Satellite Hyperspectral Mapping of Farmland Soil Organic Carbon in Yuncheng Basin Along the Yellow River, China

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Consider elaborating in slightly more detail the argument for why environmental variables themselves are important for capturing subtle changes in SOC and how hyperspectral data addresses this limitation, although it is already noted.

All three main conclusions are consistently and directly supported by the results presented in the respective “Results” sections and in the accompanying tables and figures. The article contains two tables labeled Table 2 (rows 244 and 355). The other tables will then need to be renumbered in the headings and in the text.

Overall, the presentation of the results is well structured, visually compelling and allows the reader to easily understand the key findings of the study. Some of the figures are documented in poorer and poorly legible quality. For example Figure 3, or some legends in Fiures 7 and 10 (f).

There is a lack (maybe I just overlooked it) of detail on how the sampling points were geographically distributed (e.g. systematic, random, stratified) and what the total number of samples taken was.

Author Response

Comment 1: Consider elaborating in slightly more detail the argument for why environmental variables themselves are important for capturing subtle changes in SOC and how hyperspectral data addresses this limitation, although it is already noted.

Reply 1: Thank you. We have made the additions .（lns 42-46, 82-88）

 

Comment 2: All three main conclusions are consistently and directly supported by the results presented in the respective “Results” sections and in the accompanying tables and figures. The article contains two tables labeled Table 2 (rows 244 and 355). The other tables will then need to be renumbered in the headings and in the text.

Reply 2: We apologize for our negligence. We have thoroughly checked and revised Table 3. （lns365, 371）

 

Comment 3: Overall, the presentation of the results is well structured, visually compelling and allows the reader to easily understand the key findings of the study. Some of the figures are documented in poorer and poorly legible quality. For example Figure 3, or some legends in Fiures 7 and 10 (f).

Reply 3: Thank you very much for your kind recognition. We have revised the following picture again. Figure 3.4.6.7.9.10 (a-f)

 

Comment 4: There is a lack (maybe I just overlooked it) of detail on how the sampling points were geographically distributed (e.g. systematic, random, stratified) and what the total number of samples taken was.

Reply 4: Thank you very much. We have made the necessary revisions. The sampling points were arranged based on the operability of local conditions, taking into account factors such as altitude changes, land cover, and transportation accessibility for collecting soil samples. Additionally, soil sampling points were appropriately densified in fragmented farmland areas and areas with relatively complex terrain (such as mountainous regions), while this also led to a slightly sparse distribution of sampling points in relatively flat terrain areas. In total, 312 soil samples were collected throughout the study area, with the sampling points distributed approximately every 10,000 acres. The main purpose of this was to make the SOC model in complex areas more accurate. （lns 42-46, 130-138）

Reviewer 2 Report

Comments and Suggestions for Authors

Lines 41–42 state “According to the soil‑forming factors theory…,” yet the original works in which this theory was first proposed are not cited:

• Dokuchaev, V. V. 1883. Russian Chernozem. (English translation: Principles of Soil Science, 1899).
• Jenny, H. 1941. Factors of Soil Formation: A System of Quantitative Pedology. McGraw‑Hill, New York.

Moreover, the superscript “2” in “factors theory 2” presumably refers to a literature citation, but it should be formatted according to the journal’s style as [2]. In addition to the five canonical soil‑forming factors, it would be appropriate to acknowledge a sixth factor—anthropogenic influence.

The traditional Dokuchaev paradigm indeed presents a simple flow: factors → properties. Early soil science assumed that the five factors fully explain soil properties. However, empirical observations have shown that identical combinations of soil‑forming factors can yield soils with different properties, and conversely that different factor combinations can produce very similar soils. To resolve such “hard” diagnostic problems, Merkurii Gilyarov developed the zoological diagnostic method of soils. This led to the non‑Dokuchaev paradigm: factors → processes → properties, with an explicit system of elementary soil‑forming processes. Theoretically, this paradigm is nearly flawless, but in practice identifying these elementary processes is extremely difficult, if not impossible. This note is intended to caution against overemphasizing the outdated factors → properties paradigm.

Lines 42–44: the sentence

“Digital soil mapping (DSM) leverages these environmental covariates to effectively predict and spatially map soil types and properties at various scales”
requires clarification, since not all of the mentioned covariates—particularly biotic and temporal ones—have standardized spatial metrics and therefore cannot be directly mapped.

Comment on lines 84–86:

“The factors that influence the spatial distribution of SOC in farmland areas located along the middle reaches of the Yellow River remain unclear. Therefore, further research is required.”
This statement conflicts with the earlier claim that the classical soil‑forming factors (parent material, climate, organisms, topography, time) suffice to explain the spatial distribution of soil properties.

Comment on lines 90–92: in the phrase

“Integrate satellite hyperspectral data, natural factors, and farmland management environmental variables into a random forest (RF) model…”
terms from different conceptual levels are mixed: “satellite hyperspectral data” (a direct surface property), “natural factors” (an undefined concept, presumably the classical factors), and “farmland management environmental variables” (a group of agronomic indicators). To maintain logical consistency, I suggest grouping them uniformly as “input predictors” and defining each category explicitly.

Positive note on the Study Area section: it is clearly and systematically presented, with geographic location, climatic characteristics, and soil conditions described, providing a solid foundation for subsequent analysis.

Comment on lines 112–114:

“The predominant soil type in this region is cinnamon soil, classified as Cambisols in the World Reference Base for Soil Resources (WRB).”
Please cite the full source for the WRB classification:
IUSS Working Group WRB. 2015. World Reference Base for Soil Resources 2014, update 2015. World Soil Resources Reports No. 106. FAO, Rome.
Also add a reference (research article or report) or present your own field data to substantiate that cinnamon soil/Cambisols dominate this region.

Comment on the caption of Figure 2: it lacks critical metadata—no mention of which satellite acquired the image or the acquisition date. I recommend augmenting the caption with the satellite platform (e.g., Sentinel‑2A) and the date of acquisition to ensure completeness and reproducibility.

Comment on lines 116–118 and Figure 1: the authors state that “312 sampling points were uniformly established across the study area (approximately one point per 10 000 mu, ~667 ha),” yet Figure 1 shows clusters of points and large gaps, contradicting the claim of uniform sampling. Please clarify the sampling algorithm or spatial deployment rules, or adjust the description to reflect the actual distribution.

Comment on the Methods section and Table 1: the text indicates that topographic factors (slope, aspect) were derived from a USGS DEM at 30 m × 30 m resolution (and earlier mentions an ISRIC DEM at 250 m × 250 m), while Table 1 lists a 1 km resolution for slope and aspect. Please verify and harmonize the DEM source and spatial resolution across the text and table to avoid inconsistencies.

Comment on Table 1: the “Environmental Variable” column uses abbreviations (MAT, MAP, EVA, NDVI, NPP, SOS, EOS, Irr, Dra, NF) without providing their full names or measurement units (e.g., °C for MAT, mm for MAP, index values for NDVI). I recommend adding the full variable names and units directly in the table or as a footnote.

Comment on Table 1, under the “Vegetation” category: variables NDVI, NPP, SOS, and EOS are spectral indices sensitive not only to vegetation status but also to soil background, moisture, and atmospheric conditions. Please include a note on these methodological limitations, and if possible describe any atmospheric or soil‑background correction procedures applied.

Comment on the Methods section listing data sources and resolutions (e.g., DEM 30 m × 30 m, SoilGrids 250 m × 250 m, NF 5 km): the paper does not describe how disparate spatial resolutions were harmonized before extracting predictor values at sampling points. It is unclear whether all layers were resampled to the finest (30 m), coarsest (5 km), or an intermediate resolution (e.g., 1 km), nor which resampling method (nearest neighbor, bilinear, majority) was used. Without this information:

1. Using both high‑resolution 30 m layers and highly aggregated 5 km layers may introduce systematic bias in SOC predictions at fine scales.
2. The analysis cannot be reproduced, since the spatial accuracy of matching 5 km × 5 km data to GPS sample coordinates is unspecified.
3. A fundamental DSM principle is violated: predictors must share a consistent spatial scale to allow meaningful interpretation of variable effects.
   Recommendations:

• Specify the target resolution for all rasters and the resampling method (e.g., “All environmental rasters were resampled to 250 m × 250 m using bilinear interpolation in ArcGIS 10.8.”).
• Justify your choice of scale (e.g., why aggregate 30 m layers to 5 km or vice versa).
• Discuss potential errors from combining layers of different resolution and the measures taken to minimize them.
• If applicable, describe intrapolygon or window‑based averaging for coarse layers (e.g., mean of all pixels in an X × X window).

Comment on Table 1 and the Methods section regarding the variable aspect: since aspect is a circular variable (0–360°), please clarify how it was encoded—whether raw degrees, sine–cosine transformations (sin(aspect), cos(aspect)), or categorical direction classes (north, east, etc.)—to ensure correct modeling and interpretability.

Comment on Figure 11: the variable EVA (actual evapotranspiration) is assigned to the “Climate” latent factor, yet it is an integrative indicator influenced by both climatic drivers (radiation, temperature, precipitation) and surface properties (soil moisture, vegetation cover). It should not be treated as a strictly climatic variable. Please rename the latent “Climate” factor (e.g., to “Climate–Hydrology” or “Energy–Water Balance”) to reflect its combined nature.

Comment on Figure 11: the latent factor labeled “Vegetable” is a misnomer (literally “edible plants”) and conceptually inappropriate for spectral and biophysical vegetation indicators (NDVI, SOS, EOS, NPP). I recommend renaming it “Spectral Vegetation Indices” or “Vegetation Biophysical Parameters” to accurately convey its content.

Finally, you first employ Random Forest (RF) to predict SOC spatially, then switch to Structural Equation Modeling (SEM) for interpretability. Since RF offers higher site‑specific accuracy but limited transferability, and SEM yields slightly lower explanatory power but better generalizability and mechanistic insight, please clarify:

1. Which specific advantages of RF motivated its initial use?
2. Why was SEM chosen as the secondary step instead of, for example, interpreting RF variable importance?
3. How do you evaluate the trade‑off between predictive accuracy and model transferability to other regions?

Answering these questions will help readers understand the rationale for your two‑stage modeling approach and its broader applicability.

 

Author Response

Comment 1: Lines 41–42 state “According to the soil‑forming factors theory…,” yet the original works in which this theory was first proposed are not cited:

• Dokuchaev, V. V. 1883. Russian Chernozem. (English translation: Principles of Soil Science, 1899).
• Jenny, H. 1941. Factors of Soil Formation: A System of Quantitative Pedology. McGraw‑Hill, New York.

Moreover, the superscript “2” in “factors theory 2” presumably refers to a literature citation, but it should be formatted according to the journal’s style as [2]. In addition to the five canonical soil‑forming factors, it would be appropriate to acknowledge a sixth factor—anthropogenic influence.

The traditional Dokuchaev paradigm indeed presents a simple flow: factors → properties. Early soil science assumed that the five factors fully explain soil properties. However, empirical observations have shown that identical combinations of soil‑forming factors can yield soils with different properties, and conversely that different factor combinations can produce very similar soils. To resolve such “hard” diagnostic problems, Merkurii Gilyarov developed the zoological diagnostic method of soils. This led to the non‑Dokuchaev paradigm: factors → processes → properties, with an explicit system of elementary soil‑forming processes. Theoretically, this paradigm is nearly flawless, but in practice identifying these elementary processes is extremely difficult, if not impossible. This note is intended to caution against overemphasizing the outdated factors → properties paradigm.

Reply 1: Thank you very much for your approval of the manuscript. We have replaced the reference [2] with the original one recommended for you, and added reference [3].（lns 41-46）

 

Comment 2: Lines 42–44: the sentence

“Digital soil mapping (DSM) leverages these environmental covariates to effectively predict and spatially map soil types and properties at various scales”
requires clarification, since not all of the mentioned covariates—particularly biotic and temporal ones—have standardized spatial metrics and therefore cannot be directly mapped.

Reply 2: We have corrected this paragraph. （ lns 42-46 ）

 

Comment 3: Comment on lines 84–86:

“The factors that influence the spatial distribution of SOC in farmland areas located along the middle reaches of the Yellow River remain unclear. Therefore, further research is required.”
This statement conflicts with the earlier claim that the classical soil‑forming factors (parent material, climate, organisms, topography, time) suffice to explain the spatial distribution of soil properties.

Reply 3: Thank you very much for your approval of the manuscript. We have corrected this paragraph. （ lns 92-94 ）

 

Comment 4: Comment on lines 90–92: in the phrase

“Integrate satellite hyperspectral data, natural factors, and farmland management environmental variables into a random forest (RF) model…”
terms from different conceptual levels are mixed: “satellite hyperspectral data” (a direct surface property), “natural factors” (an undefined concept, presumably the classical factors), and “farmland management environmental variables” (a group of agronomic indicators). To maintain logical consistency, I suggest grouping them uniformly as “input predictors” and defining each category explicitly.

Reply 4: Thank you very much for your valuable opinions and suggestions. They have been of great help to my thesis. It has been revised. （ lns 95-97 ）

 

Comment 5: Positive note on the Study Area section: it is clearly and systematically presented, with geographic location, climatic characteristics, and soil conditions described, providing a solid foundation for subsequent analysis.

Comment on lines 112–114:

“The predominant soil type in this region is cinnamon soil, classified as Cambisols in the World Reference Base for Soil Resources (WRB).”
Please cite the full source for the WRB classification:
IUSS Working Group WRB. 2015. World Reference Base for Soil Resources 2014, update 2015. World Soil Resources Reports No. 106. FAO, Rome.
Also add a reference (research article or report) or present your own field data to substantiate that cinnamon soil/Cambisols dominate this region.

Reply 5: We have cited the recommended reference [24] and have also added reference [25]. （lns 119 ）

 

Comment 6: Comment on the caption of Figure 2: it lacks critical metadata—no mention of which satellite acquired the image or the acquisition date. I recommend augmenting the caption with the satellite platform (e.g., Sentinel‑2A) and the date of acquisition to ensure completeness and reproducibility.

Reply 6: 2.3 section provides the details. （ lns 145-149 ）

 

Comment 7: Comment on lines 116–118 and Figure 1: the authors state that “312 sampling points were uniformly established across the study area (approximately one point per 10 000 mu, ~667 ha),” yet Figure 1 shows clusters of points and large gaps, contradicting the claim of uniform sampling. Please clarify the sampling algorithm or spatial deployment rules, or adjust the description to reflect the actual distribution.

Reply 7: Thank you very much. We have made the necessary revisions. The sampling points were arranged based on the operability of local conditions, taking into account factors such as altitude changes, land cover, and transportation accessibility for collecting soil samples. Additionally, soil sampling points were appropriately densified in fragmented farmland areas and areas with relatively complex terrain (such as mountainous regions), while this also led to a slightly sparse distribution of sampling points in relatively flat terrain areas. In total, 312 soil samples were collected throughout the study area, with the sampling points distributed approximately every 10,000 acres. The main purpose of this was to make the SOC model in complex areas more accurate. （lns 42-46, 130-138）

 

Comment 8: Comment on the Methods section and Table 1: the text indicates that topographic factors (slope, aspect) were derived from a USGS DEM at 30 m × 30 m resolution (and earlier mentions an ISRIC DEM at 250 m × 250 m), while Table 1 lists a 1 km resolution for slope and aspect. Please verify and harmonize the DEM source and spatial resolution across the text and table to avoid inconsistencies.

Reply 8: Thank you very much. We have corrected the content of Table 1. The 250 meters × 250 meters resolution ISRIC digital elevation model is the source of soil texture data, and the DEM data is from the United States Geological Survey (USGS; website: www.usgs.gov) (resolution 30 meters × 30 meters). The climate factor data is from the "Fine Mapping of Mountain Environments" project data (resolution 30 meters × 30 meters). （lns 209-214）

 

Comment 9: Comment on Table 1: the “Environmental Variable” column uses abbreviations (MAT, MAP, EVA, NDVI, NPP, SOS, EOS, Irr, Dra, NF) without providing their full names or measurement units (e.g., °C for MAT, mm for MAP, index values for NDVI). I recommend adding the full variable names and units directly in the table or as a footnote.

Reply 9: The units have been added in Table 1.

 

Comment 10: Comment on Table 1, under the “Vegetation” category: variables NDVI, NPP, SOS, and EOS are spectral indices sensitive not only to vegetation status but also to soil background, moisture, and atmospheric conditions. Please include a note on these methodological limitations, and if possible describe any atmospheric or soil‑background correction procedures applied.

