the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 22 Jul 2026
Spatial Downscaling of Pollen-Based Vegetation Reconstructions into Paleo Land Cover Maps
Abstract. Past land cover dynamics provide important context for understanding long-term ecosystem responses to climatic change, but paleoecological records are typically sparse, irregular, and site-based. Here, we present a spatial downscaling framework that transforms pollen-derived vegetation reconstructions into gridded paleo land cover maps. The pipeline first translates pollen-based vegetation composition into land cover class frequencies, then distributes these classes within pollen source areas using environmental similarity, and finally applies a U-Net-based interpolation approach to generate spatially continuous reconstructions. We apply the framework to Alaska and western Canada, producing 300 m land cover reconstructions across 15 land cover classes and 119 time slices. The source-area product preserves the direct link to pollen records and contains approximately 2.56 billion reconstructed pixels, while the U-Net-based product extends these reconstructions to continuous spatial coverage across the full study region. Reconstructed land cover trends show broad stability during the late Holocene, with stronger changes around 10 ka BP, including declining evergreen needle-leaved forest cover toward older time slices and corresponding changes in sparse vegetation and shrubland. Relative uncertainty estimates indicate higher uncertainty in forested regions and increasing uncertainty with age, reflecting both class ambiguity and reduced paleo-record support. An example analysis of elevational forest dynamics demonstrates the potential of the dataset for spatially explicit paleoecological applications. The framework provides a bridge between site-based pollen records and landscape-scale analyses, while emphasizing that reconstructed maps represent model-based estimates rather than direct observations of past environments.
- Preprint
(5706 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-4084', Anonymous Referee #1, 20 Aug 2026
-
RC2: 'Potentially novel methods paper but needs more detail', Anonymous Referee #2, 26 Sep 2026
Review of
Spatial Downscaling of Pollen-Based Vegetation Reconstructions into Paleo Land Cover Maps
By Laura Schild, et al.
This manuscript describes a new methodology to create spatially continuous land cover maps based on pollen spectra recovered from sediments. The method has three main parts: 1) creating a statistical relationship between modern pollen spectra and contemporary satellite remote sensing-based land cover maps, 2) using this statistical model with a fossil spectrum and raster fields of bioclimatic variables to create an estimate of land cover in the “source area” of the paleoecological archive, and 3) using another statistical model based on machine learning to both interpolate and extrapolate reconstructed land cover from a source area to create a spatially continuous land cover reconstruction over a defined spatial domain.
This study represents a novel and ambitious attempt to address a longstanding goal of the palaeoecological community, that is to create continuous reconstructions of land cover based on paleoecological records. While the authors succeed, admirably noting a range of caveats in their discussion, the presentation of the method and results of this study are so flawed it is impossible to have any confidence in the final result.
Because the reconstruction results are presented only superficially – only five time slice maps are included in the manuscript and the time series reconstruction does not include dynamic ice or sea level extent – this paper would probably be better suited for a journal more focused on geoscientific modeling methods as opposed to Climate of the Past.
For these reasons, I cannot recommend this manuscript for further consideration for publication.
General comments
A critical flaw in this study is the lack of information about the modern pollen data used to establish the central relationship between land cover and pollen spectra. It is possible that the authors did a thorough job of considering the time of collection and pollen source area, but this is impossible to tell from reading the current manuscript. The phrase “source area” is used at least 28 times in the manuscript, but nowhere is it described how this was determined. As far as I could tell, information on basin type and size, and therefore pollen source area, come from the LegacyVegetation dataset (Schild et al., 2025). It is not possible from reading the paper describing that dataset to immediately see the definitions of the pollen source area for every site; nevertheless, this key information should be included in the current manuscript for every site that was used in the reconstruction. In any case, I am highly skeptical that the pollen “footprints” shown e.g., in figure 3 can extend to what appears to be >100km from individual pollen archives. This is reinforced by a look at Figure 6 in Schild et al. (2025), which shows that several sites in Alaska and northwest Canada in the LegacyVegetation dataset are not represented by “at least one large lake (≥ 50 ha) or several smaller basins” and therefore must represent the broad scale vegetation patterns more than 100km from the site. In fact, for peatlands and small lakes, the pollen spectrum might be representative of a very small area indeed, perhaps dominated by the vegetation in an area only a few 10s of meters from the archive and therefore could only be represented by a single pixel in the satellite land cover map.
