Forest Cover in the Congo Basin: Consistency Evaluation of Seven Datasets

Abstract

Tropical forests play an essential role in the carbon and water cycles of terrestrial ecosystems, but they are increasingly threatened by human activities and climate change. For places where ground observations are scarce, like in Equatorial Africa, remote sensing is a key source of information for monitoring the temporal and spatial dynamics of forests over large areas. Several Earth Observation-based global maps were developed in recent decades using different definitions of the land-use/land-cover (LULC) classes. While such products are widely used for monitoring land use and planning land management, the consistency of these LULC maps for the Congo Basin has never been analyzed and quantified at the ecosystem level. Here, we selected seven of the most-used global maps and analyzed their consistency over the Congo Basin. After reclassification into forest/non-forest masks and spatial resampling, we assessed the agreement and disagreement percentage across the different tropical ecoregions of Africa, from moist forest to miombo, including savanna. The datasets showed differences in forest area as a function of spatial resolution, with higher forest area levels at coarser resolutions (e.g., from 74.1% to 88.5% forest cover when upscaling the GLCLU from 30 m to 1 km over the Congo Basin). A higher agreement between the datasets was found for forest area over moist forest (between 88.18% and 99.38%) in comparison to savanna (32.82%–99.84%) and miombo (53.83%–99.7%). These discrepancies led to large differences in forest cover, varying from a net loss of 205,704 km² to a net gain of 50,726 km² over 2001–2019 depending on the dataset used. This study draws attention to the uncertainty associated with these products with regard to forests, particularly in regions of biological importance, such as the miombo and savanna regions, which remain poorly understood. Indeed, the two major uncertainties affecting the quality of LULC products are related to the different spatial resolutions and biological definition of “forest” adopted by each product.

1. Introduction

Tropical forests are key components of the global and water carbon cycles, regulating the climate system by storing almost 50% of the terrestrial carbon [1] and by storing and recycling a large part of the precipitations [2]. However, these ecosystems and the services they provide are increasingly threatened by climate change (e.g., drought) and anthropogenic pressures (e.g., agricultural clearance) [3,4], potentially impacting the future resilience of forests. Land-cover change is among the main anthropogenic disturbances [5], with forests being converted to agricultural land, mining areas, and energy infrastructure [6]. In particular, Africa was subject to a net forest loss of 3.9 million ha per year between 2010 and 2020, higher than deforestation in South America or Asia over the same period [7].

Compared to South America and Southeast Asia, which are mainly affected by large-scale disturbances caused by agricultural expansion (e.g., soy and palm oil), forest loss in Africa mostly occurs at a fine scale due to shifting agriculture [6]. In most of the cases, small plots are burned and cleared, and then they are cultivated for a few years before being left fallow for more or less long periods [8,9], allowing a secondary forest to grow and causing rapid land-cover dynamics, which are difficult to monitor [10].

To understand the dynamics of carbon fluxes in African tropical forests, it is essential to understand the dynamics of the land-use/land-cover changes. If the African carbon sink was stable until 2010, carbon losses have been found to increase in the last decade due to increasing tree mortality, deforestation, and forest degradation [4,11]. In this context, international programs such as Reducing Emissions from Deforestation and Forest Degradation (REDD+) were developed to help countries reduce their carbon emissions caused by deforestation [11]. To this end, countries need to accurately monitor forest gains and losses, especially in Africa, where information is scarce [12]. Accurate measurements of aboveground biomass and forest dynamics are provided by ground-based forest inventories. However, they are limited to local scales, as they require high sampling density, human resources, and forest accessibility [13]. As a promising alternative, remote-sensing observations allow us to cover a whole region/country, especially in remote places [11].

Over the past decades, Earth Observation (EO) data have been intensively used to monitor the land-cover dynamics. Two kind of datasets are available, namely (i) land-use/land-cover (LULC) maps, which include forest classes [14,15,16,17], and (ii) products dedicated to tree cover or forest monitoring [18,19]. Even if all of these products theoretically allow us to monitor either the loss/gain in canopy cover or the transition from forest to other land classes, they are likely to suffer from limitations that are inherent to the remote-sensing sensors and the methodology used in their development. These limitations are particularly frequent in Equatorial Africa, where some regions of tropical forests are almost permanently covered by clouds, thus highly impacting the quality of maps developed with optical observations [12]. The low availability of non-cloudy observations, combined with the low resolution of some sensors, does not allow researchers to properly capture the fine-scale disturbances induced by shifting agriculture and selective logging specific to this region [12,20]. Several remote-sensing products for forest monitoring have emerged over the last few decades. However, a clear assessment of their consistency and limitations is necessary for policy makers to adequately select the most relevant map for their needs.

In this study, we intercompared seven commonly used LULC and forest-monitoring datasets in order to show their similarities and differences. We also identified the key aspects that need to be addressed before generating new land-use maps for forest monitoring. We first homogenized the definition of forest classes across products that were resampled at the same spatial resolution. We then compared seven different global products to each other for the years 2005, 2010, and 2015, which is the period shared by all products, to identify the regions and the land-use classes for which we observed the main inconsistencies. The analysis was carried out over the Congo Basin and covered different ecoregions representative of the main tropical forest types in Equatorial Africa, including (1) moist forest, which is a region of Central Africa that is highly affected by cloud [21]; (2) savanna forests; and (3) miombo forests.

2. Materials and Methods

2.1. Studied Areas

2.1.1. Congo Basin Studied Area

The Congo Basin region encompasses six countries: (i) Cameroon, (ii) Gabon, (iii) the Republic of the Congo, (iv) the Democratic Republic of the Congo, (v) Equatorial Guinea, and (vi) the Central African Republic (Figure 1). The Congo Basin region is located between 8.4° and 30.0° E in longitude and between 13.5° S and 6.0° N in latitude. Maps were compared over the entire zone, i.e., an area of 3.48 million km², which has different climates. Based on the Köppen–Geiger classification, the region is mostly classified as “tropical rainforest”; “tropical monsoon”; “tropical savanna”; and “temperate dry winter, hot summer” [22].

Three ecoregions were defined with the Terrestrial Ecoregions of the World (TEOW) [23] classification, which is based on biogeographic, phytogeographic, and zoogeographic data:

(i)

Miombo, which contains the TEOW regions with “miombo woodlands”.

(ii)

Moist forest, which contain the TEOW regions with “costal”, “lowland”, or “swamp” forests.

(iii)

Savanna, which contain the TEOW regions with “forest–savanna mosaic” (

Table A2. Definition of the ecoregions based on the Terrestrial Ecoregions of the World.

