Evaluation of Agricultural Bare Soil Properties Retrieval from Landsat 8, Sentinel-2 and PRISMA Satellite Data

Round 1

Reviewer 1 Report

This manuscript comprehensively evaluated the potential of simulated and real spectral data from PRISMA, Sentinel-2 and Landsat 8 in mapping soil organic carbon and soil texture at a field scale. The result indicated a promising accuracy of real and resampled PRISMA in predicting soil properties and thus paves the way for future application in a greater scale. This manuscript is overall well-written with clear objectives, solid methodology & data, and meaningful discussion. This study clearly addresses the current knowledge gap and the outcome of this study provides a nice reference for mapping soil information using hyperspectral satellite. I suggest that this manuscript can accepted for publication after minor revision that addresses my following concerns.

Detailed comments are listed below:

Lines 63-64: Indeed, numerous studies have been conducted for predicting soil properties using laboratory from local to global scales. Please provide correct statement here.

Lines 178: What are the sampling time for two sites?

Table 2: The date of Landsat 8 data for mapping was quite different from others, is it hard to find ideal one in 2021?

Lines 294-295: It is necessary to cite caret package in a proper way. Take my side for example, the reference is Max Kuhn (2021). caret: Classification and Regression Training. R package version 6.0-90. https://CRAN.R-project.org/package=caret.

Line 300: Please also provide the equation for R².

Lines 347-348: I have a feeling that the use of apparent electrical resistivity for the 0–0.5 m topsoil might not be appropriate for two reasons: (1) the soil data was sampled at 0-0.1 m, so there is a big mismatch of target depth interval between soil properties and apparent electrical resistivity; (2) for the use of apparent electrical resistivity as independent data, there is a strong assumption that apparent electrical resistivity is highly corrected to SOC and soil texture, is there any reference to support it? Please carefully think about these two concerns.

Figure 3: I suggest to indicate this field in Figure 2a for better understanding.

Lines 361-362: Again, SpatialPack package should be properly cited with a reference, not only a link.

Line 385: cubist should be correct as Cubist for consistency throughout the manuscript.

Table 3: RRMSE was never used in the results, do you think is it still necessary to include in the table?

Figure 4: I have carefully checked all the points in these figures, I do not understand that why three remote sensing data had different maximum values for the same soil property. For example, the maximum observed silt for Sentinel-2 was greater than Landsat 8 and PRISMA, and the minimum observed SOC for Sentinel-2 was also greater than Landsat 8 and PRISMA. Do you have any explanation for it?

Lines 440: CR is short for continuum removal in this study not for Cubist Regression. Please also correct the same typo in line 452.

Figure 6: Why Landsat 8 have so many missing pixels compared to PRISMA?

Lines 528-530: This statement did not provide useful information, please revise.

Line 550: spectral can be corrected as spectra.

Line 577-579: Once applying into broad-scale mapping, in addition to satellite spectra, it would be of great importance to include other environmental covariates as well for more robust modelling. A recent global review from Chen et al. (2022) summarized the importance of environmental covariates in spatial prediction of different soil properties at a broad-scale.

Chen, S., Arrouays, D., Mulder, V. L., Poggio, L., Minasny, B., Roudier, P., ... & Walter, C. (2022). Digital mapping of GlobalSoilMap soil properties at a broad scale: A review. Geoderma, 409, 115567.

Author Response

RC = Reviewer Comments

OR = Our Responses

NL = New Line

Reviewer 1:

RC: Lines 63-64: Indeed, numerous studies have been conducted for predicting soil properties using laboratory from local to global scales. Please provide correct statement here.

OR: Thank you for the comment. The sentence was rewritten in a correct way.

RC: Lines 178: What are the sampling time for two sites?

OR: The sampling dates for both sites are: 11 November 2019, 28 January 2020 and 17 February 2020 for Maccarese, and 20 January 2020 for Pignola. Sampling time for the two sites was added to the manuscript (NL = 184 - 185).

RC: Table 2: The date of Landsat 8 data for mapping was quite different from others, is it hard to find ideal one in 2021?

OR: The mapping image used for Landsat 8 dates of 20^th April 2018, compared to PRISMA and Sentinel-2 images acquired on mid-spring 2021, as it was not possible to find useful cloud-free Landsat 8 image on 2021. For this reason, we selected an image acquired in the same time of the year (mid-spring) in 2018 and showing similar field conditions. An explanation regarding the choice of the dates was added to the manuscript (NL = 392 - 396).

