UAV-Based Land Cover Classification for Hoverfly (Diptera: Syrphidae) Habitat Condition Assessment: A Case Study on Mt. Stara Planina (Serbia)

Round 1

Reviewer 1 Report

Authors have used UAV to map a land cover for object-based land cover classifications for assessing the Trend of Hoverfly (Diptera: Syrphidae) Species Richness. It is a great improvement on simple UAV based LULC mapping with good implications. The effort by the authors is very much appreciated. People who have beginner to intermediate knowledge in image processing and RS system can understand the article.

Following are the comments:

1. Abstract need one or two-liner from the 1.1 section relating the habitat degradation and Hoverfly.
2. The introduction of the article sufficiently provides the problems but the last para in section 1.1. does not link those problems but suddenly jumps on the tasks. As there is no actual data for quantification of habitat loss, then how this work links it? Is it data production, quantification, or new method development?
3. Again, objectives are on point and are focused on the problem tackled in the article.
4. OBIA is not new (author states Large-Scale Mean-Shift (LSMS)-segmentation algorithm became the focus of remote sensing community since being introduced by Fukunaga and Hostetler in 1975 and complemented by Commaniciu in 2002), it has been quite old now and the tools/algorithms that you are using are also developed long ago.
5. Both sections are written well, but they are not linking well. Yes, UAV is great but has lots of its problems, high resolution is not that easy as authors are presenting. It is easy for one feature extraction but a lot of pain in making such seven classes. Please try to fuse two sections and point out problems, other works that provide a solution, your identified gap and the solution with the objectives like in 1.1. last paragraph.
6. Everything done here is related to LC mapping and classes are linked with hoverfly species. Although stated” According to the list of potential hoverfly richness (Table S13) the number of hoverfly species in this study site”, there is no clear description of how authors linked several species of hoverfly in the region with only land cover information. Please explain how these species of plants influence hoverflies in the intro.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In the manuscript entitled „UAV - Object Based Land Cover Classification for Assessing the Trend of Hoverfly (Diptera Syrphidae) Species Richness A Case Study on Mt. Stara Planina (Serbia)” submitted to journal remote sensing by Ivošević et al. authors set 4 research goals: to obtain very high resolution (VHR) land cover maps base on the framework by De Luca et al. (2019), to determine the habitat degradation in the study areas, to compare potential and recent richness of hoverfly species and finally to present tools for spatial and landscape ecology research.

The applied methodology (unmanned aerial vehicle (UAV) for (VHR) imagery acquisition, Support Vector Machine (SVM), Random Forest (RF), and k-nearest neighbours (k-NN) for land cover classification) presented understandably. But the assessing of Hoverfly Species Richness issues should be better clarified, especially the use of land cover data. I find the paper acceptable for publication with major revision.

Some remarks & questions:

In the article great emphasis is put on the issues related to the acquisition of very high resolution (VHR) land cover maps. The description of method of obtaining data, from the acquisition of images to the evaluation of the results of land cover classification, is presented in great detail. Undoubtedly, the first and fourth goals set by the authors have been achieved.

The third goal was also achieved, although with some reservations. Page 13 rows 432– 433 "The material on which the list of recent hoverfly species richness was conducted was the result of around thirty years of research on three selected study sites". The results of more than 30 years of research have been summarized in the article within 7 lines. There is no reference to the literature describing the research methods. Only a reference to the research on potential hoverfly species richness was given. Hoverfly species richness has probably changed for 30 years. Giving only the numerical results is a great simplification of the topic. Are the data describing the number of hoverfly species (page 13, rows 438–439) averaged.

My biggest doubts are related to the definition of habitat degradation. Why are certain types of land cover described as degraded, such as forest patches, road, meadow, and shrubs. Probably these areas are under anthropopressure, some are also strongly transformed, but degradation is a very "strong" term meaning that their quality was destroyed or spoiled. Should all areas where human activities took place be considered ecologically degraded? Is classifying an area as degraded on the basis of a comparison of potential and tested hoverfly species richness too categorical?

The proportions of the article are disturbed. The topic related to the acquisition of land cover data dominates, and the issues related to the hoverfly species richness are rather briefly presented. I recommend changing the title. As the title suggests, that the obtained land cover maps were used to assess the richness of the hoverfly species. However, in the text, the number of hoverfly specie was obtained on the basis of other studies, without using land cover data. In the title of the article, the authors included "Assessing the Trend of ..." in a statistical sense, a trend is an increase or decrease in the value of a variable over time. The article does not present changes in time for both land cover and the number of hoverfly species.

