Review Reports - Soil Reflectance Composite for Digital Soil Mapping in a Mediterranean Cropland District

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper addresses a critical issue in precision agriculture—creating high-resolution soil maps cost-effectively. This aligns with sustainable development goals, particularly in soil conservation and agricultural optimization. The application of Sentinel-2 Soil Reflectance Composite (SRC) and terrain-based covariates is robust and innovative for digital soil mapping at a localized scale. The use of the HISET methodology for selecting bare soil pixels is a highlight. The detailed comments are:

1. The normalization of SOC, TN, and CaCO₃ had minimal impact on improving prediction accuracy, except for SOC. This suggests deeper issues with covariate relationships that merit additional exploration.

2. While the study mentions previous work on SRC and related indices, a more detailed comparison with alternative methodologies (e.g., machine learning or other geostatistical methods) would enrich the discussion.

3. The accuracy of STU classification varies significantly, with REO soils showing poor classification accuracy. Including more robust sampling for rare STUs or advanced machine-learning approaches could enhance these results.

Author Response

Dear reviewer, thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions/corrections highlighted/in track changes in the re-submitted files.

Comment: The paper addresses a critical issue in precision agriculture—creating high-resolution soil maps cost-effectively. This aligns with sustainable development goals, particularly in soil conservation and agricultural optimization. The application of Sentinel-2 Soil Reflectance Composite (SRC) and terrain-based covariates is robust and innovative for digital soil mapping at a localized scale. The use of the HISET methodology for selecting bare soil pixels is a highlight. The detailed comments are:

The normalization of SOC, TN, and CaCO₃ had minimal impact on improving prediction accuracy, except for SOC. This suggests deeper issues with covariate relationships that merit additional exploration.

REPLY: Thanks for your comment. In the discussion, we tried to better explain this part, explaining that only SOC_norm showed a significant improvement in terms of the final map accuracy. However, the statistical relationships between SOC, TN, CaCO3 and covariates were significant but moderate, and normalization cannot improve such relationships.

While the study mentions previous work on SRC and related indices, a more detailed comparison with alternative methodologies (e.g., machine learning or other geostatistical methods) would enrich the discussion.

REPLY: In this new version, the relative merits of machine learning and other geostatistical methods have been mentioned in the Discussion (line 647-658). In the introduction chapter, we added some sentences (Line 131-140) about the selection of interpolation methods and the sampling size, reporting some references. In the Discussion session, we incorporated references to studies that applied these methods in similar contexts, highlighting similarities, differences, and the respective strengths and limitations.

The accuracy of STU classification varies significantly, with REO soils showing poor classification accuracy. Including more robust sampling for rare STUs or advanced machine-learning approaches could enhance these results.

REPLY: Unfortunately, REO soil type shows a peculiarity that it is difficult to map with remote sensing, because it is characterized by a lithological discontinuity. Increasing sampling points could help to improve the results a little, but it is out of our budget. In any case, we don’t think that the REO STU prediction increases significantly, increasing the number of sampling. Such soil type is characterized by an important lithological discontinuity in depth (between 50 and 80 cm), that shows sandy or sandy-loam texture under a topsoil from clay-loam to loamy texture. In addition, the spatial distribution of this STU is very discontinuous and probably follows ancient paleo-channels active during reclamation activities. We tried to better explain this issue in the results (Line 588-598) and in the discussion (line 693-699).

Reviewer 2 Report

Comments and Suggestions for Authors

This study explores the development of accurate soil property maps using regression kriging, terrain-based covariates, and multispectral Sentinel-2 imagery to support digital agriculture, with a focus on mapping key soil attributes such as clay, sand, organic carbon, and cation exchange capacity (CEC). However, the manuscript lacks significant innovation and has limited applicability to other regions. To improve both the novelty and the broader application of the study, I recommend that the authors explore the possibility of directly mapping soil typological units (STUs) using Sentinel-2 imagery and terrain indicators. Additionally, the authors should assess the prediction accuracy of the model and consider whether it could be improved by incorporating known conditions, such as soil organic carbon (SOC). Rather than simply validating the SCMaP methodology (Rogge et al., 2018), the study's objective could be expanded to identify the optimal approach for soil typological mapping, which could be applied to other regions as well. This would increase the potential impact and generalizability of the results. Furthermore, certain writing issues should be addressed to improve clarity and precision.

