# Reviews

> Original submission was not changed.

## Reviewer 1

The work is interesting and worthy of publication. However, it is **worth reducing it somewhat**, since such a volume of information is difficult to understand.
For this purpose, it is worth **removing the application**. If there is a link to the software, this is unnecessary. Surely all the information is on the site.

> Dear reviewer,
>
> Thank you for reviewing our article. All reviewers provided valuable feedback.
>
> We understand your concern about the volume of the publication. However, we would have liked you to be more specific about which parts should be excluded. Since the application will continue to improve, we feel it is important to have its core features in their current form documented in the paper.
>
> Best regards,
>
> Martin Schröder, Rainer Kiko and Reinhard Koch

## Reviewer 2

The authors present a new tool (MorphoCluster) to classify and identify images of plankton objects based on unsepervised clustering that improves the performance of other available tools. The authors show convincingly that MorphoCluster saves time, reduces classification biases due to dependence of training data and low sensitivity to detect new groups and increases the precision in classification. The manuscript is well written and I congratulate the authors for their good work. I have only a few minor comments that might improve the text:

**Line 122. It would be useful to indicate where and when this set of data was generated.**

> - [x] Include details about dataset

**Line 180. Perhaps a value of 'm' should be indicated here to illustrate what a "large value" is.**

> - [x] Done

Lines 229. I guess that this is the reason that 1192 clusters were validated (Table 2) but only 280 classes were identified (Table 3). In some way, this part of the process validates the previous cluster selection since both phases depend on the same criterion (similarity or dissimilarity between two groups of objects). **It would be interesting to analyze if the distance among merged clusters is lower than certain threshold.** If this is the case, I will suggest that this **agglomerative clustering technique is used to establish a distance threshold that in turn can be applied to assignation of new clusters in the previous step** as a way to optimize the later application of MorphoClusters to other samples.

> While this is a nice idea, we would expect such a threshold to be highly data-dependent. It could therefore only be calculated after doing the merges manually and would therefore be of no use during this process.

**Figure 4. The y-axes of both graphs are not exactly equal that distort slightly the comparisons.**

> - [x] Scale equally

Line 274 (and other parts of the manuscript). **I wonder if it is possible to indicate how much time takes the equivalent analysis of this sample with ZooImage.**

> - [x] Cite ZooImage
>
> We don't know.

Table 3. I am surprising that **146 classes of non-living materials were identified** (this amount is even higher than that of the living material). It indicates that there is a fairly high diversity within of this particle group. However, I am **not sure which the ecological or handy significance of this classification is**. I recognizes that this aspect is mentioned in the Discussion (line 515), although perhaps it requires some more extensive explanation.

> - [ ] Elaborate?
>
> Nearly 90% of all objects in the data set are non-living, therefore it is expected that a large number of clusters emerges and a reviewer can spot significant differences between them.

Line 329. This result is expected for Lo as there were only 11 non-living categories.

Figure 6. **Please, to explain what 'artefact' means in this context.**

> It does not make much sense to explain at this point what the individual supercategories contain. Maybe a table at the end that shows the individual included labels?
>
> - [x] Include list of groups in appendix.

**Lines 374-392. I think that these lines are repetitive (similar questions are exposed in the Introduction). The manuscript is large and possibly these paragraphs can be deleted.**

> *This is the introduction of the discussion.*
>
> - [x] Refuse and defend.
>
> As many readers might skip directly to the discussion (all the more so when one considers the length of the paper) we found it important to recall the motivation. As actually only 374-385 repeat the motivation for the work, the savings in space would be insignificant.

Line 574. Please, give some **example about which environmental information would be included.**

> - [x] Done

On overall, I miss some paragraphs in the Discussion devoted to **discuss how this tool would perform in analyzing plankton samples with other features (for instance, lower abundance and/or diversity of objects) in comparison with alternative tools.** In other words, **would be MorphoCluster more suitable for some particular type of samples?**

>MorphoCluster is not practical for very small data sets, discovery of very small classes, or small numbers of very diverse objects. It is well suited for data sets similar to the one used in the experiments. Around 100 similar images are required to somewhat reliably be retrieved as a cluster. If the taxonomic units of interests are expected to contain less than ~100, MorphoCluster will not be of much use.
>
>- [ ] Elaborate on that in the discussion.

