A Methodological Literature Review of Acoustic Wildlife Monitoring Using Artificial Intelligence Tools and Techniques
Round 1
Reviewer 1 Report
This paper is a review of artificial intelligence (AI) used for the acoustic monitoring of wildlife. I was excited to see a review of this topic, especially across taxa, and I think it is an important review to support the ongoing development of acoustic monitoring. While this is a valuable topic, the paper was not written in a way that would be useful to the majority of the practitioners in the field it is writing about. This paper feels like it was written by engineers that have not worked with biologists in the field they are writing about. While I would like to see this paper published, I think there is significant work to be done to make it a paper that will contribute to the field. My first suggestion is to find a collaborator and additional author with extensive experience in acoustic monitoring. My next main recommendation is to expand the review so that this paper could be a reference for people transitioning to using AI for their acoustic data processing. It would be good to know which algorithms are the most accurate or the advantages/disadvantages and what are good resources for applying these methods. What does a practitioner need to consider? Please consider the audience of this review – since it was submitted to Sustainability and not an engineering or computer science journal, it needs considerable work to be helpful to the biological sciences.
Additional feedback:
**Title The title does not capture what this paper is about
**Abstract
Ln 19-20 “bioacoustic wildlife monitoring” – for simplicity, I recommend just referring to the topic as “acoustic monitoring” throughout.
**Introduction
Ln 29 remove the word “breeding,” acoustic monitoring is used for learning about a broad range of behaviors well beyond breeding behaviors.
Ln 41 The barrier to acoustic processing data is the large number of recordings collected, but “large number of datasets” is not a clear way to say this. We usually say that acoustic recordings are “cheap,” and processing those recordings to turn them into meaningful data is time-consuming and requires a lot of training.
Ln 46-47 I am unsure if this statement about AI in bioacoustics first being used for marine environments is accurate, but these are not appropriate papers. Cite the first paper to use AI in the field.
Ln 49 “forested wildlife species” is a very poor descriptor to summarize non-marine organisms. Use “terrestrial organisms” instead.
Ln 50 “large sound recordings” is poor wording. Ruff et al. (and any other paper discussing AI for bioacoustics) would primarily be focused on how to process many files (large dataset), not how to deal with a “large” recording in filesize or duration.
**Methods
Ln 73. Delete the sentence about Fig. 1. Should not talk about the figure directly in the text, should only reference it. Add a reference to Fig. 1 at the end of the sentence starting on ln 57.
Ln 104 The review should include all bioacoustics that use AI. Why limit it to papers after 2015? A key paper from the bat literature using neural networks is Parsons & Jones 2000 (https://doi.org/10.1242/jeb.203.17.2641).
Broadening the citations included in the review would make this paper more impactful.
**Results
Ln126 missing word – hundreds to millions of recordings or files
Ln130 be consistent with capitalization – Convolutional Neural Network (and ln 162)
Ln144 “song meter” is not a general term to be used in this way. Throughout should refer to detectors in a broad way. Song Meters are a brand by Wildlife Acoustics (https://www.wildlifeacoustics.com/products). There is a lot to say about the algorithms and ways to choose a detector. Don’t get into talking about specific products in this paper.
Ln204 What is an F1-score?
Ln230-238 This is a helpful paragraph, and would like to see more like this in the paper
Ln239-245 remove the paragraph. See comments above for ln144
Ln264 Background noise is not much of an issue anymore because of the acoustic software capabilities and AI
Ln269 multi-species calls due to call overlap. This is one of the biggest issues with processing. Unclear if the authors are referring to when two individuals are overlapping in the same recording (e.g., singing at the same time) or for species that have similar acoustic profiles and are very difficult to distinguish from one another. This is the most valuable place where AI plays a role.
A last thing to consider adding to a revision is a discussion of differences between processing echolocation (active sensing) versus social calls (passive sensing). The acoustic signals between these two types of sensing are considerably different. Is there an AI algorithm that is more accurate for one or the other?
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This ms represents a needed study due to the growth of the use of AI in bioacoustic studies. The general methodology for identifying useful studies to analyze seems reasonable, although more details are needed. For example, when listing the search terms that were used to find the articles, the researchers specify that “koalas” was one of the search terms with no explanation of why that particular species was chosen. This was much more specific than the other search terms, which suggests that the authors were seeking to find a particular citation to include in the study, which is problematic. I also think it would be good to include supplemental information with a full listing of the articles that were considered and why they were rejected.
While I believe the general idea is a good one, I find the presentation of the information to be vague and difficult to follow. For example, each section of the results section is quite short with very little detail. Much of the information in the text appears to refer to the tables for details, but the tables are also limited in terms of the information they provide. In particular, the “Performance” column of table 1 is really not useful as written. Some of this is because the different numbers aren’t explain, so they are difficult to interpret. In addition, because the numbers represent many different types of measurements and no attempt is made to put them on the same scale, there is no easy way to tell how they compare.
I think section 3.2.1 is a good idea to include, but it needs more explanation to be useful. For example, LDFC is said to be ineffective when the species only calls in the morning, but no reason is given. Some of these ideas might be points that are raised in the original journal article, but at least some of that information needs to be in this article.
Section 3.2.2 is also too short – a great deal more information is needed to really understand the ideas.
Lines 239-245 aren’t clear. I wasn’t able to determine what the paragraph was trying to say. Likewise, figure 3 is low-resolution and prints out as jagged, to the point where it is difficult read.
I liked section 3.5, because it does provide a good summary of ideas, but with more details in the earlier sections, this would be more useful.
Section 4 has some issues with the explanation of ideas, such as the claim that AT is automatically predicting the outcomes of datasets as this is not how the tools are being used in these types of study. They are being used for analysis after the fact, not prediction.
In the end I think this is a useful start, but it needs to be significantly expanded before it will provide significant benefits to the journal’s readers.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
If the article is for the general reader it would be better to clarify the abbreviations like CNN, CRNN, HMM, F1-score and others, and describe the AI method$
Lines 134-138, as well as lines 154-156, 158-169 looks like introduction;
I did not see the advetage disadventage of AI method for marine animals, just general statement;
Line 225
| like Simple Minded Audio Classifiers |
repited twice,
Line 372-379 repited introduction
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Thank you for the thoughtful and thorough responses to the reviewer comments. While it looks like there have been substantial improvements. There is still another round of major revisions to be done.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
While I appreciate the additions to the text, I still find what has been included is too vague to really get a good overview of the topic. I feel that the reader needs to go and read most, if not all, of the literature you cited to be able to get a good understanding. A review article needs to give a more clear summary of the topic if it is going to be helpful to the reader.
And while I appreciate the idea of having a "performance" column in table 1, I cannot agree that it provides valuable information given that there are so many different units being used and there is no clear way to convert them to get a meaningful comparison. Without that, the fact that one method gives a higher number than another won't be useful without context.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
I think this revision addresses my major concerns to the point that it seems suitable for publication.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
