Overall

This manuscript needs further proofreading and editing to due to spelling and grammatical errors and awkward phrasing in English. Detailed grammatical edits have not been specified in this review as a complete editing of the manuscript is required. There are also undefined acronyms.

While I understand the challenges the authors faced with sample collection. I do not think nested LOOCV is sufficiently rigorous. Especially since the feature selection was done using LOOCV, and in effect then the entire data set, there is no truly independent test set. LOOCV is an insufficiently rigorous test method, particularly when it is used for both hyper parameter tuning and feature selection. There is not a true independent evaluation of model performance in this study.

The authors should be using LOOCV on a separate training data set for feature selection and hyperparameter optimization, with the test data set withheld for final validation only. The data should either be split into relatively small sample sets given only 79 samples are available, or more data should be acquired and added to the study.

The R² values using the LOOCV are low enough that a true independent validation might have no relationship at all between the sensors and the parameters of interest. The results with an independent test set might also be very similar. There is no way to tell given the analysis approach which will have introduced an unknown amount of positive bias into the validation results.  The author’s conclusions that it is possible to model soil attributes satisfactorily is not warranted given the analysis approach.

Scientific Questions

Line 24: Make clear how validation was done in the abstract. Cross validation? Independent test dataset?

Line 57: Why did you choose to focus on gamma-spectrometry, magnetic susceptibility and apparent electrical conductivity. The authors need to add a justification. The explain the value of these approaches, by not why they selected them vs other methods.

Line 147: What correction factor was used for the Walkley-Black method? Was a soil specific correction factor available and used? Elemental analysis by dry combustion would be a better analysis option, however I understand due to cost is may not always be possible. However, the under consumption of organic matter during Walkley black and associated correction factor should be discussed.

Line 228: While I understand the challenges the authors faced with sample collection. I do not think nested LOOCV is sufficiently rigorous.

Technical Corrections

Line 25: Please state actual best r2 values not just greater than 0.2. That is unclear

Line 91: While I agree that the best models are those that use the smallest number of variables, the authors need to explain why and not just state it is better as a self-evident truth.

Line 198: Table 1 is introduced much earlier than it is discussed.

Line 422: Elaborate on what the relationships are between the VisNIR and XRF sensors.

Line 427: Add more specifics about what pedogenesis and soil attributes are being accounted for with these sensors

Line 433: State performance metrics not just satisfactory performance. Satisfactory is contextual and subjective. Just stating not greater than 0.5 is still vague.

Line 451: It can justify the low R² values obtained is very unclear.

Line 452: Further justification and explanation needs to be added that 0.2 to 0.5 is satisfactory. I don’t necessarily disagree, but a more rigorous argument needs to be made rather than just stating a convention. Why does this present informative results?

Line 454: It is unclear what point the authors are trying to make with the statement about standardized laboratory conditions

Line 461: Add more explanation why, not just that Cracknell and Reading (2003) states it.

Line 465: Explain why this is satisfactory not just that it is

Line 471: The discussion about NULL_RMSE and NULL_MAE is valuable. This needs to be explained earlier and used to justify the results as satisfactory.

Line 532: Based on the nested LOOCV approach, the results are not truly unbiased estimates

Line 544: The authors cannot truly conclude that modelling results was satisfactory with the validation approach they used. Soil attributes is very general and they should be specific about which attributes they could actual model.

Figure 3: Spelling mistakes

Figure 4 - 8: Define acronyms in caption so figure is stand alone. Write out x axis label to make it easier to read.

Figure 9: Spelling mistakes

The manuscript aims to provide a novel methodological framework for combining terrain attributes and data from geophysical sensors with machine learning algorithms in order to understand the pedosphere system and model soil attributes. An analysis of the importance of pedoenvironmental variables in predictive modelling is also presented.

Although the study has scientific significance to it, it is undermined by the writing issues present in the manuscript. The manuscript contains numerous grammatical and spelling errors (Eg - misspelt hyperparameters in 244 etc.). Furthermore, some ideas are not well fleshed out which hinders the understanding. For example, the discussion about the null model could be further elaborated while discussing the results.

Additionally, I agree with the comments that the LOOCV is not a rigorous method of validating the true performance of the modelling since it lacks a true test set to evaluate the model against. It is recommended that the authors use a separate test set for validation purposes.

Some of the specific issues are highlighted below:

• Grammatical errors and spelling mistakes need to be fixed.
• The methodological flowchart - In Pearson's test 95% threshold branching, at least one branch must have 'yes'.
• The mathematical notations in the paper need to be more consistent. For Eg, in eq 4, change RMSE_NULL to NULL_RMSE.
• Many of the symbols used have not been clearly explained earlier. For example, it is not specified what the values O_m and O_i mean in Eq 4 (I believe these must be Q_m and Qobs_i, respectively). Similarly, in eq 5, it is not clear what Q_train means.

Overall, the presentation of the manuscript needs to be improved significantly in further revisions, along with a diligent model validation approach.