RC: Lines 294-295: It is necessary to cite caret package in a proper way. Take my side for example, the reference is Max Kuhn (2021). caret: Classification and Regression Training. R package version 6.0-90. https://CRAN.R-project.org/package=caret.

OR: The reference of the caret package is updated and cited on the manuscript following your indication.

RC: Line 300: Please also provide the equation for R².

OR: The equation for R² is added on the manuscript (NL = 339).

RC: Lines 347-348: I have a feeling that the use of apparent electrical resistivity for the 0–0.5 m topsoil might not be appropriate for two reasons: (1) the soil data was sampled at 0-0.1 m, so there is a big mismatch of target depth interval between soil properties and apparent electrical resistivity; (2) for the use of apparent electrical resistivity as independent data, there is a strong assumption that apparent electrical resistivity is highly corrected to SOC and soil texture, is there any reference to support it? Please carefully think about these two concerns.

OR: Thank you for your comment. Indeed, electricity is conducted through soil pores filled with moisture, but water retention and flux properties are related with some basic soil properties that are more often and more easily measured in agricultural fields, like clay and SOC. Several studies have investigated the relationship between resistivity/conductivity and clay or SOC, finding meaningful results (some examples are Williams and Hoey 1987, Sudduth et al., 2005, Triantafilis and Lesch, 2005, Werban et al., 2009 and Casa et al., 2013). It is also not uncommon to use soil sampling depths that do not correspond to the geoelectric survey (see e.g., among the ones above: Sudduth et al. 2005, Triantafilisi and Lesch, 2005 and Casa et al. 2013). Indeed, there is a practical interest in relating the mapping of topsoil properties, that for optical remote sensing can only refer to the very surface of the soil, to the average soil properties across the plowed layer. We have added these references in the section you were referring to.

In addition, we would like to stress that this is not a “validation” in the strict sense of the word. For this comparison, we have purposely avoided terms like “validation” and “ground-truth” in Section 2.5.3 and throughout the entire paper, to prevent potential readers to be misled in its interpretation. The purpose of this examination is not to validate the predictions, but to provide an additional and external comparison of spatial processes, one of them coming from an independent field campaign.

RC: Figure 3: I suggest to indicate this field in Figure 2a for better understanding.

OR: Thank you for your suggestion. We added the borders of the field used in Figure 3, to Figure 2a.

RC: Lines 361-362: Again, SpatialPack package should be properly cited with a reference, not only a link.

OR: Thank you for the comment. The SpatialPack package is properly cited with a reference in the manuscript.

RC: Line 385: cubist should be correct as Cubist for consistency throughout the manuscript.

OR: Thank you for the comment. cubist is updated as Cubist throughout all the manuscript.

RC: Table 3: RRMSE was never used in the results, do you think is it still necessary to include in the table?

OR: RRMSE results were deleted from Table 3 and Table 4, and were replaced by RPIQ. which however was used and referred to on the results.

RC: Figure 4: I have carefully checked all the points in these figures, I do not understand that why three remote sensing data had different maximum values for the same soil property. For example, the maximum observed silt for Sentinel-2 was greater than Landsat 8 and PRISMA, and the minimum observed SOC for Sentinel-2 was also greater than Landsat 8 and PRISMA. Do you have any explanation for it?

OR: We provided an explanation about these differences in section 2.2 (NL = 190 – 196). Basically, the soil properties ranges are the same for L8/OLI and PRISMA due to the same sampling area (30 m x 30 m) within the ESU, while for S2/MSI the sampling area was smaller according the spatial resolution of the ESA’s sensor (see Table 1), consequently, maximum and minimum values can slightly differ between S2/MSI and L8/OLI or PRISMA.

RC: Lines 440: CR is short for continuum removal in this study not for Cubist Regression. Please also correct the same typo in line 452.

OR: Thank you for your comment. Cubist Regression citation is corrected within the manuscript.

RC: Figure 6: Why Landsat 8 have so many missing pixels compared to PRISMA?

OR: Unfortunately, we were not able to retrieve a cloud-free L8/OLI image close enough to the same acquisition date of the other two satellite (see table 3), thus we used L8/OLI image acquired in the same time of the year (mid-spring) but in 2018. However, this entailed a slightly different bare soil availability according the NBR2 index threshold adopted in this work. Consequently, for the soil properties maps derived by L8/OLI we obtained some empty areas especially close to the borders and between two sub-parcels, where some vegetation residues were detected. We added few sentences to explain this choice in section 2.5.3 (NL = 392 - 396).