Reply 10: Thank you very much. NDVI, NPP, SOS and EOS are indeed sensitive to soil background, moisture and atmospheric conditions, and may have cross-effects. These pieces of information may reflect SOC from different perspectives. However, when the random forest (RF) model is applied to predict SOC, it usually does not require the deliberate removal of highly correlated features (collinearity) as in linear models (such as linear regression, logistic regression). This is mainly because the modeling mechanism of RF fundamentally differs from linear models (such as linear regression, logistic regression), making it extremely robust (Robustness) to collinearity. In soil science, many environmental variables often have some degree of interaction effects, and these pieces of information may cross-effect the soil organic carbon status. RF can effectively utilize these redundant but potentially useful information to capture more complex nonlinear relationships and improve prediction accuracy. Therefore, NDVI and other indicators were not corrected. Other related articles did not perform any related processing. For example, the following literature:

Wang, R.; Du, W.; Li, P. et al. High-resolution mapping of cropland soil organic carbon in Northern China. Agron 2025, 15(2), 359.

Tian, H.; Zhang, J.; Zheng, Y. et al. Prediction of soil organic carbon in mining areas. Catena 2022, 215, 106311.

 

Comment 11: Comment on the Methods section listing data sources and resolutions (e.g., DEM 30 m × 30 m, SoilGrids 250 m × 250 m, NF 5 km): the paper does not describe how disparate spatial resolutions were harmonized before extracting predictor values at sampling points. It is unclear whether all layers were resampled to the finest (30 m), coarsest (5 km), or an intermediate resolution (e.g., 1 km), nor which resampling method (nearest neighbor, bilinear, majority) was used. Without this information:

1. Using both high‑resolution 30 m layers and highly aggregated 5 km layers may introduce systematic bias in SOC predictions at fine scales.
2. The analysis cannot be reproduced, since the spatial accuracy of matching 5 km × 5 km data to GPS sample coordinates is unspecified.
3. A fundamental DSM principle is violated: predictors must share a consistent spatial scale to allow meaningful interpretation of variable effects.
   Recommendations:

• Specify the target resolution for all rasters and the resampling method (e.g., “All environmental rasters were resampled to 250 m × 250 m using bilinear interpolation in ArcGIS 10.8.”).
• Justify your choice of scale (e.g., why aggregate 30 m layers to 5 km or vice versa).
• Discuss potential errors from combining layers of different resolution and the measures taken to minimize them.
• If applicable, describe intrapolygon or window‑based averaging for coarse layers (e.g., mean of all pixels in an X × X window).

Reply 11: Thank you very much for your valuable opinions and suggestions. To address the differences in coordinate reference systems and resolutions, all datasets were resampled to 30 m × 30 m raster data using the nearest-neighbor resampling method in ArcGIS 10.8, and the data were projected onto the EPSG:3857—WGS 84/World Mercator coordinate system.（lns 233-236）

 

Comment 12: Comment on Table 1 and the Methods section regarding the variable aspect: since aspect is a circular variable (0–360°), please clarify how it was encoded—whether raw degrees, sine–cosine transformations (sin(aspect), cos(aspect)), or categorical direction classes (north, east, etc.)—to ensure correct modeling and interpretability.

Reply 12: Thank you very much. The aspect we are using is the raw degree.

 

Comment 13: Comment on Figure 11: the variable EVA (actual evapotranspiration) is assigned to the “Climate” latent factor, yet it is an integrative indicator influenced by both climatic drivers (radiation, temperature, precipitation) and surface properties (soil moisture, vegetation cover). It should not be treated as a strictly climatic variable. Please rename the latent “Climate” factor (e.g., to “Climate–Hydrology” or “Energy–Water Balance”) to reflect its combined nature.

Reply 13: Thank you very much. Environmental variables are generally combined with the soil factor theory and soil landscape models. We draw on the articles of others, such as the following ones. If any changes are needed, then proceed with the subsequent modifications.

 Zhang T, Huang L M, Yang R M. Evaluation of digital soil mapping projection in soil organic carbon change modeling[J]. Ecological Informatics, 2024, 79: 102394.

Tian, H.; Zhang, J.; Zheng, Y. et al. Prediction of soil organic carbon in mining areas. Catena 2022, 215, 106311.

 

Comment 14: Comment on Figure 11: the latent factor labeled “Vegetable” is a misnomer (literally “edible plants”) and conceptually inappropriate for spectral and biophysical vegetation indicators (NDVI, SOS, EOS, NPP). I recommend renaming it “Spectral Vegetation Indices” or “Vegetation Biophysical Parameters” to accurately convey its content.

Reply 14: Thank you very much. Environmental variables are generally combined with the soil factor theory and the soil-landscape model. We have drawn on the articles of others, such as the following ones. If any changes are needed, proceed with the next step of modification.

 

Comment 15: Finally, you first employ Random Forest (RF) to predict SOC spatially, then switch to Structural Equation Modeling (SEM) for interpretability. Since RF offers higher site‑specific accuracy but limited transferability, and SEM yields slightly lower explanatory power but better generalizability and mechanistic insight, please clarify:

1. Which specific advantages of RF motivated its initial use?
2. Why was SEM chosen as the secondary step instead of, for example, interpreting RF variable importance?
3. How do you evaluate the trade‑off between predictive accuracy and model transferability to other regions?

Answering these questions will help readers understand the rationale for your two‑stage modeling approach and its broader applicability.

 

Reply 15: Thank you very much for your approval of the manuscript. We have revised the content and added reference [23]. （lns 82-88）

Q1: Which specific advantages of RF motivated its initial use?​​

We employed Random Forest (RF) as the primary modeling approach for spatial prediction of Soil Organic Carbon (SOC) due to three key advantages aligned with our study objectives:

1. Non-linear Handling Capability: RF effectively captures complex interactions between hyperspectral features (e.g., GF-5 bands processed via FOD-DWT-sCARS) and environmental covariates (Section 2.7). This was critical given the non-linear relationships between spectral indices (e.g., NPP, NDVI) and SOC observed in our preliminary analysis.
2. High Spatial Accuracy: RF minimized overfitting (mtry = 5, num. trees = 500) while achieving superior site-specific prediction accuracy (R² = 0.89, RMSE = 1.14 g/kg; Table 4), essential for fine-resolution SOC mapping (Research Objective 1).
3. Robustness to Noise: The ensemble approach reduced uncertainty in SOC predictions (SD range: 0.08–0.58 g/kg; Figure 9), particularly valuable for regions with heterogeneous farmland management practices.

These strengths directly supported our goal of generating high-precision SOC spatial maps (Section 3.4).

 

Q2: Why was SEM chosen as the secondary step instead of interpreting RF variable importance?

While RF variable importance (Section 2.8) quantified individual factor contributions (Figure 10g), SEM provided three critical enhancements for mechanistic interpretation:

1. Pathway Analysis: SEM revealed direct/indirect drivers (e.g., nitrogen fertilizer [NF] → soil properties → SOC [path coefficient = 0.14]; Section 3.6), which RF’s permutation importance cannot disentangle.
2. System Validation: SEM’s goodness-of-fit (χ²/df = 1.129, SRMR = 0.06) confirmed that the hypothesized structure of five environmental categories (climate, management, etc.) plausibly explains SOC dynamics (Figure 11).
3. Generalizable Framework: SEM generated transferable causal relationships (e.g., farmland management’s total effect = 0.38; Section 3.6), addressing Research Objective 3 on driving mechanisms.

Thus, SEM complemented RF by converting localized predictions into a basin-scale mechanistic model.

 

Q3: How do you evaluate the trade-off between predictive accuracy and model transferability?

Our two-stage approach explicitly balanced these needs:

1. RF for Precision: RF prioritized local accuracy using region-specific variables (e.g., Yuncheng’s cropping index [CI] phenology; Section 3.2), achieving the highest possible spatial fidelity (Figure 8).
2. SEM for Generalization: By abstracting RF outputs into categorical drivers (topography, management, etc.), SEM created a portable framework (Figure 11). This allows adapting the model to other basins by recalibrating locally measurable variables (e.g., replacing CI with region-specific rotation indices).
3. Empirical Validation: The SEM’s R² (0.56) indicates moderate generalizability, but we acknowledge that transferability requires external validation—a limitation noted in Section 5. Future work will test this framework in the Yellow River’s upper/lower reaches.