the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 28 Jan 2026
Soil science-informed neural networks for soil organic carbon density modelling under scarce bulk density data
Abstract. Soil organic carbon (SOC) density is a key variable for quantifying soil carbon stocks, yet its modelling is challenged by sparse and inconsistent measurements of bulk density and coarse fragments relative to SOC content. Conventional digital soil mapping approaches typically model SOC density as a single target variable, thereby underutilising abundant SOC content data and overlooking physical relationships among soil properties. This study evaluates a soil science-informed neural network for SOC density prediction that explicitly constrains the SOC–BD relationship, and compares it with univariate and multivariate neural network architectures. Across sparsely sampled target variables, including SOC density, bulk density, and coarse fragments, the soil science-informed model achieves comparable or slightly improved prediction accuracy relative to multivariate and univariate models. Although it yields lower accuracy for SOC content, the soil science-informed model better preserves physically plausible SOC–BD joint distributions and generates smoother, more temporally stable SOC density trajectories. Overall, the results demonstrate that incorporating soil physical constraints into machine learning models adds value beyond univariate accuracy, improving robustness, plausibility, and temporal coherence of SOC density predictions under sparse data conditions. Moreover, the latent parameters inferred by the soil science-informed model improve model interpretability and offer additional soil science relevant insights beyond predictive accuracy.
- Preprint
(4454 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 19 Jun 2026)
- RC1: 'Comment on egusphere-2026-229', Anonymous Referee #1, 23 Feb 2026 reply
-
RC2: 'Comment on egusphere-2026-229', Anonymous Referee #2, 08 Jun 2026
reply
Overall assessment: The manuscript addresses an important challenge in DSM, i.e. estimating SOC density when bulk density and coarse fragment observations are scarce. The proposed SiNN is an interesting attempt to introduce soil knowledge into a ML framework through explicit SOC-BD relationships. The manuscript is generally well written and the idea is relevant to current efforts in physics-informed and process-informed ML for environmental sciences. However, the current manuscript does not yet fully demonstrate that the proposed architecture provides substantial advantages over simpler alternatives, nor does it fully establish the significance of the inferred latent parameters.
Major comments
Introduction: Authors propose a soil science-informed neural network, but the distinction between "soil science-informed", "process-informed", "physics-informed", and "mechanistic" modelling remains unclear throughout the paper.The proposed framework appears to introduce prior soil relationships through latent variables and constraints rather than through underlying physical equations.
I encourage the authors to more clearly position SiNN within the broader literature on process-informed and physics-informed ML and clarify whether the main contribution is the incorporation of a soil mixing relationship into a multitask NN architecture.
Authors should clarify (1) what makes SiNN fundamentally different from standard process/physics-informed options and (2) is the contribution primarily the embedding of a soil mixing equation?. Currently, the manuscript appears closer to a constrained multi-task NN than to a genuinely physics-informed framework.
It is difficult to determine whether the methodological innovation lies primarily in the latent parameterisation, the multitask learning strategy, or the incorporation of the SOC-BD relationship.
Section 2.2: The manuscript includes 362 covariates from 15 groups. If the objective is to improve scientific interpretability through soil-informed modelling, why is the model driven by such a large and complex predictor space? Were there some form of importance analysis, feature attribution or group-level covariate importance. Otherwise, interpretability gains from latent soil parameters may be offset by opacity in the predictor space.
Section 2.3: The current formulation appears closer to a multitask NN with embedded empirical constraints than to a genuinely process-informed or physics-informed framework.
Section 2.4: The manuscript reports performance metrics but does not provide information regarding fold-to-fold variability, confidence intervals, or statistical significance of differences among models. Given that many of the reported performance gains are relatively small, additional statistical evidence would help determine whether the observed differences exceed expected cross-validation variability.
Section 3: The reported improvements in SOC density prediction are marginal. The manuscript argues that plausibility and temporal consistency are the main benefits of SiNN, which may be true, but this shifts the contribution away from prediction accuracy. I recommend reportng explicitly on quantifying effect sizes of the improvements, significance of differences between models and variability across folds. Without those, it appears unclear whether the observed differences exceed normal cross-validation variability.Latent parameters remains speculative. A key claim is that latent parameters of oBD and mBD across land covers improve interpretability. However, interpretability is asserted rather than demonstrated. Do latent parameter estimates correspond to known soil processes and the estimated mBD values realistic across soil types? like the estimated oBD values, while assessing spatial coherency between them
Section 3.3: SOC changes can be abrupt or gradual depending on land-use change or peatland disturbance. Smoother predictions could arise simply from stronger regularization. Authors should discuss this distinction more carefully and avoid presenting smoothness as inherently desirable without validation against known temporal dynamics.
Section 4.2: It is argued that inferred oBD and mBD values provide additional insight into soil physical processes and improve interpretability. However, because no independent observations of these latent parameters are available, most interpretations remain hypothetical. The manuscript would be strengthened by demonstrating that estimated values fall within expected ranges across soil types and by providing additional evidence that the inferred spatial patterns correspond to known pedological gradients.
Section 4.3. The discussion links spatial patterns in latent parameters to land-cover and land-use effects. While these interpretations are plausible, they are currently not independently validated.
Similar concerns apply to the subsequent derivation of porosity and compaction indicators. The manuscript presents these products as evidence of added soil-science value, yet they are derived from latent variables that themselves have not been independently validated.
The manuscript states that latent parameter estimates enhance interpretability and improve transparency. This conclusion currently appears stronger than the presented evidence supports in the paper.
Section 4.4: I appreciate that the authors acknowledge the lack of direct validation for temporal SOC density changes and discuss uncertainty quantification as future work. However, uncertainty may be particularly important in this study because the predictive gains achieved by SiNN are relatively modest. Demonstrating uncertainty reduction could potentially provide a stronger justification for the proposed framework than improvements in predictive accuracy alone.