RC: Lines 528-530: This statement did not provide useful information, please revise.

OR: The statement was rewritten on the text.

RC: Line 550: spectral can be corrected as spectra.

OR: Spectral was corrected as spectra (NL = 530).

RC: Line 577-579: Once applying into broad-scale mapping, in addition to satellite spectra, it would be of great importance to include other environmental covariates as well for more robust modelling. A recent global review from Chen et al. (2022) summarized the importance of environmental covariates in spatial prediction of different soil properties at a broad-scale.

Chen, S., Arrouays, D., Mulder, V. L., Poggio, L., Minasny, B., Roudier, P., ... & Walter, C. (2022). Digital mapping of GlobalSoilMap soil properties at a broad scale: A review. Geoderma, 409, 115567.

OR: Thank you for the suggestion. The comment was taken into consideration

Reviewer 2 Report

General Comment on the paper “ Evaluation of Topsoil Properties Retrieval from Landsat 8, Sen-2 tinel-2 and PRISMA Satellite Data”. This paper shows an interesting application of hyperspectral remote sensing (lab and image data) for soil properties analysis. Partial least-squares regression (PLSR), cubist regression, and random forest algorithms  was used to predict the SOC, clay, sand, and silt content. The dataset with 150 PRISMA reflectance bands, after correction  with the application of PLSR to the spectral variable selection is found as a valuable tool for improving accuracies of the obtained estimates.  Even though the paper is technically correct, the narrative is in many part very hard to follow,  in my opinion the ms.  should be better focus on the chemometric approach.  Publication after mayor  revision is recommended. 

Comments:

The title and abstract seems to be very generic about what properties of the soil it is possible to  predict with satellite imagine, and perhaps it is also too optimistic considering that on average the soil is covered by vegetation. I suggest to indicate in the title  that the research focus  only on bare soil. Also I suggest to use prudently the RDP for assessing the ability of NIR spectra to predict soil properties. Be sure to give the credit the past prediction NIRS methodologies by updating the literature. I suggest to add Bellon-Maurel, V., Fernandez-Ahumada, E., Palagos, P., Roger, J-M., McBratney, A.B., 2010. Critical review of chemometric indicators commonly used for assessing the quality of the prediction of soil attributes by NIR spectroscopy. TrAC Trends in Analytical Chemistry 29, 1073-1081. Chang, C.-W., Laird, D.A., Mausbach, M.J., Hurburgh C.R., 2001. Near-infrared reflectance spectroscopy - Principal components regression analyses of soil properties. Soil Science Society of America Journal 65, 480-49. The authors must highlight their regression model results in the abstract. In keywords, I would recommend to add PLSR o prediction  regression analysis .

1) Introduction. Lines 38-77: the authors go a long way on the  strength of the relationships between spectral features and soil properties but the prevision models of soil properties were not  well addressed. The introduction should improve to predictive  models used in this research. In particular the back ground   of the  prediction of  SOM in the Vis-NIR range with chemometric approach using laboratory soil reflectance spectroscopy and link this with available satellite hyper-spectral images (L117). As is written the introduction is very unbalanced on satellite images description  (L97-128) and  poorly  linked with the previous. If possible, I suggest to the authors to add a new section detailing state of the art in SOM model predicion. In this section, authors have to describe the relevant related work in which estimation of organic carbon or other soil characteristics with remote sensing and soil proximal sensing is described. After describing those work, authors have to identify the gap in the existing published work and detail how their paper will cover part of this gap.

2) I do not understand the statement in L67-68  “non-linear relationships between soil properties and spectra as well as increasing variances of soil properties that lead to larger prediction errors”  this referee to geostatists or chemometric models? In any case the linear models were not introduced and their limits even those not linear.  

3) Cutanic Luvisol. L 162.174 The soil classification (L163) is not  described at all, please indicate the classification with the references if is FAO WRB. Please add a soil classification for Pignola area,  the criteria fixed by FAO WRB to identify luvisols si that  they have a clay  illuviation and in many case a clay horizons to match those with their weathering . I also recommend to describe if this soil showed a surface carbonate top soils. I find the description of the soil very poor, how do explain the  variability of this soil?   

