Using Geospatial Information to Map Yield Gain from the Use of Azospirillum brasilense in Furrow

Round 1

Reviewer 1 Report

Line 55 - The statement the existing research on yield gains “hardly presents a statistical difference” needs to be expanded. Is the lack of statistical difference due to the methods used in the research or is this referring to the absolute size of yields not being significant? Please add cites for this as well.

Line 68 - A multilayer perceptron is typically abbreviated as MLP, and machine learning is abbreviated as ML. Please clarify and make sure these abbreviations are consistent / accurate throughout the paper.

Line 79 - A number of different yield estimation approaches are covered in the introduction up to this point, yet it is not made clear why Kriging or spectral models are more suitable for this application or predicting yields in general over another approach such as neural networks or random forests.

Line 82, 86 - It still seems a bit unclear if you are actually trying to just generate two separate models to map untreated yield and treated yield (as sort of an ex-post theoretical yield gain) vs a single model that can be used to predict future potential yield gain.

Line 104 - Needs to be more clear that “image-based” is referring to the spectral model approach.

Line 136 - The use of the word “treatments” here is used to refer to both the varying approaches including “treating” with A. brasilense (which is then referred to as “treated) and “treating” without A. brasilense (“control). This is very confusing, and I would suggest not referring to both of these as treatments.

Line 179 - This is a very generalized citation and needs more context. As it is, it makes it sound like geostatistics are magic that can just make accurate predictions anywhere.

Line 192 - Need a sentence or so to explain the role of the semivariogram and covariances are in Kriging for those not familiar with the approach.

Line 208 - These preliminary tests seem significant to the work. Can you provide some more details on the testing used to justify the phase selected?

Line 222 - I suggest detailing the point selection process for the “lower volume” training data used for the spectral model. Is this necessary due to the resolution of the imagery or are limited training points being artificially enforced to replicate some real world scenario?

Line 225 - “PLANET satellite” should be “PlanetScope imagery”

Line 232 - Suggested change: “...the prediction models for Br1 and Br2 were separate and not used to estimate yield outside from which the training samples were extracted”

Line 334 - The variation between treated and untreated due to biotic and abiotic factors seems a significant aspect to modeling yield gain, and was not included in the introduction or methods (I am assuming here that this variability was existing knowledge). I would suggest at the very least incorporating more context around this variation in the discussion, and potentially adding this into the introduction and possibly methods (if it impacted decision making regarding modeling approach; if it didn’t impact the yield gain modeling approach, I’d mention why in the discussion).

I would suggest adding more about the applicability and suitability of these methods in other scenarios (i.e., real world applications for yield gain model) to at least the discussion. Would models for Br1 and Br2 work elsewhere? A clear opportunity for this research was to test the suitability of the Br1 model in Br2, and vice versa. As the approach is, my general feeling is that these models are going to be overfit, given that they use data from a single field with heavily interspersed training and validation data. That said, this approach may still be reflective of real world applications. For example, if it is common to be able to sample yield every X meters, but in your data collection with above average equipment you sampled every X/10 meters. This ties into the broader discussion around how these methods are to be applied in real world scenarios, which is currently lacking.

Also related to application of methods: What if the models are impacted by conditions that change each year? If so, wouldn’t it be impossible to calibrate and use a model in one year before it’s useless the next year (i.e., models only work on the data they are trained on - overfitting essentially). Or is the purpose only to estimate yield gain after crops are collected to justify the use of A. brasilense? I don’t think solutions to all of these potential issues, if they apply, are necessary, but they should at least be presented as things which need to be considered before your approach could be implemented more broadly; and/or the intended use case be laid out more clearly.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The study presents an approach to monitor biomass gain by remote sensing and geostatistics. This is a timely and relevant topic. Nevertheless, I see some major flaws in the manuscript which need to be solved to make the manuscript suitable for publication.

1)     I have some doubts if it is useful to combine a comparison of geospatial observation methods with the object under observation (i.e. yield change by the application of Azospirillum). Therefore, I suggest to focus clearly either on the comparison of the geospatial methods or the yield gain assessment. From the methods section, it seems as the comparison of geospatial methods is the focus.

     2) Closely related to 1: the exact aim of the study remains unclear for me from the introduction. Is the aim i) to assess the different methods for biomass gain (kriging vs. spectral modeling), ii) assess yield gain from the treatment with Azospirillum or iii) a combination of both? I suggest to re-write the two final paragraphs of the introduction (lines 79-89) and make the statement more clear for the reader. As mentioned in 1, I have the impression that the focus is on the geospatial methods. This should also become clear from the aims.

3)     I suggest to re-structure the methods section to make it more clear for the reader what has been done for the comparison of the geospatial assessment methods.

4)  In the discussion, I miss the point why remote sensing is necessary at all when dense point data is available.

Specific comments:

Title: I think, that the title is slightly misleading as the focus is on a comparison of the methods for creating spatially continuous datasets and not so much on the effect of the specific treatment. See also comment 1.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper reports how effectively Kriging and machine learning-based spectral models can estimate crop's yield where Azospirillum brasilense was applied.

The manuscript is clear as to what to study and thus, it reads well. Nevertheless the authors need to enunciate 1) what is the innovation of this work and 2) what kind of broader (theoretical or practical) conclusion the authors can draw from the work.

- Abstract: provide some conclusion at the end.

- lines 79-84: too long a sentence. Moreover, one sentence constitutes a paragraph, which is not recommended. Consider dividing into multiple sentences.

- line 89: Clear state what this work contributes to the literature. Highlight the difference of this work from previous ones. In addition, provide explanation on why you chose such method?

- lines 368-385: Consider removing hyphens. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf