PeatCover: Towards a peat mapping at 90 m using satellite remote sensing and a priori inventories

Chen, Man; Ciais, Philippe; Hugelius, Gustaf; Aires, Filipe

doi:10.5194/essd-2026-254

Preprints

https://doi.org/10.5194/essd-2026-254

Preprints

17 Jul 2026

| 17 Jul 2026

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

PeatCover: Towards a peat mapping at 90 m using satellite remote sensing and a priori inventories

Man Chen, Philippe Ciais, Gustaf Hugelius, and Filipe Aires

Abstract. Peatlands cover just 3–4 % of Earth's land surface, yet store an estimated 600–700 Pg carbon (PgC), approximately one-third of Earth's soil carbon, making them critical regulators of the global carbon cycle. However, their spatial extent remains highly uncertain, particularly at fine spatial scales and in data-sparse regions. Existing global peatland datasets rely on inventories and regional products of heterogeneous quality, leading to inconsistencies and uncertainties in both estimated peat area and its spatial distribution. These limitations hinder accurate assessments of peatland-climate feedbacks, carbon budgets, policy development and restoration efforts. Here we propose a machine learning framework to map Northern peatland (boreal and subarctic) at a spatial resolution of 90 m that combines a priori information from existing peat databases (PEATMAP, Global Peatland Database, and CORINE Land Cover) with satellite observations (Landsat optical indices, topographic and hydrological data). The model first generates a continuous Peatland Index (PI) at 3 arc-second (~90 m) resolution, that can be combined with user-defined thresholds to obtain a binary peat classification. In regions with reliable coarse resolution peat cover information, the PI is used to downscale this peat fraction and obtain a high resolution classification. Our peatland map at 90 m was evaluated through both quantitative and qualitative approaches. Fully independent validation using the Peat-DBase field dataset (over 180,000 peat and non-peat observations) demonstrates an overall accuracy of 69.0 % and an F1-score of 0.81. Regional comparisons with published maps shows an overall accuracy of 69.9 % (F1=0.82) in Eurasia and of 63.9 % (F1=0.74) in North America. Qualitative spatial evaluation against Google Earth map across multiple case-study regions reveals that our map successfully captures fine-scale spatial details absent in existing inventories, including explicit delineation of open water bodies, river networks, and topographic constraints on peatland distribution. This work provides a spatially coherent, high-resolution peatland dataset spanning the Northern Hemisphere, offering improved capabilities for carbon stock estimation, hydrological modeling, and monitoring peatland degradation. The PeatCover dataset is available at https://doi.org/10.5281/zenodo.21031869 (Chen et al., 2026).

Received: 09 Apr 2026 – Discussion started: 17 Jul 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

Man Chen, Philippe Ciais, Gustaf Hugelius, and Filipe Aires

Status: open (until 17 Nov 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on essd-2026-254', Mathias Bellat, 13 Oct 2026 reply
Foreword
I would like to thank the authors for submitting their paper to Earth Solid System Data. Thank you also to the editors for the opportunity to review this article.
I am not a specialist in soil geochemistry or in frozen environments. I am trained in soil micromorphology and basic component (e.g., texture, SOC, pH, ...) analysis, digital soil mapping for machine learning methods, and their interpretation, especially in Mediterranean to arid contexts.
Relevance of the manuscript
Short summary
This manuscript presents an analysis of peatlands at high-latitude locations in Eurasia and America. It combines previous datasets with a modern Neural Network modelling approach. Based on a few predictors, it relies on the homogenisation and adaptation of the previous dataset. The final product is a binary map of peatlands, which has been evaluated against a fully independent set of peatland observations. The finer spatial resolution (90m) and homogenisation of the peatland areas provide a truly valuable piece of information for earth sciences. The potential uses are numerous, but current global warming pressures particularly affect these regions and areas of the world that need better understanding.

Title
I would suggest changing to explicitly mention the spatial limitation of your dataset by adding some words like „high latitude“ and „northern hemisphere“.

Abstract
No comment

Article
Major comments:
L. 49-50 How do you justify that Histosols or Histels are all peatlands? Are there not peatlands outside this soil class?

§4.4 The entire paragraph needs to be rewritten, as it is not clear to the reader why the description of your evaluation process belongs to the results part. You should more clearly separate the parts where you describe your evaluation method from the results of this evaluation process. All the material is here and understandable for sure. But for a non-expert reader, the current structure is not adequate. Furthermore, please add a few words on the Kolmogorov–Smirnov references and purpose in your case.

Providing the code used for the analysis would enhance the reproducibility and transparency of the study.

Minor comments:
L.164-167 While I do understand this limitation based on remote sensing coverage, I do not understand why it will not be applied to the two databases mentioned above? Since you described them as heterogeneous, it is possible that some local surveys were also supplied by aerial imagery.

L. 190-195. How did you calculate the merged values for July of each year (average, median, total ...). Second question: how were the values of July alone „better“ than the ones combined from June, August, and September?

L.268-269 What validation method did you use: Leave-One-Out Cross-Validation, cross-validation (how many folds)? Also, why did you not use a spatial splitting strategy such as BlockCV (Schratz et al., 2019; Valavi et al., 2019)?

L.351-354 This sentence belongs to the discussion or § 4.2 below but not to this part of the results.

§5.4 needs to come before §5.3.

§5.3 is too limited;; please provide more extensive possibilities for your dataset regarding current global and regional policies on soil or peatland management (Canadian, European, Russian, USA). This should not be exhaustive but at least result in 5-10 sentences.

§5.4 Another limitation you did not mention in this paragraph is the limited spatial coverage of your study, reduced to high northern latitudes.

Language
L.171 „Change Service (C3S), 2019)provides “ a space between the „)“ and „provides“

Graphics, images, equations and tables
Figure 2: Please insert „a priori “ in italics.

Figures 3 and 7: Please consider using colourblind-friendly colours. The viridis package in R provides all needed adaptations for discrete and continuous values (https://sjmgarnier.github.io/viridis/).

Equations 1-6: Please describe the characters in a consistent way below each equation. While you provide an extended description of the equation function, the purpose of some characters iis not clear.

Figure 4. This figure is welcome and provides a valuable evaluation of the three products presented. One question remains: why do you observe a peak at 100° for the PEATMAP product?

Table A1 I don‘t understand this table. Is it related to equation 1-6? In which case they should be set out in the main text.

References:
The authors have provided all necessary references regarding the topic. Below are the references mentioned in the comments above.
Schratz, P., Muenchow, J., Iturritxa, E., Richter, J., & Brenning, A. (2019). Hyperparameter tuning and performance assessment of statistical and machine-learning algorithms using spatial data. Ecological Modelling, 406, 109–120. https://doi.org/10.1016/j.ecolmodel.2019.06.002

Valavi, R., Elith, J., Lahoz-Monfort, J. J., & Guillera-Arroita, G. (2019). blockCV: An r package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models. Methods in Ecology and Evolution, 10(2), 225–232. https://doi.org/10.1111/2041-210X.13107

Reply
Citation: https://doi.org/10.5194/essd-2026-254-RC1
RC2:
'Comment on essd-2026-254', Tomislav Hengl, 16 Oct 2026 reply
The "PeatCover: Towards a peat mapping at 90 m using satellite remote sensing and a priori inventories" article by Chen et al. is well written and main conclusions and recommendations reflect the work objectives. "PeatCover" presents predictions of the northern peatland extent at 90 m resolution. Authors use a priori inventories and predictions of peat extent (PEATMAP, GPD, CORINE Land Cover) combined with Landsat 8 optical imagery, MERIT DEM/Hydro terrain variables, Water Table Depth (WTD), and an ESA CCI Water Bodies mask to produce a “Peatland Index (PI)” and the final binary peat classification map for Northern territories (Europe, northern Asia and North America). Visual comparisons across representative case studies (including blanket bogs in Scotland, fragmented peatlands in Finland and Estonia-Latvia, wetland networks in the West Siberian Lowlands, and coastal complexes in the Hudson Bay Lowlands and Alaska) confirm substantial improvements in spatial detail over existing inventories. The modeling framework used is, however, potentially biased and definitively sub-optimal. Authors have also chosen a somewhat complicated framework (Fig. 2) where they first fit a model (to produce a Peatland Index) then merge these predictions with training covariate data to adjust final values (circular reasoning problem). In addition, the training data “3.1.1 Training database” is a mixture of synthetic data (random draws from PEATMAP) and ground-truth observations which is sub-optimal. I recommend instead simply using for training only actual ground-truth data (Peat-DBase, PDD-Can-1) and a small proportion of pseudo-absences/-occurrences to avoid producing biased estimates. The authors also miss numerous public / open data sets that could help increase accuracy of peatland extent predictions and make this work more convincing.
I therefore recommend major revision of this work with recomputation of models and layers, using only actual ground observations for training and rewriting some sections, then resubmission. We would also appreciate it if you could share your training points and code used to fit models and produce predictions so we can validate all steps. In addition, the maps attached to the publication (https://doi.org/10.5281/zenodo.21031869) show only binary results (0/1) / no probability maps are shared although these are shown in Fig. 3. I recommend sharing also your fine-grain outputs including all auxiliary image products so that everyone can check and validate your outputs.
Here are the major objections to the work presented:
The work ignores multiple open, global and publicly available data sets that can be added to modeling and analysis. Covariate layers not used in the analysis:

