Optimizing Herbicide Use in Fodder Crops with Low-Cost Remote Sensing and Variable Rate Technology

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Title: Optimising Herbicide Use in Fodder Crops with Low-Cost Remote Sensing and Variable Rate Technology

Authors: Luís Alcino Conceição 1,2, Luís Silva 1,3,*, Susana Dias 1, Benvindo Maçãs 4, Adélia Sousa 5,
Costanza Fiorentino 6 Paola D’Antonio 6, Sofia Barbosa 3,7 and Salvatore Faugno 8

Submitted to MDPI Applied Sciences.

---------------------------------------

Summary:

If I understand the article correctly, the main objective of the reported research is
to design a method to map weeds with digital images that decreases the cost to the farmer
and reduces the amount of herbicides released in the environment. This is my reading
of lines 103--107.

In PLOT I where the herbicide was not applied the % of weeds was greater than
in PLOTs II and III where the herbicide was applied. PLOTs II and III also showed
higher crop production. While the weed population's reduction in PLOT II was
observed, the crop production in PLOT II (Variable Rate of Herbicide Application)
was lower than in PLOT III (Fixed Rate of Herbicide Application). This may be a subject
for a future investigation.

A fixed rate application is not as costly as a smartphone application, i.e., no images
are collected and stored in the cloud or locally and no energy is used to analyze the
data on a computer and maintain the required opensource software in the cloud.

Another potential problem for the proposed VRT method is that the reported weed recognition
is lower than the RF and KNN. As I suggest in my last specific comment below there may be
a risk of not applying a herbicide due to a failure of not recognizing weeds on a given
plot. If this is correct, you may want to address this problem.

I was not clear on how the images were collected? Was that collection done manually by
the team members?

The materials and methods sections is adequate. I was able to follow the data analysis
methods. The results are tied to the materials and methods. In general, the article is
well structured.

Overall, I enjoyed reading and reviewing your article. Please, I apologize if I
misunderstood some of the points that I address in my specific comments below.

Specific Comments:

1) Figure 1 is helpful. Is there any reason why the number of sample points in the
three plots is different? If my counts are correct, Plot I has 6; Plot II -- 8; and
Plot III -- 7. Is there a logistical reason for this?

2) Line 154--157: First soil was distinguished from the vegetation by the application of
a filter of 0 or 1 for ExG Index values below and above 50, respectively; secondly weeds
and crop were distinguished from values below and above 120, respectively (Figure 3).

How camera-dependent is your approach (i.e., the filter in the first step and
the separation of the crop and the weeds in the second step)? Do you think that if
you had used a different camera (different than the 12 MP Xiaomi Redmi 8), these
parameters would have been different? This would add to the cost of your approach,
because changing parameters for each smartphone is software engineering labor.

3) Line 54: ... technologies capable of reduce the amount of herbicide used with the
environmental and

Change "capable of reduce" to "capable of reducing." "Of" after "capable" is followed
by a gerund.

4) Lines 62--64: Even though to a post-emergence control the discrimination of green
plant material and background inter-species, separation is needed mostly based on
artificial intelligence methods [8,9].

I am not sure I understand this sentence. Are you saying that that the separation
between green plant material and background inter-species is needed for AI to
be effective? Please clarify.

5) What is the purpose of Table 1 at the bottom of p. 2 and the following equations (i.e.,
Equations 1 -- 6) on p. 3? Do you want to show how the Kappa Coefficient works, because you
used it in your data analysis or do you want to suggest that your image-based VRT
method streamlines it? Or both?

6) Is N=150 in Table 3 the total number of images you have taken?

7) Lines 195-198: Considering the multiplication of the total area of the Plots II
and III by the applied amounts of herbicide, an average reduction of 0.22 l ha-1 was
achieved with the VRA in the Plot II. These results mean a 30 % reduction in the
herbicide used in the fixed dose parcel.

Good! The reduction of herbicide is 30%. This is significant.

8) Lines 253--258: Comparable to other studies using different classification methods,
our value of kappa of 78% is similar to those with 77% in Object-based algorithms [26]
or 73% in the Random Forest (RF) model [27]. Regard to the accuracy of weeds identification
our method reaches 45%, less than Islam et al., [28] that reached values of 96% and 94%,
for the RF and Support Vector Machine models, and 63% for the K–Nearest Neighbours model.

