| 16 Feb 2026
Quantifying soil organic carbon stocks above the treeline in the Swiss Alps
Abstract. More than 90 % of the total carbon (C) in alpine ecosystems is stored belowground, yet spatial estimates of soil organic carbon (SOC) stocks remain scarce due to limited accessibility and the demanding nature of SOC stock estimates in rocky alpine terrain. By combining new measurements at 144 sites across the Swiss Alps with data from existing inventories, we compiled a comprehensive dataset on SOC stocks totalling 307 sites from treeline to the permafrost region (1750 m – 3100 m a.s.l.). We predicted the spatial distribution of SOC by linking stock measurements to environmental covariates using Quantile Regression Forests (QRF) and produced a SOC stock map at 25 m resolution illustrating the spatial SOC variability in alpine terrain. Our results show that SOC stocks average 7.3 ± 3.3 kg m⁻² in alpine grasslands and 1.8 ± 1.7 kg m⁻² in partly vegetated areas around and above the vegetation line. Overall, the alpine region of Switzerland, which covers one-third of the total country area, stores an estimated amount of 47.6 Mt SOC, representing a non-negligible share of the Swiss greenhouse gas inventory. Vegetation productivity, represented by the Normalized Difference Vegetation Index (NDVI) and topo-climatic covariables, together with vegetation-derived indicators of humus content and soil pH, were highly informative for spatial predictions. This study identifies hotspot regions of SOC storage and influential spatial predictors of its distribution, providing a quantitative baseline for assessing the status-quo and future changes in alpine SOC stocks under continued climate and land-use change. The observed increase in SOC stocks with increasing NDVI suggests that climate change-driven greening at high elevations, where vegetation cover is currently sparse, may enhance SOC storage, although the rates and magnitude of these changes require further investigation.
Competing interests: One of the coauthors, Frank Hagedorn, is a member of the editorial board of Biogeosciences.
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.
Overview
This study quantifies soil organic carbon stocks above the treeline in the Swiss Alps by combining 307 field sites with pedotransfer-based subsoil extrapolation and spatial modelling using Quantile Regression Forests. It estimates 47.6 Mt SOC stored across alpine terrain and identifies NDVI, vegetation-derived humus indicators, temperature, and topography as key predictors of SOC spatial variability.
Overall, the manuscript presents a valuable and timely contribution to SOC analysis in alpine ecosystems. It is well-written, provides an important baseline for SOC stocks and fits very well the scope of the journal. However, clarifications are needed before the manuscript could be considered for publication, especially regarding the robustness of PTF for subsoil extrapolation, the predictor circularity, the coarse fragment integration, the pseudo-zero weighting, and the full uncertainty propagation.
Main concerns
1- First, a major methodological concern relates to the use of pedotransfer functions (PTF) to estimate subsoil SOC stocks. Subsoil values are extrapolated from only 17 soil profiles stratified by bedrock type and elevation bands, yet these relationships are applied to 307 sampling sites across the alpine region. Given the well-known heterogeneity of alpine soils, including high skeletal content, cryoturbation, discontinuous soil depth, and highly variable pedogenic pathways, the robustness and transferability of these PTF require more rigorous justification and uncertainty assessment. Another option is to move all this analysis into supplementary information. In particular, the procedure of selecting the “best-fitting” profile for each site may introduce overfitting and underestimate the structural uncertainty associated with subsoil extrapolation. Related to this, the manuscript does not clearly explain how soil depth is estimated or constrained across the study area. Without explicit constraints on soil depth, extrapolation of SOC stocks below the measured 0–20 cm layer becomes difficult to evaluate. Moreover, soil types such as Umbrisols, Podzols, Rankers, Rendzinas, and Cambisols can occur over short spatial distances and exhibit very different vertical SOC distributions and bulk density patterns.
2- Second, Landolt-H and Landolt-R are partly circular predictors: both are vegetation-derived proxies likely co-varying with SOC. Their strong importance may reflect shared response rather than mechanistic control, potentially inflating predictive performance while limiting process inference. Moreover, the uncertainties of these variables are not integrated in the modelling approach, and I assume that these uncertainties are spatially distributed.
3- Third, the treatment of coarse fragments and bare-rock areas raises methodological concerns that may bias SOC stock estimates. The sampling protocol did not allow the quantification of coarse fragments larger than 5 cm, which are common in alpine soils. In addition, the treatment of bare-rock areas through the introduction of “pseudo-zero” SOC values based on expert judgement lacks sufficient methodological justification. This assumption directly influences SOC predictions in sparsely vegetated, high-elevation zones and may significantly affect regional SOC stock estimates. It is unclear why bare-rock areas were not explicitly excluded from SOC calculations, unless the study explicitly considers petrogenic organic carbon or organic matter stored in lithic substrates.
4- Finally, although the geostatistical framework appears robust and overall well implemented, the treatment of uncertainty remains incomplete. The reported total SOC stock estimate does not appear to fully propagate the different sources of uncertainty inherent to the methodology. In particular, uncertainties associated with pedotransfer functions, covariate uncertainties, and model structural uncertainty are not explicitly integrated into the final confidence intervals. In addition, the assessment of spatial autocorrelation in the model outputs appears to rely primarily on visual inspection and not on formal statistical evaluation (variogram analysis or Moran’s I statistics on residuals).
Specific comments
L11: “… of the total [organic] carbon…”, because in carbonated context, some of carbon is include in carbonates.