Abstract: The abstract should be condensed, focusing on the main objectives of the research and the SCMaP methodology adapted from Rogge et al. (2018).
Sentinel-2 Image Dates: The manuscript does not specify the dates of the Sentinel-2 images used. Given that the images differ by scanning date, this could influence soil property mapping. It is important to clarify this aspect of the methodology.
Regression Kriging and Overfitting: The regression kriging model demonstrates good overall performance, but it may be prone to overfitting, especially with the inclusion of multiple covariates like multispectral data. Overfitting could reduce the model's ability to generalize to new data or different soil regions. To reduce the potential for overfitting, repetitive random selections for validation could be considered. In this study, the validation datasets were randomly selected (15%). Since the accuracy of the predicted maps largely depends on the validation dataset, I recommend averaging predictions over multiple random selections to improve the robustness of the results.
Methodological Framework: A more detailed methodological framework is needed in Sections 2.3 to 2.5 to clearly present the steps and rationale behind the process.
Figures 2 and 3: I believe Figures 2 and 3 may not be essential to the manuscript and could potentially be removed or replaced to streamline the presentation.
Exploration of New Prediction Methods: New prediction methods should be explored and compared to determine which approach might yield the best results for soil property mapping. This would help validate the current methodology and possibly improve its predictive accuracy.

Author Response

COMMENT: This study explores the development of accurate soil property maps using regression kriging, terrain-based covariates, and multispectral Sentinel-2 imagery to support digital agriculture, with a focus on mapping key soil attributes such as clay, sand, organic carbon, and cation exchange capacity (CEC). However, the manuscript lacks significant innovation and has limited applicability to other regions.

REPLY: Thanks for your comment, however we would argue against the statement that our work would lack innovation and would have a limited applicability. The use of multi-temporal satellite images and in particular the inclusion of the new Soil Reflectance Composite (SRC) products to map soil features is an element of novelty. The first paper on the methodology of SCMAP model and Sentinel-2 images was very recent (Heiden et al., 2022), and was dedicated to large scale soil mapping. To the best of our knowledge no previous paper has tested the use of these products in a high resolution local digital soil mapping context. Moreover, the approach used in our work can be easily replicated in other regions, since SRC maps can be obtained in other regions (e.g. for whole Europe: https://doi.org/10.1016/j.geoderma.2024.117113) by using the SCMAP model, as well as DEM and DEM covariates which are widely available. Therefore, the method proposed, focusing on regression kriging (with stepwise regression) and the use of soil legacy data, can indeed be used elsewhere. Of course, the relationships between soil features and covariates, as well as the map accuracy will be site-specific and will depend on the sampling frequency, the soil variability etc. This is common for all the digital soil mapping approaches.

To improve both the novelty and the broader application of the study, I recommend that the authors explore the possibility of directly mapping soil typological units (STUs) using Sentinel-2 imagery and terrain indicators. Additionally, the authors should assess the prediction accuracy of the model and consider whether it could be improved by incorporating known conditions, such as soil organic carbon (SOC). Rather than simply validating the SCMaP methodology (Rogge et al., 2018), the study's objective could be expanded to identify the optimal approach for soil typological mapping, which could be applied to other regions as well. This would increase the potential impact and generalizability of the results. Furthermore, certain writing issues should be addressed to improve clarity and precision.

REPLY: Thanks for the suggestions. In this new version of the paper, we carried out three different approaches to generate STUs maps, using the same methods of supervised classification (Winner takes all), but with different covariates: 1) only terrain attributes and SRC bands; 2) only predicted maps of soil features, both topsoil and subsoil for sand and clay; 3) terrain attributes, SRC bands and predicted maps, all together. We added the fig. 6 of the three STUs maps and the tab.8 with the comparison of the confusion matrices. As you can see, we obtained the most accurate prediction of STUs using all the covariates.

SPECIFIC COMMENTS:

Abstract: The abstract should be condensed, focusing on the main objectives of the research and the SCMaP methodology adapted from Rogge et al. (2018).

