Annual forest cover maps in Africa during 2015&ndash;2023 by analyses of PALSAR-2, Landsat, and GEDI LiDAR datasets with knowledge-based algorithms

Yao, Yuan; Xiao, Xiangming; Zhang, Chenchen; Tang, Hao; Qin, Yuanwei; Meng, Cheng; Pan, Baihong; Pan, Li; Wang, Jie

doi:10.5194/essd-2026-269

Preprints

https://doi.org/10.5194/essd-2026-269

Preprints

24 Jun 2026

| 24 Jun 2026

Status: this preprint is currently under review for the journal ESSD.

Annual forest cover maps in Africa during 2015–2023 by analyses of PALSAR-2, Landsat, and GEDI LiDAR datasets with knowledge-based algorithms

Yuan Yao, Xiangming Xiao, Chenchen Zhang, Hao Tang, Yuanwei Qin, Cheng Meng, Baihong Pan, Li Pan, and Jie Wang

Abstract. According to the Food and Agriculture Organization of the United Nations (FAO) Global Forest Resources Assessment (FRA) 2020 report, Africa has the highest annual net forest loss rate worldwide during 2010–2020, approximately 50 % higher than that of South America. Multiple high-resolution forest cover data products derived from optical and/or microwave remote sensing data show large discrepancies in forest area estimates and spatial distribution in Africa. To date, few studies have evaluated these datasets using the FAO forest definition and consistent assessment data. Here, we generate annual forest cover maps in Africa at 30 m resolution from 2015 to 2023 using Phased Array type L-band Synthetic Aperture Radar-2 (PALSAR-2), Landsat imagery, and knowledge-based algorithms. We compare the resulting PALSAR-2/Landsat forest/non-forest (FNF) maps with four widely used forest datasets: (a) Landsat tree canopy cover from Global Forest Watch (Landsat-GFW), (b) PALSAR/PALSAR-2 Forest/Non-Forest Map from the Japan Aerospace Exploration Agency (JAXA FNF4), (c) global map of forest cover 2020 from the European Commission Joint Research Centre (JRC GFC2020 v2), and (d) the FAO FRA 2020 forest statistics. Using the criteria of FAO forest definition (canopy height > 5 m; canopy cover > 10 %), we assess four satellite-based forest cover products for 2020 using canopy height and canopy cover measurements derived from spaceborne light detection and ranging (LiDAR) measurements in the NASA Global Ecosystem Dynamics Investigation (GEDI) mission. We find that PALSAR-2/Landsat FNF, JAXA FNF4, and JRC GFC2020 v2 have high and consistent overall accuracy (OA; approximately 90 %), whereas Landsat-GFW (tree cover > 10 %) has substantially lower accuracy (69 %). Forest area estimates for 2020 from PALSAR-2/Landsat (8.8 × 10⁶ km²), JAXA FNF4 (8.9 × 10⁶ km²), and JRC GFC2020 v2 (7.6 × 10⁶ km²) are larger than the FRA 2020 statistics (6.4 × 10⁶ km²). Forest area and spatial distribution from PALSAR-2/Landsat FNF are most consistent with those from JAXA FNF4, followed by JRC GFC2020 v2. Landsat-GFW (16.7 × 10⁶ km²) differs substantially from the other three products. Our annual forest cover maps complement FRA reporting and support a better understanding of the magnitude, dynamics, and drivers of forest gain and loss across Africa.

Received: 09 Apr 2026 – Discussion started: 24 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 2132 KB)

Supplement (1975 KB)

Download & links

Yuan Yao, Xiangming Xiao, Chenchen Zhang, Hao Tang, Yuanwei Qin, Cheng Meng, Baihong Pan, Li Pan, and Jie Wang

Status: final response (author comments only)

Subscribe to comment alert

RC1:
'Comment on essd-2026-269', Anonymous Referee #1, 21 Jul 2026
This manuscript presents annual 30 m forest cover maps of Africa from 2015 to 2023 derived from PALSAR-2 backscatter and Landsat NDVI, and evaluates four satellite forest products against GEDI canopy height and cover under the FAO forest definition. The topic is well suited to ESSD, the continental coverage and nine year time series are valuable, the data and code are openly archived, and the use of a very large GEDI sample to build complete confusion matrices is a meaningful improvement over previous one sided assessments. However, several issues concerning the independence of the reference data, the internal consistency of the accuracy table, the validity of the time series, and the fairness of the product comparison need to be resolved before the accuracy claims can be regarded as robust. Overall, I have to recommend a reject.

Lines 227 to 230, 380 to 387: JRC GFC2020 v2 incorporates the GEDI and Landsat derived global canopy height layer among its inputs. Evaluating it against GEDI canopy height therefore does not constitute an independent assessment, and its leading overall accuracy may partly reflect shared input data. This dependency should be stated and its effect on the ranking of products discussed.

