This is a review report for the manuscript, entitled: “Soil organic carbon projections and climate adaptation strategies across Pacific Rim Agro-ecosystems” by Syu et al.

Soil organic carbon (SOC) is highly relevant with C cycle and grain production. Thus, the responses of SOC to climate changes should be assessed for adaption management. This is a well-structed manuscript. The aims are clear and the methods are proper. This study applied a lot of associated methods (including machine learning and geostatistical approach) to evaluate the spatial-temporal SOC responses under different circumstances. So, the results are valuable for managers. Although the predictions can be estimated based on the environmental covariates under climate changes, I still hope the authors can consider or make an assumption regarding soil-forming time. Soil formation is a time-dependent process, and it appears that an inherent steady-state assumption has been made in this study, but not explicitly mentioned. Below are the specific comments.

Response:

We sincerely thank the reviewer for their kind suggestions and constructive comments, which have improved the structure, clarity, and quality of our manuscript.

Comments:

1. Line 28-33: In abstract, this sentence is clear, but too long. Please split it into two or more sentences.

Response: We thank the reviewer for pointing this out. We have revised the sentence to make it more concise.

Revised text: “Under SSP1-2.6, SOC stocks were projected to decline by up to 20.9%, especially in uplands, due to erosion driven by extreme rainfall (R95p, R99p). In contrast, SSP2-4.5 and SSP5-8.5 predicted SOC stock increases of 7.9% and 58%, respectively, particularly in mountainous areas where higher TNx and TXx enhanced productivity.” (Line 28-32).

1. Line 34: climate-topography interaction is an innovative finding in this ms. Try to clarify it in abs, not just mention the interaction.

Response: We thank the reviewer for pointing this out. We have revised the sentence to clarify the climate-topography interaction.

Revised text: “Partial least squares path modeling revealed a climate–topography interaction and explicitly quantified their contributions to SOC stocks, dominated by topography and followed by prolonged dry spells (CDD). This interaction is more pronounced in uplands than forested areas, where topography mitigates temperature extremes and their effects on SOC stocks. Extended CDD may decrease organic inputs by reducing vegetation growth and soil moisture, thereby enhancing carbon losses (Line 32-39).”

1. Line 46: Using “many scales” as the subject sounds a bit awkward.

Response: We thank the reviewer for pointing this out. We have revised the sentence to refine the sentence.

Revised text: “Multiple studies have examined whether carbon storage in agricultural soils can offset global warming, and various frameworks have been developed for evaluating the dynamics of SOC. Among these, the landscape scale has enabled researchers to consider the interplay between natural processes, human patterns, and SOC dynamics (Viaud et al., 2010) (Line 47-50).”

1. Line 62-83: This paragraph is quite wordy. The real point that the authors want to address is “meta-analytical evaluation” with a geostatistical approach. Please reorganize the paragraph and point out the points in this paragraph.

Response: We thank the reviewer for pointing this out. We believe that all previous studies and literature mentioned are essential to emphasize the approaches. Therefore, we have revised the sentence structure to highlight the focus on 'meta-analytical evaluation' using a geostatistical approach.

Revised text: “In digital soil mapping (DSM), soil properties in unsampled areas are predicted using statistical or machine learning models that relate soil observations to environmental factors (Grunwald, 2009). Spatial variability is key to model construction (Zhu and Lin, 2010). Two major approaches are (1) non-geostatistical methods based on the SCORPAN model (Jenny, 1941; McBratney et al., 2003), such as multiple linear regression (MLR), generalized additive models, Cubist, and random forest (RF) (Quinlan, 1992; Breiman, 2001); and (2) geostatistical methods accounting for spatial autocorrelation, including ordinary, simple, and universal kriging. Machine learning algorithms (e.g., MLR, RF, Cubist) are widely applied for mapping SOC content and stocks (Lamichhane et al., 2019; Siewert, 2018; Rudiyanto et al., 2018). Model performance depends on spatial scale, observation density), and terrain (Zhu and Lin, 2010; Tsui et al., 2016; Keskin and Grunwald, 2018). This study used regression kriging, a hybrid approach combining regression models (e.g., Cubist, RF) with spatial interpolation of residuals (Ma et al., 2017). Although regression kriging often improves RF and Cubist predictions, its advantage is not universal, highlighting the need for meta-analytical evaluations (Vaysse and Lagacherie, 2015; Ma et al., 2017; Lamichhane et al., 2019). (Line 64-78).”