Miombo	Moist Forest	Savanna
Angolan miombo woodlands	Northwestern Congolian lowland forests	Northern Congolian forest–savanna mosaic
Central Zambezian miombo woodlands	Northeastern Congolian lowland forests	Western Congolian forest–savanna mosaic
	Central Congolian lowland forests	Southern Congolian forest–savanna mosaic
	Atlantic Equatorial coastal forest
	Cross Sanaga–Bioko coastal forests
	Western Congolian swamp forests
	Eastern Congolian swamp forests

).

These regions are located across Equatorial Africa (Figure 1) and were selected to be representative of the different forest types and climates in this region. These different ecoregions are defined as follows:

2.1.2. Miombo Forest

The miombo woodland zone that we studied is located between a longitude range of 17.45°–29.98° E and a latitude range of 4.05°–13.46° S (Figure 1) and covers a surface area of 438,808 km². It is composed of grassland, savanna, and shrubland. Temperatures are between 9 and 18 °C for the minimum and between 24 and 27 °C for the maximum. The precipitation, 1000 to 1200 mm per year, is concentrated between November and March [24,25].

2.1.3. Moist Forest

The moist forest zone that we studied is located between a longitude range of 8.42°–29.98° E and a latitude range of 6.04° N–5.74° S (Figure 1) and covers a surface area of 1.8 million km². It is composed of three principal types of forests: (i) swamp forest; (ii) lowland forest, and (iii) coastal forest. This region has high humidity, with precipitation around 2000 mm per year, and a low seasonality, with temperatures between 21 and 27 °C [26].

2.1.4. Savanna Forest

The savanna forest zone that we studied is located between a longitude range of 10.07°–29.98° E and a latitude range of 6.04° N–9.90° S (Figure 1) and is covered by a surface area of 1.1 million km². It is composed of tropical forest, savanna, and shrublands. The climate is reflective of that of a savanna, with precipitation around 1400 mm per year. The temperatures are between 18 and 21 °C for the minimum and between 27 and 30 °C for the maximum [27]. Rainfall is concentrated during the wet season. The rest of the year is dry [28].

2.2. Land-Cover Data

We used five LULC datasets in our intercomparison study: (1) the Global Land Cover and Land Use (GLCLU) [14]; (2) the land-cover product of the European Space Agency Climate Change Initiative (ESACCI) [15]; (3) the MODIS Land Cover Type 1 Product (MODIST1) [16]; (4) the MODIS Land Cover Property 2 (MODISP2) [16]; and (5) the HIstoric Land Dynamics Assessment + (HILDA+) [17]. Two forest-monitoring maps were also used: (6) the Global Forest Watch (GFW) [18] and (7) the Tropical Moist Forest Dataset (TMF) [19] (

Table 1. Summary of the datasets used in the intercomparison study, along with their main characteristics. The characteristics include spatial resolution, temporal resolution, temporal coverage, sensors, classification methodology, and number of classes.

	HILDA+	MODIST1	MODISP2	ESACCI	GLCLU	GFW	TMF
Spatial resolution	1000 m	500 m	500 m	300 m	30 m	30 m	30 m
Temporal resolution	Annual	Annual	Annual	Annual	Every 5 years	Annual	Annual
Temporal coverage	1960–2019	2001–2022	2001–2022	1992–2020	2000–2020	2000–2022	1990–2022
Sensors	Multi-source	MODIS	MODIS	AVHRR, MERIS, SPOT-VGT, PROBA-V, S3-OLCI	Landsat GEDI	Landsat	Landsat
Classification methodology	Supervised	Supervised	Supervised	Unsupervised	Supervised	Supervised	Supervised
Number of classes	6	18	12	36	110	Continuous tree cover between 0 and 100% Year of forest loss	6

). They are all freely available (

Table A1. Data availability.

	Web Address Dataset Download	Date of Download	Reference Publication
HILDA+	https://doi.pangaea.de/10.1594/PANGAEA.921846?format=html#download	6 December 2023	Winkler et al., 2021 [17]
MODIST1	https://appeears.earthdatacloud.nasa.gov/	21 March 2024	Sulla-Menashe et al., 2019 [16]
MODISP2	https://appeears.earthdatacloud.nasa.gov/	21 March 2024	Sulla-Menashe et al., 2019 [16]
ESACCI	https://cds.climate.copernicus.eu/datasets/satellite-land-cover?tab=overview	4 December 2023	ESA. “Land Cover CCI Product User Guide Version 2”. 2017 [15]
GLCLU	https://storage.googleapis.com/earthenginepartners-hansen/GLCLU2000-2020/v2/download.html	1 December 2023	Potapov et al., 2022 [14]
GFW	https://storage.googleapis.com/earthenginepartners-hansen/GFC-2022-v1.10/download.html	6 December 2023	Hansen et al., 2013 [18]
TMF	https://forobs.jrc.ec.europa.eu/TMF/data	1 March 2024	Vancutsem et al., 2021 [19]

).

Compared to LULC maps, GFW and TMF only provide changes in forest coverage. GFW is composed of one map for the year 2000 reporting the tree-cover percentage (global forest cover, GFC), and a second map containing the year of forest loss [18]. For TMF, we used the annual change maps, which include deforestation and degradation from 1990 to 2022 [19]. Compared to GFW, TMF provides annual forest regrowth after deforestation [19], while GFW only provides information about forest regrowth that occurred between 2000 and 2012 [18]. Since the TMF is dedicated to tropical moist forests, its uncertainty was explored across the entire region, with a closer look at moist and savanna forests, as these are biomes that encompass the forests targeted by the TMF.

The LULC maps are based on observations from different sensors, mainly optical ones. MODIST1 and MODISP2 used the MODIS sensor (Moderate Resolution Imaging Spectroradiometer) [16]. ESACCI is based on several sensors: AVHRR (Advanced Very High Resolution Radiometer), MERIS (Medium-Resolution Imaging Spectrometer), SPOT-VGT (Satellite Pour l’Observation de la Terre-Vegetation), PROBA-V (Project for On-Board Autonomy, with the V standing for Vegetation), and S3-OLCI (Sentinel-3 Ocean and Land Colour Instrument) [15]. GFW and TMF are based on Landsat imagery, just like GLCLU, which also incorporates GEDI (Global Ecosystem Dynamics Investigation) LiDAR data [14,18,19]. For HILDA+, remote sensing-based LULC and statistical data are merged [17]. These different datasets are classified using supervised or unsupervised techniques to generate a number of classes, and this number varies from one product to another, from 6 classes for HILDA+ to 110 classes for GLCLU. This heterogeneity in number of classes is partly linked to the definition of forests that is based on different tree-height and tree cover-percentage criteria that vary among products. The selected databases have different spatial resolutions, from 30 m (TMF, GLCLU, and GFW) to 1 km for HILDA+. All products provide land classes on an annual basis, except for GLCLU, with a five-year temporality. Due to their different levels of temporal coverage, we restrained our analysis to the years 2005, 2010, and 2015, which are shared by all products.

