Lavender Field Detection via Remote Sensing and Machine Learning for Optimal Hive Placement to Maximize Lavender Honey Production
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsArticle: Lavender Field Detection via Remote Sensing and Machine Learning for Optimal Hive Placement to Maximize Lavender Honey Production
General Comments:
The article is methodologically thorough and well-structured, representing a solid contribution to the field of precision apiculture and agricultural remote sensing.
The integration of classical machine learning and cutting-edge deep learning techniques is particularly commendable.
However, increased clarity on some methodological choices, validation efforts against ground truth honey production, and uncertainty quantification would improve the robustness and applicability of the findings.
The manuscript could better highlight novel contributions versus previous work, especially given the number of cited studies using UAV and remote sensing for crop detection.
Title and abstract:
Both the title and abstract effectively convey the essence of the work, balancing technical detail with accessibility, and highlight the innovative integration of remote sensing, machine learning, and beekeeping optimization. Therefore, they are well chosen for their purpose.
Introduction
The introduction, while thorough, is very dense and could benefit from more structured subsections or paragraphs focusing separately on the problem statement, background, and objectives.
It touches briefly on climatic challenges but does not explicitly link how the study addresses these in the research design.
The cited beekeeping statistics could be updated or expanded with more recent global trends related to climate change or pollinator decline to reinforce urgency.
- Material and Methods
Although the description of the methods is detailed, the discussion on the choice of parameters for each classification method (e.g., kernel type for SVM, tree numbers for RF) is absent, which is critical for reproducibility.
The metadata and preprocessing steps of the orthophotos (radiometric corrections, atmospheric corrections) were not discussed.
The explanation of accuracy assessment metrics is good but could be complemented by cross-validation strategy descriptions, if any.
There is no mention of potential overfitting in the deep learning training and how it was addressed (e.g., validation datasets, early stopping).
Details on masking or exclusion masks applied to reduce shadow misclassifications could be expanded.
- Results and discussions
The discussion would benefit from including quantitative uncertainty or sensitivity analysis in melliferous potential estimation.
There is no validation of predicted optimal hive locations with actual beekeeper data or field performance.
The potential impacts of interspecific competition (other pollinators) and anthropogenic disturbances are acknowledged but not quantitatively considered.
More direct comparison with other remote sensing resolutions (e.g., UAV imagery) is discussed only briefly; empirical comparisons could strengthen claims.
The assumption that all lavender within a 3 km radius contributes uniformly to honey potential may oversimplify forage quality and accessibility.
Conclusion
Conclusions mostly restate what was discussed without emphasizing key novel scientific contributions or specific recommendations for beekeeper practice or policy.
There is an opportunity to better address how this methodology can be generalized across different regions or crop types.
Suggestions for future validation and interdisciplinary collaboration (e.g., between remote sensing specialists and apiculturalists) could be more detailed.
Author Response
See the attached file.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript employs remote sensing and machine learning for lavender field detection to optimize hive placement for maximizing lavender honey production. The research topic demonstrates high practical value.
1.Overall, the results analysis lacks depth and contains no valuable discussion.
2.The manuscript fails to provide accuracy trends of the training and validation sets during model training, making it impossible to assess potential overfitting.
3.No actual yield data from local beekeepers are referenced to compare with model predictions after obtaining results, leaving the accuracy of the potential map unverified.
4.The explanation for the 'low learning values of lavender types 1 and 2' is unclear and lacks correlation with sample characteristics.
5.The analysis uses only 2020 remote sensing imagery, ignoring impacts caused by temporal changes in lavender field layouts.
Author Response
See the attached file.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe study is very interesting and useful for both beekeepers and lavender producers. Below are some suggestions that the authors should consider to improve the manuscript:
In general, the manuscript has certain issues related to long sentences that are not fully understood by the reader. I recommend breaking them into shorter sentences. Some paragraphs mix results, discussion, and recommendations, which could confuse the reader about which parts are supported by data and which are inferences or practical recommendations.
