Review Reports - UAV-Borne RGB Imagery and Machine Learning for Estimating Soil Properties and Crop Physiological Traits in Peanut (<i>Arachis hypogaea</i>): A Low-Cost Precision Agriculture Approach

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article, titled "Digital Transformation in Agriculture: Impact of Drones and Advanced Sensors on Monitoring and Sustainable Productivity in Peanut Cultivation (Arachis hypogaea)" presents an important approach and a significant contribution for the use of aerial imagery based on precision agriculture tools. Some suggestions are presented in the appendix to clarify important points of the research.

Comments for author File: Comments.pdf

Author Response

Please see the attachment. A detailed point-by-point response to the reviewer’s comments has been provided in the attached report.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The article focuses on monitoring selected soil and crop properties using unmanned aerial vehicles (UAVs) with RGB images. This method makes it possible to reduce the number of samples taken and increase the efficiency of agricultural machinery use. The crop grown is peanuts on a targeted experimental field in the province of Manabi, Ecuador. The experiment involved evaluating four peanut varieties, two different planting densities, and two different farming methods. The results showed significant correlations in some places with either soil or plant properties. In terms of the methods applied, the Random Forest model proved to be more advantageous. The application of RGB can therefore be an effective method for predicting key soil and crop parameters. The application of these methods can also reduce operating costs.

I consider the topic to be original because it brings new information to the field of study. The application of drones will, in the future, and to some extent already now, contribute to more effective monitoring of soil and crop properties. The contribution of the above-mentioned methodologies and their application can improve production efficiency and reduce the number of physical soil or plant samples taken. If further research proves the application of this method, it will clearly be beneficial.

Overall, the methodology also describes methodological procedures for determining soil properties, which are already described in the relevant known standard. Therefore, it is not necessary to describe it in detail, but only to refer to the relevant standard.

The authors used many cited sources, focusing on and pointing out, in contrast to other sources, the Random Forest application and its advantages over the nearest neighbor test. In terms of methodology, there is room for improvement in the article. It is not necessary to describe the basic analyses in detail. Otherwise, the procedures are described in detail. The authors obtained basic aerial data from individual flight missions, from which the resulting orthomaps were generated. The conclusions presented evaluate the results achieved and point out the advantages of the individual indices applied. The individual measured soil properties were correlated with the spectral indices obtained. Bulk density, field water capacity, and organic matter create a certain RGB reflectance, which the authors describe in detail. They therefore represent a contribution to practice and, through their statistical evaluation, to science. The advantages of RGB lie precisely in their application as a non-destructive alternative. The benefits of the publication are those defined by the authors in the article as the possibilities for the reliable use of RGB-based indices as fast and effective tools for estimating properties.

The image quality could be higher in some places. The authors applied up to 80 different sources, the number of which could be reduced. However, on the other hand, their number is related to the extensive discussion of each subchapter and introduction. Other comments are provided in the following description.

Supplementary statistical analyses do not need to be part of the main article.

Strengths:

- The article is quite extensive, including appendices

- Introduction – provides essential information on the topic

- Methodology – sufficiently detailed, although the determination of some physical parameters is overly detailed

- Results and discussion – detailed description of results linked to discussion

- Conclusion – evaluation of results achieved; rapid and non-destructive methods will clearly be applicable in the future

Weaknesses:

- Methodology – the determination of some physical parameters is too detailed, e.g. SBD, FC, PWP, SOM, etc. Only the method, or more precisely the standard applied, according to which the procedure was carried out (name), could have been stated, and it is not necessary to state the entire procedure

Other comments:

Methodology – In the submitted methodology, I cannot find the depth profile from which the samples were taken or in which the properties were determined, e.g., 0-20 cm, etc. – therefore, I suggest adding the required parameters. In addition to the above, it is also important to specify the number of repetitions of individual measurements.

Figure 1 – you define several fields, the question is how many monitoring points there were for sampling (Figure 1) – did each field represent its own area of interest?

145-149 – in the lines mentioned, you use parameter designations with different indices. Is this correct? I suggest modifying them: w, aqua

Fig. 4 – The figure is mentioned twice in the submitted article – please modify.

Author Response

Please see the attachment. A detailed point-by-point response to the reviewer 2 comments has been provided in the attached report.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear authors, thank you for your submission. The article addresses a current topic of great interest to precision agriculture.

Below are some comments regarding improvements needed to improve the article before it can be accepted:

1) Considering that this is not a review article, the title must be changed; in particular, ‘Digital Transformation in Agriculture’ must be removed; furthermore, the use of an RGB sensor does not allow for the inclusion of ‘advanced sensors’ in the title.

2) Line 18, Maybe you could put the corresponding authors' email here.