1. Line 293: Why were the slope, aspect, and flow accumulation not taken into consideration?

Response: We thank the reviewer for pointing this out.  The covariate selection for the Cubist model was consistent with that of the Random Forest model, where all covariates were included in the analysis. The results indicated that aspect, curvature, and flow accumulation exhibited relatively low variable importance. Accordingly, we have revised the manuscript to clarify this point.

Revised text: “The results of the Cubist model indicated that the importance of aspect, curvature, and flow accumulation was relatively low, thus they exhibited either no usage or very low usage frequency (Fig. 3a). (Line 30-304)”

1. Materials and Methods: Consider making a figure to demonstrate all the data input and model processing.

Response: We thank the reviewer for the helpful suggestion. We have added a flow chart to demonstrate the methodology developed in this study, including data preparation, modelling, and prediction (Fig. S2).

Revised figure: The flow chart of methodology is illustrated in Fig. S2. (Supplementary file Line 6-8)

Fig. S2. Methodology flow chart of data preparation, modelling, and prediction developed in this study

1. In Fig. 5: The landscape region, in fact, is classified by elevation (<100m, 100m – 1000m, and >1000m). Therefore, “Forested region” is improper. It is quite confusing with the land cover. Besides, for the category of land cover in the x-axis, the “upland” should be upland farming.

Response: We thank the reviewer for pointing this out. To avoid any confusion regarding the title of the graph, we have changed the “Forested region” to “Mountainous region”. We also sincerely appreciate the reviewer’s suggestion; however, we believe that maintaining the current “upland” category in the x-axis could help the readers to see the figure better. Therefore, we added the additional sentence in the figure caption to clarify it.

Revised text: “Fig. 5. Boxplots of topsoil (0–30 cm) soil organic carbon (SOC) stocks for various land cover: (left) plain regions (<100 m in elevation), (middle) slopeland regions (100–1000 m in elevation), and (right) Mountainous regions (>1000 m in elevation) at Zhuoshui River watershed and Laonong River watershed. The upland land cover represents the upland farming” (Line 898-901).

1. Line 326: explain what is dry farming area? I can’t find it in fig. 5.

Response: We thank the reviewer for pointing this out. The dry farming was intended to emphasize the upland farming. We have changed the term “dry farming” to “upland farming” to avoid misunderstanding and help the reader to understand it better. 

Revised text: “As shown in Fig. 5, in the lowlands of the ZRW, the lowest average SOC stock was identified in upland farming areas (1.93 kg m⁻²), whereas the highest average SOC stock was identified in paddy fields (3.08 kg m⁻²) (Line 335-338)”.

1. Line 383-384: The inference (citation) should not appear in the result section, basically.

Response: We thank the reviewer for the helpful suggestion. We have moved all of the citations mentioned in the Results section to the Discussion section and checked again carefully.

1. Line 434-436: The same as above. The two citations are suggested to move to the discussion section, where the authors can compare their studied area with others.

Response: We thank the reviewer for the helpful suggestion. We have moved all of the citations mentioned in the Results section to the Discussion section and checked again carefully.

1. Line 522: The sentence is unclear to me.

Response: We appreciate the reviewer's attention to this matter. We have rephrased the sentence to improve clarity.

Revised text: “Furthermore, forests cover the majority of the landscape in mountainous areas and represent the primary reservoirs of SOC stocks (Line 533-535).”

1. Line 572-574: The authors should describe this limitation of the modeling work. Wildfires are important for soil formation and ecosystems.

Response: We thank the reviewer for this helpful suggestion. We have added a paragraph to further describe the limitation of the modeling work and improve the clarity of the manuscript.