Line 179–185: The authors should clarify in the methodology how the information from the techniques used to detect and classify lavenders was quantitatively integrated. (Was it weighted? Was it simply merged through spatial union?) It would be useful to clarify whether there was a priority criterion when discrepancies occurred between methods.
Line 189–208: The text describes production and location, but mixes annual data without specifying whether the figures are averages or from a specific year (e.g., the 5 tons of lavender honey). It would be relevant to clarify whether the total lavender area mentioned in the study area matches the surface detected in the image analysis; otherwise, any differences should be discussed.
Line 231–261: The authors should clarify why they chose July 2020 imagery and whether this date coincides with the lavender peak flowering period in the region (this affects reflectance and classification accuracy). The spectral band description is correct, but they could mention whether any atmospheric correction was performed prior to classification. The use of “circular regions” for lavender sampling is interesting, but the authors should justify why this shape improves classification compared to irregular polygons.
Line 278–294: The value of 150 kg·ha⁻¹·year⁻¹ for lavender comes from the literature, but the text does not indicate whether it was empirically verified in the study area. This is important because local conditions (climate, soil, management) can reduce or increase that potential.
Line 318: The RF method achieves higher accuracy, but it would be interesting for the authors to discuss whether this result could be due to overfitting in the training set, especially given that there are only four classes.
Line 586–621: The text states that the total number of detected lavenders is “slightly lower than the actual value” but does not explain how the real number was estimated for this comparison.
Line 620–622: When mentioning the area range (0.2 m² to 38 m²), it would be advisable to specify whether this refers to the horizontal canopy projection or the ground surface occupied.
Line 648: The transition between the Mapping Melliferous Potential of Lavenders and Optimal Hive Placement sections is somewhat abrupt. A short introductory sentence could be added to explain the shift from quantifying potential to evaluating hive placement strategies.
Line 657: The text mentions that the estimate of 17,000 kg/year is theoretical, but does not explain how that value is distributed among hives or what the actual expected production would be, considering collection efficiency and other factors.
Author Response
See the attached file.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsMajor Revisions
-
Methodological Rigor:
-
Address class imbalance via inverse frequency weighting:
weight_k = Total_Pixels / (Class_Pixels_k × Class_Count) -
Validate RF superiority with McNemar’s test (report χ²/p-values).
-
Formalize melliferous potential:
MPLP(i) = 150 × (Lavender_Area_i / 100) × Coverage_Fraction_i
-
-
Results Depth:
-
Explain 140-fold yield gap using:
-
Pixel coverage stats (e.g., 22.7% mean lavender cover).
-
Heatwave impact (35°C vs. 12°C baseline) [71].
-
-
Propose empirical validation:
*"Deploy 30 hives across PBLE zones. Monitor yield + DNA metabarcoding for monofloral purity. Model:
Honey_obs = 0.87·PBLE + 1.2·Temp - 0.3·Management + ε"*
-
-
Reproducibility:
-
Include missing figures (study area map, classifier comparisons, MPLP/PBLE maps).
-
Share code/data: GitHub repository with DOI.
-
Minor Revisions
-
Fix typos: "Iand" → "Land"; "Lund" → "Land".
-
Convert URLs to formal citations (e.g., [73] Ministry of Agriculture, 2022).
-
Clarify sample discrepancies (Sec 2.3: 3,750 samples vs. Table 1: 2,612).
Comments for author File: Comments.pdf
The English could be improved
-
Typos: "Iand" (p.1), "Lund" (pp. 3,5,6) → "Land".
-
Awkward phrasing: "Lavenders identified by ML and SVM methods cover much larger areas than they actually do" (p.13).
Author Response
See the attached file.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsI hope that the comments provided have contributed to improving the quality of the article. I would like to express my best wishes for your future academic and professional achievements.
