the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 09 Jun 2022
Weathering intensities in tropical soils evaluated by machine learning, clusterization and geophysical sensors
Abstract. Weathering is widely used for pedogenesis and soil fertility studies, once it affects several soil attributes. Understanding the intensities of weathering can provide answers for environmental issues, soil and geosciences studies. Recently, there are available geotechnologies (such as geophysics and machine learning algorithms) that can be applied in soil science to provide pedosphere information. In this research, we performed a method to evaluate weathering intensity in a heterogeneous tropical area by proximal remote sensing data acquired by geophysical and satellite images respectively. The area is located in southwest Brazil, with 184ha and we sampled 79 sites (all with soil analysis) using toposequence knowledge. Afterwards, the principal component analysis and the ideal number of clusters was determined. Then, we determine and used the ideal number of clusters, weathering index, as input data in four modelling (prediction and spatialization) algorithms to infer different weathering intensities in soils formed from the same soil parent material. The results showed that the best model performance was for the random forest reaching 3 clusters as the ideal number. The surface pixel reflectance acquired from a Synthetic Soil Image, the terrain surface convexity and digital elevation model were the covariates that most contributed to modelling processes. The model’ specificity was greater than sensitivity. The East areas over diabase such as the Nitisol presented greater weathering intensity than the Nitisol over West diabase areas. The areas over siltite/metamorphosed siltite and Lixisols presented moderate weathering rates. The relief and topographic position strongly affected the weathering, once they controlled the hydric dynamics. The geophysical variables were related to soil attributes and weathering, which contributed to modelling and clusterization processes. The different weathering rates are mainly modulated by geomorphic processes that relief, topographic position, and the associated soil types control water dynamic at the landscape and directly affect the weathering intensities.
- Preprint
(1312 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
CC1: 'Comment on soil-2022-17', Sara Ramos Santos, 26 Jul 2022
Dears colleagues
I have fill comments to say/ask about the work.
1. The use of geotecnology associated to marching learning is fundamental for the future of pedology, due to increasing the prediction process in great areas. Congratulations!
2. Which variables do you used to calculate the intemperism index? Are the variables obtained with sensors our in laboratory?
3. I think that would be good if this article will have a section in R&D correlationing intemperism to mapping soil fertility in this area.
Thanks in advance!
Congratulations for the work!
Citation: https://doi.org/10.5194/soil-2022-17-CC1 -
CC2: 'Reply on CC1', Danilo Mello, 01 Aug 2022
Dear Sarah,
1. The use of geotecnology associated to marching learning is fundamental for the future of pedology, due to increasing the prediction process in great areas. Congratulations!
Thank you for your comment and your congratulations for us related to this manuscript.
2. Which variables do you used to calculate the intemperism index? Are the variables obtained with sensors our in laboratory?
In this work we used the weathering index obtained with laboratory data, as explained in section 2.3. Weathering rates. We then combined this index with data obtained via geophysical sensors (in the field) and satellite (SYSI), which correspond to soil attributes generated by different intensities of weathering and, consequently, pedogenesis. Then we de-correlated the variables (Table 3). Subsequently, the PCA was performed using the kmeans method, where clustering indices corresponding to the different weathering indices were generated.
3. I think that would be good if this article will have a section in R&D correlating weathering to mapping soil fertility in this area.
The fertility issue was not addressed, because it was not the focus of this work. In addition, to properly assess soil fertility we would need soil fertility management data (liming and fertilization) by the company working in the area (since it is an agricultural, private area and company data is restricted). Therefore, it will not be possible nor feasible to hold a session on this topic in this work, which does not prevent this from being done in future works (by the way, we thank you for the idea and we will try to incorporate this in another work).
Regards,
Citation: https://doi.org/10.5194/soil-2022-17-CC2
-
CC2: 'Reply on CC1', Danilo Mello, 01 Aug 2022
-
CC3: 'Comment on soil-2022-17', Guilherme Oliveira, 31 Aug 2022
This article represents an excellent application of Machine learning techniques in geophysical-soil science data. The manuscript brings an excellent approach to the geophysical relationships with processes that occur in the soil (weathering) and variables of easy acquisition.
The methodological flowchart is very well defined, However, I had doubts in some specific parts.
First of all, the English writing needs to be improved. I suggest submitting to a specific proofreading company.
Why did you use the F1-score test instead of the Kappa metric or accuracy metric?
Is the data balanced? The accuracy is a good metric when the data is balanced. This is very important because all of these algorithms have a bias with unbalanced data. Moreover, the algorithms used (excepted for RF) required standardized data.
Why didn't you employ the covariates in the unsupervised clustering? I dont understand why you compare the number of PCA dimensions with the number of clusters.
With proper revisions, I believe it to be an important contribution to the journal.Citation: https://doi.org/10.5194/soil-2022-17-CC3 -
AC1: 'Reply on CC3', Danilo Mello, 20 Oct 2022
Guilherme Oliveira, 31 Aug 2022
This article represents an excellent application of Machine learning techniques in geophysical-soil science data. The manuscript brings an excellent approach to the geophysical relationships with processes that occur in the soil (weathering) and variables of easy acquisition.
A: Thank you for acknowledging our research.
The methodological flowchart is very well defined; However, I had doubts in some specific parts.
A: Thanks for the suggestions.
First of all, the English writing needs to be improved. I suggest submitting to a specific proofreading company.
A: We sent the for a general review of English (American English) to a specialized company, where a geoscience specialist also reviewed the entire manuscript (Proofreading service). A certificate attesting to the new revision of the manuscript was inserted in the "supplementary material" field.
Why did you use the F1-score test instead of the Kappa metric or accuracy metric?
A: The accuracy is an appropriate and good metric to evaluate the model’s performance, when the data is balanced. That is not our case. As the geophysical data used, as well as the data from the physicochemical analysis of the soils, are unevenly distributed in terms of the number of samples, over the different geologies and soil classes, our data were considered unbalanced.
Imbalanced data always becomes one of the classification issues. Imbalanced dataset occurs when one or several classes have less sample (the minority); while another/other classes are the majority, such as agricultural fields. Accuracy is the most intuitive evaluation index to show the performance of a model. However, when the data classes are imbalanced, a supplementary index is required. That index is the F1-Score, which is a harmonic average of Precision and Recall. The F1-Score can accurately evaluate the performance of the model when the data is imbalanced (Lee and Park, 2021; Wardhani et al., 2019). To further support this assertion, we have inserted two citations from recent works in good journals in the text that detail the use of the F-1 Score in terms of accuracy.
Is the data balanced? The accuracy is a good metric when the data is balanced. This is very important because all of these algorithms have a bias with unbalanced data. Moreover, the algorithms used (excepted for RF) required standardized data.
A: No, the geophysical and soil physico-chemical data are unbalanced. We understand the reviewer's point of view and we agree. Due to that, the results and discussion section were based on F-1 Score metrics, instead of accuracy. We only presented the accuracy and kappa for readers to compare. We used the F-1 Score metric due to criticisms and even recommendations from some other reviewers on articles previously submitted for publication in this and other journals. For this same reason, we also used 2 other parameters to complement the F-1 Score (sensitivity and specificity of the models).
Why didn't you employ the covariates in the unsupervised clustering? I dont understand why you compare the number of PCA dimensions with the number of clusters.
A: We did not employ the covariates in the unsupervised clustering because they were used to map the groups of evaluated attributes. If these covariates were used in clustering, they would act as predictor and predicted variables at the same time, generating overestimated performance results that do not match reality.
We did not compare the optimal number of clusters with the cps results. We use the two data together to create the groups using the k-means method. The use of cps instead of pure data is because the cps are not correlated with each other. On the other hand, pure data can present correlation between them.
With proper revisions, I believe it to be an important contribution to the journal.
A: Thank you for all the suggestion. We made a few changes to the text to better clarify future readers and they were certainly great contributions to our work.
Citation: https://doi.org/10.5194/soil-2022-17-AC1
-
AC1: 'Reply on CC3', Danilo Mello, 20 Oct 2022
-
CC4: 'Comment on soil-2022-17', Rafael Siqueira, 05 Sep 2022
I’ve read the pre-print version of this manuscript with great interest, due to the novelty associated to the application of modern technologies to investigate soil weathering and pedogenetic development, a very welcome initiative in Soil Science. The authors successfully combine two powerful approaches or tools constantly applied in the Pedometrics field: computational statistical techniques (machine learning and multivariate analysis) and geophysics proximal sensing data, aiming to understand in a quantitative manner how the soils develop in a tropical landscape. Nonetheless, I would like to write down some comments as reflections and suggestions:
I am not a native speaker, but a new look at the English writing for corrections would be important to improve the quality of the text. Furthermore, some parts of the text are a bit clumsy and a review would be welcome.
Abstract
Line 17 - I am not quite sure if you could call machine learning as a “geotechnology”, for the reason it was not originally developed for geosciences.
Line 19 – Satellite imagering is not a proximal sensing technique, but a remote sensing technique.
Line 21 – You must explain in the Abstract why you used the PCA and clusters analysis. At the same time, I recommend to cite what information have you used to create the clusters, and moreover, what they really mean, in the Abstract.
Line 22 – Change “we determine and used the ideal number of clusters” for “we used the ideal number of clusters”. You’ve already cited you determined the number of clusters before.
Line 27 – “The Nitisol over East diabase presented greater weathering intensity than”.
The last sentence of the Abstract is clumsy.
Introduction
The Introduction is very well written, making a brief literature review about weathering and the geophysical sensors.
Line 94 – Would not be better if you used just “model weathering intensity using combined data from geophysical sensors” than “model weathering index using combined data from geophysical sensors”?
Line 97 – I suggest to remove the objective 4, since you did not explore it in your results.
Line 99 – I would remove the citation about satellite here.
Material and methods
Line 109 - I guess the name of the municipality cited in the text is with just one “f”.
Line 135 - You could standardize the Embrapa citation. Sometimes you cite Embrapa, 2011 and sometimes Embrapa, 2017.
Line 135 – In the laboratory analysis topic, you make a complete description about physical and chemical analysis of soils according to Embrapa. However, you do not use any of them, unless the oxides contents. I suggest to remove all analyses which you did not apply in this paper.
Line 146 - I think it’d be productive you write one or two more lines explaining better how the oxides contents, the only analysis you really used, are obtained. Moreover, you cite Fe2O3 content but you also do not use it, only the SiO2 and TiO2 contents.
Line 150 - If you consider that index (Wl) used in the text as sufficient for your goals, I suggest you to better characterize this index in particular and its meaning about weathering. Moreover, I see in many parts of the text you using “weathering indexes” when in this paper you use just one index.
Line 197 - You are very correct in saying that the use of the IGC data allowed the creation of a more detailed topographic data than orbital sources, such as SRTM, could do. But just a hint for future works: with the scale of IGC database you claimed (1:10.000) you’d be able to obtain an MDE with a way finer resolution than you obtained, which would be more appropriate to the goals that you wanted to achieve.
Table 1 – very nice table explaining each morphometric variable. It will be a good reference for other authors.
Line 217 - It will be interesting you cite in an only place the “7 parameters derived from geophysical sensors data” which you used. According to your descriptions in the “Geophysical data collection” and the Table 3, you used 5 parameters from the geophysical sensors, not 7.
Line 228 - “argiluviation or ferralitization indexes”. Again, why two indexes if you supposedly used just one? Furthermore, it is the SiO2/TiO2 index that you described previously one of them? If so, you should have to indicate that before.
Line 229 - “These groups characterize themselves by present similar values within the groups, but different values between one group to another”. You should refine the description about the statistical clustering.
Line 275 - Very good insight about the combined use of the PCA and k-means analysis as well as the use of the Nested LOOCV for handling few samples.
Line 233 - “This result was used to extract the values of covariates (morphometric and geological data)”. Not only these covariates, but also the Synthetic Soil image (SYSI).
Line 236 – “base database”?
Line 250 – “increases the performance of machine processing algorithms”. Not necessarily.
Line 260 – the remotion by correlation is not only associated to “reduce computational time” but also to minimize the problem of multicollinearity.
Line 272 – I think it is “repeatecv”.
Line 350 – You cite here that you used the Kruskal Wallis test to choose the best model. But in your results, you only showed the Kruskal Wallis being used to differentiate the weathering clusters. You need to synchronize your material and methods with your results in this regard.
Line 359 – You don’t need to cite RF twice in the sentence.
Line 360 – Before “The RF algorithm presented equivalent performances than other algorithms”, insert “In other studies,”.
Line 381 – “performance precision”? Redundancy.
Line 381 – Paragraph shows some clumsy sentences.
Figure 5 - I’ve seen that the authors explain for the first time what the clusters really mean (weathering intensity) in a clear way only on the Figure 5. I suggest the authors to anticipate this important explanation. At the same time, I suggest the authors to label the clusters with the information of weathering, at last, the numbers 1,2,3 that the authors arbitrarily chose are just categorical, not expressing by themselves the weathering degree.
Line 451 – According to the rest of your text, the cluster 3 has higher weathering, followed by cluster 1 and then the cluster with lesser weathering, cluster 2. In this sentence, you have been confounded the order of the clusters.
Line 458 – “West Rodic Nitisol” – I think the authors confounded the terms. Here, the correct would be “East Rodic Nitisol”.
Line 465 – “For the weathering index”.
Table 3 – interesting table. Could explain in a paragraph which of the parameters explain better the weathering, according with your data? For me, the most important were the W1, plus magnetic susceptibility and ECa.
Line 494 – East diabase
Line 495 – West diabase and West Rhodic Nitisol
Line 514 – free drainage
Line 523 – but there is higher K40 in the supposedly more weathered Lixisols (cluster 1)
Line 526 – decreasing or increasing ECa? You must explain that.
Line 539 – you do not model with the nested LOOCV, but you evaluated it.
My suggestions aim to improve the quality of this excellent paper. I congratulate the authors for their work and hope to read more on this very interesting topic.
Citation: https://doi.org/10.5194/soil-2022-17-CC4 -
AC2: 'Reply on CC4', Danilo Mello, 20 Oct 2022
Rafael Siqueira, 05 Sep 2022
I’ve read the pre-print version of this manuscript with great interest, due to the novelty associated to the application of modern technologies to investigate soil weathering and pedogenetic development, a very welcome initiative in Soil Science. The authors successfully combine two powerful approaches or tools constantly applied in the Pedometrics field: computational statistical techniques (machine learning and multivariate analysis) and geophysics proximal sensing data, aiming to understand in a quantitative manner how the soils develop in a tropical landscape. Nonetheless, I would like to write down some comments as reflections and suggestions:
A: Thank you for taking the time to read our work and recognize the importance of our research.
I am not a native speaker, but a new look at the English writing for corrections would be important to improve the quality of the text. Furthermore, some parts of the text are a bit clumsy and a review would be welcome.