---

> Dear reviewer,
>
> Thank you for reviewing our article. All reviewers provided valuable feedback. We would like to thank you in particular for your positive words and the time you obviously invested to read the paper very carefully.
>
> As requested, we did the following:
>
> - Report the spatial and temporal extent of the data set
> - Report an example value for the "m" parameter in the methods.
> - Fix figure 4 so that both y-axes are equal.
> - Include examples of environmental data that could be used in addition to images.
>
> To explain what "artefact" (and other broad supercategories) in Figure 6 mean, we included a complete list of all supercategories in the appendix.
> Artefacts are e.g. air bubbles that are imaged at the surface and are caused by the mixing of the water column when the CTD enters the water, or grey images that are collected in very turbid waters. All images and classifications can be found in the published dataset (https://doi.org/10.17882/73002).
>
> Your suggestions to optimize the process sound interesting and should be discussed in detail.
>
> A comparison to ZooImage would surely enrich the paper, but we are not aware of any published estimates regarding the sorting speed when using ZooImage. ZooImage works similar to EcoTaxa and we would therefore expect a similar performance if manual curation of all images is requested. Processing the dataset we used here with Zooimage to obtain a performance estimate would take a lot of time and resources, which we currently can not provide to only obtain such a performance estimate. We note that the lack of such estimates actually hampers the comparison of these tools and the development of more effective tools.
>
> Regarding your concern about the large number of non-living classes:
> Our article aims to describe the benefits of the MorphoCluster application. It is true that we find many more classes than normally distinguished, but we think that it would exceed the scope of this article to speculate about the ecological significance of certain classes. This would require an in depth analysis of e.g. distribution patterns of these novel classes in an environmental context. We of course will conduct such analysis, but this would add considerable material to this article. We therefore prefer to keep the current formulation.
>
> We understand your concern about the first paragraphs of the discussion being repetitive. It is true that similar arguments are raised in the Introduction, but many readers do not consider Introduction or Material & Methods when they first explore an article (all the more so when one considers the length of the paper). We therfore found it important to recall the motivation at the beginning of the discussion. And as actually only lines 374-385 repeat the motivation for the work, the savings in space would be insignificant. Please excuse that we would rather keep the structure as it is.t
>
> We included a small paragraph
>
> Best regards,
>
> Martin Schröder, Rainer Kiko and Reinhard Koch

## Reviewer 3

Are the results clearly presented? : Can be improved

Are the conclusions supported by the results? : Can be improved

The authors presents interesting material reporting their activities aimed at creating an improved software for marine plankton sorting. The manuscript is relatively extensive - the authors present their activities against the background of current technology and knowledge in relation to identification software of plankton species composition.

The authors express their thoughts rather in a **fairly hermetic language** - mathematical and IT. Of cause this is understandable due to the subject matter discussed, but it can narrow down the number of readers.

I don't find any errors at work – therefore in principle, it could be published in its current form.

I have following comments only. The content of the manuscript proves that the authors know this topic very well and that they want to convey as much detailed content as possible. However **I miss showing how to use the methodology presented by the authors in an operational way**. But I do not urge authors to expand the content of the manuscript - the manuscript is already very long at this time, which may not encourage some scientists to read.

> What is meant by that? Contradicts Rev1.

Did the authors notice that the term **“MorphoCluster” is repeated over 80 times**? Maybe they are able reduce this number?

> - [x] Check which mentions could be removed.

