Deep Edge IoT for Acoustic Detection of Queenless Beehives

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a well-structured and methodical study on IoT-based acoustic detection of queenless beehives using edge computing. The authors effectively integrate sensor technology, signal processing, and machine learning to achieve real-time classification on resource-constrained hardware. The experimental design is robust, covering data acquisition, feature optimization, model evaluation, transfer learning, and hardware deployment. However, the following comments still need to be addressed.

(1) The study uses data from only two hives (m11, m12), with queenless states artificially induced. How do you address potential bias from limited hive diversity or artificial induction?

(2) The pre-processing includes a 60Hz band-pass filter for electrical noise, but no tests validate performance under real-world acoustic interference. Can you provide evidence of the system’s noise resilience or suggest mitigations?

(3) PCA is used to rank MFCC band importance (Table 3), but MFCCs are inherently non-physical. Why not directly analyze frequency bands linked to known bioacoustic markers? Does PCA’s variance-based ranking align with biological relevance?

(4) The 4-layer NN (Figure 8) was selected after Pareto tuning, but no ablation studies justify its depth/width. Were simpler architectures or model compression techniques explored to further optimize edge compatibility?

(5) The NN outperforms KNN/SVM in accuracy and TL adaptability, but no comparison is made with state-of-the-art edge-compatible models (e.g., MobileNet, TinyLSTM). How does the proposed NN fare against such architectures in accuracy-resource trade-offs?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Add some references to the equations.

Explain why are selected features most dominant.

Show some other parameters:levels..maybe spectrogrma of recordes signalas...time domain signals..

add some parameters of used equipment for monitoring the sound...sensitivity of microphone. A/D convertor resolution...electronic noise..

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The work deals with the use of electronic systems as tools for monitoring the presence or absence of queen in a hive. To this end, the authors use sound recordings to create a dataset with the two possible states. They also study the implications of different parameters, both from acquisition and processing point of view of, on the accuracy of the system, comparing the results provided by different Neural Network and Deep Learning algorithms and comparing their performance.

The manuscript addresses a useful issue and the document is generally well structured. However, some aspects need to bee improved.

• Certain parts of the document can be summarised.
• The abstract of the paper is contained, almost in full, in the introduction. The authors should take the time to rewrite it in other words.
• The references must follow a sequential order. In line 43, references 6 and 10 are included, without first indicating 7, 8 or 9.
• The structure of the introduction is somewhat confusing. The paragraph between lines 35 and 48 would fit better after line 74, once the importance of sound for bees has been introduced.
• Line 64, remove the reference.
• The title of section 2.1 is somewhat confusing. Perhaps the ‘experimental’ part should be deleted and only indicate that it refers to the procedure for obtaining the dataset.
• Figure 1. A more general view of the position of the box in the hive could be included.
• Regarding the set-up used, where some absorbing material is included, can this influence the acquired sound? (Higher absorption at high frequencies, ‘enclosure’ modes).
• Line 153. The use of reference 31 is not justified.
• Lines 155 to 157. The notation used for the recording sequence is somewhat confusing, as the last C lacks an orphan. It may confuse the reader.
• Line 166. ‘...to multiple signal processing techniques including filtering, normalization, ...’. It would be better to be more specific and put all of them or just say ‘... to multiple signal processing techniques’, as they are included after.
• Lines 173 to 176. Make reference to table 1 so as not to leave the sampling frequencies used in the air.
• Line 189 to 190. Reference should be made to Table 1.
• Line 197 to 189. It would be useful to include some reference to the Mel spectrogram. This paragraph could be shortened by including such a reference.
• Figure 3 does not provide relevant information to the document, especially if some reference is included.
• Line 224 – Include reference to Principal Component Analysis.
• Line 233. The abbreviation PCA does not appear earlier in the text. Specify it on line 224.
• Paragraph between lines 243 and 252. Some sentences are repetitive.
• Lines 263 and 264. Include references to the parameters used.
• Line 270. Include reference or duly justify the exclusive use of F1-score.
• Paragraph 2.5. The system should be described in more detail. How is audio acquisition carried out on the Arduino system, other components? Maximum sampling frequency, frequency response of the system, self-noise or SNR, ...
• Line 295. A system sampling frequency of 4 kHz is given. This is an important limitation taking into account the Nyquist theorem.
• Figure 5. Is this a measurement of background noise? and what is the frequency response of the microphones used? Visualising the frequency response of the system may provide more information. At high frequency, has self-noise been achieved?
• Lines 321 to 323 can be deleted.
• Section 3. Results. Much information is redundant, provided in graphical, table and text format (section 3.1 and 3.2). Reduce to avoid overloading the work.
• Figure 6. Font size larger than the rest of the figures. The outer border could be removed.
• Line 335 to 338. Some more information on the algorithms used could be included.
• Line 371. Missing space between connected and (FC).
• Figure 7. The font size is smaller than in the other figures. Use the figure caption to better describe the content (a, b and c).
• Paragraph between lines 379 and 388. Some reference could be included.
• Figure 9. By changing the range of the ordinate axis, the difference between methods could be better visualised.
• Line 450. Sample rate 8192Hz. Wasn't it 4000 Hz? Clarify these aspects.
• Section 4. Discussion. Mention is made of the choice of configuration 3, whereas in the rest of the text it is configuration 2.
• The discussion and conclusions sections could be integrated into a single section.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

All the comments are revised properly. This manuscript can be accepted now.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Authors have improived the manuscript , senistizvity of mic is given in mV/Pa with referent value..so it should be written.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have improved the work by modifying or clarifying almost all recommendations. Some aspects that could have left some doubts have been correctly solved. I would just like to point out some minor aspects that need to be improved. Firstly, the resolution of figure 5 should be improved, as it is clearly lower than the rest. Similarly, the zoom on Figure 9 is somewhat confusing. It could be enough with the zoom. Finally, authors should try to use the same format in all the tables.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf