the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 Aug 2026
Spatially explicit seasonal modelling reveals the impact of vegetation–soil coupling on soil organic carbon prediction in bimodal-rainfall agroecosystems
Abstract. Soil organic carbon (SOC) underpins soil fertility, climate regulation, and ecosystem resilience. Yet high-resolution SOC data remain scarce in many African agroecosystems. In bimodal rainfall systems, phenological phases may imprint distinct spectral signals associated with SOC processes. However, these temporal patterns are rarely exploited in digital soil mapping. We hypothesise that SOC spatial variability is most detectable when vegetation–soil coupling is strongest and that geographically weighted machine learning enhances the ecological interpretability of SOC models in data-scarce settings. To test these hypotheses, we develop a spatially explicit, seasonally-informed framework for SOC prediction in bimodal-rainfall agroecosystems. Using environmental covariates and 11-year seasonal Landsat 7 and 8 composites to capture the full phenological cycle in Tanzania (March–May 'Long Rains', June–August 'dry season', October–December 'Short Rains'), we compare global Random Forest (RF) with Geographical Random Forest (GRF) across each seasonal window, using an annual composite as our climatological baseline.
Feature selection using Boruta reduced the covariate set by 34–50 % per temporal window without compromising accuracy. Season-specific GRF models achieved R² values of 0.292–0.405 based on spatial cross-validation, whereas random cross-validation inflated performance to 0.444–0.471 (average overestimation = 19 %). October–December covariates provided the highest predictive power using only 13 predictors. Across the temporal windows, near-infrared reflectance, green band (October–December), red band (March–May), and topsoil clay emerged as the most informative predictors. SHapley Additive exPlanations (SHAP) values extracted from the global RF were represented spatially to map the directional influence of these predictors on global predictions and subsequently assessed alongside local feature importance maps to evaluate the extent of spatial non-stationarity. This two-stage interpretability approach confirmed that the GRF models adapted dynamically to localised eco-pedological realities, such as soil texture-driven storage capacity and canopy-driven phenological masking. By demonstrating when, where and how SOC is most observable, the framework provides an operational blueprint for local organisations to design leaner, seasonally targeted soil surveys and remote sensing protocols that strengthen climate smart land management policies and national carbon accounting systems.
- Preprint
(5311 KB) - Metadata XML
-
Supplement
(4602 KB) - BibTeX
- EndNote
Status: open (extended)
-
RC1: 'Comment on egusphere-2026-4622', Anonymous Referee #1, 25 Sep 2026
reply
-
CC1: 'Reply on RC1', Ana Cristina Costa, 04 Oct 2026
reply
Thank you very much for your thoughtful and positive feedback on our manuscript. The comments provided improve the clarity, quality, and contribution of our work. We are preparing a point-by-point reply to address all comments and revising the manuscript accordingly.
We reformulated the first two operative questions:
- We replaced "When to sample?" with “When is SOC best predicted?”.
- We replaced "What to measure?" with “Which covariates predict SOC best?”.The three questions now map cleanly onto the actual analytical contributions: seasonal timing → covariate selection/importance → modelling strategy. Most importantly, the first question asks when the available seasonal satellite information enables the most accurate SOC prediction, which is what our analysis actually evaluates. Moreover, the second question avoids claiming that SOC itself is directly sensed by satellite and accurately describes our comparison of season-specific remote-sensing covariates.
Citation: https://doi.org/10.5194/egusphere-2026-4622-CC1 -
AC1: 'Reply on RC1', Annette Achieng', 06 Oct 2026
reply
We would like to thank the reviewer for the constructive comments which lead to an improved manuscript. Besides reformulating the first two operative questions, we addressed all comments as follows.
