therefore, the most effective approach is to utilize multiple datasets at different scales that capture different aspects, preferably with varying map specifications, such as different levels of spatial detail (spatial resolution) and those generated through different satellite sensors or methodologies. Nevertheless, very high-resolution maps may not be accessible everywhere due to their large storage requirements, high computational demands, and time-consuming nature. As a result, relying on globally available, open-source maps becomes a necessary alternative.

While this study is comprehensive and extensive, it may not cover all available mapping possibilities. Moreover, it does not discuss region-specific issues in depth that may further explain the presented variability of map coverages across regions. Moreover, the use of geospatial reference information in EUDR processes is optional. Nevertheless, the findings of this study can serve as a guideline when opting for the use of remotely sensed map products in EUDR-related measures. Further research should focus on quantifying uncertainty in EUDR enforcement in practical scenarios such as dry runs. Additionally, detailed analyses of specific cases, such as agroforestry systems, as well as hotspots of EUDR-relevant land use change, are needed to validate sources of regional uncertainty and better understand factors contributing to false positives and false negatives. Such analyses are essential to reveal in more detail opportunities and limitations when relying on remote sensing reference maps for EUDR compliance purposes. Moreover, they provide critical guidance to national competent authorities, operators, and other stakeholders involved in enforcement and monitoring processes.