I appreciate your honesty on reporting lower weed identification accuracy. I also appreciate
the fact that your method is not as sophisticated as the RF, SVM, and KNN. Your reported
weed recognition appears to be significantly lower than that of the RF, SVM, and KNN. Is
that a potential problem for your method? Is there a risk of not applying a herbicide due
to a failure of not recognizing weeds? Will this make it less attractive than the Fixed Rate
method which is easier to administer?

Comments on the Quality of English Language

It's good. Some minor spelling mistakes and grammar revisions are required.

Author Response

Response to Reviewer 1

 

Summary
If I understand the article correctly, the main objective of the reported research is
to design a method to map weeds with digital images that decreases the cost to the farmer and reduces the amount of herbicides released in the environment. This is my reading of lines 103-107.

Response: You are correct in understanding that a key objective of our research is to propose a methodology for mapping weeds using digital imagery that aims to be cost-effective and efficient for farmers, while also reducing the environmental impact of herbicides.

However, we would like to clarify that an additional aim of the study was to assess the efficiency of a variable rate herbicide application in a fodder crop, which is also part of our evaluation. To better align the manuscript with your reading and to avoid any potential misinterpretations, we propose revising lines 103–107 for greater clarity as follows:

"These methodologies of weed control may benefit farmers by increasing grower profits and reducing the amount of herbicide released into the environment. Therefore, the objective of this trial was twofold: (1) to propose a methodology for mapping weeds through digital imagery in a sufficiently rapid, low-cost, and autonomous manner to the farmer, and (2) to evaluate the efficiency of variable rate herbicide application in a fodder crop."

We hope this revision clarifies the dual objectives of the research and aligns with your understanding.

 

In PLOT I where the herbicide was not applied the % of weeds was greater than
in PLOTs II and III where the herbicide was applied. PLOTs II and III also showed
higher crop production. While the weed population's reduction in PLOT II was
observed, the crop production in PLOT II (Variable Rate of Herbicide Application)
was lower than in PLOT III (Fixed Rate of Herbicide Application). This may be a subject
for a future investigation.

Response: Thank you for your detailed insights and constructive suggestions. We completely agree that the difference in crop production observed between PLOT II (Variable Rate of Herbicide Application) and PLOT III (Fixed Rate of Herbicide Application) is an intriguing result. But it should be noted that the production or productivity of a plot is not only related to the presence or absence of weeds. Many other factors may be leading to these results. We completely agree that this is a valuable observation, and we will highlight this as an area for future research in the revised manuscript.

 

A fixed rate application is not as costly as a smartphone application, i.e., no images
are collected and stored in the cloud or locally and no energy is used to analyze the
data on a computer and maintain the required opensource software in the cloud.

Response: We appreciate your thoughtful analysis. You raise an important point regarding the cost implications of the Variable Rate Technology (VRT) compared to Fixed Rate Application. While VRT may incur higher initial costs due to the need for digital imagery, data storage, and analysis infrastructure, it is designed to optimize herbicide use over time, potentially leading to long-term cost savings and huge environmental benefits. We acknowledge that these trade-offs were not explicitly addressed in the manuscript, and we will include a discussion of this aspect in the “4.3. Practical insights and future trends” of revised version to provide a more balanced perspective.

 

Another potential problem for the proposed VRT method is that the reported weed recognition is lower than the RF and KNN. As I suggest in my last specific comment below there may be a risk of not applying a herbicide due to a failure of not recognizing weeds on a given plot. If this is correct, you may want to address this problem.

Response: We appreciate your thoughtful analysis and constructive suggestions, which will help us improve the clarity and impact of our manuscript. Regarding the concern about weed recognition accuracy and the potential risk of herbicide application failures, we acknowledge that this is an important limitation of the proposed methodology. While the reported recognition accuracy of the proposed method is lower than RF and KNN, the methodology prioritizes rapid and low-cost processing, which could be advantageous for real-time field applications. However, we agree that improving recognition accuracy is crucial to minimize the risk of untreated weed patches. We will address this limitation explicitly in the revised manuscript and discuss the potential implications of recognition errors. Additionally, we will include a suggestion that future studies should explore strategies to enhance weed recognition accuracy while maintaining cost-effectiveness.