The distinction between GRF and a conventional RF has been clarified in the Introduction section of the revised manuscript: “While conventional global models, such as the standard RF, fit a single model to the entire study area and therefore assume that predictor–response relationships are spatially stationary, GRF accounts for spatial heterogeneity by fitting location-specific models using observations from surrounding neighbourhoods (Georganos et al., 2021; Xiao et al., 2026). This allows the relationships between SOC and its environmental covariates to vary across space, potentially capturing local controls on SOC that a global RF model may overlook.” Furthermore, section 2.5.2 (previously, 2.4.2) in the Methodology now explains better how PyGRF adapts to local environmental conditions to address spatial non-stationarity:
“Whereas global RF assumes a uniform target–predictor relationship across the landscape, GRF explicitly models spatial non-stationarity by calibrating local RF models using nearby observations within a specified spatial bandwidth. Predictions from the global model and the local sub-models are then combined through a weighting scheme that accounts for spatial proximity. The bandwidth hyperparameter 𝜆 defines the number of neighbours contributing to each local model, while the local weight hyperparameter 𝛼 ranges from 0 to 1, assigning greater importance to the local model as values approach 1 and increasing the contribution of the global model as values approach 0. Sun et al. (2024) proposed using the Incremental Spatial Autocorrelation (ISA) technique to determine the bandwidth 𝜆 and the local weight 𝛼 of PyGRF models. The rationale behind this is that a stronger spatial autocorrelation, which denotes a higher similarity among local soil observations, has a higher likelihood of resulting in local models that are more robust. This theory-driven approach is not necessarily intended to maximise predictive performance, but rather to reduce the need for exhaustive hyperparameter tuning by providing an informed starting point.”
Thank you for alerting us about using the term "eco-pedological zones". We have amended the manuscript accordingly and we no longer use the term "eco-pedological zones". Moreover, the term “eco-pedological relationships” was clarified at the beginning of section 2.5: “A multi-faceted approach was employed to interpret the spatial and temporal interactions between vegetation and soil processes (i.e., eco-pedological relationships) driving SOC predictions in northern Tanzania.”
All comments in the supplement PDF were addressed in the revised manuscript. The following is a short summary of the introduced changes. The page numbers below refer to the supplement PDF.
Abstract: We changed the following to explain what "vegetation–soil coupling" refers to: “Dynamic feedback between canopy growth and topsoil properties, hereafter termed “vegetation–soil coupling”, generate seasonal spectral patterns that can provide information on SOC variability. However, in bimodal rainfall systems, these seasonally varying vegetation–soil interactions remain largely underexploited in digital soil mapping.”
Phenological phases (line 53): We have rephrased the statement to better reflect our analysis and discussion. “First, when is SOC best predicted? Since the efficiency of feature-space sampling strategies depends on capturing a diagnostic covariate feature space (Castaldi et al., 2019a), identifying the season when vegetation–soil interactions provide the strongest expression of SOC variability in satellite-derived covariates can establish a more informative feature space for guiding sampling design. This maximises information yield per sample, and in turn increases map accuracy at a constant budget or reduces field sampling requirements for a given accuracy level.”
Interaction of climate and soil characteristics (lines 66-68): We have expanded the text to explicitly detail the biophysical and optical mechanisms governing the interaction between seasonal climate, phenological dynamics, and topsoil properties. Specifically, we elaborate on how seasonal rainfall triggers distinct phenological phases which dynamically alter surface reflectance and govern the spectral observability of SOC. The text now reads:
“Climatic factors such as seasonal rainfall generate distinct phenological phases, with pulses of biomass production and residue inputs that differentially influence SOC formation and stabilisation across depths (Piao et al., 2019; Van der Putten et al., 2013). In bimodal rainfall systems, the Long Rains (March–May) and Short Rains (October–December) drive distinct combinations of plant growth, residue retention and soil moisture conditions that dynamically alter surface reflectance and dictate the spectral observability of SOC (Lobell and Asner, 2002; Yang et al., 2015). This interactive biophysical feedback loop—whereby soil properties such as texture and moisture retention capacity regulate biomass production, while seasonal vegetation canopy changes serve as a proxy for underlying SOC accumulation and microbial activity—can be referred to as vegetation–soil coupling.”