4) L190- 192. Please provide a Table that  resume the chemical properties of this soil surface sample. You have to show  the chemical composition of the sampled soil, all this information will be desirable, if you would like to compare predicted soil properties with measured data, in this case you need to detail what are the main soil properties (pH, texture, carbonate etc.)   please include the  deviation standard and indicate the references of soil analysis.

5) Spectral measurements (L200-205) there is very low information on how the soil spectra were acquired, the optical set up of the instrument,   the spectra reflectance are converted in absorbance (-log(R)? Please describe the optical calibration of your instrument, what is the angle and the distance of the soil sample from the sensor?  Explain the distance and the angle of the artificial  illumination. Please, note that the problems of sample preparation, kind of illumination,   and calibration have a great influence in soil reflectance measurements. 

6) 2.5. Soil Properties estimation (284-321) This section is very poorly written and need more detail. How the soil samples are divided into calibration and validation datasets in PLSR? Please explain better the calibration and validation for the other models. Why do you use only 10-fold cross validation technique for all models, this is not correct for RF? It is not clear how do you  optimize and validate all the models with the same LOOCV? Please explain better, this means that was applied  a Leave one out cross-validation procedure also for cubist? However, there is no description how was achieved in terms of dividing the data up on the validation and calibration  set. How many times this division actually took place? Was the LOOCV  done randomly? For example  the presence of  overlies  that can caused many limitation of the carbon predictive models were considered? The methodology to measure the spatial variability of your data must be better defined in each model. 

7) 2.5.1. Laboratory and resampled spectra. This section (L313_320) is not clear and need to be expanded,  how do you optimize the resampling process, why do you apply the Gaussian model, why do you use a halogen lamp? The most important point is why the authors did not apply the portable spectroradiometer directly for the field measurements,  I don’t understand why they prefer to use using a halogen lamp that is  far from the solar measurements.  All this gives the impression that the authors rushed to do all kinds of complicated satellite image  analyses to predict the organic matter but neglected checking if the soils measurement spectra on the sites that they chose would be appropriate for making such approach.

8) L323 For each soil sample collected in the Maccarese and Pignola areas,… something is missing.

9) 3.1. Results. According with the section 3.1, L367 “ for clay (RPD: 3.78), sand (RPD: 3.58) and silt (RPD: 3.36), and lower for SOC (RPD: 1.81)” the prediction accuracy reached in this research would be expected in realistic cases only if several external samples from the same area (to be predicted) are present in the calibration set of the model. This results is affected by autocorrelation, clay silt and send,   all this parameter are strictly correlated between them. So, please, state clearly why do you followed this approach, otherwise, the results might be overoptimistic.   The discussion  on the prediction model should be improved with independent soil variable  for the PLSR model. The information in Table 4 are insufficient (see Bellon-Maurel). The classification of RPD to justify that your models are excellent or poor, it is no

different than using R2 and there is no basis for this classification. It is all relative!  The important measure is how uncertain is the prediction, or what is the prediction interval. This is rarely calculated.

9) 4. Discussion.  L510-512. Also the SOM has many bands in the NIR and are due to the presence of hydroxyls vibration at 1400 and 1900 nm.  Please indicate which band of the spectra are related to functional groups of SOM or N such as COO, C–H, N–H and O–H-. There are many strong signal in the range between 1400-2200 nm due to  several soil component that have high affinity with water.  To confirm your data for example you can add in the discussion the bands of  the loadings weight  that are most important for the model and used in the PLSR. I suggest to add more discussion and bibliography about physical meaning of these absorptions. L 523-561. Please explain this section, it is very hard to follow this explanation.  There is same link between the  geological substrate? How   influence the distribution of calcium carbonate in the soil and how can influence your results? Finally, in the discussion, authors must detail why do they selected PLSR and Cubist and discharged other artificial intelligence methods which sometimes offer more accurate results such as Support Vector Machines, no discussion was showed for  RF. Fig. 4 and 5 refer to PLSR model?  Authors can consider in the discussion the possible effect of other sources of information such as texture, pH, conductivity, organic matter, carbonate etc. In the case that authors have this information of analyzed soil, they can include it in section 2.2. Field and soil sampling.

10) The conclusion, I would recommend authors to think about models with a clearer basis (supporting results not only on the basis of prediction results (PLSR), especially because leave-one-out cross-validation was used, with a large chance of overoptimistic results). The conclusion should be resume the important of NIR and SWIR regions  sensitive to clay minerals and SOM, explaining why this region results that might be correlated with the soil properties  and satellite image dataset. At the end of the conclusion section, authors have to define future work.