Digging deeper into the raw LegacyVegetation dataset, which I should not have needed to do, I see that the “pollen source radius” for sites in the North America sub-dataset range from 262m to more than 500km. At a bare minimum, the manuscript needs to go into much more detail about the spatial representativeness of every site used in the model calibration, use a different source area for every site based on the characteristics of the archive, and all of this metadata needs to be included as part of the manuscript.
Furthermore, the period of collection of the “modern” pollen sample must be clarified. It is noted in the discussion of Schild et al. (2025) that “…pollen-based estimates are derived from records that span a much longer timescale than the modern forest cover data available, even though modern time slices are used for validation.” This key information also needs to be included in the current manuscript. In fact, it is highly likely that nearly all the pollen data used in the model calibration represents deposition that occurred decades to centuries (even a millennium) before the year 2020 CCI land cover map used in this study. Since the CCI-C3S land cover maps cover the period 1992-2022, it begs the question why the authors didn’t at least take a temporal average of land cover over the available period instead of a single timeslice. Or multiple calibration models against every individual year of CCI land cover maps.
While large-scale potential natural vegetation in much of the study area may have been relatively stable over the last century, disturbance, including widespread wildfires, nearly certainly changed land cover at the 300m pixel resolution used in the study. Again, at a bare minimum the manuscript must provide transparent, detailed information about the age of each modern pollen spectrum used in the model calibration step and some discussion and analysis of how the temporal mismatch between land cover map and pollen age may have affected the model outcome.
Likewise, the CHELSA bioclimatic variables represent a climatological mean over 1981-2010 (Table 1) and so are mismatched in time with the 2020 CCI land cover map.
The Copernicus CCI land cover maps are provided on an unprojected lon-lat grid of 10 arc-seconds. This means that the pixel size changes with latitude and the pixels are not equal-area; how was this handled in the model calibration step? Was a circular area interest defined? This is how it appears on the maps, although the given the size of the spatial footprint, the land cover pixel count would be different with latitude and so a circular search area of fixed radius would be inappropriate. It needs to be clarified how the CCI land cover fractions were adjusted to account for varying pixel area.
In the discussion section, no mention of edaphic controls on land cover is made. Soil type, bare bedrock, permafrost, and other landscape features have an important control on the biogeography of the study area. Why were variables including soil texture, slope, or landscape age not included among the environmental variables used to inform the regression and interpolation analyses?
Some notes on presentation
- The manuscript is so full of jargon and technical language that the reader might be persuaded to skip the methods section, accept the authors’ approach at face value, and concentrate on the results. This is wrong for any manuscript that introduces a novel method. While one might accept the study’s results, as a novel methodology the approach taken here needs to be described in a much more pedantic way. Many interested readers of Climate of the Past and ultimately data users will not be experts in geostatistics and machine learning and so it is important for the credibility of the authors to be more transparent and clear about what they did and how they did it. For example, technical terms such as “centered log ratio transformation”, “RDA”, “PCA”, “Hellinger distance”, “network simplex algorithm”, “softmax weighting”, “Aitchison distance”, etc. could be placed in a table similar to Table 1 with a name, symbol or acronym, and units if applicable, etc.
- For every equation in the manuscript, all symbols need to be described, with units included as appropriate.
- The maps shown in Figure 3 are far too small to interpret. If space in the journal is an issue, a link to full-resolution maps must be provided. I recommend splitting this large figure into several components and reproducing all maps with much higher resolution.
Specific comments
Line 85
It is noted that the LegacyVegetation dataset was used in the current study. Given the importance of this dataset for the current study, a summary of how LegacyVegetation was produced should be included in the manuscript. This includes more details about how the REVEALS method was used to reconstruct vegetation cover from pollen taxa. Was any standardization or other manipulation of the pollen data done when making the REVEALS reconstructions? For example, setting thresholds for taxa, deleting low-quality sites as low resolutions, few taxa or records, etc.?