Revised text: “Although the Cubist, Random Forest, and regression kriging models in this study demonstrated reasonable predictive performance, several limitations should be acknowledged. First, these models are data-driven and have limited extrapolation ability in regions beyond the range of the training data (Meyer and Pebesma, 2021). In addition, the uneven distribution of soil samples represents a major source of uncertainty; the scarcity of sampling points in mountainous areas results in higher prediction uncertainty (Jien et al., 2025). These limitations should be considered when interpreting the model results and the spatial distribution of SOC. (Line 602-608).”

Reference:

Meyer, H., & Pebesma, E.: Predicting into unknown space? Estimating the area of applicability of spatial prediction models. Methods in Ecology and Evolution, 12(9), 1620-1633, https://doi.org/10.1111/2041-210X.13650, 2021.

Jien, S. H., Minasny, B., Yang, B. J., Liu, Y. T., Yen, C. C., Ocba, M. A., ... & Syu, C. H.: Enhancing Soil Carbon Storage: Developing high-resolution maps of topsoil organic carbon sequestration potential in Taiwan. Geoderma, 459, 117369, https://doi.org/10.1016/j.geoderma.2025.117369, 2025.

1. Line 604: remote sensing parameters? Or “NDVI” is more specific.

Response: We appreciate the reviewer's attention to this matter. We have rephrased the sentence to improve clarity of the sentence.

Revised text: “This study established effective DSM models for predicting SOC stock, achieving an R2 range of 0.43–0.50, and identified key environmental covariates, such as topography, climate, NDVI, and the prediction interval maps for identifying areas not covered in the sampling distribution (Line 624-627).”.

1. Conclusion: So far, this conclusion is somewhat like a summary. Please try to draw some implications from this study in order to inspire the readers.

Response: We appreciate the reviewer's attention to this matter. We have revised the conclusion and emphasized more on the implications to inspire and facilitate better understanding by the reader.

Revised text: “This study established effective DSM models for predicting SOC stock, achieving an R2 range of 0.43–0.50, and identified key environmental covariates, such as topography, climate, NDVI, and the prediction interval maps for identifying areas not covered in the sampling distribution. The findings suggest that even with observed predictive disparities, the Regression Kinging with Cubist and RF models can still be relied upon for overall predictive ability. The results demonstrated spatio-temporal variability in projected topsoil SOC under different emission scenarios, with clear sensitivity to landscape type and climate extremes. Under a severe emissions scenario, SOC dynamics were highly sensitive to extreme climate events, with land type also playing a key role. These effects pose both location- and time-specific challenges for SOC management in studies on mid- to late-century time points. In the low-emission scenario (scenario SSP1-2.6), extreme rainfall events are predicted to induce a significant reduction in SOC stocks through erosion in upland areas. However, in the moderate- and high-emission scenarios (scenarios SSP2-4.5 and SSP5-8.5), warming (TNx and TXx) and extreme rainfall events (R95p and R99p) may simultaneously increase biomass input and increase soil erosion risks.

These findings indicate that SOC management strategies should be highly specific to the site and time. In forested areas (mountainous areas) of the ZRW and LRW, even though the SOC stock dynamics in forested areas are likely to be affected by extreme rainfall events, heat waves, and prolonged droughts, future mitigation strategies should focus on reducing warming, preventing wildfires, and promoting heat-tolerant tree species. In upland areas, SOC stock changes in ZRW and LRW are predicted to be mainly driven by R95p, R99p, and CWD, where significant SOC losses will occur in certain upland areas for all emission scenarios. Therefore, management strategies should emphasize soil and water conservation to ensure that excess rainfall can be infiltrated into the soil without triggering erosion. These strategies should include the implementation of eco-engineering techniques on slope lands, maintaining vegetation cover and soil permeability, and establishing effective drainage systems. Overall, clarifying the interactions between climatic extremes, land types, and SOC stocks to develop site-specific management practices is key to enhancing soil’s resilience to sustain ecosystem functions in a changing climate.”(Line 624-652).