 

I was not clear on how the images were collected? Was that collection done manually by
the team members?

Response: Sorry if this evidence is not clearer. To improve the manuscript, the term "manually" has been introduced in the first paragraph of subchapter “2.2. Image segmentation and classification”. It now reads: "...obtained manually at 0.80 m height from the ground by an RGB sensor...".

 

The materials and methods sections is adequate. I was able to follow the data analysis
methods. The results are tied to the materials and methods. In general, the article is
well structured.

Overall, I enjoyed reading and reviewing your article. Please, I apologize if I
misunderstood some of the points that I address in my specific comments below.

Response: Thank you for your positive feedback and for taking the time to carefully review our manuscript. We are pleased to hear that you found the materials and methods section adequate, the results well tied to the methods, and the overall structure clear.

We also appreciate your thoughtful engagement with the article and your specific comments, which have helped us identify areas for clarification and improvement. We understand that some points may have been open to interpretation, and we hope that our responses to your specific comments address any misunderstandings or ambiguities in the manuscript.

Thank you again for your kind words and constructive feedback. Your suggestions will help us enhance the clarity and rigor of the manuscript.

 

Specific Comments:

Point 1: Figure 1 is helpful. Is there any reason why the number of sample points in the
three plots is different? If my counts are correct, Plot I has 6; Plot II -- 8; and
Plot III -- 7. Is there a logistical reason for this?

Response 1: You're absolutely right about the plots having different numbers of sampling points. This is because the points were created in GIS software according to a regular grid of one point per hectare, but the plots have an irregular geometry. As a result, we didn't get the same number of points on all the plots.

 

Point 2: Line 154-157: First soil was distinguished from the vegetation by the application of a filter of 0 or 1 for ExG Index values below and above 50, respectively; secondly weeds
and crop were distinguished from values below and above 120, respectively (Figure 3).

How camera-dependent is your approach (i.e., the filter in the first step and
the separation of the crop and the weeds in the second step)? Do you think that if
you had used a different camera (different than the 12 MP Xiaomi Redmi 8), these
parameters would have been different? This would add to the cost of your approach,
because changing parameters for each smartphone is software engineering labor.

Response 2: Thank you for raising this important question regarding the camera dependency of our approach. We agree that the parameters used in the filtering and separation steps (e.g., thresholds for the ExG Index) could be influenced by the specific camera used for image acquisition, including factors such as sensor type, resolution, and color calibration.

The specific threshold values we used (e.g., 50 and 120 for the ExG Index) were calibrated for the Xiaomi Redmi 8 camera used in this study. If a different camera were used, these parameters might indeed require recalibration. To minimize the need for recalibration, future work could investigate the use of standardized color calibration techniques or normalization algorithms to reduce variability between cameras. By standardizing the input images regardless of the camera used, the same parameters could potentially be applied across devices, reducing the need for extensive software adjustments. We will include this suggestion in the discussion section as an avenue for further research. We agree that the need for recalibration or reengineering for each type of smartphone adds to the overall cost and may pose a challenge for practical implementation.

Thank you again for this insightful comment, which has helped us identify an important aspect that requires further elaboration.

 

Point 3: Line 54: ... technologies capable of reduce the amount of herbicide used with the
environmental and Change "capable of reduce" to "capable of reducing." "Of" after "capable" is followed by a gerund.

Response 3: Thank you very much for your correction! The changes are in the revised manuscript.

 

Point 4: Lines 62--64: Even though to a post-emergence control the discrimination of green plant material and background inter-species, separation is needed mostly based on artificial intelligence methods [8,9]. I am not sure I understand this sentence. Are you saying that that the separation between green plant material and background inter-species is needed for AI to be effective? Please clarify.

Response 4: Thank you for pointing out the lack of clarity in the original sentence. To address your concern, we have revised the sentence to improve its readability and ensure that the intended meaning is conveyed clearly. The revised sentence now reads:

"Post-emergence control often benefits from artificial intelligence methods, as they can effectively distinguish between the green plant material of different species and the background."

We believe this revision resolves the ambiguity and better communicates the role of artificial intelligence in distinguishing plant species and background in post-emergence control. We appreciate your feedback, which helped us improve this section of the manuscript.