Isik, M. S., Parente, L., & Hengl, T. (2026). Global open water and inundated vegetation occurrence layers from OPERA DSWx-S1 at 30m resolution (Version v1) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.20021037

Hengl, T. (2026). Global ensemble peatland, bare rock and bare sand extent at 100 m to 1 km spatial resolution based on multisource data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.19731872 (see layer “peatland.extent_multi_p_1km_s_2000_2020_go_epsg4326”)

van Jaarsveld, B., Wanders, N., Otoo, N. G., Sutanudjaja, E. H., Verkaik, J., Zamrsky, D., and Bierkens, M. F. P. (2026). Global hyper-resolution modeling of historical and future groundwater dynamics, Earth Syst. Dynam., 17, 1201–1236, https://doi.org/10.5194/esd-17-1201-2026.

Point data-sets with peatland occurrence or peatland thickness not used in the analysis:
A peat-depth database for Canada (PDD-Can-1) described in Dalton, Garneau, & Melton (2026);

UK LandIS points (https://portal.landis.org.uk/);

Netherlands BRO-BHR-P data set (https://portal.landis.org.uk/);

Peat scans (trajectories) in Norway described in Vollering et al. (2025);

All these data sets are also described in detail in: https://codeberg.org/openlandmap/Peatlands_extent.
By adding these additional covariate data sets (to aggregate from 30 to 90 m we recommend using average resampling), authors could potentially increase accuracy of the model and significance of the work. We have run a similar type of analysis and our results (see Glen et al., 2026; https://doi.org/10.5281/zenodo.20121128) indicate that layers such as OPERA DSWx-S1 (based on Sentinel-1) often come highest on the variable importance analysis for mapping peatland thickness etc.
I find the use of the term “Peatland Index (PI)” (a new “measure”) unnecessary because your predictions are basically output of a binary variable classification model. I recommend simply using “predicted peatland extent probability”.

From Fig. 2 it appears that you use PEATMAP and GPD as inputs to the model, then again to adjust “Our results” which could be a case of circular reasoning. Personally I find this unnecessary complication (I assume the rationale to use a hybrid method is due to different extents for regional and global products) and I would recommend instead using the ensemble estimate of extent (e.g. “peatland.extent_multi_p_1km_s_2000_2020_go_epsg4326”) which already integrates all regional and global products without any manipulation. Likewise you mention on P15L310 “Visual interpretation using Google Earth Pro indicated that PEATMAP substantially underestimates peatland extent, whereas GPD maps the spatial distribution more accurately, though it may overestimate total area due to its coarser effective resolution.” I understand the reasoning but it is, in my opinion, a ‘slippery slope’ to just look at Google Earth Pro and think that you can detect peatlands — I do not think it is possible to detect all peatlands directly from 50cm high resolution images / optical images at all. In fact, we were often surprise to find out that something that looks like a peatland on satellite images is in fact not a peatland at all. Likewise, PEATMAP is not the ground-truth of the peatland extent, but also only a mixture of bottom-up estimates and predictions. The only true global ground-truth references for peatlands are PeatDB (https://doi.org/10.5194/essd-17-7313-2025), PDD-Can-1 (https://doi.org/10.5194/essd-2026-169) and similar.

In the introduction section / Table 1, I recommend that you add additional column that explains which peatland data-sets are produced as “bottom-up” (digitize peatland on the ground, then merge areas/countries such as in The EuropeanPeatland map 2025 “EPM2025”, USDA organic soil map and Corine CLC), “predicted/inferred” or “top-down”; and which are “dedicated peatland mapping projects” and which are marginal/partial peatland mapping products (e.g. ESA CCI Land cover). At the moment it looks like all these are equally comparable peatland data sets (they are not).