A: We sent the for a general review of English (American English) to a specialized company, where a geoscience specialist also reviewed the entire manuscript (Proofreading service). A certificate attesting to the new revision of the manuscript was inserted in the "supplementary material" field. In addition, we reviewed the entire text to locate and modify the parts that were unclear.
Abstract
Line 17 - I am not quite sure if you could call machine learning as a “geotechnology”, for the reason it was not originally developed for geosciences.
A: We agree with the reviewer and modify the sentence as suggested.
Line 19 – Satellite imagering is not a proximal sensing technique, but a remote sensing technique.
A: We agree with the reviewer and modify the sentence as suggested.
Line 21 – You must explain in the Abstract why you used the PCA and clusters analysis. At the same time, I recommend to cite what information have you used to create the clusters, and moreover, what they really mean, in the Abstract.
A: We used the cluster and PCA concomitantly to reduce the number of variables, which were later used in the cluster analysis, allowing the choice of uncorrelated groups, improving the performance of the analyses. The information used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.
We modified a little a sentence in abstract and material in methods (section 2.7 Principal component analysis and clusterization) better explain this step.
Abstract: “Afterwards, the principal component analysis and the ideal number of clusters was determined, in order to reduce the number of variables, which were later used in the cluster analysis. The data used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.”
Material and methods: “... (PCA) was applied to the 6 parameters derived from geophysical sensors data (eU, eTh, K40, κ and, ECa), and the weathering indexes.”
Line 22 – Change “we determine and used the ideal number of clusters” for “we used the ideal number of clusters”. You’ve already cited you determined the number of clusters before.
A: We agree with the reviewer and modify the sentence as recommended.
Line 27 – “The Nitisol over East diabase presented greater weathering intensity than”.
A: We do not understand this query. In the text the sentence is complete and correct. Please verify:
... the Nitisol presented greater weathering intensity than the Nitisol over West diabase areas.”
The last sentence of the Abstract is clumsy.
A: We agree with the reviewer and modify the sentence as recommended.
Introduction
The Introduction is very well written, making a brief literature review about weathering and the geophysical sensors.
A: Thank you for read.
Line 94 – Would not be better if you used just “model weathering intensity using combined data from geophysical sensors” than “model weathering index using combined data from geophysical sensors”?
A: We agree with the reviewer and modify the sentence as recommended.
Line 97 – I suggest to remove the objective 4, since you did not explore it in your results.
A: We agree with the reviewer and modify the sentence as recommended.
Line 99 – I would remove the citation about satellite here.
A: At this point we judged to keep the "satellite" in the sentence due to the use of SYSI, generated from satellite images in the survey.
Material and methods
Line 109 - I guess the name of the municipality cited in the text is with just one “f”.
A: We agree with the reviewer and modify the sentence as recommended.
Line 135 - You could standardize the Embrapa citation. Sometimes you cite Embrapa, 2011 and sometimes Embrapa, 2017.
A: We agree with the reviewer and modify the sentence as recommended.
Line 135 – In the laboratory analysis topic, you make a complete description about physical and chemical analysis of soils according to Embrapa. However, you do not use any of them, unless the oxides contents. I suggest to remove all analyses which you did not apply in this paper.
A: At this point, we decided to keep all the analyzes that were carried out, because data such as granulometry, fertility and others, allow us to correlate the intensity of weathering with the environment where the samples were collected, and relate them to geology and soil type, which was one of the approaches of the article.
Line 146 - I think it’d be productive you write one or two more lines explaining better how the oxides contents, the only analysis you really used, are obtained. Moreover, you cite Fe2O3 content but you also do not use it, only the SiO2 and TiO2 contents.
A: On this specific point, we disagree with the reviewer, as these analyzes are already traditionally well known in soil science and are already duly referenced. We apply the principle of objectivity in the writing of the work while ensuring the reproducibility of the methodology.
Line 150 - If you consider that index (Wl) used in the text as sufficient for your goals, I suggest you to better characterize this index in particular and its meaning about weathering. Moreover, I see in many parts of the text you using “weathering indexes” when in this paper you use just one index.
A: The term was better explained following the reviewer's recommendations. The manuscript was revised and the passages where we cited "weathering indexes" were corrected to "weathering index". The follow sentence was added in the final of this section:
“...This index is based on the principle that during chemical weathering, there is an intense leaching of mobile elements such as basic cations and silica (SiO2) and residual concentration of less mobile and soluble oxides such as Fe2O3, A2O3 and TiO2, mainly in tropical environments. In other words, these oxides gradually increase with increase in weathering intensity while all other elements are gradually reduced. This is the basis of calculating the weathering indices of soil which actually indicate the degree of weathering (Wani et al., 2016).”
Line 197 - You are very correct in saying that the use of the IGC data allowed the creation of a more detailed topographic data than orbital sources, such as SRTM, could do. But just a hint for future works: with the scale of IGC database you claimed (1:10.000) you’d be able to obtain an MDE with a way finer resolution than you obtained, which would be more appropriate to the goals that you wanted to achieve.
A: We thank the reviewer for this valuable tip. We are already adopting this suggestion in other work that is being developed.
Table 1 – very nice table explaining each morphometric variable. It will be a good reference for other authors.
A: We thank you for the recognition.
Line 217 - It will be interesting you cite in an only place the “7 parameters derived from geophysical sensors data” which you used. According to your descriptions in the “Geophysical data collection” and the Table 3, you used 5 parameters from the geophysical sensors, not 7.
A: We have amended the sentence as per the reviewer's suggestion. In fact, there were 6 parameters, since we consider each gamma channel as a parameter. We add the following snippet to the sentence:
“... (PCA) was applied to the 6 parameters derived from geophysical sensors data (eU, eTh, K40, κ, ECa) and the weathering indexes...”
Line 228 - “argiluviation or ferralitization indexes”. Again, why two indexes if you supposedly used just one? Furthermore, it is the SiO2/TiO2 index that you described previously one of them? If so, you should have to indicate that before.
A: In this sentence we have made a mistake. Therefore, the terms argilluviation and ferralitization were excluded from the sentence.
Line 229 - “These groups characterize themselves by present similar values within the groups, but different values between one group to another”. You should refine the description about the statistical clustering.
A: Sorry, but on this particular point, we disagree with the reviewer. We believe that this sentence is well explained. This explanation was the same given in other previously published works. However, we added the word "statistics" to the sentence, as follows:
“These groups characterize themselves by present statistical similar values within the groups, but different values between one group to another.”
Line 275 - Very good insight about the combined use of the PCA and k-means analysis as well as the use of the Nested LOOCV for handling few samples.
A: We thank the reviewer for recognizing the work.
Line 233 - “This result was used to extract the values of covariates (morphometric and geological data)”. Not only these covariates, but also the Synthetic Soil image (SYSI).
A: The sentence was adjusted following the reviewer recommendation.
Line 236 – “base database”?
A: The sentence was adjusted following the reviewer suggestion.
Line 250 – “increases the performance of machine processing algorithms”. Not necessarily.
A: Thanks for the review, however on this point we disagree with the reviewer. We cite and find works where other researchers have found that there really are increases in the performance of machine processing algorithms. Furthermore, we verified in this and other works by our group that this phenomenon actually occurs. If the reviewer sends us references with contrary scientific convictions, we will review this point, otherwise, we decide to keep it.
Line 260 – the remotion by correlation is not only associated to “reduce computational time” but also to minimize the problem of multicollinearity.
A: We add this in the sentence following the reviewer recommendation.
Line 272 – I think it is “repeatecv”.
A: We corrected the sentence following the reviewer suggestion.
Line 350 – You cite here that you used the Kruskal Wallis test to choose the best model. But in your results, you only showed the Kruskal Wallis being used to differentiate the weathering clusters. You need to synchronize your material and methods with your results in this regard.
A: Thanks for the review, however we decided to keep the results more relevant and easier for readers to explain.
Line 359 – You don’t need to cite RF twice in the sentence.
A: We adjusted the sentence following the reviewer suggestion.
Line 360 – Before “The RF algorithm presented equivalent performances than other algorithms”, insert “In other studies,”.
A: We adjusted the sentence following the reviewer suggestion.
Line 381 – “performance precision”? Redundancy.
A: We adjusted the sentence following the reviewer suggestion.
Line 381 – Paragraph shows some clumsy sentences.
A: The paragraph was revised by proof-reading service and was improved.
Figure 5 - I’ve seen that the authors explain for the first time what the clusters really mean (weathering intensity) in a clear way only on the Figure 5. I suggest the authors to anticipate this important explanation. At the same time, I suggest the authors to label the clusters with the information of weathering, at last, the numbers 1,2,3 that the authors arbitrarily chose are just categorical, not expressing by themselves the weathering degree.
A: We improve the figure as the reviewer suggested and, also, we added a short paragraph explaining for what the clusters were created and what they mean, in material and methods section, as follow:
“It is important to realize that, the clusters were generated in order to demonstrate the different degrees of weathering on the different lithologies and soil types.”
Line 451 – According to the rest of your text, the cluster 3 has higher weathering, followed by cluster 1 and then the cluster with lesser weathering, cluster 2. In this sentence, you have been confounded the order of the clusters.
A: We adjusted the sentence following the reviewer suggestion.
Line 458 – “West Rodic Nitisol” – I think the authors confounded the terms. Here, the correct would be “East Rodic Nitisol”.
A: We corrected the sentence following the reviewer's recommendations.
Line 465 – “For the weathering index”.
A: We corrected the sentence following the reviewer's recommendations.
Table 3 – interesting table. Could explain in a paragraph which of the parameters explain better the weathering, according with your data? For me, the most important were the W1, plus magnetic susceptibility and ECa.
A: A new paragraph has been added, following the reviewer's recommendations:
The greatest statistical differences and, consequently, the weathering intensity were better evidenced for magnetic susceptibility, ECa and radiometric data from gamma-ray spectrometry, respectively, evidencing the differences between the different weathering index.
Line 494 – East diabase
A: We corrected the sentence following the reviewer's recommendations.
Line 495 – West diabase and West Rhodic Nitisol
A: We corrected the sentence following the reviewer's recommendations.
Line 514 – free drainage
A: We corrected the sentence following the reviewer's recommendations.
Line 523 – but there is higher K40 in the supposedly more weathered Lixisols (cluster 1)
A: Dear reviewer, we believe that there was a mistake in the interpretation. All K40 values for area were low (less than 1), including for Lixisols areas.
Line 526 – decreasing or increasing ECa? You must explain that.
A: Dear reviewer, we added the follow sentence in this paragraph:
“...however, it is harder to explain how environmental variable contribute more and how they affect the ECa values during weathering intensity, because they all pedoenvironmental variable act concomitantly to produce ECa values (parent material, pedogenetic and weathering processes, mineralogy, texture, water dynamics on landscape, magnetic susceptibility, texture, CEC and base saturation) (Mello et al., 2022).”
Line 539 – you do not model with the nested LOOCV, but you evaluated it.
A: Dear reviewer, thanks for the review. We understand the reviewer's concern, however at this point we are referring that nested-LOOCV was suitable for modeling as a whole process, for small sample sets. The modeling in this work was carried out in three stages: training, validation and test. The nested-LOOCV inner loop is done in training and validation.
“The nested-LOOCV method is a double loop process, where in the internal loop, the model is trained with a data set of size n-1, using the LOOCV for the optimization of the final model. On the other hand, the external loop corresponds to the test. In this loop, the remaining sample is predicted using the final model calculated in the inner loop. This prediction result is stored with the observed value of the remaining sample and later used to calculate the algorithm’s performanc(Jung et al., 2020; Neogi and Dauwels, 2019). The two loops are run n times (n = total number of samples, in our case 71). All samples are inserted into the outer loop, where the values predicted by the final model of each algorithm are calculated with the predicted and observed values of each sample. Then, the final result of the machine learning algorithm's performance will be obtained by predicted and observed values stored in the external loop. This is a robust method to evaluate the algorithm’s performance and detects possible samples with problems in the collections or outliers. The training set generated in each loop went through the process of selecting covariates for importance and subsequent training”.
My suggestions aim to improve the quality of this excellent paper. I congratulate the authors for their work and hope to read more on this very interesting topic.
A: Thank you for the excellent reviews that contributed to our work and for the recognition.
Citation: https://doi.org/10.5194/soil-2022-17-AC2
-
AC2: 'Reply on CC4', Danilo Mello, 20 Oct 2022
-
EC1: 'Comment on soil-2022-17', Nicolas P.A. Saby, 05 Sep 2022
I made an in depth reading of the paper and I have several major concerns about the work and the paper.
The abstract is quite vague and not very informative. Please explain the overall goal, the assumption, the approach implemented and give precise results and findings.
The English does not seem to be ok.
The work is not enough clearly presented. A lot of questions arise when reading the actual version.
The overall goal of this study is not enough clearly presented in the introduction. First (i), modelling with ML of weathering index (WI) is quite vague. But for what ? Mapping, monitoring, statistical application? In (ii), it is said that importance of the covariates will be discussed but what is these covariates. This is not clear if these are the one for the digital soil mapping approach or the observed ones using sensor data. It is also a general comment about this paper where a confusion is made between these 2 types of information. The description of the proximal and remote sensing data should be better explained. The first ones are not spatially explicit.
The way geophysical data are used is rather difficult to understand and does not seem at all appropriate. It is quite difficult to understand why these data are mixed with the WI in the PCA step. This means that Geophysical data are not used as covariates but as a covariable. The PCA is not a method to evaluate (as it written in the abstract) but a multivariate analysis to explain the multivariate relations between variables. Moreover, the results of the PCA are not presented. It is not then possible to understand the signification of the principal components.
Why did you use a clustering approach ? We could expect a more quantitative digital soil mapping approach of the raw value of WI instead. Why are you creating clusters of WI and sensor data to map them after? Again, as the results of the PCA and the clustering are not presented, it is difficult to understand the signification of the cluster. It could happen that no correlation occurs between these data and the clusters represents only one of 2 variables. The results table 3 seem to validate this assumption as the discrimination between the different cluster are not very significant. However, an extensive discussion is provided about the signification of these clusters based on an interpretation of the dsm maps and by expertise. However, all the maps are wrong. Maybe a point map of the clusters would be better to be interpreted first. Finally, it is difficult to understand the adding value of the geophysical data.
In the 2.1, two sets of data are presented (16 and 79 points) but only the 79 points are used for the study. Could you explain ? finally, it is topsoil map of WI ?
The equation 1 is not explained enough. What is SIO2 and TIO2 ?
In 2.7, two approaches are presented for the clustering, eg scott lethod and k-means. Which one did you use? Do you mean that a kmeans approach was implemented and the optimal number of clusters was selected using a scott approach ?
The 2.8.2 should be titled "validation of the map".
The explanation of the nested approach is quite difficult to follow. Do you mean that there is a tuning of the ML algorithm at each step of the LOOCV? This is not indicated in the fig 3. How this tuning is done ? Bye cross validation?
