the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Sep 2026
Dynamic vegetation–soil–water feedbacks as an under-recognised source of structural uncertainty in digital soil mapping
Abstract. Digital soil mapping (DSM) produces spatially continuous soil information for land-use planning, carbon accounting, and Earth system analysis, resting on the SCORPAN framework and its use of remote-sensing and terrain covariates as indicators of soil-forming processes. Similar approaches estimate ecological and hydrological conditions from vegetation, moisture, and climate data. Two assumptions receive little scrutiny: that soil-covariate relationships stay constant in space and time, and that covariates act as independent predictors. In reality vegetation, soil, and water form a coupled system in which each influences the others, and the strength and even the direction of these interactions vary across landscapes and shift over time. We argue that these dynamic relationships are an important but under-recognised source of structural uncertainty, distinct from the sampling, measurement, and model-fitting error usually partitioned in DSM. Drawing on wetlands, drylands, sloping farmland, burned watersheds, drought-affected forests, and thawing permafrost, we show how the feedbacks break the assumptions behind DSM, why the same problem affects land-surface, ecological, and hydrological models, and how mapping might be made feedback-aware.
- Preprint
(521 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-4516', Anonymous Referee #1, 30 Sep 2026
-
AC1: 'Reply on RC1', Odunayo David Adeniyi, 07 Oct 2026
Response to Reviewer 1
We thank the reviewer for the careful assessment of our manuscript. We appreciate the recognition of the central premise and the comments that helped us clarify the distinction between feedback-driven structural uncertainty and other established sources of uncertainty in digital soil mapping (DSM). We have revised the manuscript accordingly while keeping the perspective focused.
- Distinction from other sources of uncertainty
We agree that the original manuscript did not sufficiently distinguish feedback-driven structural uncertainty from omitted variables, model misspecification, sampling limitations, and transferability. We have therefore clarified the definition in Section 2. We now describe feedback-driven structural uncertainty as arising when the soil–covariate relationship depends on the state or history of the coupled soil–vegetation–water system, while that state or history is not fully represented by the covariates used for mapping.
Section 4 now also distinguishes this mechanism from other uncertainty sources. Omitted-variable uncertainty concerns missing explanatory factors, model misspecification concerns an inadequate functional form, and sampling limitations concern insufficient representation of the application domain. These mechanisms can overlap, but feedback-driven structural uncertainty concerns the changing soil–covariate relationship across system states and histories.
- Sampling and additional observations
We have also clarified when additional observations may reduce this uncertainty. Additional samples can be useful when they capture different feedback regimes or system states, whereas increasing sampling density within an already represented regime primarily reduces sampling uncertainty. Section 6 now therefore emphasises targeting underrepresented feedback regimes using indicators such as hydrological position, phenological variability, and recent disturbance.
- Implications for feedback-aware mapping
Following the reviewer's suggestions, we have clarified the role and limitations of the approaches proposed in Section 6. Time series and phenological metrics are presented as ways of representing temporal state and history that single-date observations miss, rather than as direct evidence of reciprocal feedbacks. Spatially and temporally varying models and soil-landscape stratification can accommodate differences between settings, but require defensible definitions of feedback regimes and transitions. We also now explicitly acknowledge that process-model outputs may provide information about otherwise unobserved system states while introducing uncertainty from the process model itself. Repeated monitoring is presented as a means of constraining temporal signals and system history rather than, by itself, demonstrating feedback causality.
- Conceptual example and future empirical testing
We agree that a concrete example would help make the proposed mechanism clearer. We have therefore added a short conceptual example based on soil organic carbon mapping. It illustrates how locations with similar current vegetation and moisture conditions may have different soil organic carbon because of different disturbance, management, or hydrological histories, and how a model trained mainly under one feedback regime may therefore transfer poorly to another.
We have not added an empirical proof-of-concept analysis, as this would require a separate study beyond the scope of the present perspective. Instead, we have clarified the four testable predictions in Section 5 as a basis for future empirical testing, including testing the relationship between structured residuals, feedback-strength indicators, temporal covariate drift, feedback-regime stratification, and covariate-importance instability.
- Overall revision
Finally, we have moderated several statements that were too absolute in the original manuscript. In particular, we no longer imply that feedback-driven structural uncertainty cannot be reduced by additional observations. Rather, we argue that it may persist when additional sampling only increases observations of the same static covariates and does not capture the relevant system state or history.
We thank the reviewer again for these constructive comments. We believe the revisions have clarified the conceptual contribution of the manuscript, its distinction from established sources of DSM uncertainty, and the practical and empirical implications of the proposed framework.
Citation: https://doi.org/10.5194/egusphere-2026-4516-AC1
-
AC1: 'Reply on RC1', Odunayo David Adeniyi, 07 Oct 2026
-
RC2: 'Comment on egusphere-2026-4516', Anonymous Referee #2, 14 Oct 2026
In their manuscript, Adeniyi et al. examine how interactions between plants, soil and water create hidden errors in digital soil maps. The main message of the paper is that standard soil mapping methods, such as the SCORPAN model, rely on fixed snapshot data and incorrectly assume that environmental factors act independently of the soil. In reality, plants, soil and water form a connected system where each factor influences the others across different landscapes and over time. Using multiple examples, the authors argue that ignoring these two-way relationships creates a distinct type of map error, termed structural uncertainty, that cannot be fixed by collecting more soil samples or using more flexible algorithms. The manuscript highlights that this same limitation affects related areas such as climate, ecology, and water modeling, and discusses ways to make future mapping dynamic and feedback-aware.
The authors address an important topic, namely how digital soil mapping approaches can avoid errors caused by treating dynamic vegetation–soil–water interactions as static and independent. The impact of the manuscript could be significantly strengthened by language editing to improve readability, as the message is often obscured by overly complex wording. For example:
- The formulas presented on lines 86 – 89 and the motivation behind the re-writing of the equations need to be explicitly explained in understandable language.
- Sentences such as ‘An endogenous model can pass random cross-validation yet degrade under transfer where the feedback state differs, since applicability diagnostics flag novel predictor values but not shifts in predictor-response relationships’ (L90-92) or ‘Local-coefficient methods respond in part, fitting a relationship per location or separating fixed from spatially varying effects, but treat non-stationarity as spatial when the feedbacks are also temporal’ (L137-139) will not be understandable to most readers and need drastic simplification.Sentences of similarly high complexity are found throughout the manuscript and need to be simplified for their meaning to be clear to readers with a varying level of familiarity with the topic.
Furthermore, many literature examples lack sufficient detail to be understandable by readers not familiar with the details of those studies. In general, the manuscript would benefit from a clearer description of many of the provided examples, to better underpin the message the authors want to bring across. For example:
- ‘Vegetation drives soil and water, […] setting infiltration’ (L97-98): wordings such as ‘drives’ and ‘setting’ need more explanation for the reader to understand in which way vegetation affects which aspects of soil and soil water.
- ‘Organic matter inputs […] set soil carbon‘ (L101-102): ‘to set’ does not explain mechanistically how organic matter inputs affect soil carbon.
- ‘Afforestation carbon change depends on […], reversing between carbon-poor and carbon-rich grasslands’ (L103-105): it’s not clear how the initial relationship looks like, and what it means if it’s reversed. Same for L134: ‘vegetation-to-soil and soil-to water relations differ by climate […], and can reverse sign’. It’s not clear what is meant by this, or which mechanism this is based on.
- ‘Adding soil layers cuts soil moisture modelling error by about a quarter’ (L114), or ‘erosion shifts Mollisol infiltration towards preferential flow’: statements like these need more information to be informative to the reader.As a solution to relying on snapshot data and one-size-fits-all relations, the authors propose using time-series data and region-specific modelling, for example in section 6. It is, however, not clear whether such data exist in sufficient quantity at a high-enough spatial resolution to resolve the issue the authors discuss. A better discussion on how the availability of such data affects the proposed solutions would strengthen the quality of the manuscript.
To improve the overall quality of the manuscript and its potential impact, I would recommend the authors to (1) perform extensive editorial editing focusing on simplifying the language and improving the readability and (2) provide clearer examples of the dynamic relations and feedbacks between commonly used predictors, beyond wording such as ‘X drives Y’, ‘X sets Y’, ‘X controls Y’, ‘X outweighs Y’, ‘X swings with Y’, etc. More concrete examples explaining the mechanisms behind selected relations will help the reader to understand the need for a better incorporation of the dynamic nature of these relations in digital soil mapping. Furthermore, (3) concrete examples of how dynamic relations can be incorporated in digital soil mapping, together with which data can be used to achieve this, would help the reader understand how feasible the proposed approach is.
Citation: https://doi.org/10.5194/egusphere-2026-4516-RC2 -
AC2: 'Reply on RC2', Odunayo David Adeniyi, 20 Oct 2026
Thank you for the comments, which has helped us improve the clarity and practical relevance of the manuscript. Following the three main recommendations, we've (1) simplified the language throughout, (2) replaced general wording such as "X drives Y" or "X sets Y" with short descriptions of the underlying mechanisms, and (3) clarified how dynamic relationships can be represented in DSM.
We now explain the notation in plain language in Section 2. The standard form S = f(x) + e treats covariates as fixed predictors of soil. However, vegetation and soil water are themselves partly shaped by soil, which we express as V = g(S, W, t) and W = h(S, V, t). The covariates used to predict soil may therefore partly reflect the soil being predicted, written as S = f(x(S, t)) + e. We also define endogeneity in this predictive sense, with an example: NDVI partly reflects soil water-holding capacity.
The two sentences highlighted by the reviewer were rewritten. The first now states that a model may perform well in random cross-validation but predict less accurately in a new area or period if the relationships between soil, vegetation, and water differ there. The second now explains that local-coefficient methods allow relationships to differ between locations but may not capture changes over time. We simplified other sentences of similar complexity throughout the manuscript.