All three competing approaches are NN variants. This creates a somewhat narrow benchmark. For DSM applications, readers would likely expect comparison with other established methods such as tree-based options togther with NN or otherwise. If computational constraints prevent full benchmarking, the authors should justify why NN only comparisons are sufficient.
Throughout the manuscript, "soil science-informed", "physically constrained", and "mechanistic" appear to be used somewhat interchangeably. These concepts are not equivalent, mechanistic models represent causal processes, physical constraints enforce admissible solutions while soil-informed models may simply encode empirical relationships.
Citation: https://doi.org/10.5194/egusphere-2026-229-RC2
Model code and software
EasyDensity Xuemeng Tian et al. https://github.com/AI4SoilHealth/EasyDensity.jl
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,220 | 740 | 100 | 2,060 | 113 | 152 |
- HTML: 1,220
- PDF: 740
- XML: 100
- Total: 2,060
- BibTeX: 113
- EndNote: 152
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
I have read the manuscript describing soil informed neural networks to predict SOC density under sparse auxiliary data by leveraging multivariate learning and a soil-relation–informed ML.
The topic is timely for the DSM/ML community, and the paper’s central idea, using constraints to improve plausibility and robustness under missing BD/CF and is potentially valuable.
However, there are several issues in the mathematical formulation, unit consistency, evaluation design, and soil-science framing that needs attention.
I outline some comments below.
-L.21 The Introduction implies SOC density as a DSM-driven stakeholder target. In practice: carbon accounting uses SOC stocks per area (depth-integrated), not only DSM.
- Depth handling (0–20 cm) and LUCAS 2018 exclusion, The study excluded 0–10 and 20–30 cm to focus on 0–20 cm. need to explain: whether 0–10 and 10–20 exist and why not used,
- Unit consistency and dimensional correctness in SOC density (Eq. 1) with SOC content in g kg⁻¹, BD in g cm⁻³, and output in kg m⁻³.
While the formula can be numerically correct if the implicit conversions cancel , it is not dimensionally transparent and is easy to misapply. I suggest Rewrite Eq. (1) with explicit conversion constants, or (b) define the variables as mass fraction and kg m⁻³ explicitly.
Also motivation says stakeholders care about “SOC density” more than content. In practice, stakeholders commonly want SOC as well as SOC stock per area (e.g., Mg ha⁻¹ for a depth interval). Please temper the claim and clarify the end-use context (accounting, monitoring, agronomy, reporting).
- SOC–BD mechanistic constraint (Eq. 2) and SOM=1.724⋅SOC content
this is incorrect, because SOC content is in g kg⁻¹ (not a fraction), so SOM becomes order 10–100+ (dimensionless), which makes (1−SOM) negative.
The Federer reference is not the origin of the equation. This equation of mixing is due to Adams WA. 1973. The effect of organic matter on the bulk and true densities of some uncultivated podzolic soils. J Soil Sci, 24 (1973), pp. 10-17
- Coarse fragments. In Lucas, CF is measured in mass basis, how did you convert to volume basis as in Eq 2
- L.85 The study promotes soil-informed ML, but at the same time uses 362 covariates from 15 groups, many of which are highly correlated, multi-scale, and partially redundant. That creates a tension between the stated philosophy and the modelling design.
- Cross-validation design likely suffers from leakage (repeated sites + spatial autocorrelation). It stated five-fold CV via random partitioning. But the dataset contains repeated measurements at the same sites across years. Random folds will almost certainly place the same site in both train and test folds (even if different years), inflating performance and plausibility diagnostics.
Additionally, DSM with dense covariates typically demands spatially blocked CV (or at least spatial buffering) to avoid optimistic estimates.
Use grouped CV by site ID so all time points for a site stay in one fold. Ideally, combine with spatial blocking (e.g., spatial k-fold) to reflect mapping/generalisation performance.
- Targets are transformed to reduce skewness, and constrain to [0,1], achieved through “log transformation and scaling using a standard scaler.” A standard scaler does not constrain to [0,1]; it standardises to mean 0, variance 1. Correct the description. Coul be Min–Max scaling?
- SOC density can only be “truly” validated where BD and CF exist, and here they exist only in 2018. The study claims robustness under sparse BD availability. But have not rigorously validated the “sparse BD reconstruction” claim. It just means better internal consistency + smoother time series. But not neccessarily correct reconstruction when BD is truly absent.
- Temporal consistency filter appears logically inconsistent (likely a typo or mis-specified threshold). It wrote assume SOC changes < 0.5 g kg⁻¹ yr⁻¹, but use a “conservative threshold of 50 g kg⁻¹ yr⁻¹ for the maximum absolute difference across measurements.” This is confusing . Justify with citations and show sensitivity (how many series removed under alternative thresholds).
- Some reported units and plausibility statements are incorrect. Example: “extreme changes exceeding 60 g cm⁻³” for SOC density trajectories. SOC density is in kg m⁻³ (or equivalently g L⁻¹), not g cm⁻³.
- Small sample sizes make stratified metrics unreliable (Table 3). In Table 3, Wetland has N = 2 but reports R² = 0.90. This is not meaningful.
-MSE and R² are fine, but heavy tails and log transforms can distort interpretation. Add residual analysis (bias by SOC quantiles, BD quantiles)
For the joint SOC–BD space: consider distance metrics or coverage metrics i.e, how much predicted mass lies outside observed support.
- Uncertainty quantification is missing, since “gains are modest”, uncertainty reduction may be the key value proposition. Provide at least one uncertainty (ensembles, MC dropout, deep ensembles) for SOC density maps.
L.109 eq 2 is not mechanistic, still empirical