283 should be titled validation not training
2.8.3 Why are you explaining that you are using a LOOCV now? This is very confusing as it is not explained in the fig 3. All the indicators explained in this section are based the result of the confusion matrix where the 4 kinds of results are computed, eg FP TP, FN and FN.
It is not explained how the different maps produced during the LOOCV are combined at the end of the process. Do you compute the dominant value? Are using a probability approach?
The "pls" is usually a regression approach and is not adapted to a classification exercise. Could you detail which algorithm you used?
The size of the dataset is quite not adapted to the use of ML algorithm like random forest. You may consider more classical algorithm like the multinomial algorithm. As you use a LOOCV, we could not exclude an overfitting of the data. A k-fold cross validation would be more adapted. The hyper parameter of the different algorithms are not described so it is difficult to understand the quality of the models.
The sampling design is also not all adapted to a digital soil mapping approach. The locations were selected by expert judgement and are not very well spread over the area. There is extensive discussion about how to collect data in the recently published paper. This not really discussed in this paper. It is only acknowledged that the sample are not enough even but the spread of the data into the covariate space should be checked to discuss the validity of the map produced.
I think this article needs to be thoroughly revised before publication. Th use of the clustering step and the geophyical data should be better justified.
Citation: https://doi.org/10.5194/soil-2022-17-EC1 -
AC3: 'Reply on EC1', Danilo Mello, 20 Oct 2022
Nicolas P.A. Saby, 05 Sep 2022
I made an in depth reading of the paper and I have several major concerns about the work and the paper.
The abstract is quite vague and not very informative. Please explain the overall goal, the assumption, the approach implemented and give precise results and findings.
A: Rereading the abstract, we agreed with the reviewer and made the necessary adjustments in order to clarify the research objectives, main results and conclusions of the study.
The English does not seem to be ok.
A: We sent the for a general review of English (American English) to a specialized company, where a geoscience specialist also reviewed the entire manuscript (Proofreading service). A certificate attesting to the new revision of the manuscript was inserted in the "supplementary material" field.
The work is not enough clearly presented. A lot of questions arise when reading the actual version.
A: Thanks for the suggestion, we agree with the reviewer. This may have occurred given the number of covariates and the explanations of the relationship between them with weathering, in addition to the clustering and modeling technique used. However, we revised the entire manuscript and some paragraphs that we thought were confused in the explanations, and/or missing some details were rewritten in order to clarify the reader. Another problem may have been English, which was also revised.
The overall goal of this study is not enough clearly presented in the introduction. First (i), modelling with ML of weathering index (WI) is quite vague. But for what? Mapping, monitoring, statistical application? In (ii), it is said that importance of the covariates will be discussed but what is these covariates. This is not clear if these are the one for the digital soil mapping approach or the observed ones using sensor data. It is also a general comment about this paper where a confusion is made between these 2 types of information. The description of the proximal and remote sensing data should be better explained. The first ones are not spatially explicit.
A: Thanks for the review. Dear reviewer, we have re-evaluated the objectives and they are clearly met. Modeling the intensity of weathering means predicting and spatializing the data (maps shown in figures 5 and 6, for example, and table 2). The applications of obtaining weathering intensities, or in other words, more and less weathered soils, were briefly explained in the introduction (paragraph x). The importance of the variables used were those used in the modeling and which were duly presented and discussed in section 3.1 (3.1 Evaluation of model's performance, uncertainty and variables importance), including a figure (figure 4) demonstrating what these variables were. We believe that these variables should not be presented in the objectives of the work, but they should be explained and demonstrated in the methodology, as well as presented and discussed in the sessions indicated in the manuscript.
We have added the following paragraph after the objectives to clarify the relationship between digital mapping, remote and near sensing, and weathering:
“In this work we use the digital soil mapping approach (using geophysical sensors and satellite imagery) to estimate different intensities of weathering in an area that are comparable in terms of geology and soil types.”
We have added a new paragraph in the introduction to clarify the concept of remote and near sensing:
“The proximal sensing is the use of field-based sensors to obtain signals from the soil when the sensor’s detector is in contact with or close to (within 2 m) the soil (e.g., geophysical sensors). The sensors provide soil information because the signals correspond to physical measures, which can be related to soils and their properties. When the distance at which a particular sensor acquired a data is greater than 2m from the target, there is remote sensing (e.g., satellite images) (Rossel et al., 2011).”
The way geophysical data are used is rather difficult to understand and does not seem at all appropriate. It is quite difficult to understand why these data are mixed with the WI in the PCA step. This means that Geophysical data are not used as covariates but as a covariable. The PCA is not a method to evaluate (as it written in the abstract) but a multivariate analysis to explain the multivariate relations between variables. Moreover, the results of the PCA are not presented. It is not then possible to understand the signification of the principal components.
A: In this work we used the weathering index obtained with laboratory data, as explained in section 2.3. Weathering rates. We then combined this index with data obtained via geophysical sensors (in the field) and satellite (SYSI), which correspond to soil attributes generated by different intensities of weathering and, consequently, pedogenesis. Then we de-correlated the variables (Table 3). Subsequently, the PCA was performed using the kmeans method, where clustering indices corresponding to the different weathering indices were generated. We used the cluster and PCA concomitantly to reduce the number of variables, which were later used in the cluster analysis, allowing the choice of uncorrelated groups, improving the performance of the analyses. The information used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.
In addition, we adjusted the abstract following the reviewer issues about PCA.
We did not show PCA because it was used only to de-correlate sensor data.
Why did you use a clustering approach? We could expect a more quantitative digital soil mapping approach of the raw value of WI instead. Why are you creating clusters of WI and sensor data to map them after? Again, as the results of the PCA and the clustering are not presented, it is difficult to understand the signification of the cluster. It could happen that no correlation occurs between these data and the clusters represents only one of 2 variables. The results table 3 seem to validate this assumption as the discrimination between the different cluster are not very significant. However, an extensive discussion is provided about the signification of these clusters based on an interpretation of the dsm maps and by expertise. However, all the maps are wrong. Maybe a point map of the clusters would be better to be interpreted first. Finally, it is difficult to understand the adding value of the geophysical data.
A: We use a cluster approach in order to separate weathering intensity into classes.
In fact, we could expect a more quantitative digital soil mapping approach with raw WI values. However, it was not the focus of this work, since there are already other works that have already done this in the area of pedometrics.
In this work we used the weathering index obtained with laboratory data, as explained in section 2.3. Weathering rates. We then combined this index with data obtained via geophysical sensors (in the field) and satellite (SYSI), which correspond to soil attributes generated by different intensities of weathering and, consequently, pedogenesis. Then we de-correlated the variables (Table 3). Subsequently, the PCA was performed using the kmeans method, where clustering indices corresponding to the different weathering indices were generated. We used the cluster and PCA concomitantly to reduce the number of variables, which were later used in the cluster analysis, allowing the choice of uncorrelated groups, improving the performance of the analyses. The information used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.
Indeed, it could happen that no correlation occurs between these data and the clusters represents only one of 2 variables. However, as the reviewer observed, we performed a non-parametric assessment of sensor data and WI between groups (the Kruskal-Wallis test) to assess this issue and found that virtually all of them had statistical differences.
We believe that the reviewer is a little mistaken, since the values in table 3 indicate differences for most classes. Regarding the issue that the maps are wrong, we would like to know based on what arguments the reviewer asserts this issue. Wrong about what? As such, we are unable to respond.
The geophysical data contributed towards having significant relationships with weathering and the soil attributes associated with them. as discussed in the work.
In the 2.1, two sets of data are presented (16 and 79 points) but only the 79 points are used for the study. Could you explain? finally, it is topsoil map of WI?
A: Yes, the reviewer is right. Sixteen soil profiles were described, as shown in figure 1. Seventy-nine points where samples were collected for physical-chemical laboratory analysis and, geophysical data. Data from the 16 soil profiles were used for soil classification, for further analysis between the intensity of weathering and the type of soil.
In fact, the weathering intensity was evaluated in the first 20cm of the soil, so yes, for the top soil, although the gamma gets values of readings of 30cm of depth.
The equation 1 is not explained enough. What is SIO2 and TIO2?
A: Thanks for the suggestion. We better explained the equation and the relationship between the elements and, add a new paragraph in material and methods section, as follow:
“The W1 index (Eq. 1) is based on the principle that during chemical weathering, there is an intense leaching of mobile elements such as basic cations and silica/ silicon oxide (SiO2) and residual concentration of less mobile and soluble oxides such as Fe2O3, Al2O3 and TiO2, mainly in tropical environments. In other words, these oxides gradually increase with increase in weathering intensity while all other elements are gradually reduced. This is the basis of calculating the weathering indices of soil which actually indicate the degree of weathering (Wani et al., 2016).”
In 2.7, two approaches are presented for the clustering, eg scott lethod and k-means. Which one did you use? Do you mean that a kmeans approach was implemented and the optimal number of clusters was selected using a scott approach?
A: The Scott test was used to determine the ideal number of clusters, while the k-means test was used to classify the samples in the group determined by scott.
The 2.8.2 should be titled "validation of the map".
A: We adjusted the sentence following the reviewer suggestion.
The explanation of the nested approach is quite difficult to follow. Do you mean that there is a tuning of the ML algorithm at each step of the LOOCV? This is not indicated in the fig 3. How this tuning is done? Bye cross validation?
A: The LOOCV (cross-validation) was used to optimize the hyperparameters of the evaluated algorithms, in each inner loop run. This phase is located inside the dotted square in figure 3. We are changing the flowchart to make this information clearer.
In addition, we tried to explain the nested-LOOCV method as follow:
“The nested-LOOCV method is a double loop process, where in the internal loop, the model is trained with a data set of size n-1, using the LOOCV for the optimization of the final model. On the other hand, the external loop corresponds to the test. In this loop, the remaining sample is predicted using the final model calculated in the inner loop. This prediction result is stored with the observed value of the remaining sample and later used to calculate the algorithm’s performance (Jung et al., 2020; Neogi and Dauwels, 2019). The two loops are run n times (n = total number of samples, in our case 75). All samples are inserted into the outer loop, where the values predicted by the final model of each algorithm are calculated with the predicted and observed values of each sample. Then, the final result of the machine learning algorithm's performance will be obtained by predicted and observed values stored in the external loop. This is a robust method to evaluate the algorithm’s performance and detects possible samples with problems in the collections or outliers. The training set generated in each loop went through the process of selecting covariates for importance and subsequent training”.
283 should be titled validation not training.
A: We adjusted the sentence following the reviewer suggestion.
2.8.3 Why are you explaining that you are using a LOOCV now? This is very confusing as it is not explained in the fig 3. All the indicators explained in this section are based the result of the confusion matrix where the 4 kinds of results are computed, e.g., FP TP, FN and FN.
A: The LOOCV is used to optimize the hyperparameters of the evaluated algorithms. It is in this item that this point is made. In figure 3 this phase is the two-color square. The performance parameters used in this work (indicators) were all taken from the confusion matrix from which the FP TP, FN and PN indices are obtained. This point is inside the dotted square in figure 3.
It is not explained how the different maps produced during the LOOCV are combined at the end of the process. Do you compute the dominant value? Are using a probability approach?
A: Dear reviewer, we believe that this question is duly explained in item 2.8.3:
“...Then, the final map was created by combining the 71 prediction maps generated for each algorithm tested. In addition, the mode value for each pixel of the final map was calculated. The prediction error map was elaborated, considering the number of times that each algorithm chose the mode value in each map pixel normalized by the number of final maps (%)....”
“...The best final map chosen by the previous statistical tests was used to extract the geophysical sensor data, weathering indexes values at the sampling points....”
The "pls" is usually a regression approach and is not adapted to a classification exercise. Could you detail which algorithm you used?
A: Thank you for this revision. In fact, this algorithm (pls) can be used as a classifier. As detailed in the text, the algorithms used, as well as their explanation, can be found in the caret manual at the link https://topepo.github.io/caret/available-models.html "6 Available Models". The pls is defined as "Partial Least Squares"- PLS. This algorithm can be used in classification and regression. The detailed explanation of the use of this algorithm is not relevant in this article, since it is an application of the algorithms (modeling) in soil science.
The size of the dataset is quite not adapted to the use of ML algorithm like random forest. You may consider more classical algorithm like the multinomial algorithm. As you use a LOOCV, we could not exclude an overfitting of the data. A k-fold cross validation would be more adapted. The hyper parameter of the different algorithms are not described so it is difficult to understand the quality of the models.
A: Firstly, we would like to have a larger number of samples than our current number and, we believe that all researchers would like to, but in our field conditions it was not possible as many other field conditions around the world. In addition, there is no minimum number of samples in the literature, indicated as a correct reference value for modeling soil and/or geophysical attributes. In addition, there are several researches published in good quality scientific journals, including Geoderma, which the authors also used a small number of samples in modeling processes (Fabijańczyk et al., 2017; Gebauer et al., 2020; Granger et al., 2017; Peukert et al., 2012; dos Santos Teixeira et al., 2021; Zhang and Zhu, 2019). In addition, LOOCV is more suitable for datasets with a small number of samples than k-fold.
In addition, we added the following sentence in material and methods:
The hyperparameters of each algorithm are described in the caret package manual in chapter 6. “Models described” available at https://topepo.github.io/caret/train-models-by-tag.html.
The sampling design is also not all adapted to a digital soil mapping approach. The locations were selected by expert judgement and are not very well spread over the area. There is extensive discussion about how to collect data in the recently published paper. This not really discussed in this paper. It is only acknowledged that the sample are not enough even but the spread of the data into the covariate space should be checked to discuss the validity of the map produced.
A: This sampling theory is valid and most used in geoesthetics. In ML, the use of conditioned Latin hypercube is being more used to define sampling sites. In addition, we would like to emphasize that during data collection, we advanced as far as possible due to the limitations of the sugarcane crop. However, we tried to carry out a distributed and representative sampling of the area, collecting data on all types of soils and lithology existing in the area, besides the topossequences.
I think this article needs to be thoroughly revised before publication. The use of the clustering step and the geophyical data should be better justified.
A: We agree with the reviewer and, we revied and explained in more detail and justification the two approaches (issues) highlighted by the reviewer.
Citation: https://doi.org/10.5194/soil-2022-17-AC3
-
AC3: 'Reply on EC1', Danilo Mello, 20 Oct 2022
-
RC1: 'Comment on soil-2022-17', Anonymous Referee #1, 15 Sep 2022
The article has an interesting general idea, but it raises some questions about the study proposal.
The first general question in the article that I didn't see answered is: What would be the real importance of evaluating soil weathering by different techniques? Soil fertility? This can be measured directly. 'environmental issues'? there are also techiniques to evaluate directly.
Listing the objectives of this study, not all of them were answered in the course of the article, nor in the conclusions section, leaving readers without a final answer to certain points raised.
Some errors in concepts and terms were noticed, such as:
-Machine learning is not a geotechnologie.
-‘proximal remote sensing’? satellite image is not proximal and geophysical is not remote.