Lines 249 to 252, 293 to 296, Table 1: The number of footprints meeting the FAO definition and the number of footprints falling within the mapped forest class of PALSAR-2/Landsat are reported as identical, and the two figures are used interchangeably in the text describing Figure 2b. As presented, the row totals of the confusion matrix and the reference class totals cannot both be correct. Please verify the counts and clarify which quantity each column and row represents.

Lines 253 to 261, 356 to 363: The dataset spans nine years, but the assessment uses GEDI data from 2020 alone. No evidence is provided that the earlier and later maps achieve comparable accuracy, yet interannual forest area values and pixel frequency statistics are reported and interpreted. Either extend the assessment to additional years using available GEDI acquisitions, or state explicitly that the time series itself remains unvalidated.

Lines 165 to 178, 356 to 358: The manuscript documents substantial year to year shifts in PALSAR-2 acquisition months and in Landsat good observation counts across regions. Because the decision rules use fixed backscatter and NDVI thresholds, these shifts can produce apparent forest gain or loss, particularly in seasonally dry savanna. The reported annual area range should be evaluated against this source of variability before it is presented as a change signal.

Lines 238 to 247: The three year window rule that rewrites improbable sequences alters the maps for 2016 to 2022 but is applied without justification of the assumption that single year transitions are implausible, and without any comparison of accuracy before and after correction. In regions with fire, shifting cultivation, or rapid regrowth, genuine single year changes will be removed. Please quantify how many pixels were altered and demonstrate that the procedure improves rather than degrades accuracy. The period given on line 244 also appears to be a typo?

Lines 222 to 226, 288 to 292, 477 to 480: Tree canopy cover exceeding ten percent is not a forest map under the FAO definition, since it excludes no height criterion, no minimum mapping unit, and no land use requirement. Reporting its low accuracy alongside the other products invites the reader to interpret a definitional mismatch as a mapping error. The alternative threshold tested in the Discussion should be integrated into the main comparison, and the framing of this product adjusted accordingly.
Citation: https://doi.org/10.5194/essd-2026-269-RC1
RC2: 'Comment on essd-2026-269', Frederic Achard, 25 Aug 2026

General Comments:
This paper presents an interesting approach regarding the mapping of tree cover in Africa around year 2020 from the use of two datasets of satellite imagery (PALSAR-2 and Landsat) along period 2015-2023 and a GEDI LiDAR dataset for year 2020.
Two main issues need attention before publication: (1) the thematic scope of the paper need to be corrected (mapping tree cover instead of forest cover) and (2) the temporal scope of the paper need to be revisited (the study does not seem to provide robust valid information on changes in tree cover).

Specific Comments:
The authors “generate annual forest cover maps in Africa at 30 m resolution from 2015 to 2023”. However they use only the structural parameters (i.e. tree cover percentage and tree height) of the FAO definition of forest. They do not consider the land use criteria of FAO definition.
Ref: FAO, 2023 FRA 2025 Terms and Definitions
“Land spanning more than 0.5 hectares with trees higher than 5 meters and a canopy cover of more than 10 percent, or trees able to reach these thresholds in situ. It does not include land that is predominantly under agricultural or urban land use.”
The FAO definition of forest excludes “land that is predominantly under agricultural or urban land use”. In particular it excludes tree plantations such as cocoa or oil palm plantations. The maps that are generated by the authors is more corresponding to a tree cover map (trees > 5m) than a forest map as defined by FAO. This can explain partly the differences with datasets such as the JRC GFC2020 map and the FAO FRA 2020 statistics. In particular in regions with large areas of tree crop plantations such as in Ivory Coast/Ghana as illustrated in site c of Figure 6 where most of the land is covered by cocoa plantations that appear as tree cover in PALSAR-2/Landsat, JAXA FNF4 and Landsat-GFW datasets whereas the JRC GFC2020 v2 displays only forest cover.
See in particular explanation of difference between forest and tree cover from WRI GFW blog: https://globalnaturewatch.org/blog/data-and-tools/global-forest-watch-and-the-forest-resources-assessment-explained-in-5-graphics-2/