You have created a training data set with 10 million pixels (5 million peatland and 5 million non-peatland; P12L265). This point data set is a mixture of synthetic (random draws from PEATMAP) and actual observations and measurements of the peatlands. I do not understand why you generated so many synthetic points from PEATMAP (not a ground-truth). Your predictions hence potentially lead to a bias as the PEATMAP product might be over-/under-estimating some areas. I recommend instead using only Peat-DBase + additional points I have listed above and a small proportion (e.g. <10%) of additional pseudo-absences / pseudo-occurrences (synthetic points). Your choice of having 50% vs 50% points is also largely arbitrary. Removing too many 0 values can also lead to bias in predictions.

The choice of the modeling framework “feedforward neural network (NN) with a single hidden layer” P12L270 is not substantiated. I recommend comparing the results you get with (at least) the Random Forest model.

In the end, I would like to say that I truly empathize with all the problems you experience with modeling peatland extent: data is scattered, there are tens of overlapping peatland extent products (which can be overwhelming) and some large countries such as Russian federation, central Africa etc have only limited bottom-up data (gaps in what we know about peatlands are huge). Not to mention that peatlands are poorly defined (what is an engineering definition of “peatland”?) and no strict FAO/ISO or similar standard exists.
Hopefully you will find the above listed recommendation useful and these will help you increase accuracy of your modeling. I am willing to revisit and review the revised version once ready.
Cited references:
Dalton, A. S., Garneau, M., & Melton, J. R. (2026). A peat-depth database for Canada (PDD-can-1). Earth System Science Data Discussions, 2026, 1–40. https://doi.org/10.5194/essd-2026-169

Glen et al. (2026). Global disturbance and emissions from organic soils: leveraging Earth observation–based geospatial data within an IPCC framework. Frontiers in Environmental Science 14: 1761716. https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2026.1761716/

Isik, M. S., & Hengl, T. (2026). Global 30 m resolution distribution of organic soils (Histosols) based on soil taxonomy and peat depth training points (Version v1) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.20121128

Tegetmeyer, T., Kaiser, M., Chobanova, M., Tanneberger, F., & Barthelmes, A. (2025). The EuropeanPeatland map 2025 (“EPM2025”) - technical description and data source. Greifswald, Germany: Greifs-wald Mire Centre.

Vollering, J., Gatis, N., Kusk Gillespie, M., Muggerud, K.-K., Nerhus, S. D., Rydgren, K., & Sparf, M. (2025). Terrain is a stronger predictor of peat depth than airborne radiometrics in Norwegian landscapes. SOIL, 11(2), 763–791. https://doi.org/10.5194/soil-11-763-2025

Reply
Citation: https://doi.org/10.5194/essd-2026-254-RC2

Man Chen, Philippe Ciais, Gustaf Hugelius, and Filipe Aires

Data sets

PeatCover: Northern Peatland Map at ~90 m resolution Man Chen, Philippe Ciais, Gustaf Hugelius, and Filipe Aires https://doi.org/10.5281/zenodo.21031869

Man Chen, Philippe Ciais, Gustaf Hugelius, and Filipe Aires

Viewed

Total article views: 1,103 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
690	279	134	1,103	165	132

HTML: 690
PDF: 279
XML: 134
Total: 1,103
BibTeX: 165
EndNote: 132

Views and downloads (calculated since 17 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	243	94	12	349
Aug 2026	74	26	7	107
Sep 2026	129	60	79	268
Oct 2026	244	99	36	379

Cumulative views and downloads (calculated since 17 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	243	94	12	349
Aug 2026	74	26	7	107
Sep 2026	129	60	79	268
Oct 2026	244	99	36	379

Viewed (geographical distribution)

Total article views: 1,055 (including HTML, PDF, and XML) Thereof 1,055 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 27 Oct 2026

Download

Preprint (21204 KB)
Metadata XML

Short summary

Peatlands are vital carbon stores, yet mapping them consistently across borders is challenging due to varying data quality and definitions. We developed PeatCover, a ~90 m resolution map that harmonizes existing inventories with Earth observation and hydro-topographic data through machine learning. PeatCover provides a spatially coherent dataset that resolves fine landscape structures. These improvements enable refined carbon assessments and methane emission models.


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