2.3. Homogenization of Datasets and Forest Definition

The processing described below was applied to each available common year of the datasets (2005, 2010, and 2015). The map labeled “reference” has the lowest original resolution. The other map labeled “compared” has the highest original resolution. To homogenize the representation of forest/non-forest classes across products, we reclassified each category based on the definition of FAO that defines a forest as an assembly of trees higher than 5 m, with a tree cover higher than 10%. Information on tree cover and height was not present in all the datasets. We tried to match the FAO definition as closely as possible. However, some compromises had to be made (

Table A3. HILDA+ reclassification, based on Winkler et al., 2021 [17].

General Class	Description	Forest Mask Classification	IPCC Class
Urban	Artificial surfaces, and urban and built-up areas, including urban parks and sports areas, green spaces, industrial areas, deposits, and extraction sites (mining, etc.).	Non-forest	Settlement
Cropland	Herbaceous and woody crops (also for hay production), including tree/shrub crops, orchards, plantations, and multiple/layered crops, incl. mosaics (with cropland area fraction > 40%).	Non-forest	Agriculture
Pasture/rangeland	Managed herbaceous plants (cover > 10%), including managed grasslands (e.g., prairies, steppes, savannas, and mosaics with tree/shrubs): grasslands or meadows used for livestock grazing, hay production with different intensities, etc.	Non-forest	Grassland
Forest	Trees with > 5 m height (cover > 10%), including forest plantation, and trees on seasonally or permanently flooded areas, including mangroves.	Forest	Forest
Unmanaged grass/shrubland	Natural herbaceous plants (cover > 10%), including grasslands (e.g., prairies, steppes, savanna, and mosaics with tree/shrubs), or natural shrub cover (>10%), including permanently or regularly flooded areas (wetlands) and (herbaceous) wetlands.	Non-forest	Shrubland
Sparse/no vegetation	Bare areas, sparse vegetation (2%–10%), snow and ice, rocks, sand, and mudflats.	Non-forest	Sparse vegetation
Water		Non-forest	Water

,

Table A4. MODIST1 reclassification, based on Sulla-Menashe et al., 2019 [16].

General Class	Description	Forest Mask Classification	IPCC Class
Evergreen needleleaf forests	Dominated by evergreen conifer trees (canopy > 2 m). Tree cover > 60%.	Forest	Forest
Evergreen broadleaf forests	Dominated by evergreen broadleaf and palmate trees (canopy > 2 m). Tree cover > 60%.	Forest	Forest
Deciduous needleleaf forests	Dominated by deciduous needleleaf (larch) trees (canopy > 2 m). Tree cover > 60%.	Forest	Forest
Deciduous broadleaf forests	Dominated by deciduous broadleaf trees (canopy > 2 m). Tree cover > 60%.	Forest	Forest
Mixed forests	Dominated by neither deciduous nor evergreen (40–60% of each) tree type (canopy > 2 m). Tree cover > 60%.	Forest	Forest
Closed shrublands	Dominated by woody perennials (1–2 m height), >60% cover.	Non-forest	Shrubland
Open shrublands	Dominated by woody perennials (1–2 m height), 10%–60% cover.	Non-forest	Shrubland
Woody savannas	Tree cover of 30%–60% (canopy > 2 m).	Forest	Forest
Savannas	Tree cover of 10%–30% (canopy > 2 m).	Forest	Forest
Grasslands	Dominated by herbaceous annuals (<2 m).	Non-forest	Grassland
Permanent wetlands	Permanently inundated lands with 30%–60% water cover and >10% vegetated cover.	Non-forest	Wetland
Croplands	At least 60% of area is cultivated cropland.	Non-forest	Agriculture
Urban and built-up lands	At least 30% impervious surface area, including building materials, asphalt, and vehicles.	Non-forest	Settlement
Cropland/natural vegetation Mosaics	Mosaics of small-scale cultivation of 40%–60%, with natural tree, shrub, or herbaceous vegetation.	Non-forest	Agriculture
Permanent snow and ice	At least 60% of area is covered by snow and ice for at least 10 months of the year.	Non-forest	Others
Barren	At least 60% of area is non-vegetated barren (sand, rock, and soil) areas with less than 10% vegetation.	Non-forest	Others
Water bodies	At least 60% of area is covered by permanent water bodies.	Non-forest	Water

,

Table A5. MODISP2 reclassification, based on Sulla-Menashe et al., 2019 [16].

General Class	Description	Forest Mask Classification	IPCC Class
Barren	At least 60% of area is non-vegetated barren (sand, rock, and soil) or permanent snow/ice, with less than 10% vegetation.	Non-forest	Others
Permanent snow and ice	At least 60% of area is covered by snow and ice for at least 10 months of the year.	Non-forest	Others
Water bodies	At least 60% of area is covered by permanent water bodies.	Non-forest	Water
Urban and built-up lands	At least 30% of area is made up of impervious surfaces, including building materials, asphalt, and vehicles.	Non-forest	Settlement
Dense forests	Tree cover > 60% (canopy > 2 m).	Forest	Forest
Open forests	Tree cover of 10–60% (canopy > 2 m).	Forest	Forest
Forest/cropland mosaics	Mosaics of small-scale cultivation 40%–60%, with >10% natural tree cover.	Non-forest	Agriculture
Natural herbaceous	Dominated by herbaceous annuals (<2 m). At least 10% cover.	Non-forest	Grassland
Natural herbaceous/cropland mosaics	Mosaics of small-scale cultivation 40%–60%, with natural shrub or herbaceous vegetation.	Non-forest	Agriculture
Herbaceous croplands	Dominated by herbaceous annuals (<2 m). At least 60% cover. Cultivated fraction >60%.	Non-forest	Agriculture
Shrublands	Shrub cover >60% (1–2 m).	Non-forest	Shrubland

,

Table A6. ESACCI reclassification, based on ESA 2017 [15].