The authors claim that their “annual forest cover maps complement FRA reporting and support a better understanding of the magnitude, dynamics, and drivers of forest gain and loss across Africa”. This statement needs to be revisited as, to my view, the study does not provide a robust assessment of changes of forest cover (gain or losses). The differences between annual maps can be due to a combination of (i) variability / error range in the individual annual maps, (ii) transition from forest to tree crops plantations (loss then gain) and (iii) changes in forest cover (mostly losses). To my view the ‘dynamic’ between annual maps (as displayed in Figure 4) is mainly due to (i), partly to (ii) and only for a very limited extent to (iii). A comparison with FAO national data on changes would be useful to indicate if the outcomes of the study could be meaningful. The authors are reporting an increase of forest (tree cover) of +0.3 10⁶ km² for Africa from 2016 to 2022 (from 8.6 × 10⁶ km² in 2016 to 8.9 × 10⁶ km² in 2022) whereas FAO FRA 2025 is reporting a loss of forest of c. 30 million ha (-0.3 10⁶ km²) for Africa over period 2015-2025. The TMF dataset is also reporting a decrease of tropical humid forests at -0.016 10⁶ km² / year over period 2015-2019 (in Vancutsem et al that can be updated to 2024 from https://forobs.jrc.ec.europa.eu/TMF/data#stats).
This opposite trend between the study results (positive trend) and FAO or TMF outcomes would need to be supported (through a robust accuracy assessment) or at least explained (as related to error range in the annual maps).
The Accuracy assessment using GEDI footprints (section 2.3.2) is very robust and valid for the assessment of the map of year 2020 (tree cover) but does not provide information on the accuracy of changes between years nor on the dynamics. A specific accuracy assessment (from a statistical sample of reference plots) would be needed to assess the accuracy of annual changes. See Olofsson et al (2014) for good practices and Vancutsem et al (2021) as example of accuracy assessment of changes (disturbances)
References:
Olofsson P et al. Good practices for estimating area and assessing accuracy of land change. Remote Sens. Environ. 148, 42–57 (2014).
Vancutsem et al., Long-term (1990–2019) monitoring of forest cover changes in the humid tropics. Sci. Adv. 2021; 7 : eabe1603;

Technical corrections
In introduction
Previous work regarding land cover mapping in Africa to be mentioned:
Mayaux, P., De Grandi, G.F., Rauste, Y., Simard, M., and Saatchi, S., 2002. Large Scale Vegetation Maps Derived from the Combined L-band GRFM and C-band CAMP Wide Area Radar Mosaics of Central Africa, International Journal of Remote Sensing, 23: 1261-1282.
Mayaux, P et al, 2003. A Land-cover Map of Africa . European Commission, EUR 20665 EN
In section 2.2.4 (Other forest products)
Use of FAO forest statistics: update the statistics of year 2020 from FAO FRA 2020 with statistics (of same year 2020) from FAO FRA 2025 (FAO. 2025. Global Forest Resources Assessment 2025. Rome. https://doi.org/10.4060/cd6709en)
Add reference to JRC Tropical Moist Forest Transition Map (Vancutsem et al 2021) when describing JRC GFC2020 v2 dataset
In Figure 6:
Colours of panels a5 and b5 to be verified (green and black seem inverted)
Order in legend to be corrected as “ The subsequent columns present forest cover maps from PALSAR-2/Landsat, JAXA FNF4, JRC GFC2020 v2 and Landsat-GFW”

Citation: https://doi.org/10.5194/essd-2026-269-RC2

Yuan Yao, Xiangming Xiao, Chenchen Zhang, Hao Tang, Yuanwei Qin, Cheng Meng, Baihong Pan, Li Pan, and Jie Wang

Supplement

https://doi.org/10.5194/essd-2026-269-supplement

Data sets

Annual forest cover maps for Africa from 2015 to 2023 at 30 m spatial resolution by analyses of PALSAR-2 and Landsat datasets Y. Yao et al. https://doi.org/10.5281/zenodo.19464248

Model code and software

The Google Earth Engine code for generating annual foerst cover map in Africa Y. Yao et al. https://doi.org/10.5281/zenodo.19477789

Yuan Yao, Xiangming Xiao, Chenchen Zhang, Hao Tang, Yuanwei Qin, Cheng Meng, Baihong Pan, Li Pan, and Jie Wang

Viewed

Total article views: 230 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
141	75	14	230	24	15	22

HTML: 141
PDF: 75
XML: 14
Total: 230
Supplement: 24
BibTeX: 15
EndNote: 22

Views and downloads (calculated since 24 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	1	0	1
Jul 2026	90	45	5	140
Aug 2026	50	30	9	89
Sep 2026	0

Cumulative views and downloads (calculated since 24 Jun 2026)

Month	HTML	PDF	XML	Total
Jun 2026	1	0	1
Jul 2026	90	45	5	140
Aug 2026	50	30	9	89
Sep 2026	0

Viewed (geographical distribution)

Total article views: 184 (including HTML, PDF, and XML) Thereof 184 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 03 Sep 2026

Download

Preprint (2132 KB)
Metadata XML

Short summary

Africa has experienced rapid forest loss, yet existing maps often disagree. Here, we produce annual forest cover maps at 30 m resolution from 2015 to 2023 using multiple satellite data and expert knowledge and compare them with other forest datasets. We find that forest area estimates of satellite-based products are larger than the Global Forest Resource Assessment statistics. Our results improve confidence in forest monitoring and support better understanding of forest change across Africa.


Total:	0
HTML:	0
PDF:	0
XML:	0