**Are the authors sure that Conclusions should be included in the structure of Discussion?**

> - [x] Change

---

> Dear reviewer,
>
> Thank you for reviewing our article. All reviewers provided valuable feedback.
>
> We understand your concern about the "fairly hermetic language". However, we believe that it is appropriate for the content of this paper.
>
> We would have liked you to be a little more specific concerning the use of the methodology in an operational way. We believe that sufficiently described how to apply our method. More detailed descriptions would go beyond the scope of this paper and are more appropriate for the documentation of the software. Like you wrote, the manuscript is already long.
>
> We realize that the name of the application was repeated very often in the manuscript. We have reduced its use by more than 10%.
>
> The conclusions have been changed to a proper section.
>
> Best regards,
>
> Martin Schröder, Rainer Kiko and Reinhard Koch

## Reviewer 4

Authors proposed novel approach for plankton images for sorting without using single image. They provided simulation code to be verified. I am not very confident that this manuscript could be related with journal of Sensors topic. However, the research background for literature search is excellent. The research method and result data are concrete. There are some English grammar issues. **I hope authors can mention that this article could be related with sensors in the abstract, discussion or conclusion sections.** Therefore, the manuscript could be minor revision. There are some minor comments as below.

> - [x] Relate article to Sensors journal somewhere.
>
> The problem is that the special issue itself does not fit well into the general scope of Sensors, nor its section Physical Sensors. However, the scope of the Special Issue is well met.
>
> > *Invitation:* The ocean has a huge impact on our way of life, and there is a need to **monitor** and protect its **biodiversity**. Additionally, the ocean industrial potential for health, minerals, and oil has promoted the need for **constant and real-time monitoring**. Because of all of these concerns, it is important to develop **efficient** and smart marine sensors to improve our knowledge of the sea environment and to support marine sustainable development. Taking advantage of the most recent progress in sensing techniques, the purpose of this Special Issue is to provide an opportunity for the engineering and biological marine community to exchange knowledge and information on the latest advances and challenges in marine sensors and their applications to monitor sea environment.

1. The **small labels in Figure 8** are need to be larger.

> Done

2. In Figure 8 label, please change to Figure 7.

3. please **fig. 1** change to Figure 1.

4. In Line 131, **please mention author name** in [27] and [35] if possible.

> Done

5. Please label Equation numbers between 251 and 252.

> Done.

6. Please put **Table 1 labels before Table**.

> - [x] Check requirements of Journal

7. In Figure 4, there are **two labels merged** such as Flota and Poeobius.

> Done.

8. please **fig. 5** change to Figure 5.

9. Please put **Table 3 labels before Table**.

10. If possible, **combine Figures 8 and 9 into one Figure**.

> Fig 8 and 9 illustrate different things, so this is not possible.

11. In MDPI reference format, please remove doi information.

> *Will be done by MDPI, I think.*

12. Delete vol. in the reference.

> *Will be done by MDPI, I think.*

13. Please use abbreviated journal names.

> *Will be done by MDPI, I think.*

14. Please **provide city and country and date of conference information**.

15. Please use abbreviated conference names.

> *Will be done by MDPI, I think.*

> Dear reviewer,
>
> Thank you for reviewing our article. All reviewers provided valuable feedback. We would like to thank you in particular for carefully checking the finer details.
>
> The paper is a feature article for the Sensors Special Issue "Sensor Applications on Marine Recognition“. The topic of our article matches the following keywords of the special issue: “Marine recognition for biology”, “Deep learning and AI methods in marine recognition”, “Computer vision and image processing methods for marine recognition” and “Data management, annotations, access, visualization, and sharing“. We therefore believe that our article is a perfect fit for the special issue and for the Sensors journal.
>
> 1, 4, 5,7,14: Done.
>
> Placement of table captions (6,9) and abbreviation of figure references (2,3,8) are a personal choice. The journal does not enforce a style.
>
> It is not possible to combine fig 8 and 9 because they illustrate different things.
>
> 11,12,13,15: According to MDPI, their production team will adjust the references after the paper is accepted. We included all information that is needed.
>
> Best regards,
>
> Martin Schröder, Rainer Kiko and Reinhard Koch