This paper presents an important study integrating digital soil mapping (DSM) and CMIP6 climate projections to assess spatial–temporal SOC stock dynamics in two contrasting Taiwanese watersheds. The manuscript is generally well structured and provides a comprehensive analysis. However, some parts require improvement.

Response:

We thank the reviewer for their kind suggestions and constructive comments, which have improved the structure, clarity, and quality of our manuscript.

Comments:

1. Line 102: Please explain why the 0–30 cm soil depth was selected. Soil changes due to temperature and rainfall are generally most pronounced within the top 10 cm.

Response: We thank the reviewer for bringing this to our attention. Although climate-induced (temperature and rainfall) changes often show the strongest effects in the 0–10 cm depth, the 0–30 cm depth was chosen based on the Intergovernmental Panel on Climate Change (IPCC) (2006, 2019) recommendation. The 0–30 cm depth represents the primary processes influencing SOC dynamics, such as root activity, litter incorporation, and microbial decomposition, where the majority of accumulation and loss of SOC occur (Food and Agriculture Organization (FAO), 2019). Therefore, to improve the clarity of the manuscript, we have added the citation from IPCC in section 2.2 Soil samples and analyses.

Revised text: “A total of 901 topsoil samples (0–30 cm, based on IPCC (2006, 2019a) recommendation) were obtained (Line 129-130).”

References:

FAO. Measuring and modelling soil carbon stocks and stock changes in livestock production systems: Guidelines for assessment (Version 1). Livestock Environmental Assessment and Performance (LEAP) Partnership. Rome, FAO. 170 pp, 2019.

IPCC. 2006 IPCC Guidelines for National Greenhouse Gas Inventories: Agriculture, Forestry and Other Land Use. Volume 4, https://www.ipcc-nggip.iges.or.jp/public/2006gl/vol4.html (last access: October 22, 2025), 2006.

IPCC. 2019 Refinement to the 2006 IPCC Guidelines for National Greenhouse Gas Inventories: Agriculture, Forestry and Other Land Use. Volume 4, https://www.ipcc-nggip.iges.or.jp/public/2019rf/vol4.html (last access: October 22, 2025), 2019.

1. Line 124: In Figure S1a, the legend for the colours is missing.

Response: We appreciate the reviewer’s comment. The different colors in Figure S1a represent township boundaries. However, since the main purpose of this figure is to illustrate the spatial distribution of sampling points, we will remove the background colors to avoid distraction from the main focus of the figure (Line 125).

Revised figure: please see attached file.

Fig. S1. The sampling sites (a), land cover (b), mean annual temperature (c), and total annual precipitation (d) of the Zhuoshui River watershed (ZRW) and the Laonong River watershed (LRW). (Supplementary file Line 1-4) 

1. Line 130: It would be helpful to indicate size in millimetres (mm).

Response: We appreciate the reviewer's helpful suggestion. We have revised the sentence and included the 35-mesh screen soil sieve opening in mm (0.5 mm).

Revised text: “After the samples had been air-dried at room temperature, they were sieved through a 35-mesh screen (0.5 mm sieve opening) and stored in plastic containers (Line 125-126).”

1. Line 134: Please provide the full name for the abbreviation TOC.

Response: We appreciate the reviewer's helpful suggestion. We have revised the sentence and included the abbreviation of TOC (Total Organic Carbon).

Revised text: “Because the LOI method typically overestimates SOC (Li et al., 2021), a correction function was applied to adjust SOC content from LOI values to those obtained using a total organic carbon (TOC) analyzer (solid TOC cube, Elementar) (Line 128-130).”

1. Line 170: More information is needed on how the resolution was changed from 1 km to 20 m.

Response: We appreciate the reviewer's attention to this matter. The climatic variables, including mean annual temperature and total annual precipitation, were originally at a spatial resolution of 1 km. To match the spatial scale of other covariates, these raster layers were resampled to 20 m resolution using the resample function from the raster package in R (Hijmans, 2022), with bilinear interpolation (method = "bilinear"). (Line 169-169)

Reference:

Hijmans, R. J. raster: Geographic data analysis and modeling. https://doi.org/10.32614/CRAN.package.raster, 2022.