Line 87-88
As noted above, the “the estimated pollen source area provided [by the] the LegacyVegetation dataset” needs to be described in much more detail in the current manuscript and should not require downloading the entire 185Mb dataset to figure out; even then it is not clear which of the North America sites in the dataset on Pangaea are actually used in the reconstruction presented in this manuscript.
Line 90
Note that the CCI land cover product has 10 arc-second resolution that varies with latitude rather than being an equal-area raster data product. This point is actually quite misleading in the original CCI documentation.
Line 96
What were the implications of keeping “the lichen and moss class …constant through time”? Why was this done?
Line 97
Isn’t a separate model constructed for each modern pollen record, as opposed to “paleo-pollen” noted here? How can a regression model between paleo-pollen and modern land cover be constructed?
Line 103
If the criteria for selecting core-top samples extends to 1000 BP, the implications of the temporal mismatch need to be discussed in further detail as noted above. Furthermore, are these “additional data” part of the LegacyVegetation dataset or do they come from somewhere else? If these data are from other sources, how was the pollen source area estimated? Many surface pollen samples are collected in soils or moss polsters that represent very localized signals.
Also, what does the 1000km buffer actually mean? Were modern pollen data from up to 1000km away from the target site used in creating the regression model? Did the buffer size also account for differences in elevation? A site even 10’s of km away from the target but 1000m higher might represent a completely different land cover. This section requires much more explanation.
Equation 1
More information should be provided about the relationship between the “clr” values. This equation needs more explanation.
Line 110
The characteristics of “Elastic Net regularized regression” should be described in more detail.
Line 136
The acronym “RDA” does not appear to be defined anywhere.
Line 137
The acronym “PCA” does not appear to be defined anywhere.
Line 147
Does the assumption that “land cover transitions are more likely to occur between environmentally similar classes…” account for transitions caused by disturbance, e.g., wildfires?
Line 175
Is “the current land cover” the 2020 CCI land cover map? Please add precision.
Line 188
All symbols in “softmax(−D/τ)” need a definition and units. See suggestion about putting these in a table.
Equation 4
This equation needs every term clearly defined.
Table 1
The line “Mean Daily Mean Near-Surface Air Temperature of the Quarter” appears to be missing something. Which quarter?
Line 235
How was the 256x256 pixel tile selected? Presumably these are tiles based on the 10 arc-second CCI raster, so a rough radius of 77km around each pollen site in the meridional direction and less than half of that in the zonal direction? These tiles would vary area within and among each tile as noted above because of the use of an unprojected raster.
Line 240
If “Ocean pixels are set to a “nodata” value using a sea mask” why does it appear that no sea mask has been applied in Figure 3c?
Line 255
How are the “observed compositions” of each pollen record location determined? Is this from the previous step? The process for quantifying uncertainties could be described in a flowchart similar to Figure 1.
Line 272-273
More details about the “source areas of the pollen records” must be included in this paper.
Line 306
If I correctly infer the methods applied in this study, land cover is resolved on a 10 arc-second grid, which is substantially smaller than 300m at the high latitudes used in this study.
Line 307
119 discrete time slices at 500-year intervals covers 59,500 years. On line 75, it is stated that the reconstruction covers “the last 20 ka”. How can we reconcile these numbers? If only fewer timeslices are available in the published dataset, this should be noted. Also, high-resolution land cover maps, both point-based and interpolated for each of the 119 time slices should be included as part of the publication, perhaps in supplementary materials.
Figure 3
The “example time slices” are not labelled. These must be included on the figure and in the caption. And how can there be perfect “pollen agreement” in areas over water?
Figure 4
Total area in km2 should be presented here instead of pixel count. The varying area of pixels with latitude makes this figure potentially misleading.
Figure 6
No legend, axis labels, or color scale are provided for panel b; what is presented here?
