Weed Detection: Innovative Hyperspectral Image Analysis for Classification and Band Selection of Site-Specific and Selective Weeding Robot
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsWeed detection is a crucial prerequisite for precise management of farmland. This research employs hyperspectral imagery for weed detection in melon fields, possessing overall research value. However, it also contains some apparent flaws, detailed as follows:
1.The article primarily focuses on the optimization of algorithms, all of which are conventional. Therefore, the innovation point of the article is not clear and should be refocused.
2.The article solely covers the aspect of weed detection. However, the first paragraph of the introduction describes weed control methods, and the second paragraph discusses weed control actuators, both of which deviate entirely from the main topic. It is recommended to rewrite these sections.
3.The current mainstream detection method is the RGB method, which generally achieves good detection accuracy. What are the reasons for adopting the hyperspectral method in this article? A single sentence in the third paragraph of the introduction is far from sufficient to explain this.
4.The fourth paragraph and subsequent content of the introduction do not summarize the existing issues in current weed identification. Among the references cited by the authors, techniques such as hyperspectral imaging have achieved excellent detection accuracy, with weed identification rates as high as 98%-100%. What, then, is the significance of this research paper? In other words, what are the challenges in weed detection in melon fields? To effectively introduce the authors' research content, a new literature review and summary should be conducted.
5.In the experimental design, were the weeds artificially planted or naturally grown? What is the basis for selecting the types of weeds? It is recommended to supplement this information.
6.In the first paragraph of Section 2.6, what is the basis for selecting four spectral bands for screening?
7.Algorithms such as Random Forest are conventional and do not necessitate detailed description or definition.
8.In the results section, specific evaluation indicators should be provided. The statement "evaluated by experts" is meaningless without objective indicators such as weed identification rates or detection times.
9.In the opening sentence of the third paragraph of the conclusion, are the spectral characteristics presented in weed identification related to the selected weed species? Different weeds may exhibit different spectral characteristics, which could impact detection accuracy. The article does not elaborate or investigate this aspect, rendering the results not universally applicable.
10.In the final paragraph of the conclusion, since the article does not include statistics on model response speed, power consumption, or other relevant metrics, how can it be determined that computational resources are minimal and costs are low?
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper summarizes the results of an attempt to distinguish between crops and weeds in watermelon fields using hyperspectral imaging, reporting that four spectral bands (480 nm, 550 nm, 686 nm, and 750 nm) are effective for this purpose. However, object detection techniques based on deep learning, which is widely used for such purposes, has not been implemented, and the usefulness of this study has not been fully evaluated. Additionally, content that should be described in "2. Materials and Methods" is included in "3. Results and Discussion," and the citation methods are incorrect, making this paper unsuitable for publication in this journal.
1. Introduction
There are some mentions of cases utilizing RGB cameras, which seems to relate to research on object detection using deep learning, but this should be clarified.
While the accuracies and processing speeds in previous studies are noted, the specific targets segmented and the metrics used for accuracy should be clarified.
2. Materials and Methods
Although labeling was performed on 572 frames, the number of pixels for each of crops, weeds, ground, and reference is not specified.
In Figure 3, rectangles are shown in purple or orange, but the reason for the color differentiation is unclear.
There is no description of the processing environment or the method for setting hyperparameters when using classification models such as machine learnings.
3. Results and Discussion
The discussion is very limited. A comparison with previous studies is essential. Additionally, legends are not displayed in the figures.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis study is highly valuable for advancing precision agriculture by addressing the critical challenge of weed detection in sensitive crops like watermelon. Using hyperspectral imaging and machine learning techniques, it identifies key spectral bands that effectively distinguish weeds from crops. The proposed methodology not only reduces the computational complexity of hyperspectral data but also provides a practical foundation for developing cost-efficient, multispectral camera-based systems. These systems can enable precise, non-chemical weeding, minimizing crop damage and labor costs while enhancing yield quality. Furthermore, the integration of this approach into robotic weeding platforms has the potential to revolutionize sustainable weed management in commercial agriculture. This study also lays the groundwork for applying similar techniques to other crops, fostering broader adoption of advanced technologies in smart farming. Although the study is very important, the manuscript still has some issues which need to be addressed before further consideration. Please find my comments below.
Here are 10-15 suggestions to improve the reviewed paper, focusing on clarity, methodology, and overall impact:
- Simplify the abstract for better readability, particularly for non-specialist audiences. Include numerical results concisely and remove excessive methodological details.
- Add more global context about the significance of hyperspectral imaging in intro particularly in agriculture to highlight its importance beyond watermelon and melon cultivation.
- Explain why the "Normalized Crop Sample Index" (NCSI) is a novel contribution. Compare it explicitly with existing indices like NDVI to emphasize its advantages.
- Please also expand on the impact of environmental conditions (e.g., lighting, soil variability) during data collection on model performance and misclassification error.
- Justify and provide more detailed reasoning for excluding bands beyond 850 nm due to a low signal-to-noise ratio. Explore whether this limitation could be addressed with improved hardware or preprocessing.
- Discuss and reason why Random Forest and XGBoost outperformed other methods under varying conditions and how this insight could guide future implementations.
- In the discussion section there should be some comparisons with similar studies focusing on different crops or environmental conditions to establish the broader applicability of the findings.
- One important aspect of this study should be the focus on how future work could be improved. Please clearly state the limitations of the study, such as small sample size or potential overfitting, and propose specific directions for future research.
Please address these suggestions and I wish you all the best for the revisions.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe author has made some adjustments in response to previous reviewer comments; however, many sections still lack sufficient revision:
1.As the author noted, a brief introduction to the weeding actuator is necessary. Since spring tines also belong to the mechanical system, these two elements can be merged for clarity.
2.In the second paragraph of the introduction, the phrase 'The sensing… the images' lacks clear prior explanation for why 'deep learning models are thus suitable for image analysis,' leaving readers confused.
3.When addressing reviewer question 10, the author claims that low computational resources and costs stem from reducing spectral bands to four. However, in response to question 6, the author states that four spectral bands are a common choice. This creates a logical inconsistency. Arbitrarily reducing the number of spectral bands cannot serve as an innovation point (as mentioned in the abstract) or a primary contribution.
4.The revision suggestion from point 8 in the previous review—requiring the inclusion of recognition rates or detection times—must be implemented. Authors should modify this content rather than adding it as a limitation in the results section.
5.The paper must include a separate discussion section instead of merging it with the results. Authors should reorganize and add this section accordingly.
Author Response
Comment 1: As the author noted, a brief introduction to the weeding actuator is necessary. Since spring tines also belong to the mechanical system, these two elements can be merged for clarity.
Response 1: We agree with the reviewer; we merged the two elements to mechanical systems in the second paragraph of the introduction chapter.
Comment 2: In the second paragraph of the introduction, the phrase 'The sensing… the images' lacks clear prior explanation for why 'deep learning models are thus suitable for image analysis,' leaving readers confused.
Response 2: The reviewer is right; the sentence did not appear as is in the original version of the manuscript and was probably a combination from 2 half sentences during the revision. Since it is not accurate and does not contribute to the paper, we removed it.
Comment 3: When addressing reviewer question 10, the author claims that low computational resources and costs stem from reducing spectral bands to four. However, in response to question 6, the author states that four spectral bands are a common choice. This creates a logical inconsistency. Arbitrarily reducing the number of spectral bands cannot serve as an innovation point (as mentioned in the abstract) or a primary contribution.
Response 3: The statement about the reduction of computational resources was related to multispectral analysis (4 bands) in comparison to the computational resources required to perform hyperspectral analysis (840 bands) and multispectral analysis (4 bands) and not between different 4 bands analysis. It is also interacted with the computational power incentive of a multispectral cameras regarded in the revised manuscript in the introduction chapter in paragraphs 3, 5 and 6 and the computational cost and time regarded in the 1st and 2nd paragraphs in section 2.6. We modified and clarified the computational resources in the 3rd paragraph from the end of the conclusions chapter in revision 2 of the manuscript (current) to increase clarity.
Comment 4: The revision suggestion from point 8 in the previous review—requiring the inclusion of recognition rates or detection times—must be implemented. Authors should modify this content rather than adding it as a limitation in the results section.
Response 4: we will try to address the reviewer comment and explain further the issue of weeding point accuracy and detection time. The investigation of the calculated weeding points revealed that all points were positioned on a weed pixel at a distance from watermelon pixels. Quantification of the accuracy of the weeding point position and its effectiveness on weed control will require a dedicated experiment which is not in the scope of the presented study. We added a clarification in the paragraph between figure 12 and figure 13. Regarding the detection times – once the algorithm is defined, after the training and testing stages, the execution time of the algorithm is few milliseconds per image and is not the limiting factor in the system. The detection times were not one of the goals of this research and were not investigated in this study.
Comment 5: The paper must include a separate discussion section instead of merging it with the results. Authors should reorganize and add this section accordingly.
Response 5: In order to improve the reading flow, the discussions in this paper was divided between a discussion on a specific result and therefore appears in the results chapter (each specific discussion point is next to the related result), and, general discussion issues that appear in the conclusions chapter. according to the reviewer comment we moved some of the discussions from the results chapter to the conclusions chapter and modified the chapter.
Reviewer 2 Report
Comments and Suggestions for AuthorsI understand the author’s argument.
I have no further comments.
Author Response
comment 1: I understand the author’s argument.
response 1: thank you
comment 2: I have no further comments.
response 2: we thank the reviwer