-Spectroscopy method is not a geotechnology
I miss a paragraph explaining how the fact of weathering affects everything you said. Bearing in mind that this process takes years and years to affect soil properties, would that make any difference now? And if you say this can be seen in the difference in soil types, then isn't it easy to study the difference between soils?
Table 2 – value of accuracy from kknn is wrong.
When comparing the algorithms, I didn't see you talking about which parameters were used in each of them.
In terms of modeling and mapping, as you mentioned, the number of samples is very low, not being a great number of samples to work in an area of almost 200 ha. In addition to using it to calibrate a model with RF. Don't you think this would affect the construction of the model? don't you think the data wouldn't be overfitted?
Speaking now about the mapping exercise, in my opinion the distribution of samples is not adequate to carry out the mapping exercise.
My main concern is: Even knowing that the amount of samples was not enough and as the arrangement of samples is not suitable for the DSM approach, you still decided to carry out the article.
I suggest improving the quality of the data and methodologies used and strongly revising the article
Citation: https://doi.org/10.5194/soil-2022-17-RC1 -
AC4: 'Reply on RC1', Danilo Mello, 20 Oct 2022
Anonymous Referee #1,
The article has an interesting general idea, but it raises some questions about the study proposal.
A: Thanks for the acknowledgment, we checked all the issues about the study proposal pointed out by the reviewer.
The first general question in the article that I didn't see answered is: What would be the real importance of evaluating soil weathering by different techniques? Soil fertility? This can be measured directly. 'environmental issues'? there are also techiniques to evaluate directly.
A: Weathering operates in a multitude of spatial (from a nanometer to a planetary scale) and temporal (from thousands to only a few years) scales and its action impacts several, if not all, Earth systems. Weathering releases solutes that nourish every terrestrial ecosystem, triggers the biogeochemical cycles of every chemical element and thus control both water (e.g., river, ocean, groundwater) and soil chemistry. From a larger scale perspective, weathering is one of the major planetary sinks for CO2, and thus, play a pivotal role on climate change and life itself on Earth. As an extremely intricate and complex process, and in response to its interface nature, weathering has been assessed, quantified, measured, and studied from innumerous techniques and perspectives; all originated from different scientific fields and backgrounds. Many of such approaches and techniques are expensive and time-consuming and thus make it difficult to gather information about large areas in an accessible, reliable, and fast way. In the current scenario, where science must tackle many Earth science challenges in which weathering is a central process, the development of novel tools to study weathering is crucial to achieve the expected goals (Egli et al., 2001; Frings and Buss, 2019; Ruiz et al., 2020, 2022).
Listing the objectives of this study, not all of them were answered in the course of the article, nor in the conclusions section, leaving readers without a final answer to certain points raised.
A: We agreed with the reviewer and added a new paragraph:
“The combined use of geophysical sensors, satellite images and morphometry, by different machine learning algorithms proved to be a robust method and were able to model different weathering intensities.”
The other parts of the objectives we believe are answered along the other parts of the conclusions.
Some errors in concepts and terms were noticed, such as:
-Machine learning is not a geotechnology.
A: We adjusted the sentences following the reviewer suggestions.
‘proximal remote sensing’? satellite image is not proximal and geophysical is not remote.
A: We adjusted the sentences following the reviewer suggestion.
Spectroscopy method is not a geotechnology
A: We adjusted the sentences following the reviewer suggestion.
I miss a paragraph explaining how the fact of weathering affects everything you said. Bearing in mind that this process takes years and years to affect soil properties, would that make any difference now? And if you say this can be seen in the difference in soil types, then isn't it easy to study the difference between soils?
A: We agree with the reviewer and add the following paragraph:
“Weathering operates in a multitude of spatial (from a nanometer to a planetary scale) and temporal (from thousands to only a few years) scales and its action impacts several, if not all, Earth systems. Weathering releases solutes that nourish every terrestrial ecosystem, triggers the biogeochemical cycles of every chemical element and thus control both water (e.g., river, ocean, groundwater) and soil chemistry. From a larger scale perspective, weathering is one of the major planetary sinks for CO2, and thus, play a pivotal role on climate change and life itself on Earth. As an extremely intricate and complex process, and in response to its interface nature, weathering has been assessed, quantified, measured, and studied from innumerous techniques and perspectives; all originated from different scientific fields and backgrounds. Many of such approaches and techniques are expensive and time-consuming and thus make it difficult to gather information about large areas in an accessible, reliable, and fast way. In the current scenario, where science must tackle many Earth science challenges in which weathering is a central process, the development of novel tools to study weathering is crucial to achieve the expected goals (Egli et al., 2001; Frings and Buss, 2019; Ruiz et al., 2020, 2022).”
Dear reviewer, in relation to the time in which weathering acts and affects soil properties, there are numerous works in the literature that report how weathering can be extremely fast, especially in the tropical environment and under more susceptible parent materials.
Egli, M., Mirabella, A., & Fitze, P. (2001). Clay mineral transformations in soils affected by fluorine and depletion of organic matter within a time span of 24 years. Geoderma, 103(3-4), 307-334.
Ruiz, F., Andrade, G. R. P., Sartor, L. R., dos Santos, J. C. B., de Souza Júnior, V. S., & Ferreira, T. O. (2022). The rhizosphere of tropical grasses as driver of soil weathering in embryonic Technosols (SE-Brazil). Catena, 208, 105764.
Ruiz, F., Sartor, L. R., de Souza Júnior, V. S., dos Santos, J. C. B., & Ferreira, T. O. (2020). Fast pedogenesis of tropical Technosols developed from dolomitic limestone mine spoils (SE-Brazil). Geoderma, 374, 114439.
Table 2 – value of accuracy from kknn is wrong.
A: The text was adjusted following the reviewer suggestion.
When comparing the algorithms, I didn't see you talking about which parameters were used in each of them.
A: We added a short sentence in text:
The hyperparameters of each algorithm are described in the caret package manual in chapter 6. “Models described” available at https://topepo.github.io/caret/train-models-by-tag.html.
Here we presented the hyperparameter:
Algorithm
Hyperparameters
Definition
RF
“mtry”
number of covariates that are chosen at random for each node in the tree. Being the only parameter optimized.
ntree
number of trees. BREIMAN, L, 2002, considers keeping the value constant at 100.
nodesize
the minimum number of data points on each terminal node. BREIMAN, L, 2002, considers keeping the value constant at 1 and 5 for classification and regression.
KKNN
kmax
Number of neighbors within the search area. Optimized.
distance
Maximum search distance from the center of the test area. Optimized.
kernel
Sample space transformation kernel. Optimized.
Partial Least Squares (PLS)
ncomp
the number of main components to be used in modeling. Optimized.
avNNet (Model Averaged Neural Network)
size
Number of units (neurons) in the hidden layer. Optimized.
decay
parameter for weight drop using neural network optimization. optimized
bag
Logical index of the use of bootstrap methodology in training optimization. Optimized.
In terms of modeling and mapping, as you mentioned, the number of samples is very low, not being a great number of samples to work in an area of almost 200 ha. In addition to using it to calibrate a model with RF. Don't you think this would affect the construction of the model? don't you think the data wouldn't be overfitted?
A: Firstly, we would like to have a larger number of samples than our current number and, we believe that all researchers would like to, but in our field conditions it was not possible as many other field conditions around the world. In addition, there is no minimum number of samples in the literature, indicated as a correct reference value for modeling soil and/or geophysical attributes. In addition, there are several researches published in good quality scientific journals, including Geoderma, which the authors also used a small number of samples in modeling processes (Fabijańczyk et al., 2017; Gebauer et al., 2020; Granger et al., 2017; Peukert et al., 2012; dos Santos Teixeira et al., 2021; Zhang and Zhu, 2019).
We are aware of the possibility of overfitting, so applying the nested-LOOCV methodology, which we believe is more suitable for a small number of samples.
The nested-LOOCV method is indicated as a set for small data sets, which other methods of evaluation of test samples would not be viable due to the low number of samples in a test samples (Ferreira et al., 2021), being more used in the field of medicine in human experiments or where the number of samples is limited, providing an unbiased estimate of the true error (Chen et al., 2017; Li et al., 2018; Xing et al., 2011; Xu et al., 2020). The Nested-LOOCV method is a double loop process, where in the first loop the model is trained with a data set of size n-1, and the test is done in the second loop with the missing sample and used to validate the test and the training performance.
Speaking now about the mapping exercise, in my opinion the distribution of samples is not adequate to carry out the mapping exercise.
A: Here, we would like to emphasize that: during data collection, we advanced as far as possible due to the limitations of the sugarcane crop. However, we tried to carry out a distributed and representative sampling of the area, collecting data on all soil types, toposequences and, lithology existing in the area.
My main concern is: Even knowing that the amount of samples was not enough and as the arrangement of samples is not suitable for the DSM approach, you still decided to carry out the article.
A: We understand the reviewer concern. However, we do have a few observations to note: Firstly, we would like to have a larger number of samples than our current number and, we believe that all researchers would like to, but in our field conditions it was not possible as many other field conditions around the world. In addition, there is no minimum number of samples in the literature, indicated as a correct reference value for modeling soil and/or geophysical attributes.
There are several researches published in good quality scientific journals, including Geoderma, which the authors also used a small number of samples in modeling processes such as: (Fabijańczyk et al., 2017; Gebauer et al., 2020; Granger et al., 2017; Peukert et al., 2012; dos Santos Teixeira et al., 2021; Zhang and Zhu, 2019)
Finally, during data collection, we advanced as far as possible due to the limitations of the sugarcane crop. However, we tried to carry out a distributed and representative sampling of the area, collecting data on all soil types, toposequences and, lithology existing in the area.
I suggest improving the quality of the data and methodologies used and strongly revising the article.
A: The entire article was revised following the reviewer suggestion.
References
Chen, X., Zhang, H., Lee, S.-W. and Shen, D.: Hierarchical high-order functional connectivity networks and selective feature fusion for MCI classification, Neuroinformatics, 15(3), 271–284, 2017.
Egli, M., Mirabella, A. and Fitze, P.: Clay mineral transformations in soils affected by fluorine and depletion of organic matter within a time span of 24 years, Geoderma, 103(3–4), 307–334, 2001.
Fabijańczyk, P., Zawadzki, J. and Magiera, T.: Magnetometric assessment of soil contamination in problematic area using empirical Bayesian and indicator kriging: A case study in Upper Silesia, Poland, Geoderma, 308, 69–77, doi:https://doi.org/10.1016/j.geoderma.2017.08.029, 2017.
Ferreira, R. G., da Silva, D. D., Elesbon, A. A. A., Fernandes-Filho, E. I., Veloso, G. V., de Souza Fraga, M. and Ferreira, L. B.: Machine learning models for streamflow regionalization in a tropical watershed, J. Environ. Manage., 280, 111713, 2021.
Frings, P. J. and Buss, H. L.: The Central Role of Weathering in the Geosciences, Elements, 15(4), 229–234, doi:10.2138/gselements.15.4.229, 2019.
Gebauer, A., Ellinger, M., Brito Gomez, V. M. and Ließ, M.: Development of pedotransfer functions for water retention in tropical mountain soil landscapes: spotlight on parameter tuning in machine learning, Soil, 6(1), 215–229, 2020.
Granger, S. J., Harris, P., Peukert, S., Guo, R., Tamburini, F., Blackwell, M. S. A., Howden, N. J. K. and McGrath, S.: Phosphate stable oxygen isotope variability within a temperate agricultural soil, Geoderma, 285, 64–75, doi:https://doi.org/10.1016/j.geoderma.2016.09.020, 2017.
Jung, Y., Lee, J., Lee, M., Kang, N. and Lee, I.: Probabilistic analytical target cascading using kernel density estimation for accurate uncertainty propagation, Struct. Multidiscip. Optim., 1–19, 2020.
Lee, J. and Park, K.: GAN-based imbalanced data intrusion detection system, Pers. Ubiquitous Comput., 25(1), 121–128, 2021.
Li, Y., Liu, J., Huang, J., Li, Z. and Liang, P.: Learning brain connectivity sub-networks by group-constrained sparse inverse covariance estimation for Alzheimer’s disease classification, Front. Neuroinform., 12, 58, 2018.
de Mello, D. C., Safanelli, J. L., Poppiel, R. R., Veloso, G. V., Cabrero, D. R. O., Greschuk, L. T., Mello, F. A. de O., Francelino, M. R., Ker, J. C. and Leite, E. P.: Soil apparent electrical conductivity survey in different pedoenvironments by geophysical sensor EM38: a potential tool in pedology and pedometry studies, Geocarto Int., (just-accepted), 1–24, 2022.
Neogi, S. and Dauwels, J.: Factored Latent-Dynamic Conditional Random Fields for Single and Multi-label Sequence Modeling, arXiv Prepr. arXiv1911.03667, 2019.
Peukert, S., Bol, R., Roberts, W., Macleod, C. J. A., Murray, P. J., Dixon, E. R. and Brazier, R. E.: Understanding spatial variability of soil properties: a key step in establishing field‐to farm‐scale agro‐ecosystem experiments, Rapid Commun. Mass Spectrom., 26(20), 2413–2421, 2012.
Ruiz, F., Sartor, L. R., de Souza Júnior, V. S., dos Santos, J. C. B. and Ferreira, T. O.: Fast pedogenesis of tropical Technosols developed from dolomitic limestone mine spoils (SE-Brazil), Geoderma, 374, 114439, 2020.
Ruiz, F., Andrade, G. R. P., Sartor, L. R., dos Santos, J. C. B., de Souza Júnior, V. S. and Ferreira, T. O.: The rhizosphere of tropical grasses as driver of soil weathering in embryonic Technosols (SE-Brazil), Catena, 208, 105764, 2022.
dos Santos Teixeira, A. F., Silva, S. H. G., Soares de Carvalho, T., Silva, A. O., Azarias Guimarães, A. and de Souza Moreira, F. M.: Soil physicochemical properties and terrain information predict soil enzymes activity in phytophysiognomies of the Quadrilátero Ferrífero region in Brazil, CATENA, 199, 105083, doi:https://doi.org/10.1016/j.catena.2020.105083, 2021.
Wani, S. A., Najar, G. R., Wani, J. A., Ramzan, M. and Hakeem, K. R.: Weathering and approaches to evaluation of weathering indices for soil profile studies–An overview, Soil Sci. Agric. Environ. Prospect., 183–198, 2016.
Wardhani, N. W. S., Rochayani, M. Y., Iriany, A., Sulistyono, A. D. and Lestantyo, P.: Cross-validation metrics for evaluating classification performance on imbalanced data, in 2019 international conference on computer, control, informatics and its applications (IC3INA), pp. 14–18, IEEE., 2019.
Xing, J., Wang, S. X., Jang, C., Zhu, Y. and Hao, J. M.: Nonlinear response of ozone to precursor emission changes in China: a modeling study using response surface methodology, Atmos. Chem. Phys., 11(10), 5027–5044, 2011.