We've rewritten Section 3 to describe mechanisms rather than only the direction of influence. Roots create channels or block pores, litter supplies organic matter, and the canopy changes how much rainfall reaches the soil. Litter amount and quality affect how much carbon enters the soil and how much is decomposed or stabilised. The afforestation example now states the direction of change explicitly. The soil-moisture example specifies that soil information was included as model input, and the erosion example explains that erosion changes soil structure and redirects water through preferential flow paths.
Section 6 now discusses how cloud cover, limited revisit frequency, crop residues, and sensor resolution reduce the number of suitable observations. Spatiotemporal fusion can partly compensate, and repeated field measurements may be too sparse to represent feedback regimes across a landscape. We also note that time series show when covariates change but do not by themselves identify feedbacks.
Section 4 now includes an SOC mapping example. Two sites with similar current vegetation and moisture, such as a recently drained wetland and a long-term cropland, can differ in SOC because of their different histories. The example shows how temporal trajectories or disturbance history could make this difference explicit, and how block cross-validation that holds out whole feedback regimes could test it. Section 6 outlines practical options: time-series covariates, models stratified by soil-landscape units, and sampling targeted at under-represented regimes using hydrological position, phenological variability, and recent disturbance.
We believe these changes have made the manuscript clearer and more useful to the DSM community, and we thank the reviewer for helping us improve it.
Citation: https://doi.org/10.5194/egusphere-2026-4516-AC2
-
AC2: 'Reply on RC2', Odunayo David Adeniyi, 20 Oct 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 290 | 80 | 36 | 406 | 42 | 44 |
- HTML: 290
- PDF: 80
- XML: 36
- Total: 406
- BibTeX: 42
- EndNote: 44
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This is an interesting perspective, and I agree with the central premise raised by the authors. Soil properties emerge from interacting processes whose influence depends on environmental setting and history, and these interactions form a feedback loop perpetuated across space and time in terrestrial landscapes. Therefore, this perspective raises a relevant concern for predicting soil properties: whether relationships learned from environmental snapshots or long-term averages remain reliable across locations, periods, and disturbances. That said, I think that the conceptual contribution and practical implications of the proposed feedback-aware framework need further clarification for practical application in spatiotemporal digital soil mapping (DSM).
My understanding is that the authors argue that commonly used covariates for predicting soil properties, under the SCORPAN framework, provide an incomplete description of a soil-plant-water system with feedbacks, legacies, and possible transitions. Consequently, similar observed covariates may correspond to different soil conditions, and a model that performs well within its training domain may become unreliable under another inference space. These concerns overlap with established challenges in model transferability and spatiotemporal DSM, including those discussed by Heuvelink et al. (2021). If I understand correctly, the authors argue that these known effects constitute a distinct class of structural uncertainty. That said, I think the proposal of structural uncertainty requires more explanation and stronger justification.
I find the idea of structural uncertainty interesting. That said, I am confused, as the manuscript does not clearly establish how structural uncertainty differs from uncertainty associated with omitted variables, inadequate model architecture, or differences in soil–covariate relationships between the conditions used to train the model (i.e., model parameterization) and those where it is applied. Limited or unrepresentative training data may further restrict model applicability and inference, particularly when relevant environmental regimes or histories are poorly sampled. I think it is important to clarify how these sampling limitations interact with the proposed feedback-related uncertainty, and under what conditions additional observations would reduce it or not.
Furthermore, I find the idea of feedbacks interesting, but the manuscript would benefit from explaining the conditions under which feedbacks create a specific practical limitation that existing DSM approaches do not adequately address. The proposed distinction between feedback-related uncertainty and existing uncertainty categories needs further clarification. I encourage the authors to explain how this contribution differs from omitted-variable effects, model misspecification, and changing training-to-application relationships.
Section 6 identifies several reasonable directions, but I feel that their connection to the proposed uncertainty mechanism needs more work. This section is important and may require some additional effort for clarification. Here are some ideas from reading this section that may help the authors improve their proposal. A) Time series and phenology metrics can represent seasonal changes and potentially some environmental history, but they do not automatically identify reciprocal feedbacks. B) Local models and stratification by soil–landscape units can accommodate differences among settings, but require a defensible approach to defining regimes and handling transitions. I like the idea of soil-landscape units, but the authors should connect it more strongly to their proposal. C) Process-model outputs may provide information about unobserved patterns, but they could also introduce uncertainty from the process model itself due to a wrong mathematical interpretation of the physical world. D) Repeated monitoring can reveal temporal change, but I think that attributing that change specifically to feedbacks requires additional explanation. Therefore, I think it is important to help readers connect these ideas and sources of uncertainty to proposed solutions.
Finally, it could be useful to provide a clearer conceptual or empirical example. This effort could connect the proposed mechanism to a practical DSM problem. For example, it could identify the target soil property, relevant covariates and history, the circumstances under which a conventional model becomes unreliable, the additional information or model structure needed, and an appropriate validation design. It could also explain how to distinguish the contribution of feedbacks from alternative explanations. I think this effort would help readers connect the ideas and see a clear explanation/application of the proposed actions. This will help clarify how an explicit focus on feedback adds to established DSM methods and would make the manuscript’s contribution more actionable.