REPLY: The abstract was reviewed and condensed. We preferred to avoid the reference (Rogge et al., 2018) in the abstract.

Sentinel-2 Image Dates: The manuscript does not specify the dates of the Sentinel-2 images used. Given that the images differ by scanning date, this could influence soil property mapping. It is important to clarify this aspect of the methodology.

REPLY: SRC was calculated by SCMaP model that uses hundreds of Sentinel-images with different dates (Valid bare soil pixels extracted per each SRC pixel: Minimum=19.225, Maximum=391.653, Mean=217.470). In 2.3.2, we reported that we used all the available Sentinel images (free of cloud) in the period between 2018 and 2022. This is the key point of the method: creating a multi-temporal composite map of Sentinel-2 images, where each pixel is the mean of reflectance values when the soil was bare. We tried to improve the explanation of the method in 2.3.2, and we added the fig.2 to report the extraction of pixels for each month.

Regression Kriging and Overfitting: The regression kriging model demonstrates good overall performance, but it may be prone to overfitting, especially with the inclusion of multiple covariates like multispectral data. Overfitting could reduce the model's ability to generalize to new data or different soil regions. To reduce the potential for overfitting, repetitive random selections for validation could be considered. In this study, the validation datasets were randomly selected (15%). Since the accuracy of the predicted maps largely depends on the validation dataset, I recommend averaging predictions over multiple random selections to improve the robustness of the results.

REPLY: To remove the overfitting of the regression we decided to use “Forward stepwise regression”, as explained in 2.4. The criterion of the inclusion of a predictor was based of the F-statistic, associated to a p-value < 0.05. In our opinion, cross validation, as you indicated, would not modify the potential of overfitting, whereas a stepwise approach is the most common method to exclude redundant covariates. Regarding validation strategy to determine the accuracy of the maps, we decided to use an external validation (15%), instead of cross-validation. The repetition of the procedure multiple times is not straightforward because of the forward stepwise regression and the ordinary kriging of the residuals, providing somehow different estimation models.

Methodological Framework: A more detailed methodological framework is needed in Sections 2.3 to 2.5 to clearly present the steps and rationale behind the process.

REPLY: We tried to draw a new figure (Fig.2) representing the procedure framework. In addition, we tried to improve the text along the methodological section.

Figures 2 and 3: I believe Figures 2 and 3 may not be essential to the manuscript and could potentially be removed or replaced to streamline the presentation.

REPLY: Ok, we removed the fig.2 and 3, and we replaced them with a single figure (new Fig.2) reporting the sketch of the procedure. This new figure should help the readers to understand the approach used for this paper.

Exploration of New Prediction Methods: New prediction methods should be explored and compared to determine which approach might yield the best results for soil property mapping. This would help validate the current methodology and possibly improve its predictive accuracy.

REPLY: The prediction methods that could be used are many, and it is difficult to test all the methods. Based on the analysis of the merits and drawbacks of different methodologies for our specific case study, in particular the limited number of samples, favouring non-stationary geostatistics rather than machine learning approaches, we decided to adopt multiple stepwise regression kriging for the reasons we explained at the end of section 2.4.

Reviewer 3 Report

Comments and Suggestions for Authors

Overall this paper is well structured and covers important aspects of the research. However addressing the following comments will enhance clarity and consistency.

General Comments

1. The manuscript lacks consistent references to figures and tables. For example, "Table 1" is mentioned, but "tab. 2" is used elsewhere. Ensure uniformity in referring to figures and tables.

2. Acronyms are explained multiple times throughout the text, which is unnecessary after their initial introduction.

3. The interchangeable use of "CaCO3" and "limestone" is problematic. CaCO3 refers to the chemical compound calcium carbonate, whereas limestone is a carbonate sedimentary rock composed primarily of CaCO3. It is recommended to use the term "Calcium Carbonate" or "CaCO3" consistently.

Specific Comments

1. Line 112: Does "LC map" stand for "Land Cover"? If so, please specify it as "Land Cover (LC)" when first mentioned.

2. Line 183: Clarify the number of soil samples used for each prediction.

3. Line 184: Provide an explanation for the large sample period. Discuss whether this extended timeframe affects the harmonization of the samples.