Xu, X., Li, W., Tao, M., Xie, Z., Gao, X., Yue, L. and Wang, P.: Effective and Accurate Diagnosis of Subjective Cognitive Decline Based on Functional Connection and Graph Theory View, Front. Neurosci., 14, 2020.
Zhang, G. and Zhu, A.-X.: A representativeness heuristic for mitigating spatial bias in existing soil samples for digital soil mapping, Geoderma, 351, 130–143, doi:https://doi.org/10.1016/j.geoderma.2019.05.024, 2019.
Citation: https://doi.org/10.5194/soil-2022-17-AC4
-
AC4: 'Reply on RC1', Danilo Mello, 20 Oct 2022
Status: closed
-
CC1: 'Comment on soil-2022-17', Sara Ramos Santos, 26 Jul 2022
Dears colleagues
I have fill comments to say/ask about the work.
1. The use of geotecnology associated to marching learning is fundamental for the future of pedology, due to increasing the prediction process in great areas. Congratulations!
2. Which variables do you used to calculate the intemperism index? Are the variables obtained with sensors our in laboratory?
3. I think that would be good if this article will have a section in R&D correlationing intemperism to mapping soil fertility in this area.
Thanks in advance!
Congratulations for the work!
Citation: https://doi.org/10.5194/soil-2022-17-CC1 -
CC2: 'Reply on CC1', Danilo Mello, 01 Aug 2022
Dear Sarah,
1. The use of geotecnology associated to marching learning is fundamental for the future of pedology, due to increasing the prediction process in great areas. Congratulations!
Thank you for your comment and your congratulations for us related to this manuscript.
2. Which variables do you used to calculate the intemperism index? Are the variables obtained with sensors our in laboratory?
In this work we used the weathering index obtained with laboratory data, as explained in section 2.3. Weathering rates. We then combined this index with data obtained via geophysical sensors (in the field) and satellite (SYSI), which correspond to soil attributes generated by different intensities of weathering and, consequently, pedogenesis. Then we de-correlated the variables (Table 3). Subsequently, the PCA was performed using the kmeans method, where clustering indices corresponding to the different weathering indices were generated.
3. I think that would be good if this article will have a section in R&D correlating weathering to mapping soil fertility in this area.
The fertility issue was not addressed, because it was not the focus of this work. In addition, to properly assess soil fertility we would need soil fertility management data (liming and fertilization) by the company working in the area (since it is an agricultural, private area and company data is restricted). Therefore, it will not be possible nor feasible to hold a session on this topic in this work, which does not prevent this from being done in future works (by the way, we thank you for the idea and we will try to incorporate this in another work).
Regards,
Citation: https://doi.org/10.5194/soil-2022-17-CC2
-
CC2: 'Reply on CC1', Danilo Mello, 01 Aug 2022
-
CC3: 'Comment on soil-2022-17', Guilherme Oliveira, 31 Aug 2022
This article represents an excellent application of Machine learning techniques in geophysical-soil science data. The manuscript brings an excellent approach to the geophysical relationships with processes that occur in the soil (weathering) and variables of easy acquisition.
The methodological flowchart is very well defined, However, I had doubts in some specific parts.
First of all, the English writing needs to be improved. I suggest submitting to a specific proofreading company.
Why did you use the F1-score test instead of the Kappa metric or accuracy metric?
Is the data balanced? The accuracy is a good metric when the data is balanced. This is very important because all of these algorithms have a bias with unbalanced data. Moreover, the algorithms used (excepted for RF) required standardized data.
Why didn't you employ the covariates in the unsupervised clustering? I dont understand why you compare the number of PCA dimensions with the number of clusters.
With proper revisions, I believe it to be an important contribution to the journal.Citation: https://doi.org/10.5194/soil-2022-17-CC3 -
AC1: 'Reply on CC3', Danilo Mello, 20 Oct 2022
Guilherme Oliveira, 31 Aug 2022
This article represents an excellent application of Machine learning techniques in geophysical-soil science data. The manuscript brings an excellent approach to the geophysical relationships with processes that occur in the soil (weathering) and variables of easy acquisition.
A: Thank you for acknowledging our research.
The methodological flowchart is very well defined; However, I had doubts in some specific parts.
A: Thanks for the suggestions.
First of all, the English writing needs to be improved. I suggest submitting to a specific proofreading company.
A: We sent the for a general review of English (American English) to a specialized company, where a geoscience specialist also reviewed the entire manuscript (Proofreading service). A certificate attesting to the new revision of the manuscript was inserted in the "supplementary material" field.
Why did you use the F1-score test instead of the Kappa metric or accuracy metric?
A: The accuracy is an appropriate and good metric to evaluate the model’s performance, when the data is balanced. That is not our case. As the geophysical data used, as well as the data from the physicochemical analysis of the soils, are unevenly distributed in terms of the number of samples, over the different geologies and soil classes, our data were considered unbalanced.
Imbalanced data always becomes one of the classification issues. Imbalanced dataset occurs when one or several classes have less sample (the minority); while another/other classes are the majority, such as agricultural fields. Accuracy is the most intuitive evaluation index to show the performance of a model. However, when the data classes are imbalanced, a supplementary index is required. That index is the F1-Score, which is a harmonic average of Precision and Recall. The F1-Score can accurately evaluate the performance of the model when the data is imbalanced (Lee and Park, 2021; Wardhani et al., 2019). To further support this assertion, we have inserted two citations from recent works in good journals in the text that detail the use of the F-1 Score in terms of accuracy.
Is the data balanced? The accuracy is a good metric when the data is balanced. This is very important because all of these algorithms have a bias with unbalanced data. Moreover, the algorithms used (excepted for RF) required standardized data.
A: No, the geophysical and soil physico-chemical data are unbalanced. We understand the reviewer's point of view and we agree. Due to that, the results and discussion section were based on F-1 Score metrics, instead of accuracy. We only presented the accuracy and kappa for readers to compare. We used the F-1 Score metric due to criticisms and even recommendations from some other reviewers on articles previously submitted for publication in this and other journals. For this same reason, we also used 2 other parameters to complement the F-1 Score (sensitivity and specificity of the models).
Why didn't you employ the covariates in the unsupervised clustering? I dont understand why you compare the number of PCA dimensions with the number of clusters.
A: We did not employ the covariates in the unsupervised clustering because they were used to map the groups of evaluated attributes. If these covariates were used in clustering, they would act as predictor and predicted variables at the same time, generating overestimated performance results that do not match reality.
We did not compare the optimal number of clusters with the cps results. We use the two data together to create the groups using the k-means method. The use of cps instead of pure data is because the cps are not correlated with each other. On the other hand, pure data can present correlation between them.
With proper revisions, I believe it to be an important contribution to the journal.
A: Thank you for all the suggestion. We made a few changes to the text to better clarify future readers and they were certainly great contributions to our work.
Citation: https://doi.org/10.5194/soil-2022-17-AC1
-
AC1: 'Reply on CC3', Danilo Mello, 20 Oct 2022
-
CC4: 'Comment on soil-2022-17', Rafael Siqueira, 05 Sep 2022
I’ve read the pre-print version of this manuscript with great interest, due to the novelty associated to the application of modern technologies to investigate soil weathering and pedogenetic development, a very welcome initiative in Soil Science. The authors successfully combine two powerful approaches or tools constantly applied in the Pedometrics field: computational statistical techniques (machine learning and multivariate analysis) and geophysics proximal sensing data, aiming to understand in a quantitative manner how the soils develop in a tropical landscape. Nonetheless, I would like to write down some comments as reflections and suggestions:
I am not a native speaker, but a new look at the English writing for corrections would be important to improve the quality of the text. Furthermore, some parts of the text are a bit clumsy and a review would be welcome.
Abstract
Line 17 - I am not quite sure if you could call machine learning as a “geotechnology”, for the reason it was not originally developed for geosciences.
Line 19 – Satellite imagering is not a proximal sensing technique, but a remote sensing technique.
Line 21 – You must explain in the Abstract why you used the PCA and clusters analysis. At the same time, I recommend to cite what information have you used to create the clusters, and moreover, what they really mean, in the Abstract.
Line 22 – Change “we determine and used the ideal number of clusters” for “we used the ideal number of clusters”. You’ve already cited you determined the number of clusters before.
Line 27 – “The Nitisol over East diabase presented greater weathering intensity than”.
The last sentence of the Abstract is clumsy.
Introduction
The Introduction is very well written, making a brief literature review about weathering and the geophysical sensors.
Line 94 – Would not be better if you used just “model weathering intensity using combined data from geophysical sensors” than “model weathering index using combined data from geophysical sensors”?
Line 97 – I suggest to remove the objective 4, since you did not explore it in your results.
Line 99 – I would remove the citation about satellite here.
Material and methods
Line 109 - I guess the name of the municipality cited in the text is with just one “f”.
Line 135 - You could standardize the Embrapa citation. Sometimes you cite Embrapa, 2011 and sometimes Embrapa, 2017.
Line 135 – In the laboratory analysis topic, you make a complete description about physical and chemical analysis of soils according to Embrapa. However, you do not use any of them, unless the oxides contents. I suggest to remove all analyses which you did not apply in this paper.
Line 146 - I think it’d be productive you write one or two more lines explaining better how the oxides contents, the only analysis you really used, are obtained. Moreover, you cite Fe2O3 content but you also do not use it, only the SiO2 and TiO2 contents.
Line 150 - If you consider that index (Wl) used in the text as sufficient for your goals, I suggest you to better characterize this index in particular and its meaning about weathering. Moreover, I see in many parts of the text you using “weathering indexes” when in this paper you use just one index.
Line 197 - You are very correct in saying that the use of the IGC data allowed the creation of a more detailed topographic data than orbital sources, such as SRTM, could do. But just a hint for future works: with the scale of IGC database you claimed (1:10.000) you’d be able to obtain an MDE with a way finer resolution than you obtained, which would be more appropriate to the goals that you wanted to achieve.
Table 1 – very nice table explaining each morphometric variable. It will be a good reference for other authors.
Line 217 - It will be interesting you cite in an only place the “7 parameters derived from geophysical sensors data” which you used. According to your descriptions in the “Geophysical data collection” and the Table 3, you used 5 parameters from the geophysical sensors, not 7.
Line 228 - “argiluviation or ferralitization indexes”. Again, why two indexes if you supposedly used just one? Furthermore, it is the SiO2/TiO2 index that you described previously one of them? If so, you should have to indicate that before.
Line 229 - “These groups characterize themselves by present similar values within the groups, but different values between one group to another”. You should refine the description about the statistical clustering.
Line 275 - Very good insight about the combined use of the PCA and k-means analysis as well as the use of the Nested LOOCV for handling few samples.
Line 233 - “This result was used to extract the values of covariates (morphometric and geological data)”. Not only these covariates, but also the Synthetic Soil image (SYSI).
Line 236 – “base database”?
Line 250 – “increases the performance of machine processing algorithms”. Not necessarily.
Line 260 – the remotion by correlation is not only associated to “reduce computational time” but also to minimize the problem of multicollinearity.
Line 272 – I think it is “repeatecv”.
Line 350 – You cite here that you used the Kruskal Wallis test to choose the best model. But in your results, you only showed the Kruskal Wallis being used to differentiate the weathering clusters. You need to synchronize your material and methods with your results in this regard.
Line 359 – You don’t need to cite RF twice in the sentence.
Line 360 – Before “The RF algorithm presented equivalent performances than other algorithms”, insert “In other studies,”.
Line 381 – “performance precision”? Redundancy.
Line 381 – Paragraph shows some clumsy sentences.
Figure 5 - I’ve seen that the authors explain for the first time what the clusters really mean (weathering intensity) in a clear way only on the Figure 5. I suggest the authors to anticipate this important explanation. At the same time, I suggest the authors to label the clusters with the information of weathering, at last, the numbers 1,2,3 that the authors arbitrarily chose are just categorical, not expressing by themselves the weathering degree.
Line 451 – According to the rest of your text, the cluster 3 has higher weathering, followed by cluster 1 and then the cluster with lesser weathering, cluster 2. In this sentence, you have been confounded the order of the clusters.
Line 458 – “West Rodic Nitisol” – I think the authors confounded the terms. Here, the correct would be “East Rodic Nitisol”.
Line 465 – “For the weathering index”.
Table 3 – interesting table. Could explain in a paragraph which of the parameters explain better the weathering, according with your data? For me, the most important were the W1, plus magnetic susceptibility and ECa.
Line 494 – East diabase
Line 495 – West diabase and West Rhodic Nitisol
Line 514 – free drainage
Line 523 – but there is higher K40 in the supposedly more weathered Lixisols (cluster 1)
Line 526 – decreasing or increasing ECa? You must explain that.
Line 539 – you do not model with the nested LOOCV, but you evaluated it.
My suggestions aim to improve the quality of this excellent paper. I congratulate the authors for their work and hope to read more on this very interesting topic.
Citation: https://doi.org/10.5194/soil-2022-17-CC4 -
AC2: 'Reply on CC4', Danilo Mello, 20 Oct 2022
Rafael Siqueira, 05 Sep 2022
I’ve read the pre-print version of this manuscript with great interest, due to the novelty associated to the application of modern technologies to investigate soil weathering and pedogenetic development, a very welcome initiative in Soil Science. The authors successfully combine two powerful approaches or tools constantly applied in the Pedometrics field: computational statistical techniques (machine learning and multivariate analysis) and geophysics proximal sensing data, aiming to understand in a quantitative manner how the soils develop in a tropical landscape. Nonetheless, I would like to write down some comments as reflections and suggestions:
A: Thank you for taking the time to read our work and recognize the importance of our research.
I am not a native speaker, but a new look at the English writing for corrections would be important to improve the quality of the text. Furthermore, some parts of the text are a bit clumsy and a review would be welcome.
A: We sent the for a general review of English (American English) to a specialized company, where a geoscience specialist also reviewed the entire manuscript (Proofreading service). A certificate attesting to the new revision of the manuscript was inserted in the "supplementary material" field. In addition, we reviewed the entire text to locate and modify the parts that were unclear.
Abstract
Line 17 - I am not quite sure if you could call machine learning as a “geotechnology”, for the reason it was not originally developed for geosciences.
A: We agree with the reviewer and modify the sentence as suggested.
Line 19 – Satellite imagering is not a proximal sensing technique, but a remote sensing technique.
A: We agree with the reviewer and modify the sentence as suggested.
Line 21 – You must explain in the Abstract why you used the PCA and clusters analysis. At the same time, I recommend to cite what information have you used to create the clusters, and moreover, what they really mean, in the Abstract.
A: We used the cluster and PCA concomitantly to reduce the number of variables, which were later used in the cluster analysis, allowing the choice of uncorrelated groups, improving the performance of the analyses. The information used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.
We modified a little a sentence in abstract and material in methods (section 2.7 Principal component analysis and clusterization) better explain this step.
Abstract: “Afterwards, the principal component analysis and the ideal number of clusters was determined, in order to reduce the number of variables, which were later used in the cluster analysis. The data used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.”
Material and methods: “... (PCA) was applied to the 6 parameters derived from geophysical sensors data (eU, eTh, K40, κ and, ECa), and the weathering indexes.”
Line 22 – Change “we determine and used the ideal number of clusters” for “we used the ideal number of clusters”. You’ve already cited you determined the number of clusters before.
A: We agree with the reviewer and modify the sentence as recommended.
Line 27 – “The Nitisol over East diabase presented greater weathering intensity than”.
A: We do not understand this query. In the text the sentence is complete and correct. Please verify:
... the Nitisol presented greater weathering intensity than the Nitisol over West diabase areas.”
The last sentence of the Abstract is clumsy.
A: We agree with the reviewer and modify the sentence as recommended.
Introduction
The Introduction is very well written, making a brief literature review about weathering and the geophysical sensors.
A: Thank you for read.
Line 94 – Would not be better if you used just “model weathering intensity using combined data from geophysical sensors” than “model weathering index using combined data from geophysical sensors”?
A: We agree with the reviewer and modify the sentence as recommended.
Line 97 – I suggest to remove the objective 4, since you did not explore it in your results.
A: We agree with the reviewer and modify the sentence as recommended.
Line 99 – I would remove the citation about satellite here.
A: At this point we judged to keep the "satellite" in the sentence due to the use of SYSI, generated from satellite images in the survey.
Material and methods
Line 109 - I guess the name of the municipality cited in the text is with just one “f”.
A: We agree with the reviewer and modify the sentence as recommended.
Line 135 - You could standardize the Embrapa citation. Sometimes you cite Embrapa, 2011 and sometimes Embrapa, 2017.
A: We agree with the reviewer and modify the sentence as recommended.
Line 135 – In the laboratory analysis topic, you make a complete description about physical and chemical analysis of soils according to Embrapa. However, you do not use any of them, unless the oxides contents. I suggest to remove all analyses which you did not apply in this paper.
A: At this point, we decided to keep all the analyzes that were carried out, because data such as granulometry, fertility and others, allow us to correlate the intensity of weathering with the environment where the samples were collected, and relate them to geology and soil type, which was one of the approaches of the article.
Line 146 - I think it’d be productive you write one or two more lines explaining better how the oxides contents, the only analysis you really used, are obtained. Moreover, you cite Fe2O3 content but you also do not use it, only the SiO2 and TiO2 contents.
A: On this specific point, we disagree with the reviewer, as these analyzes are already traditionally well known in soil science and are already duly referenced. We apply the principle of objectivity in the writing of the work while ensuring the reproducibility of the methodology.
Line 150 - If you consider that index (Wl) used in the text as sufficient for your goals, I suggest you to better characterize this index in particular and its meaning about weathering. Moreover, I see in many parts of the text you using “weathering indexes” when in this paper you use just one index.
A: The term was better explained following the reviewer's recommendations. The manuscript was revised and the passages where we cited "weathering indexes" were corrected to "weathering index". The follow sentence was added in the final of this section:
“...This index is based on the principle that during chemical weathering, there is an intense leaching of mobile elements such as basic cations and silica (SiO2) and residual concentration of less mobile and soluble oxides such as Fe2O3, A2O3 and TiO2, mainly in tropical environments. In other words, these oxides gradually increase with increase in weathering intensity while all other elements are gradually reduced. This is the basis of calculating the weathering indices of soil which actually indicate the degree of weathering (Wani et al., 2016).”
Line 197 - You are very correct in saying that the use of the IGC data allowed the creation of a more detailed topographic data than orbital sources, such as SRTM, could do. But just a hint for future works: with the scale of IGC database you claimed (1:10.000) you’d be able to obtain an MDE with a way finer resolution than you obtained, which would be more appropriate to the goals that you wanted to achieve.
A: We thank the reviewer for this valuable tip. We are already adopting this suggestion in other work that is being developed.
Table 1 – very nice table explaining each morphometric variable. It will be a good reference for other authors.
A: We thank you for the recognition.
Line 217 - It will be interesting you cite in an only place the “7 parameters derived from geophysical sensors data” which you used. According to your descriptions in the “Geophysical data collection” and the Table 3, you used 5 parameters from the geophysical sensors, not 7.
A: We have amended the sentence as per the reviewer's suggestion. In fact, there were 6 parameters, since we consider each gamma channel as a parameter. We add the following snippet to the sentence:
“... (PCA) was applied to the 6 parameters derived from geophysical sensors data (eU, eTh, K40, κ, ECa) and the weathering indexes...”
Line 228 - “argiluviation or ferralitization indexes”. Again, why two indexes if you supposedly used just one? Furthermore, it is the SiO2/TiO2 index that you described previously one of them? If so, you should have to indicate that before.
A: In this sentence we have made a mistake. Therefore, the terms argilluviation and ferralitization were excluded from the sentence.
Line 229 - “These groups characterize themselves by present similar values within the groups, but different values between one group to another”. You should refine the description about the statistical clustering.
A: Sorry, but on this particular point, we disagree with the reviewer. We believe that this sentence is well explained. This explanation was the same given in other previously published works. However, we added the word "statistics" to the sentence, as follows:
“These groups characterize themselves by present statistical similar values within the groups, but different values between one group to another.”
Line 275 - Very good insight about the combined use of the PCA and k-means analysis as well as the use of the Nested LOOCV for handling few samples.
A: We thank the reviewer for recognizing the work.
Line 233 - “This result was used to extract the values of covariates (morphometric and geological data)”. Not only these covariates, but also the Synthetic Soil image (SYSI).
A: The sentence was adjusted following the reviewer recommendation.
Line 236 – “base database”?
A: The sentence was adjusted following the reviewer suggestion.
Line 250 – “increases the performance of machine processing algorithms”. Not necessarily.
A: Thanks for the review, however on this point we disagree with the reviewer. We cite and find works where other researchers have found that there really are increases in the performance of machine processing algorithms. Furthermore, we verified in this and other works by our group that this phenomenon actually occurs. If the reviewer sends us references with contrary scientific convictions, we will review this point, otherwise, we decide to keep it.
Line 260 – the remotion by correlation is not only associated to “reduce computational time” but also to minimize the problem of multicollinearity.
A: We add this in the sentence following the reviewer recommendation.
Line 272 – I think it is “repeatecv”.
A: We corrected the sentence following the reviewer suggestion.
Line 350 – You cite here that you used the Kruskal Wallis test to choose the best model. But in your results, you only showed the Kruskal Wallis being used to differentiate the weathering clusters. You need to synchronize your material and methods with your results in this regard.
A: Thanks for the review, however we decided to keep the results more relevant and easier for readers to explain.
Line 359 – You don’t need to cite RF twice in the sentence.
A: We adjusted the sentence following the reviewer suggestion.
Line 360 – Before “The RF algorithm presented equivalent performances than other algorithms”, insert “In other studies,”.
A: We adjusted the sentence following the reviewer suggestion.
Line 381 – “performance precision”? Redundancy.
A: We adjusted the sentence following the reviewer suggestion.
Line 381 – Paragraph shows some clumsy sentences.
A: The paragraph was revised by proof-reading service and was improved.
Figure 5 - I’ve seen that the authors explain for the first time what the clusters really mean (weathering intensity) in a clear way only on the Figure 5. I suggest the authors to anticipate this important explanation. At the same time, I suggest the authors to label the clusters with the information of weathering, at last, the numbers 1,2,3 that the authors arbitrarily chose are just categorical, not expressing by themselves the weathering degree.
A: We improve the figure as the reviewer suggested and, also, we added a short paragraph explaining for what the clusters were created and what they mean, in material and methods section, as follow:
“It is important to realize that, the clusters were generated in order to demonstrate the different degrees of weathering on the different lithologies and soil types.”
Line 451 – According to the rest of your text, the cluster 3 has higher weathering, followed by cluster 1 and then the cluster with lesser weathering, cluster 2. In this sentence, you have been confounded the order of the clusters.
A: We adjusted the sentence following the reviewer suggestion.
Line 458 – “West Rodic Nitisol” – I think the authors confounded the terms. Here, the correct would be “East Rodic Nitisol”.
A: We corrected the sentence following the reviewer's recommendations.
Line 465 – “For the weathering index”.
A: We corrected the sentence following the reviewer's recommendations.
Table 3 – interesting table. Could explain in a paragraph which of the parameters explain better the weathering, according with your data? For me, the most important were the W1, plus magnetic susceptibility and ECa.
A: A new paragraph has been added, following the reviewer's recommendations:
The greatest statistical differences and, consequently, the weathering intensity were better evidenced for magnetic susceptibility, ECa and radiometric data from gamma-ray spectrometry, respectively, evidencing the differences between the different weathering index.
Line 494 – East diabase
A: We corrected the sentence following the reviewer's recommendations.
Line 495 – West diabase and West Rhodic Nitisol
A: We corrected the sentence following the reviewer's recommendations.
Line 514 – free drainage
A: We corrected the sentence following the reviewer's recommendations.
Line 523 – but there is higher K40 in the supposedly more weathered Lixisols (cluster 1)
A: Dear reviewer, we believe that there was a mistake in the interpretation. All K40 values for area were low (less than 1), including for Lixisols areas.
Line 526 – decreasing or increasing ECa? You must explain that.
A: Dear reviewer, we added the follow sentence in this paragraph:
“...however, it is harder to explain how environmental variable contribute more and how they affect the ECa values during weathering intensity, because they all pedoenvironmental variable act concomitantly to produce ECa values (parent material, pedogenetic and weathering processes, mineralogy, texture, water dynamics on landscape, magnetic susceptibility, texture, CEC and base saturation) (Mello et al., 2022).”
Line 539 – you do not model with the nested LOOCV, but you evaluated it.
A: Dear reviewer, thanks for the review. We understand the reviewer's concern, however at this point we are referring that nested-LOOCV was suitable for modeling as a whole process, for small sample sets. The modeling in this work was carried out in three stages: training, validation and test. The nested-LOOCV inner loop is done in training and validation.
“The nested-LOOCV method is a double loop process, where in the internal loop, the model is trained with a data set of size n-1, using the LOOCV for the optimization of the final model. On the other hand, the external loop corresponds to the test. In this loop, the remaining sample is predicted using the final model calculated in the inner loop. This prediction result is stored with the observed value of the remaining sample and later used to calculate the algorithm’s performanc(Jung et al., 2020; Neogi and Dauwels, 2019). The two loops are run n times (n = total number of samples, in our case 71). All samples are inserted into the outer loop, where the values predicted by the final model of each algorithm are calculated with the predicted and observed values of each sample. Then, the final result of the machine learning algorithm's performance will be obtained by predicted and observed values stored in the external loop. This is a robust method to evaluate the algorithm’s performance and detects possible samples with problems in the collections or outliers. The training set generated in each loop went through the process of selecting covariates for importance and subsequent training”.
My suggestions aim to improve the quality of this excellent paper. I congratulate the authors for their work and hope to read more on this very interesting topic.
A: Thank you for the excellent reviews that contributed to our work and for the recognition.
Citation: https://doi.org/10.5194/soil-2022-17-AC2
-
AC2: 'Reply on CC4', Danilo Mello, 20 Oct 2022
-
EC1: 'Comment on soil-2022-17', Nicolas P.A. Saby, 05 Sep 2022
I made an in depth reading of the paper and I have several major concerns about the work and the paper.
The abstract is quite vague and not very informative. Please explain the overall goal, the assumption, the approach implemented and give precise results and findings.
The English does not seem to be ok.
The work is not enough clearly presented. A lot of questions arise when reading the actual version.
The overall goal of this study is not enough clearly presented in the introduction. First (i), modelling with ML of weathering index (WI) is quite vague. But for what ? Mapping, monitoring, statistical application? In (ii), it is said that importance of the covariates will be discussed but what is these covariates. This is not clear if these are the one for the digital soil mapping approach or the observed ones using sensor data. It is also a general comment about this paper where a confusion is made between these 2 types of information. The description of the proximal and remote sensing data should be better explained. The first ones are not spatially explicit.
The way geophysical data are used is rather difficult to understand and does not seem at all appropriate. It is quite difficult to understand why these data are mixed with the WI in the PCA step. This means that Geophysical data are not used as covariates but as a covariable. The PCA is not a method to evaluate (as it written in the abstract) but a multivariate analysis to explain the multivariate relations between variables. Moreover, the results of the PCA are not presented. It is not then possible to understand the signification of the principal components.
Why did you use a clustering approach ? We could expect a more quantitative digital soil mapping approach of the raw value of WI instead. Why are you creating clusters of WI and sensor data to map them after? Again, as the results of the PCA and the clustering are not presented, it is difficult to understand the signification of the cluster. It could happen that no correlation occurs between these data and the clusters represents only one of 2 variables. The results table 3 seem to validate this assumption as the discrimination between the different cluster are not very significant. However, an extensive discussion is provided about the signification of these clusters based on an interpretation of the dsm maps and by expertise. However, all the maps are wrong. Maybe a point map of the clusters would be better to be interpreted first. Finally, it is difficult to understand the adding value of the geophysical data.
In the 2.1, two sets of data are presented (16 and 79 points) but only the 79 points are used for the study. Could you explain ? finally, it is topsoil map of WI ?
The equation 1 is not explained enough. What is SIO2 and TIO2 ?
In 2.7, two approaches are presented for the clustering, eg scott lethod and k-means. Which one did you use? Do you mean that a kmeans approach was implemented and the optimal number of clusters was selected using a scott approach ?
The 2.8.2 should be titled "validation of the map".
The explanation of the nested approach is quite difficult to follow. Do you mean that there is a tuning of the ML algorithm at each step of the LOOCV? This is not indicated in the fig 3. How this tuning is done ? Bye cross validation?
283 should be titled validation not training
2.8.3 Why are you explaining that you are using a LOOCV now? This is very confusing as it is not explained in the fig 3. All the indicators explained in this section are based the result of the confusion matrix where the 4 kinds of results are computed, eg FP TP, FN and FN.
It is not explained how the different maps produced during the LOOCV are combined at the end of the process. Do you compute the dominant value? Are using a probability approach?
The "pls" is usually a regression approach and is not adapted to a classification exercise. Could you detail which algorithm you used?
The size of the dataset is quite not adapted to the use of ML algorithm like random forest. You may consider more classical algorithm like the multinomial algorithm. As you use a LOOCV, we could not exclude an overfitting of the data. A k-fold cross validation would be more adapted. The hyper parameter of the different algorithms are not described so it is difficult to understand the quality of the models.
The sampling design is also not all adapted to a digital soil mapping approach. The locations were selected by expert judgement and are not very well spread over the area. There is extensive discussion about how to collect data in the recently published paper. This not really discussed in this paper. It is only acknowledged that the sample are not enough even but the spread of the data into the covariate space should be checked to discuss the validity of the map produced.