Best regards

Author Response

RC = Reviewer Comments

OR = Our Responses

NL = New Line

Reviewer 2:

RC: The title and abstract seem to be very generic about what properties of the soil it is possible to predict with satellite imagery, and perhaps it is also too optimistic considering that on average the soil is covered by vegetation. I suggest to indicate in the title that the research focus only on bare soil.

OR: We took your suggestion into consideration. We changed the title indicating that the research focuses only on the agricultural bare soil, and we added the soil properties predicted in this research in the abstract.

RC: Also, I suggest to use prudently the RDP for assessing the ability of NIR spectra to predict soil properties. Be sure to give the credit the past prediction NIRS methodologies by updating the literature. I suggest to add Bellon-Maurel, V., Fernandez-Ahumada, E., Palagos, P., Roger, J-M., McBratney, A.B., 2010. Critical review of chemometric indicators commonly used for assessing the quality of the prediction of soil attributes by NIR spectroscopy. TrAC Trends in Analytical Chemistry 29, 1073-1081. Chang, C.-W., Laird, D.A., Mausbach, M.J., Hurburgh C.R., 2001. Near-infrared reflectance spectroscopy - Principal components regression analyses of soil properties. Soil Science Society of America Journal 65, 480-49.

OR: Thank you for your suggestion. We verified the use of the RPD for assessing the ability of the NIR spectra in predicting the main soil properties in the manuscript. And, in this regard, we added the results of the Ratio of Performance to Inter-Quartile distance (RPIQ) as proposed by Bellon-Maurel et al. (2010).

RC: The authors must highlight their regression model results in the abstract. In keywords, I would recommend to add PLSR o prediction regression analysis.

OR: As suggested, the results of the regression models are added to the abstract. Both regression models used in this study, i.e., PLSR and Cubist are added in the keywords.

RC: 1) Introduction. Lines 38-77: the authors go a long way on the strength of the relationships between spectral features and soil properties, but the prevision models of soil properties were not well addressed. The introduction should improve to predictive models used in this research. In particular, the background of the prediction of SOM in the Vis-NIR range with chemometric approach using laboratory soil reflectance spectroscopy and link this with available satellite hyper-spectral images (L117). As is written the introduction is very unbalanced on satellite images description (L97-128) and poorly linked with the previous. If possible, I suggest to the authors to add a new section detailing state of the art in SOM model prediction. In this section, authors have to describe the relevant related work in which estimation of organic carbon or other soil characteristics with remote sensing and soil proximal sensing is described. After describing those work, authors have to identify the gap in the existing published work and detail how their paper will cover part of this gap.

OR: Thank you for your suggestions. The introduction was directed more to hyperspectral satellite images description in comparison with laboratory spectroscopy, considering that the main focus of the paper is to explore the potential of hyperspectral satellite imagery on estimating agricultural bare soil properties. As mentioned in the manuscript, the spectral laboratory dataset was included in this work as it might represent the best conditions for soil properties estimation, thus allowing to evaluate the asymptotic capability of the spectral resolution of the PRISMA, S2/MSI and L8/OLI sensors, i.e., without disturbing noises affecting the satellite signal.

Regarding the absence of referring to the soil properties prevision models, we updated the Introduction by addressing separate parts related to the background of the use of chemometric approaches for soil properties prediction using both laboratory and hyperspectral satellite datasets. Additionally, we also presented citations and results of previous studies computed in this regard.

We discussed the different existing published work on this theme. It was highlighted that all the cited publications, in the case of the new launched hyperspectral remote sensing, were based on resampled and/or simulated datasets, but never using actual imagery. On NL = 136 – 146 we present how our research will cover this gap by computing the study considering chemometric approaches for the main soil properties (sand, silt and clay) and SOC content, using real new launched hyperspectral imagery.

RC: 2) I do not understand the statement in L67-68 “non-linear relationships between soil properties and spectra as well as increasing variances of soil properties that lead to larger prediction errors” this referee to geostatists or chemometric models? In any case the linear models were not introduced and their limits even those not linear.

OR: Thank you for the comment. The sentence was removed from the text as we understood that it can confuse the reader.

RC: 3) Cutanic Luvisol. L 162.174 The soil classification (L163) is not described at all, please indicate the classification with the references if is FAO WRB. Please add a soil classification for Pignola area, the criteria fixed by FAO WRB to identify Luvisols is that they have a clay illuviation and in many cases a clay horizon to match those with their weathering. I also recommend to describe if this soil showed a surface carbonate top soils. I find the description of the soil very poor, how do explain the variability of this soil?

OR: We have improved the description of the soils of the two study areas by adding the FAO WRD classification and reference and the soil description as reported in the available Regional pedological maps (both at 1:250,000 scale). The Maccarese plain (NL = 162 - 166) corresponds to a deltaic area characterized by spatial variations related to the delta and sea-level dynamics (see the reference in the text). The Pignola area (NL = 176-178) is instead a fluvic lacustrine local basin surrounded by marl and clayey schists formations belonging to the Lagonegro flysch formations and carbonate complex (see the reference in the text).

Regarding the presence of carbonate topsoil, both agricultural areas don’t have surface carbonate in the collected soils samples.

RC: 4) L190- 192. Please provide a Table that resume the chemical properties of this soil surface sample. You have to show the chemical composition of the sampled soil, all this information will be desirable, if you would like to compare predicted soil properties with measured data, in this case you need to detail what are the main soil properties (pH, texture, carbonate etc.) please include the deviation standard and indicate the references of soil analysis.

OR: Thank you for your comment. We added a table (Table 1, NL = 202) presenting the descriptive statistics (minimum, maximum, average, and standard deviation) of the measured soil properties at the two sampling sites. The presented data is the only data measured in the laboratory after soil sampling as we were not interested in this study to measure other soil properties such as pH, CEC, chemical composition, or elemental concentration.

RC: 5) Spectral measurements (L200-205) there is very low information on how the soil spectra were acquired, the optical set up of the instrument, the spectra reflectance are converted in absorbance (-log(R)? Please describe the optical calibration of your instrument, what is the angle and the distance of the soil sample from the sensor?  Explain the distance and the angle of the artificial illumination. Please, note that the problems of sample preparation, kind of illumination, and calibration have a great influence in soil reflectance measurements.

OR: Thanks for your helpful comment. A detailed description of the laboratory-controlled spectral measurements has been added in the text in section 2.3 (NL = 217 - 233). A new Figure 3 has also been added to clarify the spectral measurements protocol followed for this study.

Please note only for clarification that we used for the laboratory measurements a FieldSpec ASD spectroradiometer equipped with a contact probe containing inside quartz-halogen lamp (50 W) (see the new Figure 3).

Moreover, sample preparation was a very delicate part of this study and as you can see from the new Figure 3, we followed strictly the protocol of Ben Dor et al. 2005 as described in the newly added part.

Last, as commonly done in the literature for this kind of remote sensing studies (not NIR spectroscopy), we performed laboratory measurements under controlled conditions to be sure to not affect the soil samples reflectance spectra, otherwise not possible in the field measurements where there are so many different environmental factors affecting the reflectance spectra as you know.

RC: 6) 2.5. Soil Properties estimation (284-321) This section is very poorly written and need more detail. How the soil samples are divided into calibration and validation datasets in PLSR? Please explain better the calibration and validation for the other models. Why do you use only 10-fold cross validation technique for all models, this is not correct for RF? It is not clear how do you optimize and validate all the models with the same LOOCV? Please explain better, this means that was applied a Leave one out cross-validation procedure also for cubist? However, there is no description how was achieved in terms of dividing the data up on the validation and calibration set. How many times this division actually took place? Was the LOOCV done randomly?

OR: Thank you for allowing us to improve this section. However, we never declared to have split the dataset into calibration and validation datasets, but we adopted a 10-folds cross-validation (we never used LOOCV in the present work). We considered this procedure suitable to compare the capability of the investigated sensors, because a lower number of folds lead to high bias, while LOOCV causes overfitting and too optimistic statistics. We improved the section according to your observations. Concerning the RF, we agree that the cross-validation is not necessary for this algorithm, however, as a result of the Reviewer 1 observations, we removed RF from the paper.

RC: For example, the presence of overlies that can caused many limitation of the carbon predictive models were considered?

OR: We are sorry, but we did not understand the question.

RC: The methodology to measure the spatial variability of your data must be better defined in each model.

OR: PLSR, RF and Cubist, for their nature, are not spatial models, consequently, they don’t consider the spatial variability. Additionally, we considered the samples collected in the different ESU as spatially uncorrelated.

RC: 7) 2.5.1. Laboratory and resampled spectra. This section (L313_320) is not clear and need to be expanded, how do you optimize the resampling process, why do you apply the Gaussian model, why do you use a halogen lamp? The most important point is why the authors did not apply the portable spectroradiometer directly for the field measurements, I don’t understand why they prefer to use using a halogen lamp that is far from the solar measurements.  All this gives the impression that the authors rushed to do all kinds of complicated satellite image analyses to predict the organic matter but neglected checking if the soils measurement spectra on the sites that they chose would be appropriate for making such approach.

OR: Section 2.5.1. has been improved and clearly described why we apply the Gaussian model. The Gaussian model was used to provide a standardized resampling method for all the satellites (using center + FWHM), considering that the spectral response function was unavailable for PRISMA. We used the function in the hsdar package in R for that, and now we have provided an updated reference in the text.

For the second part of your comment, instead, as replied to your comment n. 5, we used for the laboratory measurements a FieldSpec ASD (spectroradiometer) equipped with a contact probe containing inside a high-intensity quartz-halogen lamp so the observation geometry is fixed according to the protocols. We followed an internationally agreed protocol based on internal soil standards (Ben Dor et al., 2005), which prescribed the use of the contact probe, to allow the comparability and transfer of results across spectral soil libraries.

Moreover, as commonly done in the literature for this kind of remote sensing studies (not NIR spectroscopy), we performed laboratory measurements under controlled conditions (following the Ben Dor et al., 2005 consolidated protocol for soil laboratory reflectance measurements) to be sure to not affect the soil samples reflectance spectra, otherwise not possible in the field measurements where there are so many factors affecting the reflectance spectra. Our aim was not to measure reflectance spectra of the soil samples directly in the field in contemporary satellite overpasses but to build a reference spectral library with a general validity for the locations of the experiment. Collecting soil samples in contemporary to satellite overpasses and preparing them (sieving, etc.) for the lab readings at controlled conditions was a rather lengthy process, as described in the new section.

RC: 8) L323 For each soil sample collected in the Maccarese and Pignola areas,… something is missing.

OR: This section 2.5.2 has been completely rewritten and improved, thanks.

RC: 9) 3.1. Results. According with the section 3.1, L367 “for clay (RPD: 3.78), sand (RPD: 3.58) and silt (RPD: 3.36), and lower for SOC (RPD: 1.81)” the prediction accuracy reached in this research would be expected in realistic cases only if several external samples from the same area (to be predicted) are present in the calibration set of the model. This result is affected by autocorrelation, clay silt and sand, all this parameter are strictly correlated between them. So, please, state clearly why do you followed this approach, otherwise, the results might be overoptimistic. The discussion on the prediction model should be improved with independent soil variable for the PLSR model.

OR: The reviewer is right, in that the calibration models have a limited validity outside the study areas being based on local models. However, our objective was not to develop models with a high extrapolation capability, but rather to compare the potential of real satellite hyperspectral and multispectral data for topsoil properties estimation.

Concerning the second point, it is true that many soil properties are more or less correlated with each other, however, we believe that autocorrelation issues did not affect our models, taking into account that each prediction model (clay, sand, silt, or SOC) is independent of the other ones and we used only spectral data, from laboratory or satellite, as predictors. It can be noted that the correlation between soil properties can also be exploited to obtain a better prediction. From our perspective, the models showed the capability of each satellite mission data to predict only a soil property at a time.

RC: The information in Table 4 are insufficient (see Bellon-Maurel). The classification of RPD to justify that your models are excellent or poor, it is no different than using R² and there is no basis for this classification. It is all relative! The important measure is how uncertain is the prediction, or what is the prediction interval. This is rarely calculated.

OR: Thank your valuable suggestion, we added RPIQ values for the models' accuracy and we avoided classifying the general model accuracy according to RPD “thresholds” throughout the manuscript. Moreover, we added the standard error of the 10 folds cv in table 5 to provide a possible error interval of the prediction.

RC: 9) 4. Discussion.  L510-512. Also the SOM has many bands in the NIR and are due to the presence of hydroxyls vibration at 1400 and 1900 nm.  Please indicate which band of the spectra are related to functional groups of SOM or N such as COO, C–H, N–H and O–H-. There are many strong signal in the range between 1400-2200 nm due to  several soil component that have high affinity with water.  To confirm your data for example you can add in the discussion the bands of  the loadings weight  that are most important for the model and used in the PLSR. I suggest to add more discussion and bibliography about physical meaning of these absorptions. L 523-561.

