PSCDR-BMPNet: A Point-Supervised Contrastive Deep Regression Network for Point Cloud Biomass Prediction
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors"PSCDR-BMP Net: A Point-Supervised Contrastive Deep Regression Network for Point Cloud Biomass Prediction"
The topic of the paper -biomass prediction using point cloud data- is relevant, and aligned with the journal's scope. The authors propose a novel supervised contrastive learning loss (PSCDR) and a tailored neural network architecture (BMP Net) that outperforms existing benchmarks on the SGCBP dataset. The work demonstrates clear relevance to precision agriculture and offers methodological innovations that can be generalized. It addresses a contemporary issue with implications for agricultural development, and public policy in the field. The topic is of clear interest to researchers and experts.
The abstract is at optimal length and reflects the content.
The composition is coherent and aligned with the stated objectives of the study. Each section logically builds upon the previous one, facilitating understanding.
The article is scientifically sound. The methodological construction is based on a rigorous literature review and uses advanced concepts from deep learning, contrastive regression, and 3D data processing.
Methodology is well described, and the integration of Chamfer Distance into a supervised contrastive loss for regression tasks represents a novel approach. The mathematical definitions are accurate, and the algorithmic implementation is clearly described. The architecture design, contrastive loss mechanism, and ablation studies seem to be well illustrated and supported by equations and figures.
The use of the SGCBP public dataset, appropriate performance metrics (RMSE, MAE, MAPE), and detailed experiments—including ablation studies and hyperparameter sensitivity analysis—demonstrates good experimental discipline.
The conclusions faithfully reflect the results obtained and contextualize the contribution to the literature.
The writing is generally clear and accessible. Minor stylistic or grammatical edits may improve the flow; no major revisions are needed.
The paper is of high quality, but for a complete and balanced review, it is important to note its weaknesses/vulnerabilities:
- The experimental validation is limited to two cereal crops (wheat and barley) from a single geographical region (Australia) and only two growth stages. This may affect the generalizability of the model.
- BMP Net combined with PSCDR is resource-intensive, potentially limiting its deployment in real-time or edge computing environments. Difficult to implement in field applications (drone, agricultural robot, edge system), which limits immediate practical applicability.
- No tools for model explanation (e.g., feature attribution, silence maps) are discussed, which could hinder user trust and adoption in practical settings. The model is a "black box"; interpretation methods (e.g. feature importance, error analysis on data segments) are not discussed. End users (farmers, agricultural engineers) may not understand the model's decisions, which affects its trust and adoption in reality.
- The paper does not include simpler, classical regression baselines to highlight the advantages of the deep learning approach more clearly.
- The model may be sensitive to changes in hyperparameters; reproducibility in other contexts may be affected.
The authors themselves also mention the limitations of the study in the paper, which increases the credibility of the research.
The paper is suitable to be published after minor revisions.
Suggestions for revision:
- Extend testing to multiple datasets and different crop types.
- Explore lightweight versions of BMP Net for real-time applications.
- Integrate with other data sources (UAV imagery, solar/climate sensors).
- Use interpretability techniques.
- More rigorous justification of hyperparameters (e.g. through automated optimization or extensive cross-validation).
In my opinion, it would increase the value of the paper because it would increase the impact of the research and increase the interest in applying the research.
I wouldn't use acronyms in the title.
I leave it up to the authors and editors whether or not to make changes to the paper, I do not insist.
Congratulations to the authors!
Further success!
Author Response
Please see the attachment.
Author Response File: 
Author Response.doc
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper presents an interesting approach with PSCDR-BMPNet. The technical aspects seem well-addressed. The minor suggestions above are primarily focused on enhancing clarity, consistency, and adherence to common academic presentation standards, which should further strengthen the manuscript. I recommend this manuscript to be accepted after some minor revisions.
Page 1, Line 12: "Predicting ground biomass" is used instead of "above-ground biomass (AGB)." Replace with "predicting above-ground biomass (AGB)." Since AGB has been defined earlier, terminology should remain consistent throughout the paper.
Page 8, Lines 304–308: Hardware setup details are thorough, but software is missing. Please Include software environment information. Providing complete hardware and software details is crucial for reproducibility.
Page 12, Table 5: Best performance values in the table are not emphasized. Highlight optimal values in bold Comment: For reader comprehension, optimal values should be clearly marked.
Page 12, Lines 453-455: The importance of τ is mentioned, but no specific guidance is provided. Please explain.
Numerous grammatical and typographical errors detract from the manuscript's clarity and professional presentation. Besides, there are numerous editorial inconsistencies throughout the manuscript, such as fluctuating tense and uneven reference formatting, which need to be corrected to meet the journal’s publication standards.
Author Response
Please see the attachment.
Author Response File: Author Response.doc
Reviewer 3 Report
Comments and Suggestions for AuthorsDear authors,
The manuscript makes an important contribution to point cloud analysis and environmental monitoring with AI. There are some recommendations for Improvement
- Proofread the manuscript for grammar and style fixes.
- Figures showing the network architecture are helpful and clearly drawn. However, the captions could provide more detailed information. Expand figure captions so they can stand alone better.
- Include brief, real-world examples or applications in the discussion section.
- Clarify key mathematical formulations with easy-to-understand descriptions. some equations (e.g., cosine similarity in the loss function) could benefit from a simplified verbal explanation to improve accessibility for non-specialist readers.
- Think about adding a visual diagram that summarizes the full pipeline of PSCDR-BMPNet.
The English language is adequate but could be improved with professional proofreading. There are occasional grammatical errors and stylistic issues that reduce clarity in places, particularly in the abstract and conclusion.
Author Response
Please see the attachment.
Author Response File: Author Response.doc
Reviewer 4 Report
Comments and Suggestions for AuthorsThe analytical methods of agricultural crop evaluation are important for increasing the effectiveness of agriculture and forecasting. Improving the methods of agricultural crop monitoring is a topical task for the development of agriculture.
The approach to choosing the tool for increasing the accuracy of predicting biomass is worth mentioning. That is the implementation of rigorous methods, to include the deep contrastive regression considering the modal differences between point cloud and 2D image data.
The area of research is relevant to the Journal subject area; tables and figures are coherent and substantiated. Literature review complies with the research.
The study could further benefit, if the authors had evaluated the accurateness of findings.
The authors point out that «Experiments on the SGCBP public dataset show that when using BMP_Net alone, the RMSE, MAE, and MAPE are 84.48, 67.31, and 0.131, respectively, outperforming BioNet. After incorporating the PSCDR method, these three key performance metrics are optimized to RMSE = 75.92, MAE = 63.19, and MAPE = 0.115». Does this improvement sufficiently impact the prediction? Is the increase in accuracy sufficient? It would be reasonable to point out the agricultural crops and territory or region suitable for the presented method.
Author Response
Please see the attachment.
Author Response File: Author Response.doc