I think this article needs to be thoroughly revised before publication. Th use of the clustering step and the geophyical data should be better justified.
Citation: https://doi.org/10.5194/soil-2022-17-EC1 -
AC3: 'Reply on EC1', Danilo Mello, 20 Oct 2022
Nicolas P.A. Saby, 05 Sep 2022
I made an in depth reading of the paper and I have several major concerns about the work and the paper.
The abstract is quite vague and not very informative. Please explain the overall goal, the assumption, the approach implemented and give precise results and findings.
A: Rereading the abstract, we agreed with the reviewer and made the necessary adjustments in order to clarify the research objectives, main results and conclusions of the study.
The English does not seem to be ok.
A: We sent the for a general review of English (American English) to a specialized company, where a geoscience specialist also reviewed the entire manuscript (Proofreading service). A certificate attesting to the new revision of the manuscript was inserted in the "supplementary material" field.
The work is not enough clearly presented. A lot of questions arise when reading the actual version.
A: Thanks for the suggestion, we agree with the reviewer. This may have occurred given the number of covariates and the explanations of the relationship between them with weathering, in addition to the clustering and modeling technique used. However, we revised the entire manuscript and some paragraphs that we thought were confused in the explanations, and/or missing some details were rewritten in order to clarify the reader. Another problem may have been English, which was also revised.
The overall goal of this study is not enough clearly presented in the introduction. First (i), modelling with ML of weathering index (WI) is quite vague. But for what? Mapping, monitoring, statistical application? In (ii), it is said that importance of the covariates will be discussed but what is these covariates. This is not clear if these are the one for the digital soil mapping approach or the observed ones using sensor data. It is also a general comment about this paper where a confusion is made between these 2 types of information. The description of the proximal and remote sensing data should be better explained. The first ones are not spatially explicit.
A: Thanks for the review. Dear reviewer, we have re-evaluated the objectives and they are clearly met. Modeling the intensity of weathering means predicting and spatializing the data (maps shown in figures 5 and 6, for example, and table 2). The applications of obtaining weathering intensities, or in other words, more and less weathered soils, were briefly explained in the introduction (paragraph x). The importance of the variables used were those used in the modeling and which were duly presented and discussed in section 3.1 (3.1 Evaluation of model's performance, uncertainty and variables importance), including a figure (figure 4) demonstrating what these variables were. We believe that these variables should not be presented in the objectives of the work, but they should be explained and demonstrated in the methodology, as well as presented and discussed in the sessions indicated in the manuscript.
We have added the following paragraph after the objectives to clarify the relationship between digital mapping, remote and near sensing, and weathering:
“In this work we use the digital soil mapping approach (using geophysical sensors and satellite imagery) to estimate different intensities of weathering in an area that are comparable in terms of geology and soil types.”
We have added a new paragraph in the introduction to clarify the concept of remote and near sensing:
“The proximal sensing is the use of field-based sensors to obtain signals from the soil when the sensor’s detector is in contact with or close to (within 2 m) the soil (e.g., geophysical sensors). The sensors provide soil information because the signals correspond to physical measures, which can be related to soils and their properties. When the distance at which a particular sensor acquired a data is greater than 2m from the target, there is remote sensing (e.g., satellite images) (Rossel et al., 2011).”
The way geophysical data are used is rather difficult to understand and does not seem at all appropriate. It is quite difficult to understand why these data are mixed with the WI in the PCA step. This means that Geophysical data are not used as covariates but as a covariable. The PCA is not a method to evaluate (as it written in the abstract) but a multivariate analysis to explain the multivariate relations between variables. Moreover, the results of the PCA are not presented. It is not then possible to understand the signification of the principal components.
A: In this work we used the weathering index obtained with laboratory data, as explained in section 2.3. Weathering rates. We then combined this index with data obtained via geophysical sensors (in the field) and satellite (SYSI), which correspond to soil attributes generated by different intensities of weathering and, consequently, pedogenesis. Then we de-correlated the variables (Table 3). Subsequently, the PCA was performed using the kmeans method, where clustering indices corresponding to the different weathering indices were generated. We used the cluster and PCA concomitantly to reduce the number of variables, which were later used in the cluster analysis, allowing the choice of uncorrelated groups, improving the performance of the analyses. The information used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.
In addition, we adjusted the abstract following the reviewer issues about PCA.
We did not show PCA because it was used only to de-correlate sensor data.
Why did you use a clustering approach? We could expect a more quantitative digital soil mapping approach of the raw value of WI instead. Why are you creating clusters of WI and sensor data to map them after? Again, as the results of the PCA and the clustering are not presented, it is difficult to understand the signification of the cluster. It could happen that no correlation occurs between these data and the clusters represents only one of 2 variables. The results table 3 seem to validate this assumption as the discrimination between the different cluster are not very significant. However, an extensive discussion is provided about the signification of these clusters based on an interpretation of the dsm maps and by expertise. However, all the maps are wrong. Maybe a point map of the clusters would be better to be interpreted first. Finally, it is difficult to understand the adding value of the geophysical data.
A: We use a cluster approach in order to separate weathering intensity into classes.
In fact, we could expect a more quantitative digital soil mapping approach with raw WI values. However, it was not the focus of this work, since there are already other works that have already done this in the area of pedometrics.
In this work we used the weathering index obtained with laboratory data, as explained in section 2.3. Weathering rates. We then combined this index with data obtained via geophysical sensors (in the field) and satellite (SYSI), which correspond to soil attributes generated by different intensities of weathering and, consequently, pedogenesis. Then we de-correlated the variables (Table 3). Subsequently, the PCA was performed using the kmeans method, where clustering indices corresponding to the different weathering indices were generated. We used the cluster and PCA concomitantly to reduce the number of variables, which were later used in the cluster analysis, allowing the choice of uncorrelated groups, improving the performance of the analyses. The information used to create the clusters were: 6 parameters derived from geophysical sensors data (eU, eTh, K40, magnetic susceptibility and, ECa), and the weathering index.
Indeed, it could happen that no correlation occurs between these data and the clusters represents only one of 2 variables. However, as the reviewer observed, we performed a non-parametric assessment of sensor data and WI between groups (the Kruskal-Wallis test) to assess this issue and found that virtually all of them had statistical differences.
We believe that the reviewer is a little mistaken, since the values in table 3 indicate differences for most classes. Regarding the issue that the maps are wrong, we would like to know based on what arguments the reviewer asserts this issue. Wrong about what? As such, we are unable to respond.
The geophysical data contributed towards having significant relationships with weathering and the soil attributes associated with them. as discussed in the work.
In the 2.1, two sets of data are presented (16 and 79 points) but only the 79 points are used for the study. Could you explain? finally, it is topsoil map of WI?
A: Yes, the reviewer is right. Sixteen soil profiles were described, as shown in figure 1. Seventy-nine points where samples were collected for physical-chemical laboratory analysis and, geophysical data. Data from the 16 soil profiles were used for soil classification, for further analysis between the intensity of weathering and the type of soil.
In fact, the weathering intensity was evaluated in the first 20cm of the soil, so yes, for the top soil, although the gamma gets values of readings of 30cm of depth.
The equation 1 is not explained enough. What is SIO2 and TIO2?
A: Thanks for the suggestion. We better explained the equation and the relationship between the elements and, add a new paragraph in material and methods section, as follow:
“The W1 index (Eq. 1) is based on the principle that during chemical weathering, there is an intense leaching of mobile elements such as basic cations and silica/ silicon oxide (SiO2) and residual concentration of less mobile and soluble oxides such as Fe2O3, Al2O3 and TiO2, mainly in tropical environments. In other words, these oxides gradually increase with increase in weathering intensity while all other elements are gradually reduced. This is the basis of calculating the weathering indices of soil which actually indicate the degree of weathering (Wani et al., 2016).”
In 2.7, two approaches are presented for the clustering, eg scott lethod and k-means. Which one did you use? Do you mean that a kmeans approach was implemented and the optimal number of clusters was selected using a scott approach?
A: The Scott test was used to determine the ideal number of clusters, while the k-means test was used to classify the samples in the group determined by scott.
The 2.8.2 should be titled "validation of the map".
A: We adjusted the sentence following the reviewer suggestion.
The explanation of the nested approach is quite difficult to follow. Do you mean that there is a tuning of the ML algorithm at each step of the LOOCV? This is not indicated in the fig 3. How this tuning is done? Bye cross validation?
A: The LOOCV (cross-validation) was used to optimize the hyperparameters of the evaluated algorithms, in each inner loop run. This phase is located inside the dotted square in figure 3. We are changing the flowchart to make this information clearer.
In addition, we tried to explain the nested-LOOCV method as follow:
“The nested-LOOCV method is a double loop process, where in the internal loop, the model is trained with a data set of size n-1, using the LOOCV for the optimization of the final model. On the other hand, the external loop corresponds to the test. In this loop, the remaining sample is predicted using the final model calculated in the inner loop. This prediction result is stored with the observed value of the remaining sample and later used to calculate the algorithm’s performance (Jung et al., 2020; Neogi and Dauwels, 2019). The two loops are run n times (n = total number of samples, in our case 75). All samples are inserted into the outer loop, where the values predicted by the final model of each algorithm are calculated with the predicted and observed values of each sample. Then, the final result of the machine learning algorithm's performance will be obtained by predicted and observed values stored in the external loop. This is a robust method to evaluate the algorithm’s performance and detects possible samples with problems in the collections or outliers. The training set generated in each loop went through the process of selecting covariates for importance and subsequent training”.
283 should be titled validation not training.
A: We adjusted the sentence following the reviewer suggestion.
2.8.3 Why are you explaining that you are using a LOOCV now? This is very confusing as it is not explained in the fig 3. All the indicators explained in this section are based the result of the confusion matrix where the 4 kinds of results are computed, e.g., FP TP, FN and FN.
A: The LOOCV is used to optimize the hyperparameters of the evaluated algorithms. It is in this item that this point is made. In figure 3 this phase is the two-color square. The performance parameters used in this work (indicators) were all taken from the confusion matrix from which the FP TP, FN and PN indices are obtained. This point is inside the dotted square in figure 3.
It is not explained how the different maps produced during the LOOCV are combined at the end of the process. Do you compute the dominant value? Are using a probability approach?
A: Dear reviewer, we believe that this question is duly explained in item 2.8.3:
“...Then, the final map was created by combining the 71 prediction maps generated for each algorithm tested. In addition, the mode value for each pixel of the final map was calculated. The prediction error map was elaborated, considering the number of times that each algorithm chose the mode value in each map pixel normalized by the number of final maps (%)....”
“...The best final map chosen by the previous statistical tests was used to extract the geophysical sensor data, weathering indexes values at the sampling points....”
The "pls" is usually a regression approach and is not adapted to a classification exercise. Could you detail which algorithm you used?
A: Thank you for this revision. In fact, this algorithm (pls) can be used as a classifier. As detailed in the text, the algorithms used, as well as their explanation, can be found in the caret manual at the link https://topepo.github.io/caret/available-models.html "6 Available Models". The pls is defined as "Partial Least Squares"- PLS. This algorithm can be used in classification and regression. The detailed explanation of the use of this algorithm is not relevant in this article, since it is an application of the algorithms (modeling) in soil science.
The size of the dataset is quite not adapted to the use of ML algorithm like random forest. You may consider more classical algorithm like the multinomial algorithm. As you use a LOOCV, we could not exclude an overfitting of the data. A k-fold cross validation would be more adapted. The hyper parameter of the different algorithms are not described so it is difficult to understand the quality of the models.
A: Firstly, we would like to have a larger number of samples than our current number and, we believe that all researchers would like to, but in our field conditions it was not possible as many other field conditions around the world. In addition, there is no minimum number of samples in the literature, indicated as a correct reference value for modeling soil and/or geophysical attributes. In addition, there are several researches published in good quality scientific journals, including Geoderma, which the authors also used a small number of samples in modeling processes (Fabijańczyk et al., 2017; Gebauer et al., 2020; Granger et al., 2017; Peukert et al., 2012; dos Santos Teixeira et al., 2021; Zhang and Zhu, 2019). In addition, LOOCV is more suitable for datasets with a small number of samples than k-fold.
In addition, we added the following sentence in material and methods:
The hyperparameters of each algorithm are described in the caret package manual in chapter 6. “Models described” available at https://topepo.github.io/caret/train-models-by-tag.html.
The sampling design is also not all adapted to a digital soil mapping approach. The locations were selected by expert judgement and are not very well spread over the area. There is extensive discussion about how to collect data in the recently published paper. This not really discussed in this paper. It is only acknowledged that the sample are not enough even but the spread of the data into the covariate space should be checked to discuss the validity of the map produced.
A: This sampling theory is valid and most used in geoesthetics. In ML, the use of conditioned Latin hypercube is being more used to define sampling sites. In addition, we would like to emphasize that during data collection, we advanced as far as possible due to the limitations of the sugarcane crop. However, we tried to carry out a distributed and representative sampling of the area, collecting data on all types of soils and lithology existing in the area, besides the topossequences.
I think this article needs to be thoroughly revised before publication. The use of the clustering step and the geophyical data should be better justified.
A: We agree with the reviewer and, we revied and explained in more detail and justification the two approaches (issues) highlighted by the reviewer.
Citation: https://doi.org/10.5194/soil-2022-17-AC3
-
AC3: 'Reply on EC1', Danilo Mello, 20 Oct 2022
-
RC1: 'Comment on soil-2022-17', Anonymous Referee #1, 15 Sep 2022
The article has an interesting general idea, but it raises some questions about the study proposal.
The first general question in the article that I didn't see answered is: What would be the real importance of evaluating soil weathering by different techniques? Soil fertility? This can be measured directly. 'environmental issues'? there are also techiniques to evaluate directly.
Listing the objectives of this study, not all of them were answered in the course of the article, nor in the conclusions section, leaving readers without a final answer to certain points raised.
Some errors in concepts and terms were noticed, such as:
-Machine learning is not a geotechnologie.
-‘proximal remote sensing’? satellite image is not proximal and geophysical is not remote.
-Spectroscopy method is not a geotechnology
I miss a paragraph explaining how the fact of weathering affects everything you said. Bearing in mind that this process takes years and years to affect soil properties, would that make any difference now? And if you say this can be seen in the difference in soil types, then isn't it easy to study the difference between soils?
Table 2 – value of accuracy from kknn is wrong.
When comparing the algorithms, I didn't see you talking about which parameters were used in each of them.
In terms of modeling and mapping, as you mentioned, the number of samples is very low, not being a great number of samples to work in an area of almost 200 ha. In addition to using it to calibrate a model with RF. Don't you think this would affect the construction of the model? don't you think the data wouldn't be overfitted?
Speaking now about the mapping exercise, in my opinion the distribution of samples is not adequate to carry out the mapping exercise.
My main concern is: Even knowing that the amount of samples was not enough and as the arrangement of samples is not suitable for the DSM approach, you still decided to carry out the article.