1. Line 171: The land-cover class and soil order variables are categorical. Were these treated as factors or numeric data?

Response: We thank the reviewer for bringing this to our attention. The variables for land-cover class and soil order are categorical and were treated as factors in the analysis. This approach ensures precise management of qualitative differences in the modeling process.

1. Line 177: It would be clearer to move Section 2.4 ("Climate data in various emission scenarios and with extreme climate indices") to the end of the Methods

Response: We appreciate the reviewer's helpful suggestion. We have moved original Section 2.4 to the end of the Methods as Section 2.7. (Line 252-269).

1. Line 214: Please clarify why 20 committees were used for the Cubist model.

Response: We thank the reviewer for pointing this out. We have added a paragraph to clarify the reason 20 committees were used for the Cubist model. For further understanding, please refer to the attached figure below.

Revised text: “We used the caret package (Kuhn, 2008) to perform hyperparameter tuning for the Cubist model, testing committee values of 1, 5, 10, 15, 20, and 25. Using 5-fold cross-validation, caret automatically evaluated the performance of each hyperparameter setting based on the root mean square error (RMSE). The tuning results indicated that setting committees = 20 produced the lowest cross-validation RMSE, suggesting that this configuration achieved the highest predictive accuracy. Therefore, committees = 20 was selected as the final model parameter.” (Line 194-200)

Reference:

Kuhn, M.: Building Predictive Models in R Using the caret Package. Journal of Statistical Software, 28(5), 1–26. https://doi.org/10.18637/jss.v028.i05, 2008.

1. Line 224: Please explain the rationale for using mtry = 7 and ntree = 500 in the Random Forest model.

Response: We thank the reviewer for bringing this to our attention. We have added a further explanation to the rationale for using mtry = 7 and ntree = 500 in the Random Forest model. Please refer to the revised text in the revised manuscript as well as the attached figure below for further understanding.

Revised text: “We used the caret package (Kuhn, 2008) to perform hyperparameter tuning for the Random Forest model. The parameter mtry was tested with values ranging from 2 to 9. Using 5-fold cross-validation, caret automatically evaluated the performance of each hyperparameter setting based on the root mean square error (RMSE). The tuning results indicated that the model achieved the lowest cross-validation RMSE when mtry = 7. The number of trees (ntree) was kept at the default value of 500, which is generally sufficient to ensure model stability (Peng et al., 2025). Therefore, the final Random Forest model was trained using mtry = 7 and ntree = 500.” (Line 209-216).” (Line 209-216)

Reference:

Peng, Y., Zhou, W., Xiao, J., Liu, H., Wang, T., & Wang, K.: Comparison of Soil Organic Carbon Prediction Accuracy Under Different Habitat Patches Division Methods on the Tibetan Plateau. Land Degradation & Development. https://doi.org/10.1002/ldr.70184, 2008.

1. Line 230: The sentence “The distribution of the two data sets is depicted in Fig. 2.” should be moved to the Results

Response: We appreciate the reviewer's helpful suggestion. We have revised the sentence and moved it to the Results section to clarify the distribution of the training data set and validation data set described in Section 3.2.

Revised text:

“3.2 Model performance in SOC stock prediction

This study constructed SOC stock predictive models using the Cubist, RF, and regression kriging with the training data set and environmental covariates. The performance of these models was evaluated using R2 and RMSE values.  Among the evaluated models, the RF model demonstrated the highest predictive performance in the training data set. The distribution of the training data set (calibration set) and validation data set is depicted in Fig. 2. Therefore, we focused on model performance in the prediction of validation data set prior to model selection. In this respect, the performance indicators of the Cubist model were R2 = 0.43 and RMSE = 0.45 kg m-2, while those of the RF model were R2 = 0.46 and RMSE = 0.43. After incorporating regression kriging, the indicators improved to R² = 0.48 and RMSE = 0.42 for the Cubist model, and remained at R² = 0.46 and RMSE = 0.43 for the RF model (Fig. 2).” (Line 282-291)

1. Line 272: Replace “Coefficient of determination (R²)” with simply R².

Response: We thank the reviewer for bringing this to our attention. We have rephrased the sentence and replaced “Coefficient of determination (R²)” with R².