 

Point 5: What is the purpose of Table 1 at the bottom of p. 2 and the following equations (i.e., Equations 1 -- 6) on p. 3? Do you want to show how the Kappa Coefficient works, because you used it in your data analysis or do you want to suggest that your image-based VRT method streamlines it? Or both?

Response 5: Thank you for your question regarding the purpose of Table 1 and Equations 1–6. The primary purpose of including these elements is to show how the Kappa Coefficient works, as it was used as a method for validating the image classification in our study. We wanted to provide sufficient detail to ensure transparency and allow readers to understand the basis of our validation process. Thank you for drawing our attention to this.

 

Point 6: Is N=150 in Table 3 the total number of images you have taken?

Response 6: Thank you for your question regarding the value of N=150 in Table 3. To clarify, N=150 refers to the number of points randomly generated within the images to validate the image classification method. These points were used as a reference to assess the accuracy of the proposed classification approach. We included this clarification in the revised manuscript to ensure that this detail is clear to readers. Thank you again for bringing this to our attention.

 

Point 7: Lines 195-198: Considering the multiplication of the total area of the Plots II
and III by the applied amounts of herbicide, an average reduction of 0.22 l ha-1 was
achieved with the VRA in the Plot II. These results mean a 30 % reduction in the
herbicide used in the fixed dose parcel.

Good! The reduction of herbicide is 30%. This is significant.

Response 7: Thank you for your positive feedback on this result. We are pleased that you found the 30% reduction in herbicide usage with the VRA to be significant. This finding highlights the potential of the proposed methodology to contribute to more sustainable agricultural practices by reducing chemical inputs while maintaining effective weed control. Thank you again for your encouraging comment!

 

Point 8: Lines 253--258: Comparable to other studies using different classification methods, our value of kappa of 78% is similar to those with 77% in Object-based algorithms [26] or 73% in the Random Forest (RF) model [27]. Regard to the accuracy of weeds identification our method reaches 45%, less than Islam et al., [28] that reached values of 96% and 94%, for the RF and Support Vector Machine models, and 63% for the K–Nearest Neighbours model.

I appreciate your honesty on reporting lower weed identification accuracy. I also appreciate the fact that your method is not as sophisticated as the RF, SVM, and KNN. Your reported weed recognition appears to be significantly lower than that of the RF, SVM, and KNN. Is that a potential problem for your method? Is there a risk of not applying a herbicide due to a failure of not recognizing weeds? Will this make it less attractive than the Fixed Rate method which is easier to administer?

Response 8: Thank you for your thoughtful comment and for acknowledging our transparency in reporting the lower weed identification accuracy of our method compared to more sophisticated algorithms like RF, SVM, and KNN. You raise an important concern regarding the potential implications of lower weed recognition accuracy.

We agree that lower weed identification accuracy could pose a risk of failing to apply herbicide in certain areas where weeds are not properly recognized. This is a limitation of the proposed method, as incomplete weed control may reduce its effectiveness compared to Fixed Rate methods. But it isn’t the bigger problem of our method, because while Fixed Rate methods are easier to administer and eliminate the risk of missing weeds, they may lead to overapplication of herbicides, which can have economic and environmental consequences. In contrast, our method, despite its lower recognition accuracy, seeks to optimize herbicide application by targeting identified weed patches and reducing chemical inputs. This trade-off between simplicity and sustainability is an important consideration.

Although, we recognize the need to improve the accuracy of our method to make it more attractive and competitive with existing VRA approaches that use other methods like RF, SVM, and KNN. Future research could focus on refining the algorithm, incorporating additional image preprocessing steps, or leveraging hybrid approaches that balance simplicity, cost, and performance.

We really appreciate your insightful questions and suggestions, which have helped us better articulate the limitations and potential applications of our method.

 

Reviewer 2 Report

Comments and Suggestions for Authors

The paper addresses an important topic regarding optimising herbicide use in fodder crops with low-cost remote sensing and variable rate technology. However, I consider that the following improvements are necessary:

- the abstract is not clear and precise enough, it should be a miniature of the paper, so that the important points of the paper can be understood: purpose, methodology, variants, determinations, results/conclusions. I consider that the abstract should be rewritten.

- the table and equations in the Introduction, lines 84-94, do not have adequate references, are these original?