Description	Forest Mask Classification	IPCC Class
Cropland rainfed	Non-forest	Agriculture
Cropland rainfed—herbaceous cover	Non-forest	Agriculture
Cropland rainfed—tree or shrub cover	Non-forest	Agriculture
Cropland irrigated or post-flooding	Non-forest	Agriculture
Mosaic cropland (>50%)/natural vegetation (tree/shrub/herbaceous cover) (<50%)	Non-forest	Agriculture
Mosaic natural vegetation (tree/shrub/herbaceous cover) (>50%)/cropland (<50%)	Non-forest	Agriculture
Tree cover—broadleaved evergreen, closed to open (>15%)	Forest	Forest
Tree cover—broadleaved deciduous, closed to open (>15%)	Forest	Forest
Tree cover—broadleaved deciduous, closed (>40%)	Forest	Forest
Tree cover—broadleaved deciduous, open (15%–40%)	Forest	Forest
Tree cover—needleleaved evergreen, closed to open (>15%)	Forest	Forest
Tree cover—needleleaved evergreen, closed (>40%)	Forest	Forest
Tree cover—needleleaved evergreen, open (15%–40%)	Forest	Forest
Tree cover—needleleaved deciduous, closed to open (>15%)	Forest	Forest
Tree cover—needleleaved deciduous, closed (>40%)	Forest	Forest
Tree cover—needleleaved deciduous, open (15%–40%)	Forest	Forest
Tree cover—mixed leaf type (broadleaved and needleleaved)	Forest	Forest
Mosaic tree and shrub (>50%)/herbaceous cover (<50%)	Forest	Forest
Mosaic herbaceous cover (>50%)/tree and shrub (<50%)	Non-forest	Grassland
Shrubland	Non-forest	Shrubland
Shrubland evergreen	Non-forest	Shrubland
Shrubland deciduous	Non-forest	Shrubland
Grassland	Non-forest	Grassland
Lichens and mosses	Non-forest	Sparse vegetation
Sparse vegetation (tree/shrub/herbaceous cover) (<15%)	Non-forest	Sparse vegetation
Sparse tree (<15%)	Non-forest	Sparse vegetation
Sparse shrub (<15%)	Non-forest	Sparse vegetation
Sparse herbaceous cover (<15%)	Non-forest	Sparse vegetation
Tree cover, flooded fresh or brackish water	Forest	Forest
Tree cover, flooded saline water	Forest	Forest
Shrub or herbaceous cover, flooded fresh/saline/brackish water	Non-forest	Wetland
Urban areas	Non-forest	Settlement
Bare areas	Non-forest	Others
Consolidated bare areas	Non-forest	Others
Unconsolidated bare areas	Non-forest	Others
Water bodies	Non-forest	Water
Permanent snow and ice	Non-forest	Others

,

Table A7. GLCLU reclassification, Potapov et al., 2022 [14].

General Class	Description	Forest Mask Classification	IPCC Class
Terra firma—true desert	3%–7% short vegetation cover	Non-forest	Sparse vegetation
Terra firma—semi-arid	11%–75% short vegetation cover	Non-forest	Grassland
Terra firma—short, dense vegetation	79%–100% short vegetation cover	Non-forest	Grassland
Terra firma—tree cover	3–5 m trees	Non-forest	Shrubland
Terra firma—tree cover	6–>25 m trees	Forest	Forest
Wetland—salt pan	3%–7% short vegetation cover	Non-forest	Wetland
Wetland—sparse vegetation	11%–75% short vegetation cover	Non-forest	Wetland
Wetland—short, dense vegetation	79%–100% short vegetation cover	Non-forest	Wetland
Wetland—tree cover	3–5 m trees	Non-forest	Wetland
Wetland—tree cover	6–>25 m trees	Forest	Forest
Open surface water	20%–100% of year	Non-forest	Water
Snow/ice	Snow/ice	Non-forest	Others
Cropland	Cropland	Non-forest	Agriculture
Built-up	Built-up	Non-forest	Settlement
Ocean	Ocean	Non-forest	Water

and

Table A8. TMF reclassification, Vancutsem et al., 2021 [19].

General Class	Description	Forest Mask Classification
Undisturbed tropical moist forest	Closed evergreen or semi-evergreen forest without any disturbance.	Forest
Degraded tropical moist forest	Closed evergreen or semi-evergreen forest (covered by existing or regrowing trees) that has been temporarily disturbed during a period of maximum 2.5 years.	Forest
Deforested land	Permanent conversion of forest into non-forested land that started, at the latest, in the current year. Disturbances were observed over more than 2.5 years, and no vegetative regrowth was detected.	Non-forest
Tropical moist forest regrowth	Pixel that has been deforested before the current year and that is currently regrowing. A minimum 3-year duration (2020–2022) of permanent moist forest-cover presence is needed to classify a pixel as forest regrowth (to avoid confusion with agriculture).	Forest
Permanent and seasonal water	This class includes permanent and seasonal water from the GWS dataset.	Non-forest
Other land cover	No data and non-TMF cover, which includes savanna, deciduous forest, agriculture, evergreen shrubland, non-vegetated cover, and afforestation.	Non-forest

). Thus, all the databases were reclassified according to forest (mask value of 1) and non-forest (mask value of 0) designations based on their original class definition (Figure 2(1a,2a)).

Each reclassified product was then resampled at lower spatial resolutions for comparison (Figure 2(1)). The spatial grids of low-resolution datasets (HILDA+, MODIST1/MODISP2, and ESACCI) were used as the reference for resampling (Figure 2(1b)). For each product, we then calculated the proportion, in percentage, of forest for each grid cell as the ratio of the sum of forest pixels over all pixels present in the grid cell (Figure 2(1c–e)). As GFW includes a continuous tree-cover percentage, mean percentage was calculated to keep the information. Finally, a new forest/non-forest mask was calculated again from the resampled maps following the FAO forest definition (tree cover > 10%; (Figure 2(2a))).

2.4. Spatiotemporal Analyses

We assessed the influence of homogenization and resampling by calculating the mean area of forest and non-forest for each product at each resolution.

We performed an intercomparison of each resampled and newly defined products to assess the agreement and disagreements between them. Products were compared two by two for each common year (2005, 2010, and 2015) to produce a confusion matrix, which is defined as follows: “NF/NF” was attributed if both products observed non-forest; “F/F” if both products observed forest; “F/NF” if the “reference” product observed forest but the “compared” product did not; and NF/F if the “compared” product observed forest but the “reference” product did not (Figure 2(2b)). We then calculated the agreement (Equation (1)) and disagreement (Equation (2)) rate based on this score for each pair of datasets, with agreement + disagreement = 100%.

$$Agreement={\displaystyle \frac{{Mean}_{NF/NF}+{Mean}_{F/F}}{{Mean}_{NF/NF}+{Mean}_{F/F}+{Mean}_{F/NF}+{Mean}_{NF/F}}}\times 100(\%)$$

(1)

$$Disagreement={\displaystyle \frac{{Mean}_{F/NF}+{Mean}_{NF/F}}{{Mean}_{NF/NF}+{Mean}_{F/F}+{Mean}_{F/NF}+{Mean}_{NF/F}}}\times 100(\%)$$

(2)

where Mean is the area averaged over the target years. Finally, we determined which land-cover (LC) classes were involved in the NF/F differences between the different pairs of databases. We grouped the LC classes of the different datasets into common classes based on the IPCC classification [29]. We defined 8 classes: agriculture, grassland, wetland, settlement, shrubland, sparse vegetation, water, and others. For most of these new classes, the class was already present in the previous definition. The class agriculture also included definitions mentioning “cropland” or “cultivation”. In the case of GLCLU, we defined shrublands as trees lower than 5 m, sparse vegetation as short vegetation with less than 10% of cover, and grassland as short vegetation with cover between 11 and 100% (Table A3, Table A4, Table A5, Table A6, Table A7 and Table A8). We calculated the mean area over the years and the percentage that each class represented.

The temporal evolution of the forest cover from the different LULC products used in this study was analyzed at their original resolution to quantify forest loss and gain. Most datasets cover the period from 2001 to 2019. However, as the GLCLU only provides information on changes over a five-year period, maps from 2005, 2010, and 2015 were used for this analysis. With GFW, we were only able to quantify forest loss, as the map does not provide information on reforestation, except in the form of a cumulative map covering 2001–2012. Where the maps allowed for calculation, we derived net changes in forest area at a yearly timescale. The cumulative sum of these net changes was then calculated to highlight the major trends in the evolution of forest cover for each dataset.

This work was carried out using Python version 3.13.5 and R Studio version 4.3.2.

3. Results

3.1. Impact of Data Homogenization and Resampling on Forest/Non-Forest Masks

3.1.1. Congo Basin

The datasets showed different spatial patterns of forest area (Figure 3). The lower the resolution, the higher the percentage of forest (Figure 4 and

Table 2. Forest and non-forest area percentage of all the datasets at each resolution and for each studied area.

	Congo Basin		Miombo		Moist forest		Savanna
Dataset	Non-Forest Percentage	Forest Percentage	Non-Forest Percentage	Forest Percentage	Non-Forest Percentage	Forest Percentage	Non-Forest Percentage	Forest Percentage	Resolution
HILDA+	30.16	69.84	47.89	52.11	8.54	91.46	56.04	43.96	1 km
MODIST1	3.16	96.84	1.96	98.04	1.01	98.99	5.70	94.30	1 km
MODISP2	2.75	97.25	1.70	98.30	0.53	99.47	5.62	94.38	1 km
ESACCI	8.20	91.80	10.88	89.12	5.29	94.71	9.46	90.54	1 km
GLCLU	11.49	88.51	20.51	79.49	2.13	97.87	21.75	78.25	1 km
GFW	0.95	99.05	1.39	98.61	0.35	99.65	0.68	99.32	1 km
TMF	27.91	72.09			2.71	97.29	42.74	57.26	1 km
MODIST1	5.29	94.71	3.60	96.40	1.95	98.05	9.69	90.31	500 m
MODISP2	4.77	95.23	3.30	96.70	1.34	98.66	9.53	90.47	500 m
ESACCI	13.93	86.07	18.46	81.54	9.25	90.75	17.12	82.88	500 m
GLCLU	14.58	85.42	25.46	74.54	3.40	96.60	27.05	72.95	500 m
GFW	1.18	98.82	1.63	98.37	0.56	99.44	0.91	99.09	500 m
TMF	30.96	69.04			4.27	95.73	49.31	50.69	500 m
ESACCI	17.02	82.98	22.20	77.80	11.30	88.70	21.55	78.45	300 m
GLCLU	16.09	83.91	27.79	72.21	4.02	95.98	29.63	70.37	300 m
GFW	1.30	98.70	1.76	98.24	0.67	99.33	1.03	98.97	300 m
TMF	32.51	67.49			5.02	94.98	52.65	47.35	300 m
GLCLU	25.90	74.10	44.63	55.37	8.13	91.87	45.42	54.58	30 m
GFW	3.63	96.37	3.90	96.10	2.90	97.10	3.50	96.50	30 m
TMF	41.40	58.60			9.94	90.06	69.54	30.46	30 m

). For example, GLCLU from 30 m to 1 km had forest proportion of 74.10% and 88.51%, respectively (Table 2). Datasets with the highest area of forest at their original resolution were MODIST1, with 94.71%; MODISP2, with 95.23%; and GFW, with 96.37% (Table 2). Datasets with the highest fraction of non-forest at their original resolution were GLCLU, with 25.9%; HILDA+, with 30.16%; and TMF, with 41.4% (Table 2).

3.1.2. Miombo Forest

Over the miombo, GFW, MODIST1, and MODISP2 detected forest cover of 96.1%, 96.4%, and 96.7%, respectively, at their original resolution (Table 2). HILDA+, GLCLU, and ESACCI provided intermediate results in terms of forest cover, with 52.1%, 55.4%, and 77.8%, respectively (Table 2). The major change with decreasing resolution was observed for GLCLU, with a forest gain of 24.1% between 30 m and 1 km (Table 2).

3.1.3. Moist Forest

All the datasets showed a very high forest cover, varying between 88.7% and 98.7%, at their original resolution (Table 2). ESACCI was the dataset with the lowest forest cover (88.7%), and TMF showed the largest increase in forest cover (7.2%) between the original resolution and the 1 km resolution (Table 2).

3.1.4. Savanna Forest

For the savanna ecoregion, the datasets with the highest proportion of forest were MODIST1, MODISP2, and GFW, with values of 90.3%, 90.5%, and 96.5%, respectively (Table 2). The datasets with lower forest cover were TMF, HILDA+, and GLCLU, with values of 30.5%, 44.0%, and 54.6%, respectively (Table 2). The change in resolution affected the proportion of forest, especially for TMF, which showed an increase of 26.8% between 30 m and 1 km resolution.

3.2. Forest/Non-Forest Agreement and Disagreement Between Products

3.2.1. The Congo Basin

For the Congo Basin, the agreement for all comparisons was between 60.5% and 99.5% (Figure 5a). The higher proportion of disagreement was observed between TMF and GFW, with 39.5% (Figure 5b), while best agreement was obtained between MODIST1 and MODISP2, with 99.48% (Figure 5a).

3.2.2. Miombo Forest

In the case of the miombo, agreement ranged from 53.83% to 99.7% across all comparisons (Figure 5c). Disagreement was high for HILDA+, at 45.62%, 45.85%, and 46.17%, when it was compared to MODIST1, MODISP2, and GFW, respectively (Figure 5d). Meanwhile, the best agreement was obtained between MODIST1 and MODISP2, with 99.7% (Figure 5c).

3.2.3. Moist Forest

For moist forest, there were few discrepancies between the datasets. Agreement was higher than 88.2% for all comparisons (Figure 5e). The higher discrepancy rate was observed between ESACCI and GLCLU or TMF, with 11.0% and 11.8%. (Figure 5f), while the best agreement was obtained between MODIST1 and MODISP2, with 99.38% (Figure 5e).

3.2.4. Savanna Forest

Agreement across the savanna ranged from 32.8% to 99.8% (Figure 5g). Higher discrepancies were observed at 1 km resolution for the majority of the datasets, ranging from 34.7% to 55.3% (Figure 5h). Only TMF showed a higher disagreement rate (67.2%) at 30 m resolution when it was compared to GFW (Figure 5h). Meanwhile, the best agreement was obtained between MODIST1 and MODISP2, with 99.84% (Figure 5g).

3.3. Identification of Land Classes Implicated in the Disagreements

The analysis of the LULC classes implicated in the NF/F differences showed different results depending on the ecoregions and the datasets.

3.3.1. Congo Basin

In the case of the Congo Basin, in terms of NF/F disagreement, HILDA+ detected mainly “shrubland” (up to 77.8%) when the second datasets detected forest (Figure 6a). MODIST1 detected mainly “grassland” (up to 79.3%) and “wetland” (up to 100%; Figure 6a). For MODISP2, the differences were mostly “grassland” (up to 92.7% for ESACCI; Figure 6a). The ESACCI dataset observed mostly “agriculture” (between 61.8% and 83.5%) instead of forest in the other datasets (Figure 6a). When TMF observed forest, GLCLU detected mostly “shrubland” (39.0%); for GFW, GLCLU detected mostly “grassland” (74.7%; Figure 6a). For TMF, the NF/F differences were related to the “deforested” class (up to 14%) and the “other land cover” class (up to 91.8%; Figure 6a).

3.3.2. Miombo Forest

In the case of miombo, HILDA+ detected mostly “shrubland” (up to 85.1%) where the second datasets observed forest (Figure 6b). Two LULC classes were involved in the NF/F differences for MODIST1 (Figure 6b): “wetland”, which accounted for between 4.5% and 100% of the differences (Figure 6b), and “grassland”, which accounted for up to 79.9% of the differences (Figure 6b). MODISP2 showed disagreement mostly for “grassland”, from 70.0% to 92.6%, when the other datasets showed disagreement for forest (Figure 6b). For ESACCI, when compared to GLCLU or GFW, most of the NF/F differences were due to the “shrubland” class (up to 30.1%), “wetland” class (up to 32.7%), or “agriculture” class (up to 34.6%; Figure 6b). When comparing GLCLU and GFW, GLCLU detected a majority of “grassland” (74.3%) when GFW observed forest (Figure 6b).

3.3.3. Moist Forest

For moist forest, HILDA+ detected mostly “shrubland” (up to 68.4%). “Wetland” was implicated in the NF/F differences when MODIST1 was the first dataset (Figure 6c), with up to 100%. MODISP2 showed disagreement mostly for “grassland”, from 76.5% to 93.8%, when the other datasets showed disagreement for forest (Figure 6c). ESACCI detected “agriculture” (>88.1%) most of the time when the other datasets detected forest (Figure 6c). In the case of TMF, the classes involved in the NF/F differences are “deforested land” (19.6%–46.7%) and “other land cover” (53.2%–79.8%; Figure 6c).

3.3.4. Savanna Forest

Across the savanna ecoregion, HILDA+ observed mostly “shrubland” (59.6%–80.5%), while the other datasets observed forest (Figure 6d). MODIST1 and MODISP2 showed similar results, with a majority of “grassland” (up to 95.9%; Figure 6d). ESACCI detected mainly “agriculture” (51%–68.0%; Figure 6d). For GLCLU, the result was split between “shrubland” (14.1%–44.2%) and “grassland” (35.1%–80.5%; Figure 6d). TMF showed a small part of “deforested land” (8.1%–15.6%) and mainly “other land cover” (84.4%–91.8%; Figure 6d).

3.4. Temporal Evolution of the Forest for Each Dataset

3.4.1. The Congo Basin

Figure 7 shows the deforestation and reforestation maps for each dataset from 2001 to 2019. GFW, TMF, GLCLU, and HILDA+ all exhibited similar patterns, highlighting large areas of deforestation in the DRC (Figure 7a, Figure 7c, Figure 7e, and Figure 7g, respectively). However, the intensity and spatial patterns of deforestation varied between datasets. Higher deforestation intensity was observed for HILDA+, with hotspots in the south, northwest, and east of the DRC (Figure 7g); and for GFW, higher deforestation intensity was observed in the north and center of the DRC, north of the Congo, Gabon, and south of Cameroon (Figure 7a). Lower intensity was observed for TMF in the northwest and southwest (Figure 7c), and the lowest intensity was observed for GLCLU, with a similar pattern to TMF (Figure 7e). ESACCI exhibited areas of deforestation of a similar intensity to GLCLU across most of the DRC, except in the southeast, northeast, and southwest, where an increase in forest area was observed. Loss/gain in forest area was also observed in the north and south of Congo (Figure 7b).

Similar patterns of forest-cover change were obtained using MODIST1 and MODISP2 (see Figure 7d and Figure 7f, respectively), with large areas of reforestation and multiple changes observed in the southern part of the area and along the Congo River and its tributaries. Figure A6 illustrates these differences in the DRC. In the south of the DRC, similar patterns of deforestation at different spatial resolutions are obtained using HILDA+, GFW, TMF, and GLCLU, while ESACCI presents a pattern of both deforestation and reforestation. The two MODIS-based products exhibit a completely different spatial distribution of changes, and this distribution seems to be partly related to flood dynamics.