I suggest improving the quality of the data and methodologies used and strongly revising the article
Citation: https://doi.org/10.5194/soil-2022-17-RC1 -
AC4: 'Reply on RC1', Danilo Mello, 20 Oct 2022
Anonymous Referee #1,
The article has an interesting general idea, but it raises some questions about the study proposal.
A: Thanks for the acknowledgment, we checked all the issues about the study proposal pointed out by the reviewer.
The first general question in the article that I didn't see answered is: What would be the real importance of evaluating soil weathering by different techniques? Soil fertility? This can be measured directly. 'environmental issues'? there are also techiniques to evaluate directly.
A: Weathering operates in a multitude of spatial (from a nanometer to a planetary scale) and temporal (from thousands to only a few years) scales and its action impacts several, if not all, Earth systems. Weathering releases solutes that nourish every terrestrial ecosystem, triggers the biogeochemical cycles of every chemical element and thus control both water (e.g., river, ocean, groundwater) and soil chemistry. From a larger scale perspective, weathering is one of the major planetary sinks for CO2, and thus, play a pivotal role on climate change and life itself on Earth. As an extremely intricate and complex process, and in response to its interface nature, weathering has been assessed, quantified, measured, and studied from innumerous techniques and perspectives; all originated from different scientific fields and backgrounds. Many of such approaches and techniques are expensive and time-consuming and thus make it difficult to gather information about large areas in an accessible, reliable, and fast way. In the current scenario, where science must tackle many Earth science challenges in which weathering is a central process, the development of novel tools to study weathering is crucial to achieve the expected goals (Egli et al., 2001; Frings and Buss, 2019; Ruiz et al., 2020, 2022).
Listing the objectives of this study, not all of them were answered in the course of the article, nor in the conclusions section, leaving readers without a final answer to certain points raised.
A: We agreed with the reviewer and added a new paragraph:
“The combined use of geophysical sensors, satellite images and morphometry, by different machine learning algorithms proved to be a robust method and were able to model different weathering intensities.”
The other parts of the objectives we believe are answered along the other parts of the conclusions.
Some errors in concepts and terms were noticed, such as:
-Machine learning is not a geotechnology.
A: We adjusted the sentences following the reviewer suggestions.
‘proximal remote sensing’? satellite image is not proximal and geophysical is not remote.
A: We adjusted the sentences following the reviewer suggestion.
Spectroscopy method is not a geotechnology
A: We adjusted the sentences following the reviewer suggestion.
I miss a paragraph explaining how the fact of weathering affects everything you said. Bearing in mind that this process takes years and years to affect soil properties, would that make any difference now? And if you say this can be seen in the difference in soil types, then isn't it easy to study the difference between soils?
A: We agree with the reviewer and add the following paragraph:
“Weathering operates in a multitude of spatial (from a nanometer to a planetary scale) and temporal (from thousands to only a few years) scales and its action impacts several, if not all, Earth systems. Weathering releases solutes that nourish every terrestrial ecosystem, triggers the biogeochemical cycles of every chemical element and thus control both water (e.g., river, ocean, groundwater) and soil chemistry. From a larger scale perspective, weathering is one of the major planetary sinks for CO2, and thus, play a pivotal role on climate change and life itself on Earth. As an extremely intricate and complex process, and in response to its interface nature, weathering has been assessed, quantified, measured, and studied from innumerous techniques and perspectives; all originated from different scientific fields and backgrounds. Many of such approaches and techniques are expensive and time-consuming and thus make it difficult to gather information about large areas in an accessible, reliable, and fast way. In the current scenario, where science must tackle many Earth science challenges in which weathering is a central process, the development of novel tools to study weathering is crucial to achieve the expected goals (Egli et al., 2001; Frings and Buss, 2019; Ruiz et al., 2020, 2022).”
Dear reviewer, in relation to the time in which weathering acts and affects soil properties, there are numerous works in the literature that report how weathering can be extremely fast, especially in the tropical environment and under more susceptible parent materials.
Egli, M., Mirabella, A., & Fitze, P. (2001). Clay mineral transformations in soils affected by fluorine and depletion of organic matter within a time span of 24 years. Geoderma, 103(3-4), 307-334.
Ruiz, F., Andrade, G. R. P., Sartor, L. R., dos Santos, J. C. B., de Souza Júnior, V. S., & Ferreira, T. O. (2022). The rhizosphere of tropical grasses as driver of soil weathering in embryonic Technosols (SE-Brazil). Catena, 208, 105764.
Ruiz, F., Sartor, L. R., de Souza Júnior, V. S., dos Santos, J. C. B., & Ferreira, T. O. (2020). Fast pedogenesis of tropical Technosols developed from dolomitic limestone mine spoils (SE-Brazil). Geoderma, 374, 114439.
Table 2 – value of accuracy from kknn is wrong.
A: The text was adjusted following the reviewer suggestion.
When comparing the algorithms, I didn't see you talking about which parameters were used in each of them.
A: We added a short sentence in text:
The hyperparameters of each algorithm are described in the caret package manual in chapter 6. “Models described” available at https://topepo.github.io/caret/train-models-by-tag.html.
Here we presented the hyperparameter:
Algorithm
Hyperparameters
Definition
RF
“mtry”
number of covariates that are chosen at random for each node in the tree. Being the only parameter optimized.
ntree
number of trees. BREIMAN, L, 2002, considers keeping the value constant at 100.
nodesize
the minimum number of data points on each terminal node. BREIMAN, L, 2002, considers keeping the value constant at 1 and 5 for classification and regression.
KKNN
kmax
Number of neighbors within the search area. Optimized.
distance
Maximum search distance from the center of the test area. Optimized.
kernel
Sample space transformation kernel. Optimized.
Partial Least Squares (PLS)
ncomp
the number of main components to be used in modeling. Optimized.
avNNet (Model Averaged Neural Network)
size
Number of units (neurons) in the hidden layer. Optimized.
decay
parameter for weight drop using neural network optimization. optimized
bag
Logical index of the use of bootstrap methodology in training optimization. Optimized.
In terms of modeling and mapping, as you mentioned, the number of samples is very low, not being a great number of samples to work in an area of almost 200 ha. In addition to using it to calibrate a model with RF. Don't you think this would affect the construction of the model? don't you think the data wouldn't be overfitted?
A: Firstly, we would like to have a larger number of samples than our current number and, we believe that all researchers would like to, but in our field conditions it was not possible as many other field conditions around the world. In addition, there is no minimum number of samples in the literature, indicated as a correct reference value for modeling soil and/or geophysical attributes. In addition, there are several researches published in good quality scientific journals, including Geoderma, which the authors also used a small number of samples in modeling processes (Fabijańczyk et al., 2017; Gebauer et al., 2020; Granger et al., 2017; Peukert et al., 2012; dos Santos Teixeira et al., 2021; Zhang and Zhu, 2019).
We are aware of the possibility of overfitting, so applying the nested-LOOCV methodology, which we believe is more suitable for a small number of samples.
The nested-LOOCV method is indicated as a set for small data sets, which other methods of evaluation of test samples would not be viable due to the low number of samples in a test samples (Ferreira et al., 2021), being more used in the field of medicine in human experiments or where the number of samples is limited, providing an unbiased estimate of the true error (Chen et al., 2017; Li et al., 2018; Xing et al., 2011; Xu et al., 2020). The Nested-LOOCV method is a double loop process, where in the first loop the model is trained with a data set of size n-1, and the test is done in the second loop with the missing sample and used to validate the test and the training performance.
Speaking now about the mapping exercise, in my opinion the distribution of samples is not adequate to carry out the mapping exercise.
A: Here, we would like to emphasize that: during data collection, we advanced as far as possible due to the limitations of the sugarcane crop. However, we tried to carry out a distributed and representative sampling of the area, collecting data on all soil types, toposequences and, lithology existing in the area.
My main concern is: Even knowing that the amount of samples was not enough and as the arrangement of samples is not suitable for the DSM approach, you still decided to carry out the article.
A: We understand the reviewer concern. However, we do have a few observations to note: Firstly, we would like to have a larger number of samples than our current number and, we believe that all researchers would like to, but in our field conditions it was not possible as many other field conditions around the world. In addition, there is no minimum number of samples in the literature, indicated as a correct reference value for modeling soil and/or geophysical attributes.
There are several researches published in good quality scientific journals, including Geoderma, which the authors also used a small number of samples in modeling processes such as: (Fabijańczyk et al., 2017; Gebauer et al., 2020; Granger et al., 2017; Peukert et al., 2012; dos Santos Teixeira et al., 2021; Zhang and Zhu, 2019)
Finally, during data collection, we advanced as far as possible due to the limitations of the sugarcane crop. However, we tried to carry out a distributed and representative sampling of the area, collecting data on all soil types, toposequences and, lithology existing in the area.
I suggest improving the quality of the data and methodologies used and strongly revising the article.
A: The entire article was revised following the reviewer suggestion.
References
Chen, X., Zhang, H., Lee, S.-W. and Shen, D.: Hierarchical high-order functional connectivity networks and selective feature fusion for MCI classification, Neuroinformatics, 15(3), 271–284, 2017.
Egli, M., Mirabella, A. and Fitze, P.: Clay mineral transformations in soils affected by fluorine and depletion of organic matter within a time span of 24 years, Geoderma, 103(3–4), 307–334, 2001.
Fabijańczyk, P., Zawadzki, J. and Magiera, T.: Magnetometric assessment of soil contamination in problematic area using empirical Bayesian and indicator kriging: A case study in Upper Silesia, Poland, Geoderma, 308, 69–77, doi:https://doi.org/10.1016/j.geoderma.2017.08.029, 2017.
Ferreira, R. G., da Silva, D. D., Elesbon, A. A. A., Fernandes-Filho, E. I., Veloso, G. V., de Souza Fraga, M. and Ferreira, L. B.: Machine learning models for streamflow regionalization in a tropical watershed, J. Environ. Manage., 280, 111713, 2021.
Frings, P. J. and Buss, H. L.: The Central Role of Weathering in the Geosciences, Elements, 15(4), 229–234, doi:10.2138/gselements.15.4.229, 2019.
Gebauer, A., Ellinger, M., Brito Gomez, V. M. and Ließ, M.: Development of pedotransfer functions for water retention in tropical mountain soil landscapes: spotlight on parameter tuning in machine learning, Soil, 6(1), 215–229, 2020.
Granger, S. J., Harris, P., Peukert, S., Guo, R., Tamburini, F., Blackwell, M. S. A., Howden, N. J. K. and McGrath, S.: Phosphate stable oxygen isotope variability within a temperate agricultural soil, Geoderma, 285, 64–75, doi:https://doi.org/10.1016/j.geoderma.2016.09.020, 2017.
Jung, Y., Lee, J., Lee, M., Kang, N. and Lee, I.: Probabilistic analytical target cascading using kernel density estimation for accurate uncertainty propagation, Struct. Multidiscip. Optim., 1–19, 2020.
Lee, J. and Park, K.: GAN-based imbalanced data intrusion detection system, Pers. Ubiquitous Comput., 25(1), 121–128, 2021.
Li, Y., Liu, J., Huang, J., Li, Z. and Liang, P.: Learning brain connectivity sub-networks by group-constrained sparse inverse covariance estimation for Alzheimer’s disease classification, Front. Neuroinform., 12, 58, 2018.
de Mello, D. C., Safanelli, J. L., Poppiel, R. R., Veloso, G. V., Cabrero, D. R. O., Greschuk, L. T., Mello, F. A. de O., Francelino, M. R., Ker, J. C. and Leite, E. P.: Soil apparent electrical conductivity survey in different pedoenvironments by geophysical sensor EM38: a potential tool in pedology and pedometry studies, Geocarto Int., (just-accepted), 1–24, 2022.
Neogi, S. and Dauwels, J.: Factored Latent-Dynamic Conditional Random Fields for Single and Multi-label Sequence Modeling, arXiv Prepr. arXiv1911.03667, 2019.
Peukert, S., Bol, R., Roberts, W., Macleod, C. J. A., Murray, P. J., Dixon, E. R. and Brazier, R. E.: Understanding spatial variability of soil properties: a key step in establishing field‐to farm‐scale agro‐ecosystem experiments, Rapid Commun. Mass Spectrom., 26(20), 2413–2421, 2012.
Ruiz, F., Sartor, L. R., de Souza Júnior, V. S., dos Santos, J. C. B. and Ferreira, T. O.: Fast pedogenesis of tropical Technosols developed from dolomitic limestone mine spoils (SE-Brazil), Geoderma, 374, 114439, 2020.
Ruiz, F., Andrade, G. R. P., Sartor, L. R., dos Santos, J. C. B., de Souza Júnior, V. S. and Ferreira, T. O.: The rhizosphere of tropical grasses as driver of soil weathering in embryonic Technosols (SE-Brazil), Catena, 208, 105764, 2022.
dos Santos Teixeira, A. F., Silva, S. H. G., Soares de Carvalho, T., Silva, A. O., Azarias Guimarães, A. and de Souza Moreira, F. M.: Soil physicochemical properties and terrain information predict soil enzymes activity in phytophysiognomies of the Quadrilátero Ferrífero region in Brazil, CATENA, 199, 105083, doi:https://doi.org/10.1016/j.catena.2020.105083, 2021.
Wani, S. A., Najar, G. R., Wani, J. A., Ramzan, M. and Hakeem, K. R.: Weathering and approaches to evaluation of weathering indices for soil profile studies–An overview, Soil Sci. Agric. Environ. Prospect., 183–198, 2016.
Wardhani, N. W. S., Rochayani, M. Y., Iriany, A., Sulistyono, A. D. and Lestantyo, P.: Cross-validation metrics for evaluating classification performance on imbalanced data, in 2019 international conference on computer, control, informatics and its applications (IC3INA), pp. 14–18, IEEE., 2019.
Xing, J., Wang, S. X., Jang, C., Zhu, Y. and Hao, J. M.: Nonlinear response of ozone to precursor emission changes in China: a modeling study using response surface methodology, Atmos. Chem. Phys., 11(10), 5027–5044, 2011.
Xu, X., Li, W., Tao, M., Xie, Z., Gao, X., Yue, L. and Wang, P.: Effective and Accurate Diagnosis of Subjective Cognitive Decline Based on Functional Connection and Graph Theory View, Front. Neurosci., 14, 2020.
Zhang, G. and Zhu, A.-X.: A representativeness heuristic for mitigating spatial bias in existing soil samples for digital soil mapping, Geoderma, 351, 130–143, doi:https://doi.org/10.1016/j.geoderma.2019.05.024, 2019.
Citation: https://doi.org/10.5194/soil-2022-17-AC4
-
AC4: 'Reply on RC1', Danilo Mello, 20 Oct 2022
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 876 | 379 | 55 | 1,310 | 26 | 28 |
- HTML: 876
- PDF: 379
- XML: 55
- Total: 1,310
- BibTeX: 26
- EndNote: 28
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