Study area (line 147): The conceptual workflow has been moved to a new section after section 2.1 Study area. Please note that this change implied that Fig. 2 and Fig. 1 switched their numbering.
When to sample (section 2.2): We have revised the opening of Section 2.2 to clarify how both the covariates collecting process and the point extraction were used to answer the first design question:
“While static covariates (soil texture, pH, terrain, and long-term climate) were held constant, dynamic surface covariates (multispectral satellite bands, vegetation indices, and actual evapotranspiration (ET)) were stratified across four distinct temporal windows: March–May (Long Rains), June–August (dry season), October–December (Short Rains), and an annual climatological baseline. By overlaying the 2,732 georeferenced soil sampling points onto each seasonal raster stack, four parallel, seasonally stratified datasets were generated through point extraction. Training and evaluating predictive models independently on each point-extracted dataset enabled a direct, systematic comparison of model accuracy across temporal windows, thereby empirically identifying the optimal phenological window where vegetation–soil coupling maximises the spectral observability of SOC.”
Causality (lines 348-350): We have rephrased the statement to reflect what can actually be achieved by DSM: “Rather than estimating operational map accuracy, these geometries were deployed purely as diagnostic stress tests to examine the structural integrity of the target–predictor relationships identified in the temporal models.”
Fig. 2 and Fig. 3: We have added references to their panels. The caption of Fig. 3c was corrected: “Heatmap comparing R² scores across SCV methods”.
Table 5 (caption and lines 463-464): The caption has been updated to clarify what the results in the table show. “Table 5. Global Moran’s I statistics for residual spatial autocorrelation in season-specific PyGRF models fitted to the full dataset.”
PyGRF local FI (lines 546-547): Thank you for this observation. We agree that using terms like “eco-pedological zones” was imprecise, as PyGRF operates as a spatially continuous algorithm rather than using predefined discrete zonal boundaries. We have revised Section 3.2.2 to clarify that PyGRF adapts its predictive weights to the continuous local biophysical context: “The PyGRF local FI maps depict how the spatially explicit model addresses spatial non-stationarity by dynamically adapting its predictive weights to the local context.”
Eco-pedological processes (line 549): We have amended the statement to refer to biophysical processes which can be verified using SHAP values: “Complementing this, the spatial SHAP maps reveal the geographical concentration of the global drivers, verifying that these localised adaptations remain grounded in valid biophysical processes.”
Fig. 10: this figure and the text referring to it in lines 684–686 were moved to section 3.2 (before sub-section 3.2.1) of the Results in the revised manuscript.
Citation: https://doi.org/10.5194/egusphere-2026-4622-AC1
-
CC1: 'Reply on RC1', Ana Cristina Costa, 04 Oct 2026
reply
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 151 | 76 | 81 | 308 | 64 | 69 | 69 |
- HTML: 151
- PDF: 76
- XML: 81
- Total: 308
- Supplement: 64
- BibTeX: 69
- EndNote: 69
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Thank you for the opportunity to review this manuscript. The study addresses relevant hypotheses and important objectives related to SOC modelling. Having these operative questions allows the reader to follow the story. However, the questions could be slightly reformulated to better reflect the results obtained. When to sense the carbon, which covariates to use, and how to model would better reflect the results, discussion, and conclusions obtained.
In addition, some concepts introduced in the Introduction would benefit from clearer definitions to fully convey the significance and context of the study. The implementation of the GRF model is a particularly novel aspect of the work and represents an innovative approach in soil mapping. However, its contribution would be more evident if the distinction between GRF and a conventional RF model were clearly introduced and explained at the beginning of the manuscript.
I also encourage the authors to be more cautious when interpreting the results in relation to eco-pedological zones. The term "eco-pedological zones" is not clearly defined in either the Introduction or the Methods section, which makes it difficult to assess the validity of some of the associated interpretations.
I have provided several additional comments in the attached PDF.
Overall, the manuscript is well written, clear, and concise, and it makes a valuable contribution to the field.