Page 4 I suggest changing the of Figure 1. composition. Reducing the map showing the area of Serbia or removing it and leaving only a small map with the location of the study areas will allow for the enlargement of very interesting detailed maps of the studied areas. The general location of the research area is well described.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this resubmission the Authors address an object-based land cover classification from UAV-derived imagery, in order to detect different land cover categories and habitat degradation, focused on how differences affects the hoverfly species richness in three different study sites over Serbia. I really appreciated the efforts made in improving the original text by all the changes made to the manuscript as suggested, denoting a hard work behind. Above all, now the work has a clear objective, in a very interesting topic in landscape ecology remote sensing and environmental management.

In my opinion, though I am not native, the authors should, above all, improve the English style and grammar. The text structure needs to be carefully checked. Several sentences are hard to read, often too long and wordy, containing unnecessary words. A knowledgeable audience might find this uncomfortable and mislead from the key contents, distracting from the interesting workflow regarding the description of the experiments and the discussion of results. This inevitably influenced my review, as the key steps of the new experiment are not clearly reported and I would reserve the right to review them once modified. As example, see section 2.5 “The Map of the Natural Vegetation of Europe (EPNV)”. Here the Authors should better address how they consider habitat "degradation", comparing the UAV-derived land cover categories and the EPMV, as defining habitat degradation is a very complex topic, even more so that in this case the results do not derive from a cover change detection approach that clearly characterizes the trend of the dynamics.

Secondly, comparing the EPNV at its own scale, with a very high-resolution land cover map, exposes the risk of getting carried away, blurring the interpretation results. What contingency plan? The authors should better address the critical limitations affecting the selected vegetation map to cast their analyses and mapping the outcomes. I think that this is the most critical point that should be addressed.

Other general issues:

Introduction, from line 61 to 76: this information is redundant with section 2.5. I suggest to move to materials and methods as it give insight of the dataset used, keeping only essential information.

The same again from line 141 to 149.

The methodology is very articulated, in order to improve the clarity, I suggest to the authors to include a flowchart able to synthetize the whole methodology, not only the UAV images processing.

Emphasis should be limited to keeping a more sober style. At least the authors do not provide further explanations.

In discussing the results, the authors have the ambition to say that the decrease in hoverfly species richness is related to the degradation of natural habitats but there is no statistical evidence to support this claim. These first results highlight a trend that should be verified although the authors already stated that this paper “represents a starting point for further similar research topics”.

In conclusion, I suggest to reconsider for publication as a scientific article on Remote Sensing after a major revision.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The author's replies and effort for revision are highly appreciated. I find authors making lots of speculative sentences rather than stating a fact with proper references. Few replies are quite biased and are just used to counter the queries such as:

1. A Few works on UAV-OBIA is false
2. No new method to quantify habitat degradation is false. Check drones journal with lots of habitats works mainly coral and coastal regions.
3. The author’s argument is human intervention but these areas are simply forest/parks… with a road. Don’t think there is any historical data or changes the determines degradation. Also what is the “negative anthropogenic effect”?
4. Quantification here is only LC mapping and integrating with potential richness, the major work is not habitat focused.
5. To understand hoverflies, their lifecycle and the area they travel during their stages should be very carefully examined, work here is a very small area and richness is simply a class of land cover. The author still misses justifying the major work of this manuscript.
6. If Hoverflies are very sensitive then the best way to do is habitat suitability with mid resolution imagery for a large area not a high resolution to a small area.
7. Could you please describe the habitat condition?
8. Three must be multiple suitability factors for the hoverfly not only LC, one example is a pollinating condition so flowering plant or season temperature could be something additional to consider.
9.  Downscaling higher to low resolution is fine but the reverse is not good for any comparison or transfer for factors.
10. UAV-Object baked LC.. not sure if this is a good term in the title or elsewhere.
11. Line 159 check 
12. The workflow is still graphical abstract but does not represent the work done, please see for good remote sensing and habitat suitability workflows.
13. Why is it necessary to have two workflows, remove 3 and graphics and make a proper workflow? These can be a graphical abstract for the website.
14. Authors state “to evaluate the potential of UAV technology in obtaining detailed classified land cover maps by taking into consideration various flight limitations and complexities of the terrain especially in the study site” but these limitations and complexities have never been introduced in problems.
15. Many labs and institutions share data openly. And if you are publicly found and doing open access that must be done. I don’t think that is an academic discussion but very much appreciate it here. Thanks for sharing all the resources.
16. Line 791 Chek blue and NIR, is it correct?
17. Line 93-94 Could not understand.
18. The added para, in conclusion, does not fit very well. Only the last paragraph is the actual conclusion to take from this reading. Need to revise the Conclusion aligning with each objection in the intro.