Line 430-431
Sea level dynamics were also very important in the study area and over the period studied (cf. Beringia). It should be noted that varying area of the continents should be included in any paleo land cover reconstruction that covers the late Quaternary.
Line 500-503
Here it is acknowledged that reconstruction of past wetland distribution would be valuable, but impossible using the method described in this paper without the inclusion of edaphic controls on vegetation cover, as noted above.
Line 522
Given the thematic resolution of the CCI maps “several forest classes [may be] environmentally plausible” because they represent different successional stages following disturbance. Limiting the training of the model to a single modern (2020) land cover map, makes detection of short-term land cover change impossible. It should be noted that considering a range of maps for the present-day land cover might be desirable as a way of dealing with this uncertainty.
Line 535
How is “absolute agreement” defined and calculated?
Line 540
Even if they don’t currently exist, it would in theory be possible to generate “independent paleo land cover” products at the spatial and temporal resolution of the pollen-based reconstructions using a forward modeling approach, with a vegetation model embedded in an earth system model or as a standalone model driven by GCM output. Ultimately a desirable use case for the spatial land cover reconstructions presented here would be to evaluate earth system model output.
Line 580
As noted above, much more needs to be said about the pollen source area. The assertion that “pollen signals typically integrate vegetation over larger spatial extents” should be backed up by a citation; isn’t this relationship highly dependent on the type and size of the archive, e.g., peatland, small lake, large lake, etc.?
Line 604-606
It is honestly acknowledged here that the method presented may not reconstruct past land cover with very much fidelity. This point should be emphasized in the abstract. Otherwise, as noted above, many data users are likely to take the results at face value, i.e., as the truth. Furthermore, the approach may not be good at representing not just “localized processes” but also large-scale disturbances such as wildfires because of the limitation of using only a single land cover map in the establishment of the modern pollen-land cover relationship.
Line 607
Suggest changing “the most probable…” to “a probable” in acknowledgement of the major limitations of the methodology.
Line 618-619
Would aDNA not also be subject to methodological challenges in translating data into land cover classes?
Line 632
Suggest changing to: “time-varying land-sea and ice masks” to acknowledge changes in coastlines as well.
Line 650-651
The dataset published on Zenodo should be broken up into a number of separate files for the point-based and wall-to-wall reconstructions and for different time slices to make it easier for people with low-bandwidth and/or limited data storage to at least be able to explore the data and better understand the methodology. The complete metadata for all sites and additional samples used for model training in this study must also be included in the freely available data for download.
Appendix A
The table here should have a table number or identifier and be called out in the main text. Please add a column with the total area (km2) of each land cover type in the study area. This will help the reader understand the regional importance of each land cover class and provide additional rationale for merging the classes.
It would be helpful to provide several tables here, with the intermediate steps showing which classes remaining after classes are merged and filtered, and a final table showing only those classes used in the paleo reconstruction.
Table C1
Provide an additional table that includes the number and description of only those numbers used in this table. It is cumbersome to have to go back to the table in Appendix A to figure out what these numbers mean.
References
Schild, L., Ewald, P., Li, C., Hébert, R., Laepple, T., & Herzschuh, U. (2025). LegacyVegetation: Northern Hemisphere reconstruction of past plant cover and total tree cover from pollen archives of the last 14 kyr. Earth System Science Data, 17(5), 2007–2033. doi:10.5194/essd-17-2007-2025
Citation: https://doi.org/10.5194/egusphere-2026-4084-RC2
Data sets
Output data for the Manuscript: Spatial Downscaling of Pollen-Based Vegetation Reconstructions into Paleo Land Cover Maps L. Schild et al. https://doi.org/10.5281/zenodo.21240822
Model code and software
PollenPixel: Source Area Reconstruction Code S. Lisovski et al. https://doi.org/10.5281/zenodo.21219588
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 277 | 139 | 34 | 450 | 33 | 31 |
- HTML: 277
- PDF: 139
- XML: 34
- Total: 450
- BibTeX: 33
- EndNote: 31
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
I've attached my comments as a PDF.