OR: In order to highlight the importance of the spectral bands for the prediction models, we added figure 8 and we provided a paragraph to comment on this figure in the discussion session. This section allowed to deepen the discussion towards the role of spectral features related to the investigated soil properties and their role on the prediction models using hyperspectral remote sensors. We considerably changed the discussion section focusing on the physical meaning of the spectral features and the importance of the spectral bands for soil properties estimation.

RC: Please explain this section, it is very hard to follow this explanation.  There is same link between the  geological substrate? How   influence the distribution of calcium carbonate in the soil and how can influence your results?

OR: Regarding the influence on the soil of geologic substrate, it should be noted that both sites are alluvial plains (fluvial-lacustrine for Pignola and coastal for Maccarese) where the geological substrate is not a dominant factor controlling the soil characteristics. Specifically, regarding the calcium carbonate in the two areas, it is very poor because mainly related to the occurrence of sparse shell fragments in the Maccarese coastal dunes and to the residual soil skeleton component in the Pignola site.

RC: Finally, in the discussion, authors must detail why do they selected PLSR and Cubist and discharged other artificial intelligence methods which sometimes offer more accurate results such as Support Vector Machines, no discussion was showed for  RF. Fig. 4 and 5 refer to PLSR model?  

OR: We tested several other machine learning (ML) approaches, including SVM, and we reported just the most performing ones for our database, taking into account that the aim of this work is not comparing the prediction algorithm but to assess the capability of the satellite sensor for soil properties estimation. PLSR and especially Cubist are widely used in the more recent papers [e.g.: https://doi.org/10.1016/j.geoderma.2018.08.006; https://doi.org/10.1016/j.geoderma.2021.115071 ], even though other approaches like CNN seem promising especially for very large dataset [ https://doi.org/10.1016/j.geodrs.2018.e00198 ].

Figures 4 and 5, now, 5, and 6, refer to the best models reported in Table 5, as explained in lines 341 – 344.

RC: Authors can consider in the discussion the possible effect of other sources of information such as texture, pH, conductivity, organic matter, carbonate etc. In the case that authors have this information of analyzed soil, they can include it in section 2.2. Field and soil sampling.

OR: We did not measure other soil properties.

RC: 10) The conclusion, I would recommend authors to think about models with a clearer basis (supporting results not only on the basis of prediction results (PLSR), especially because leave-one-out cross-validation was used, with a large chance of overoptimistic results). The conclusion should be resuming the important of NIR and SWIR regions sensitive to clay minerals and SOM, explaining why this region results that might be correlated with the soil properties and satellite image dataset. At the end of the conclusion section, authors have to define future work.

OR: Many thanks for your notes. As explained in the previous replies, we did not use LOOCV for our tests. We rewrote the conclusion section according to your suggestion.

A mention of future work was added at the end of the conclusion section as suggested.

Round 2

Reviewer 2 Report

The Authors did a valuable effort to reply to all the points evidenced in the previous revision step. The discussion  is now sounder than the previous one, and the misleading points previously evidenced have now been satisfactorily addressed. I appreciate the detailed and thoughtful responses of the authors. The new version of the manuscript now was more clear, but however I am not completely satisfied with the validation of the model, also because two very different measurements are compared, one with a distance of 290 km form the land with a natural light and one with ASD contact probe (CP) using an artificial light.

I still feel that the authors should discuss better some significant limitations of the validation process after generating the final products (SOM maps): for example, the application of an internal soil standard with large difference with SOM form 1 to 10 % and so on.

RC: For example, the presence of overlies that can caused many limitation of the carbon predictive models were considered?

OR: We are sorry, but we did not understand the question.

Sorry I made mistake I would like to know the “Outliers”  I would like to know if you detect SOM value  that deviate from the predicted data , they may indicate a variability in a measurement, which is what I would expect when trying to analyze the SOM distribution. This is a very crucial  stage because through the processing stage many other limitations occurred (limited number of samples in the modeling stage), and the ground sampling of the soils that does no well represent well the pixel). The presence of other physical factors. The only way to prove that the models applied  by the authors is really working   is to show  (after all the manipulation and processing stages took place)   that the SOM maps  consist of a reliable values as examined on the ground after all processing stage have been conducted.  This could be a prosecution of your next  paper.

Bets regards