Replies are good but they must be well incorporated in the manuscript so that no readers should raise the same question as the reviewer did. All the very best.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The authors have added extra cases and made specific objectives. However, the LC mapping with such 3 cm GSD is very dependent on given training data, and getting such accuracy up to 0.96 is highly doubtful. The algorithms are well known and they do not need much experimentation but the implication here is not useful. VHR imagery is good for field verification and 3D models, not LC maps unless there are recurring patterns such as trees and buildings, roads and rivers, etc. Here the class with meadow, tree, agriculture, etc has minor differences and can vary each season and condition. 

Reviewer 2 Report

no particular comments.

Reviewer 3 Report

„UAV-Object Based Land Cover Classification Using Machine Learning Algorithms A Case Study on Mt. Stara Planina” is a well-written, interesting article with a suitable structure and proportions. Authors set three research goals: to upgrade and complement previously published research by De Luca et al. (2019), to evaluate the possibility of using UVA technology for land cover maps creation and to creating a set of guidelines on the use of remote sensing tools to obtain data for research in the field of landscape ecology.

The applied methodology (unmanned aerial vehicle (UAV) for (VHR) imagery acquisition, Support Vector Machine (SVM), Random Forest (RF), and k-nearest neighbours (k-NN) for land cover classification) is adequate and presented understandably. The results are elaborated in detail and in line with the aim of the paper. I find the paper acceptable for publication with minor revision.

Some remarks & questions:

I suggest remodelling research objectives and resigning from formulating the first goal as: “upgrade and complement previously published research by De Luca et al. (2019)” or take a different approach to this goal. The conducted research aims to develop a method of obtaining land cover data using UAVs and methods of supervised object classification. It seems that this is a broader approach than upgrade and complement of previously published research. Research by De Luca et al. (2019) is the basis for discussion. I believe that the part of the manuscript containing the discussion should be completed. Because presented research and studies carried out by De Luca et al. (2019) are similar, it is necessary to bring more emphasis to what was brought by the application of an additional classification algorithm (k-NN), testing various parameters for RF and enlarging the area for which the data was obtained.

Page 1 row19 „Our research outstands previous studies” I would rather say they complement, extend or makes it more detailed.

Page 2 rows 52-56, are such studies available? Examples would be very helpful.

Page 4 I suggest changing the of Figure 1. composition. Reducing the map showing the area of Serbia or removing it and leaving only a small map with the location of the study areas will allow for the enlargement of very interesting detailed maps of the studied areas. The general location of the research area is well described.

Page 4 “2.2. Data Acquisition and Processing Outputs” what was the time of data acquisition? were all missions performed in the same season of the year, with similar weather conditions? If not, could it somehow affect the quality of the data? Should it not be treated as a limitation of this method, which was not included in the study?

Reviewer 4 Report

The Authors address an investigation on the use of object-based classification, coupled with different machine learning algorithms, to detect different land cover categories from very high resolution imagery, acquired by an unmanned aerial vehicle (UAV). Based on different input variable such as RGB bands, digital surface model (DSM) and normalized difference vegetation index (NDVI) they performed image segmentation and classification using the open-source software ORFEO Toolbox (OTB) and QGIS. The proposed methodology is tested and applied in three different study sites from Mt. Stara Planina, in the Eastern Serbia, with the specific aim of evaluating classification performance, providing a workflow to obtain detailed land cover maps valuable for ecological research.

I have read this paper with interest and despite it is focused on an interesting topic, I’ve highlighted several major drawbacks that limit its relevance to the international readership. Among all the considerations, the most important is that I don’t find any novelty in this work nor it added valuable informations to the already present scientific literature. This is the purpose of a contribution to a journal of proven reputation. The overall paper structure is not so well organized. Methodological section presents critical issues that should be investigated. For example, one of the key point of the research was to expand and implement a prior study methodologies, but from what has been described, it seems that it only applied to a different land cover, without making substantial changes to what had already been presented; an analysis is made per se and there is no attempt to open a broad scientific debate.