Revised text: “The performance of these models was evaluated using R² and RMSE values (Line 283-284).”

1. Lines 294–296: This section requires further explanation, as it is currently difficult to understand.

Response: We appreciate the reviewer's helpful suggestion. We have revised the structure of the sentence to help the reader understand it better.

Revised text: “The results of the Cubist model indicated that the importance of aspect, curvature, and flow accumulation was relatively low, thus they exhibited either no usage or very low usage frequency (Fig. 3a). Among the covariates included, more than half of the data incorporated covariates such as elevation (98%), annual mean temperature (62%), NDVI (59%), TRI (54%), K-value (53%), and slope (52%). These results indicated that climatic and topographic factors strongly contributed to model performances. In summary, the RF and Cubist models identified soil order, elevation, and annual mean temperature as the factors representing the influence of soil, topography, and climate, respectively, on the SOC stock in the study areas. (Line 301-309).”

1. Line 301: When creating the SOC map, did you use only the 70% training data or the entire dataset (100%)?

Response: We thank the reviewer for pointing this out. The SOC map was created using 70% of the training data set. The remaining 301 samples (30%) were used as the validation data set (validation set) to determine the model’s predictive performance. We have also described that in Line 219-221.

1. Lines 335–345: This section should be moved to “3.7 Extreme climate index parameter estimates in three emission scenarios.”

Response: We thank the reviewer for the helpful suggestion. We have moved the paragraph to Section 3.7 to better emphasize the results under three emission scenarios and facilitate better understanding by the reader.

Revised text: “In all emission scenarios, major spatial heterogeneity and temporal increases were found in SOC stocks (Table 3, Figs. 6 and 7), particularly under high-emission conditions. These findings underscore the importance of modifying the management practices of land use in the future, especially if climate change is severe. In forested areas in both watersheds, significant SOC accumulation was predicted. Areas with an SOC accumulation value of >15 Mg C ha⁻¹ were expected to exhibit an increase in SOC accumulation from <5% (2020, baseline) to more than 25% by 2100 in scenario SSP5-8.5. By contrast, lowland agricultural zones are expected to maintain relatively low SOC stocks (<9 Mg C ha⁻¹), with minor gains across scenarios. Scenario SSP5-8.5 was found to result in the greatest projected increase in SOC stocks as a result of elevated CO2 and potential biomass input, although spatial disparities are expected to increase, particularly in erosion-prone or intensively cultivated lands (Fig. S3).”  (Line 353-363).

1. Line 481: It would strengthen the discussion to compare the SOC maps produced in this study with existing SOC maps from other publications.

Response: We thank the reviewer for the helpful suggestion. On the Section 4.2 Effects of environmental covariates on SOC stocks, we have included citations from previous studies regarding SOC maps to compare and support the results of our study. Therefore, this section remained as it was (Line 493-517).

1. Figure 3: there are two “fig 3”, so remove one. Most samples appear concentrated in croplands, and future work could include a more balanced sampling across different land types (e.g., forest).

Response: We thank the reviewer for bringing this to our attention. We have deleted one of the “Fig. 3” in the figure caption (Line 890-892). We also appreciate the reviewer’s suggestion regarding future work. We believe that including a more balanced sampling across different land types will provide more comprehensive results. We will take this suggestion into consideration for our future study.

1. Figure 6: Please specify which climate scenario (e.g., CWD) is displayed.

Response: We thank the reviewer for bringing this to our attention. The Figure 6 shows the spatiotemporal predictions of SOC stocks (kg m⁻²) and SOC sequestration rates (kg m⁻² per year) relative to the 2020s under three emission scenarios. The climate scenarios were list on the top of the figure.