When the temporal evolution was considered, two different behaviors emerged among the datasets. The first behavior corresponded to datasets that exhibited a net forest gain by the end of the study period (i.e., around 2019). The second behavior corresponded to datasets showing a net forest loss by the end of the period. The ESACCI, MODISP2, and MODIST1 products detected a net forest gain ranging from 7063 km² (ESACCI) to 50,726 km² (MODIST1) by the end of the period (Figure 8a). In contrast, the HILDA+, GFW, TMF, and GLCLU products detected a net forest loss ranging from 70,601 km² (TMF) to 205,704 km² (HILDA+; Figure 8a).

3.4.2. Miombo Forest

Over the miombo region, the MODISP2 and MODIST1 datasets observed net forest gains of 439 km² and 2248 km², respectively (Figure 8b). All of the other datasets observed net forest loss (Figure 8b). Notably, HILDA+ revealed a total net forest loss of 80,884 km² between 2001 and 2019 (Figure 8b), with significant variability in the intensity of this loss across the years.

3.4.3. Moist Forest

Over moist forests, MODISP2 and MODIST1 recorded net forest gains, peaking in 2003 at 4684 km² and in 2007 at 12,505 km², respectively (Figure 8c). The other datasets observed net forest loss over the period (Figure 8c). Notably, GFW showed a total net loss of 97,218 km² (Figure 8c) by the end of the timeline, with an acceleration of deforestation between 2013 and 2019.

3.4.4. Savanna Forest

Three datasets recorded a net forest gain over the considered period in savanna forests: ESACCI (up to 23,451 km²), MODISP2 (up to 36,479 km² between 2001 and 2014), and MODIST1 (up to 38,580 km² between 2001 and 2014). The dynamics were different between ESACCI and the MODIS products, with the latter showing net gain in forest cover for some years. TMF, HILDA+, and GFW showed net forest loss over the same period (Figure 8d). The same dynamic was observed for HILDA+ and GFW datasets, which showed net forest loss throughout the period. In contrast, TMF observed net forest gain between 2017 and 2019 (Figure 8d).

4. Discussion

The comparison of forest/non-forest masks required a common definition of forest. In this study, we matched the FAO definition, which considers a pixel to be covered with forest when (i) having a tree cover of more than 10% and (ii) a tree height of more than 5 m [30]. This definition has the advantage of being based on two criteria that are regularly present in the definition of LULC classes in the databases that were used in this study. The two criteria were respected as much as possible, but the first reclassification could not meet all the requirements (Table A3, Table A4, Table A5, Table A6, Table A7 and Table A8). This leads to differences in the forest area between the datasets. Considering all the resolutions over all the studied regions, GFW, MODIST1, and MODISP2 presented a larger forest area than the other dataset. In the case of GFW, this could be explained by an overestimation of tree cover. A lack of distinction between tropical forest, plantations, and herbaceous cultivation in this product was reported over tropical regions [31]. Furthermore, the tree-cover map is only available for the year 2000. It is possible that other changes (degradation by selective logging or climate events) which occurred between 2000 and 2015 were not included in the analysis. This difference could also be explained by the 10% threshold used to define forests. A previous study showed that the value of 70% tree cover was closer to reality than a value of 30% for Gabon [32]. The choice of the tree-cover threshold for forest definition, in the case of the GFW map, is crucial and must be taken into account in the analysis. The larger forest area of MODIST1 and MODISP2 could be explained by the low resolution of 500 m. Since forest disturbances in this region are smaller than in other parts of the world, areas where there is non-forest could be missed due to the low resolution. Furthermore, in their definition of forest, the MODIS products include trees higher than 2 m. This leads us to consider a larger proportion of the savanna as forest, given that the trees are shorter than those in the moist forest or miombo woodlands. Observations made in tropical Asian forests reported that MODIS land-cover products are not able to distinguish plantations from natural forests [33], and this fact could account for the probable overestimation of forest area also in Africa. The datasets which presented the smaller forest area for the studied Congo Basin region were HILDA+, ESACCI, GLCLU, and TMF. Concerning HILDA+, an explanation could be found in the tree-cover threshold of 25%, which is associated with its low resolution and the lack of statistical information on Africa. Regarding ESACCI, the result could be due to the dataset’s difficulty in assigning the mosaic classes to the forest/non-forest masks and the use of satellites with different resolutions [34]. GLCLU also presented a smaller forest area in the majority of the ecoregions. Probably due to the definition of a forest mask only based on tree height, savanna with a low tree height was classified as non-forest. Even if GEDI offered more precision regarding the tree height, it still had limitations. Notably, in case of high slope and dense forest canopy [35,36], which can explain part of the results. The TMF showed a low forest area over all of the Congo Basin (58.6% at 30 m resolution) and savanna (30.46% at 30 m resolution). This is most likely due to the fact that TMF provides information on tropical moist forests, which represent a smaller proportion of forest cover in the savanna and miombo regions (generally over 50% in the other products in these two regions).

Most of the differences in forest area could be explained by the definition of forest. These differences in definition lead to difficulties in reclassification and are likely to account for most of the differences between the databases. The reclassification of forests in GLCLU is based on tree height (>5 m), while that of GFW is based on tree cover (>10%). At the same resolution, the datasets disagreed by 24.57% over the Congo Basin and by 44.01% over savanna forest. This contrasts with the case of MODIST1 and MODISP2, which exhibited a disagreement of 0.52% at their original resolution over the Congo Basin (Figure 5). Their reclassification is very similar, based on both tree cover (>10%) and tree height (>2 m). The only difference lies in the definition of “permanent wetland” in MODIST1, which includes vegetation cover of more than 10%, without specifying the type of vegetation (Figure 6 and Table A4). These results demonstrate that when the same definition of forest is used, including both tree cover and height, the consistency between the datasets can be high. Our work highlights the urgent need for a common definition of forests in LULC classifications to facilitate the use and the interpretation between products. Two pieces of information need to be included in the definition of the classes made up of trees: tree cover and tree height. Sannier et al. [32] reported the influence of the region on the choice of the tree-cover threshold. A threshold adapted to the forest type could allow for a better definition of the forest/non-forest mask [32]. As shown by the MODIS products and the GLCLU over the savanna, the threshold for tree height also impacts the results (Table 2, Table A4, Table A5 and Table A7). By incorporating tree cover and height into the definition of future maps, users can select an appropriate threshold for their needs and the ecoregion of interest. The choice of resolution also affects the estimated area of forest/non-forest. This area varied due to the resampling, regardless of the datasets. The lower the resolution, the greater the amount of forest, regardless of the ecoregion considered. For example, GLCLU presented forest areas of 54.58% and 91.87% at a resolution of 30 m for savanna and moist forest, respectively. At 1 km, these percentages increased to 78.25% for savanna and 97.87% for moist forest. This result shows that a low resolution can lead to an overestimation of forest cover. Resolution also has an impact on the detection of land-cover change. Deforested areas in Africa are small and difficult to detect using remote-sensing techniques [20]. It is essential to work at a high resolution when it comes to Africa and use satellites with better spatial resolution [37].