3) Introduction. In this section, the integration of remote sensing and machine learning must be explained better. Here are some references you can include: https://doi.org/10.3390/rs16244784 ; https://doi.org/10.3390/agriengineering8010009

4) Materials and Methods: Why did you choose vNDVI? What does it mean? How does it differ from standard NDVI?

5) Lines 301-309, The implementation of machine learning models needs to be explained much better. It is unclear, both in terms of hyperparameter optimisation and the use of cross-validation.

6) All tables need to be correctly formatted.

7) Table 4. What do these metrics refer to? Cross-validation? Testing? This needs to be clearly stated. Same for the following tables.

8) The quality of the figures is too low

9) line 512, please correct "Ed"

Author Response

Please see the attachment. A detailed point-by-point response to the reviewer 3 comments has been provided in the attached report.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors made the suggested corrections and improved the manuscript; therefore, the article is accepted.

Thank you.

Author Response

Thank you for your positive evaluation and for acknowledging the improvements made to the manuscript. We sincerely appreciate your constructive feedback throughout the review process, which has helped strengthen the quality and clarity of our work.

Reviewer 3 Report

Comments and Suggestions for Authors

I thank the authors for their revision and for addressing most of the points raised in the first round; nevertheless, some of the original points remain only partially addressed, and I believe that some additional methodological aspects should be clarified before the manuscript can be accepted.

1. Introduction – integration of remote sensing and machine learning.

The new paragraph added at lines 66–76 is appreciated, but in its current form, it remains rather generic and does not yet provide the reader with a substantive overview of how RS and ML are jointly used in operational precision agriculture. I would encourage the authors to expand this paragraph by discussing (i) the role of multi-temporal spectral features as predictors, (ii) how ensemble learners handle the non-linearity and collinearity typical of vegetation indices, and (iii) the issue of model transferability across sites and seasons. Several recent works combining UAV/satellite RS with ML within different cropping systems would strengthen this discussion — for instance, studies applying Random Forest and related ensemble methods to multi-temporal Sentinel-2 or UAV imagery for crop and soil monitoring (e.g., Giannico et al., 2024 -- https://doi.org/10.3390/rs16244784)

2. Machine learning implementation.

The revised paragraph (lines 323–334) is clearer regarding the 70/30 split and the use of 5-fold cross-validation for hyperparameter tuning, but several details essential for reproducibility are still missing:

-The actual hyperparameter grid tested for both RF (range of n_estimators, max_depth, min_samples_leaf) and KNN (range of k, distance metrics) should be explicitly reported, together with the optimal values selected.
-It should be stated clearly whether the 70/30 partition was performed at the level of experimental plots (n = 48) or at the level of pixels extracted from the orthomosaics. This distinction is necessary: a pixel-level split would introduce strong spatial autocorrelation between training and test sets and could substantially inflate the reported R² values. If the split was performed at the pixel level, I would strongly recommend repeating the evaluation with a plot-level (or spatially blocked) cross-validation scheme.

-Given the limited number of independent experimental units, a brief discussion of the risk of overfitting and of the effective sample size used for training would be appropriate.

3. Consistency between Spearman and Pearson analyses.

-Section 2.4 justifies the use of Spearman correlations on the basis of non-normality, yet Section 3.3.1 reports a "Pearson correlation analysis" between spectral indices and the crop variables CL and NW. According to Table C1, CL and NW passed the Shapiro–Wilk test, which would indeed justify Pearson, but this rationale is never made explicit in the main text. Please clarify this point so that the reader does not perceive an inconsistency between the methods section and the results. In this way, the reader will understand that the choice of statistical method was made on a variable-by-variable basis, depending on the distributional properties of each.

4. Out-of-range index values at the harvest stage.

-The authors note that vNDVI reaches values up to 1.70 at harvest (Section 3.2.1). Since vNDVI is, by construction, a normalised index that should remain bounded, values exceeding 1 suggest either an issue in the band scaling (e.g., 8-bit DN values used directly without normalisation to [0,1]) or in the implementation of the index formula in QGIS Raster Calculator. The authors should verify the calculation pipeline and clarify in the text how RGB bands were scaled before computing the indices, as this directly affects the interpretability of all subsequent correlation and ML analyses.

5. Figure 6.

-The caption describes a "Pearson correlation matrix", but the figure displays only the correlations of two variables (CL, NW) against the spectral indices, in the form of a bar chart rather than a matrix. Please update the caption accordingly.

6. Minor.

-Section 3.3 is titled "Spectral indices associated with soil properties", but actually deals with crop physiological variables — please check the heading.
Reference [16] in the revised manuscript has a typo (double opening bracket "[[16]").

-Several figures still appear at moderate resolution; please verify that the final exported versions meet the journal's DPI requirements.

-Check the references section.

Once these points have been addressed, I believe the manuscript will be substantially strengthened and may be accepted.