L21: Using soil pH as explanatory variable is contestable, as SOC content (organic acid) decreases soil pH.
L23-25: I agree with this statement for the upper alpine grasslands, but what about alpine grasslands near the treeline, containing the highest SOC stocks?
L35: “due to the difficulty of accessing as well the high heterogeneity” → “due to the difficulty of access as well as the high heterogeneity”.
L42: Greening of alpine area is only one consequence of climate change on carbon dynamics, influencing quantity. Other consequences could be the change in land cover and thus change in SOC chemical composition, influencing also its stability.
L58: “suppress” à “slow, reduce,…”
L75-76: The end of the sentence does not follow on from the preceding sentences.
L90: Why you mention the “time” factor above in the introduction (L53), and you didn’t include it in the analysis?
L97: Given the elevation of study sites, could some areas be recently deforested? e.g. at 1750 m, potentially leading to legacy effect on SOC stocks not integrated by your model.
L105: The mosaic of parent material leads to wider variations in soil conditions than soil pH only, with some of them that could influence SOC stocks (Fe, Al, Ca, clay, sand contents, etc.). Your study didn’t really integrate the “parent material” factor. Is the integration of geologic properties could increase model performance?
L107: Can you give here an idea of the study area surface?
L122: It is not clear why there is 20 sites from GLORIA. 6 summits x 4 cardinal directions = 24 sites?
L129: mean [annual] precipitations I supposed?
L140: It is always problematic to assess the bulk density. Given the available data of the three datasets you have, your approach seems to be the best, although the difference in soil volume could introduce bias on the coarse fragment content estimation. Because in IMIS dataset, coarse fragments >2cm were not integrated in bulk density assessment, while in BDM only coarse fragments >4,8cm (and in GLORIA for coarse fragments >10cm) were not integrated. And given the importance of such coarse fragments in alpine soils (as we can see on the soil profile pictures in appendix B), this omission could have great importance on SOC stock calculation. Did you try to compare soil coarse fragment content between the three datasets? And the bulk density?
L145: “replica” → “replicate”?
L170: In this paragraph, it is not clear how do you predict spatially the soil depth.
L171: It is not clear for me which PTF you assign for each pixel of your study area. If I have well understood your study, you first spatialized SOC stocks in the upper 0-20cm horizon using QRF model, then you used PTF to assess SOC stocks for the whole profile, right? First, I suggest adding the use of PTF in the workflow of figure 1. Second, the stratification by bedrock type and elevation is debatable. Are these two parameters the only ones that control soil properties and stratification? Coarse content is more dependent on geomorphic processes rather than bedrock and elevation.
L177: Why did you choose linear regression for FE density while you used exponential decline for SOC content?
L218: “humus-indictor map” → “humus-indicator map”.
L260: What does the activation of this option mean?
L276: Did you exclude the pixels out of AOA from your analysis?
L294: Did you predict SOC in forests?
L298: I am not familiar with such studies, but did you see: https://doi.org/10.21425/F5FBG61746 ?
L300: This pseudo-zero method needs further clarification. How did you choose the site? Did you attribute a stock of 0?
L303: Did the forest soil data be used for model calibration?
L324: No comma or conjunction after 0.49
L329: “negligable” → “negligible”)
L329: “among [the covariates]”?
L377: Can you give in supplementary the elevational density of your samples? Some elevations may be underrepresented in your dataset.
L399: The discussion needs more in-depth mechanisms explanation and not just a repetition of the results with a comparison of other studies.
L403: Can you give an estimated value of SOC stocks at national scale for comparison?
L415: What do you [they] mean by “managed” grassland?
L419: Yes, illustrating the crucial role to integrate the coarse fragments content, especially rocks and stone.
L425: And what about petrogenic organic matter?
L432-436: Did you try to compare the model results in high elevation with and without integrating pseudo-zero samples?
L451: Landolt-R is not really an estimation of the parent material properties.
L451-460: What are the uncertainties of Landolt-R and Landolt-H variables? Did you integrate these uncertainties into your model?
L471: Are there any bog and mire samples in your training dataset?
L497: pH conditions on carbonate bedrock could also limit microbial decomposition if they are too alkaline. See “optimal windows” from https://doi.org/10.1007/s10533-017-0410-1
L509: yes, but plant communities also depend of SOC content…
L515: Beware of circular reasoning: you predicted SOC using a variable that indicates high SOC content…
L530: And also the uncertainties due to the covariates themselves.
L549: Yes but this estimation remains highly uncertain, you didn’t estimate the soil depth.
L560: The spatial distribution of residuals was only assessed through visual representation. Did you use Moran’s I test to assess spatial autocorrelation?
L568: Did you sample bogs and fens in your study?
L595: This is only true for upper grasslands. What about grasslands just above the treeline?
L599: I agree but the time scale between soil formation and SOC accumulation could be very different.
L609: You mean that MAOC could be saturated in alpine soils? This seems rather unlikely, given that such results are rarely observed at low elevations.
L622: The conclusion is concise, clear and summarizes well the main findings of the study.
Figure 1: Why are “forest” and “water/ice” include as cover class although they are not analyzed in the study?
Figure 2: It is debatable and not common to predict the rock fraction using pedotransfer function. The relationship between coarse content and depth is not obvious and it is dependent on the coarse fragment considered (stone, blocks, pebbles, …).
Figure 6: SOC decreases the soil pH, so the more the SOC content, the lower the soil pH. Thus, SOC could also explain the soil pH and the vegetation preference, and not the opposite.