4. Line 258: Given that bare soil in the Mediterranean region is more prevalent during winter, why was data for the entire year used? Justify this choice.

5. Lines 294-295: Why were only these three properties analyzed in this manner? Explain why the approach was not applied to all properties.

6. Line 490: In this section and the conclusion, emphasize the high RMSE values, particularly in relation to the dataset’s standard deviation. Additionally, consider incorporating metrics such as the Ratio of Performance to Interquartile Range (RPIQ) or Mean Absolute Error (MAE) for better insight.

7. Lines 657-674: The conclusion lacks a discussion on the study's limitations and potential directions for future research. Include these for a more comprehensive overview.

Comments on Figures and Tables

1. Figure 1: This figure is not cited within the manuscript. Additionally, the text does not explain the presence of 98 calibration points. Since this figure represents a map, it would be helpful to include latitude/longitude grid lines.

2. Figure 3: Clarify whether the legend corresponds to DN (Digital Number) or reflectance. Additionally, ensure that the north arrow style is consistent across all figures.

3. Figure 4: The CaCO3 map uses a smaller scale than the other maps. Standardize the scales for better comparison.

4. Figures 5-6: It is standard practice to use the same range (e.g., Y-axis: 0.5–3; X-axis: 0.5–3) and scale (e.g., 1 cm = 1 g*kg⁻¹) in scatterplots. This makes the graphs easier to interpret and compare.

5. Table 7: The "Unit" column is unclear. Visually distinguish that "g*100g⁻¹" refers to Clay through to CaCO3 for better understanding.

Author Response

COMMENT: Overall this paper is well structured and covers important aspects of the research. However addressing the following comments will enhance clarity and consistency.

General Comments

The manuscript lacks consistent references to figures and tables. For example, "Table 1" is mentioned, but "tab. 2" is used elsewhere. Ensure uniformity in referring to figures and tables.

REPLY: Ok, thanks. We have checked the references of figures and tables along the text.

Acronyms are explained multiple times throughout the text, which is unnecessary after their initial introduction.

REPLY: we checked the manuscript and we removed the unnecessary explanations of acronyms.

The interchangeable use of "CaCO3" and "limestone" is problematic. CaCO3 refers to the chemical compound calcium carbonate, whereas limestone is a carbonate sedimentary rock composed primarily of CaCO3. It is recommended to use the term "Calcium Carbonate" or "CaCO3" consistently.

REPLY: Thank you, it’s our mistake. We modify “total limestone” with “total lime” when we introduce the CaCO3 for the first time, then we have used CaCO3.

Specific Comments

Line 112: Does "LC map" stand for "Land Cover"? If so, please specify it as "Land Cover (LC)" when first mentioned.

REPLY: Yes, it means Land Cover. We added in the text

Line 183: Clarify the number of soil samples used for each prediction.

REPLY: The number of soil observations used for each prediction is now better explained in the chapter 2.4.

Line 184: Provide an explanation for the large sample period. Discuss whether this extended timeframe affects the harmonization of the samples.

REPLY: This is a frequent issue when we have to work with legacy data. Such differences in soil sampling period, and probably different laboratories for analysis, can make legacy data not harmonized. On the other hand, we cannot avoid using legacy data during soil mapping activity. We added a sentence in section 2.2 acknowledging this issue.

Line 258: Given that bare soil in the Mediterranean region is more prevalent during winter, why was data for the entire year used? Justify this choice.

REPLY: The use of Sentinel-2 images of the whole year was adopted in the composite approaches, both with SCMaP and with SYSI models, in order to maximize the amount of data available in different soil moisture conditions. These models are based on the selection of all the pixels considered bare using a well-tested and advanced thresholding methodology (HISET). In our study area, tilled field free of crops can be found during September-October or November for winter crops like wheat, in February-March for spring-summer crops like corn or sunflower. Some farmers plough the soils during August, and then prepare the seedbed in October, others prepare the soils directly in October-November. So, it’s very complicated to find the best period to find bare soil, and the choice made in the multitemporal approach is probably the most appropriate to be generally applied over vast areas such as the whole of Europe.

Lines 294-295: Why were only these three properties analyzed in this manner? Explain why the approach was not applied to all properties.