The LULC classes mostly implicated in the disagreement between the datasets were “agriculture”, “grassland”, and “shrubland”. This result highlights the difficulties faced by remote sensing when making distinctions between different vegetation types, notably in tropical regions. These results could be explained by the signal saturation of remote-sensing data. In the case of dense vegetation or high biomass, the optical or Synthetic Aperture Radar (SAR) signals may saturate [38,39], leading to a lack of change detection or differentiation between different types of vegetation [38]. The study of vegetation phenology can be helpful for vegetation-type discrimination [40]. The ecoregions with the higher level of disagreement between datasets were miombo and savanna, which presented the higher seasonality, highlighting the importance of monitoring throughout the year. However, given the high cloud cover in the tropics, optical remote sensing cannot offer a sufficient number of cloud-free images to meet the requirements of continuous monitoring [10]. This limitation could be overcome by adding other sensors to create complementarity between data sources. Optical remote sensing provides information on vegetation photosynthesis activity, while Light Detection and Ranging (LiDAR) and Synthetic Aperture Radar (SAR) offer new, relevant insights into forest structure and vegetation water content [41,42,43,44]. Furthermore, SAR sensors are less sensitive to saturation and can penetrate cloud cover, thus increasing the number of images acquired per year. Recent forest cover-monitoring products, which are not included in this study, as they do not cover the same period, are based either on SAR images or on the fusion of SAR and multispectral images [42,45]. They were found to have a higher level of accuracy than products based solely on multispectral information [46] and to identify deforestation sooner, as well as to be able to detect both deforestation and degradation in equatorial areas [42,44]. Forest-height products based on merging multispectral and SAR images with LiDAR-based tree heights are likely to provide unique information on forest cover and regrowth [47,48,49].

Temporal analysis of the forest cover revealed that datasets with low resolution tend to show a net gain across the study region and period (MODIST1, MODISP2, and ESACCI). This is probably due to the sensors’ resolution not allowing small changes to be detected. The only exception is HILDA+, which has a resolution of 1 km and tends to observe a net loss of forest over the period. This could be explained by the algorithm used to generate the HILDA+ product, which took into account multiple remote-sensing datasets, as well as statistical data from the FAO. Since this dataset is based on statistics, it is less impacted by the above-mentioned image-processing issues and took advantage of multiple sources of information to provide a robust estimation [17]. Furthermore, this dataset has already been shown to demonstrate good accuracy over the tropical forests of Ecuador compared to TMF, even with a resolution of 1 km [50]. The forest gain was most significant in the savanna ecoregion, with a net gain of 36,827 km² at the end of the period for MODIST1. It is well known that the woody cover of the savanna in this region is highly unstable due to the mean annual precipitation [51]. This ecosystem characteristic, coupled with the low resolution of the sensors and the limited availability of optical data, could also account for the large alternance between forest gain and loss from one year to another observed in the products. GFW observed higher rates of forest loss over the moist ecoregion, with a net loss of 97,217 km² by the end of the period. The high resolution of Landsat (30 m) enables the detection of small-scale deforestation and degradation, which is typically observed in tropical Africa. Some datasets recorded gains and losses in the same region over the study period. This is notably the case for MODIS products (Figure 7, Figure A6 and Figure A7). Despite the incorporation of the hidden Markov model (HMM) in Collection 6 to stabilize classification over time, some pixels appear to be subject to high dynamics [16]. Areas of 11,746 km² and 12,045 km² were detected as having experienced at least one loss-and-gain event for MODISP2 and MODIST1, respectively (Figure A7). Analysis of the classes at these locations showed that they were mostly “savanna” and “grassland” for MODIST1, and “open forest” and “natural herbaceous” for MODISP2 (Figure A8 and Figure A9). These classes have very similar definitions in terms of tree cover or height, and these similarities in definitions could explain why a small difference in the signal could result in a different class attribution (Table A4 and Table A5).

5. Conclusions

In this study, seven of the most-used forest-cover datasets (HILDA+, MODIST1, MODISP2, ESACCI, GLCLU, GFW, and TMF) were compared across different ecoregions of the Congo Basin. Our goal was not only to provide an intercomparison but also to identify in which classes they differ. This required homogenization of the forest definition in order to resample the datasets at the same spatial resolution. In this context, a forest was defined as an area with tree cover of more than 10% and trees of more than 5 m in height. However, information on these two parameters is not always available in the class definition. Through this methodology, we have emphasized the importance of these two criteria and of taking the threshold into account, depending on the ecoregion being studied. The change in resolution influenced the forest area, with a probable overestimation at coarser resolution. For example, the GLCLU forest cover increased from 74.10% to 88.51% at 30 m and 1 km resolution, respectively. This result can be explained by the small areas of deforestation in Africa that are not detected at low resolution. Comparison of the different datasets across ecoregions showed higher levels of inconsistency for miombo (around 24.3%) and savanna (around 32%). Further analysis showed that the majority of LULC classes involved in the differences between datasets across ecoregions were related to “shrubland” (up to 77.8%), “agriculture” (up to 83.5%), and “grassland” (up to 92.7%). This illustrates the difficulty in distinguishing vegetation types using remote-sensing data. These results could be explained by the use of optical sensors, which provide few images per year due to cloud cover. This prevents the monitoring of phenology, which is a key information for vegetation type. The inclusion of new information from SAR images and LiDAR canopy heights could improve the results of future maps and provide a new understanding of African forest dynamics. For now, the presented datasets exhibited different behavior with regard to the evolution of forest cover. While other datasets observed a net loss of up to 205,704 km², MODIS products and ESACCI exhibited a net gain of up to 50,726 km² during the period of 2001–2019. This trend was apparent across all ecoregions, reflecting inconsistencies between the datasets. This work demonstrated that different conclusions about the evolution of forest cover can be drawn depending on the product used. Even without ground-truth data, the results concerning forest cover dynamics and the evaluation of inconsistencies between these products can provide valuable insights for conservation planning. This study quantifies the uncertainty and confidence that can be attributed to each product depending on the ecoregion, helping to select the most suitable land-use and land-cover (LULC) maps.