- lines 134-136, say: The collected samples – one sample made up of three sub-samples were weighed to determine the FM production per hectare, and subsamples in small paper bags placed in 65 °C until constant weight to determine forage moisture content and DM content - are you not using standard methods? is it your own method? why are you not using references?

- the figures must be corrected, the unit of measurement used is l/ha, different from the way of expression in the text.

- in the figures, the comma must be replaced with a period in the decimal separator.

- line 212 and line 224 - quotation marks are not necessary.

- the discussions contain too few references and many own statements.

- the conclusions must be developed, it must be specified whether the research hypothesis is confirmed and if there are advantages for farmers and CAP, also if there is a perspective for new research.

- the accuracy and precision of the paper must be improved.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Response to Reviewer 2

 

Comments and Suggestions for Authors

The paper addresses an important topic regarding optimising herbicide use in fodder crops with low-cost remote sensing and variable rate technology. However, I consider that the following improvements are necessary:

Point 1: the abstract is not clear and precise enough, it should be a miniature of the paper, so that the important points of the paper can be understood: purpose, methodology, variants, determinations, results/conclusions. I consider that the abstract should be rewritten.

Response 1: We apologise if the abstract didn't seem clear enough, but after a few changes we believe it fits your suggestions:

1st and 2nd sentence - purpose

3rd and 4th sentence - methodology, variants and determinations

5th and 6th sentence - results and conclusions

7th sentence - future trends.

Thank you for your thoughtful suggestion to improve this important section of the manuscript.

 

Point 2: the table and equations in the Introduction, lines 84-94, do not have adequate references, are these original?

Response 2: Thank you very much for your attention to this particular case. The reference for this information is given at the beginning of this information, and is reference [17] (Cohen, J. A Coefficient of Agreement for Nominal Scales. Educ Psychol Meas 1960, 20, 37-46, doi:10.1177/001316446002000104.). To avoid this feeling of lack of reference on the part of readers, the citation of the authors has been reinforced.

Point 3: lines 134-136, say: The collected samples – one sample made up of three sub-samples were weighed to determine the FM production per hectare, and subsamples in small paper bags placed in 65 °C until constant weight to determine forage moisture content and DM content - are you not using standard methods? is it your own method? why are you not using references?

Response 3: Sorry for our fault. We used a standardised method, the Kjeldahl method. The reference was introduced in the revised manuscript. Thank you very much for your attention!

 

Point 3: the figures must be corrected, the unit of measurement used is l/ha, different from the way of expression in the text.

Response 3: Thank you very much for your attention! The figures have been improved to standardise the document. l/ha has been replaced by l ha-1.

 

Point 4: in the figures, the comma must be replaced with a period in the decimal separator.

Response 4: Thank you very much for your suggestion. Changes have been included in the figures.

 

Point 5: line 212 and line 224 - quotation marks are not necessary.

Response 5: Thank you very much for your suggested change. Your suggestion has been followed, and the changes are included in the revised manuscript.

 

Point 6: the discussions contain too few references and many own statements.

Response 6: Thank you for your valuable feedback regarding this important section. Due to his suggestions and those of the second reviewer, the discussion was changed, improved and the references cited were increased.

 

Point 7: the conclusions must be developed, it must be specified whether the research hypothesis is confirmed and if there are advantages for farmers and CAP, also if there is a perspective for new research.

Response 7: Thank you for your valuable feedback regarding the conclusions section. We agree that the conclusions should be expanded to clearly address whether the research hypothesis was confirmed, the potential advantages for farmers and the CAP (Common Agricultural Policy), and the perspectives for future research. We hope this revised conclusion addresses your concerns and provides a more comprehensive summary of the study’s findings and implications. Thank you for your thoughtful suggestion to improve this important section of the manuscript.

 

Point 8: the accuracy and precision of the paper must be improved.

Response 8: Thank you for your feedback regarding the need to improve the accuracy and precision of the manuscript. We understand the importance of ensuring that our work is presented with clarity and scientific rigor. In response, we reviewed the presentation of the results, ensuring all numerical values, statistical analyses, and comparisons are accurate, clearly defined, and transparently reported. To enhance the accuracy and precision of the manuscript, we incorporated the feedback from your, and the second reviewer, specific comments. We are committed to improving the manuscript in line with your suggestions and ensuring it meets the highest standards of scientific accuracy and precision. Thank you for highlighting this important aspect.