REPLY: As was specified in the text, the other variables already had a parametric distribution and there was no need to normalize them.

Line 490: In this section and the conclusion, emphasize the high RMSE values, particularly in relation to the dataset’s standard deviation. Additionally, consider incorporating metrics such as the Ratio of Performance to Interquartile Range (RPIQ) or Mean Absolute Error (MAE) for better insight.

REPLY: Ok, thanks. We added RPIQ in the fig. 5 and 6, to take into account the RMSEP relative to the standard deviation of the data.

Lines 657-674: The conclusion lacks a discussion on the study's limitations and potential directions for future research. Include these for a more comprehensive overview.

REPLY: We added some comments at the end of “Conclusions”, to discuss about the strengths and the limitations of the approach and the use of SRC.

Comments on Figures and Tables

Figure 1: This figure is not cited within the manuscript. Additionally, the text does not explain the presence of 98 calibration points. Since this figure represents a map, it would be helpful to include latitude/longitude grid lines.

REPLY: We added the citations. We have also included latitude and longitude of the central area in the caption, and a simplified grid into the map of fig.1.

Figure 3: Clarify whether the legend corresponds to DN (Digital Number) or reflectance. Additionally, ensure that the north arrow style is consistent across all figures.

REPLY: we removed fig.2 and fig.3 according to the request of another reviewer, and we replace them with the new fig.2, which summarizes the workflow.

Figure 4: The CaCO3 map uses a smaller scale than the other maps. Standardize the scales for better comparison.

REPLY: we have corrected the scale

Figures 5-6: It is standard practice to use the same range (e.g., Y-axis: 0.5–3; X-axis: 0.5–3) and scale (e.g., 1 cm = 1 g*kg⁻¹) in scatterplots. This makes the graphs easier to interpret and compare.

REPLY: Thanks for the suggestion, we modified the figures, using the same range of Y and X axis, to make easier to understand. In addition, we added the value of RPIQ and we also modified R², because, in the previous version, they were incorrect.

Table 7: The "Unit" column is unclear. Visually distinguish that "g*100g⁻¹" refers to Clay through to CaCO3 for better understanding.

REPLY: Corrected

Reviewer 4 Report

Comments and Suggestions for Authors

The manuscript titled “Soil reflectance composite for digital soil mapping in a Mediterranean cropland district” applied the regression kriging method and incorporating terrain-based covariates and multispectral images of bare soil as covariates to create comprehensive maps of primary soil characteristics, which is of great significance for obtaining precise and accurate soil maps at digital agriculture approaches, with low costs.

However, there are an issue need to be more explained or discussed, 1)” Legacy data from 12 soil profiles, collected from 1980 to 2015, ……” (line 184) were used to interpolate or analyze soil features, while “ For generating the SRC, all Sentinel-2 scenes processed to surface reflectance by the MAJA algorithm available in a five-year period between 2018 and 2022 covering the Rieti plain are considered.” (line 257-259). Does the use of remote sensing data and soil data at different times acquired influence the results of the study (such as Table 6)?

Author Response

COMMENT: The manuscript titled “Soil reflectance composite for digital soil mapping in a Mediterranean cropland district” applied the regression kriging method and incorporating terrain-based covariates and multispectral images of bare soil as covariates to create comprehensive maps of primary soil characteristics, which is of great significance for obtaining precise and accurate soil maps at digital agriculture approaches, with low costs.

REPLY: Thanks, this was the main objective of our work.

REPLY: Thanks, this observation is correct, we have tried to explain better in the text the significance of this issue. In this kind of approach we cannot avoid to use soil data and remote sensing data from different periods. This is a rather common approach when using legacy data, although as the reviewer points out this could pose problems for soil features that can have high temporal variability, like SOC and, in particular, TN. This is the limitation of the approach, that we also discuss in the final part of discussions and conclusions sections. However, the innovative aspect of this paper is the use of SRC covariate to spatialize soil features. Therefore, the same approach of this paper can be used elsewhere, using a target soil sampling.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

This is the revised version of the manuscript that I previously reviewed. The authors have done a commendable job of incorporating my suggestions and improving the readability and presentation of the manuscript.