Indeed, the results may serve as a useful tool to increase landscape knowledge and management practices in the area, however, they are barely useful for an international audience. I really appreciated the willingness to share research data and the idea of disseminating knowledge in the form of videos, but this cannot be configured as an innovative aspect or even an acquired result. It is now common practice for most of the published works in the remote sensing sector to share source data, developed computational codes and results. Replicability must exist regardless of this. Beside this general concern, I have also several specific comments and suggestions which I summarize below.

Abstract

Line 19-20: what the Authors mean with this sentence? Thus, it seems that land cover classification studies or comparisons in heterogeneous landscapes have never been carried out, which is absolutely not the case. Please clarify.

I suggest that this should be rewritten considering the possibility of revising the objectives of the work. Also, as long as all the rest of the text, I would also suggest that Authors to perform a deep editing of English language, period are often unclear.

Introduction

Line 57-59: pay attention to the historical information that is given. In 1985 it was the Council of the European Communities that launched the COoRdination of INformation on the Environment program, when the Copernicus program did not yet exist (it was born in 2001). The production of Corine Land Cover data continues to be ensured today within the Land thematic area of the Copernicus program.

Line 64-66: this sentence has no sense.

Line 67-69: what do the authors mean? I think they are saying that such high resolution cannot be reach by free satellite imagery because commercial satellites (such as Worldview-3) have these capabilities. Written like this is not clear.

The general problem of the introduction is that it is limited to giving generic information without ever entering into the merits of the state of the art of the use of GEOBIA in the Land cover classification and clearly defining which gap is planned to fill with this work. At least add some more reference.

Materials and methods

Figure 2: what do the different colours of the arrows mean? is there a particular meaning? if so, it must be explained (could it be process and output?).

Line 198-199: how were the GCPs acquired? with what accuracy? I believe it is relevant information that could be provided as well as information about the NDVI camera.

Paragraphs 2.4 and 2.5: these sections have redundant descriptive information in the text which is well known to the Journal readers without giving a true explanation of why they chose to use these machine learning algorithms. A reader does not know the reasons in order to choose one instead of another so, in my opinion this election must be properly justified. On the basis of what were the training and validation points chosen? and how were the different classes assigned? (of which a complete description is missing?) Have ground truths been acquired or only through on-screen photointerpretation?

The most important thing is that the authors report that for the classification step they modified only the RF parameters to obtain the best result, leaving the SVM and k-NN classifier unchanged. What does it mean? By doing so, it is not possible to compare the performances of the different classifiers as the results are not optimal for all a priori. Consequently, results of the test procedures are likely wrong and cannot be trusted. If you want to compare these results then optimal parameters setting for all classifiers is likely the first option here.

Results and discussion

Line 369-373: once again the land cover classes are reported without explaining what they are. For example what is the difference between trees and forest patches?

Figures 4, 5 and 6: if you want to highlight the different results obtained, I would have opted to show the maps, even of a single site, obtained using the different classifiers and not only by RF. But this is a matter of preference.

Result and Discussion section appears very confusing. It is not easy to identify all the steps of the proposed image processing chain. The reader is lost in the results data that often do not find a correct explanation in the text of the discussion with a lot of redundant informations. At its actual state, discussion is not properly a discussion section that, beyond reporting the results, substantially repeats the difficulties encountered, already reported in the introduction and results section. In this case a comparison between the methods used with an abundant and relevant international literatures is missing. To improve reading some parts could also be merged.

My overall recommendation is that the current manuscript is a draft which needs a throughout revision in order to be accepted as a scientific article.

Other general issues:

Supplementary materials

I find it useless to compare a very high resolution orthomosaic with a CORINE map. To show what? I think it is evident that the comparison is impossible given the characteristics of the two datasets. Rather it would make more sense to highlight the potential of the UAV-derived imagery against the more current, freely and widely used high resolution layers of the Copernicus Land Monitoring program (https://land.copernicus.eu/pan-european/high-resolution-layers).

I I would also suggest that Authors perform a deep editing of English language, period are often unclear.

The paper has potential but there are some critical issues that should be addressed before the manuscript is ready to be published.

In conclusion, I suggest that the manuscript should be rejected in its current form.