Evaluating Tree Species Mapping: Probability Sampling Validation of Pure and Mixed Species Classes Using Convolutional Neural Networks and Sentinel-2 Time Series
, Susanne Karel 1, Markus Loew 1, Ursula Knieling 1, Kevin Kopecky 1, Christoph Bauerhansl 1, Ambros Berger 1, Stephan Graeber 1 and Lukas Winiwarter 2,3Round 1
Reviewer 1 Report (New Reviewer)
Comments and Suggestions for AuthorsDear Authors,
I bring some necessary considerations for the greatest impact of the article, as per the comments below and are also highlighted in the text of the attached article.
1. In the Introduction section, the authors discussed the representation of forests for ecological, economic and social reasons. As well as the potential and great advantages of using current technologies, such as the applicability of using remote sensing techniques via satellite and machine learning training.
However, I suggest reporting the motivation for this study, what are the problems in these regions? For example, are they issues of environmental degradation, deforestation in the face of climate and anthropogenic changes?
2. In table 1, the acronyms for the vegetation indices were highlighted. However, the full names that represent each acronym/index were not found in the text. I suggest entering the index names in full.
3. Still in table 1, the equations show their multispectral bands, which in this aspect I suggest presenting their respective wavelengths, considering that the opportunity to replicate the present study will mainly facilitate applicability to other satellites.
4. Immediately afterwards, lines 230 and 231, equations are presented, but they do not have numerical identification. Therefore, it is important to present the Equation number and call it in the text!
5. Regarding the use of satellite data, Sentinel-2, more details are needed on its characteristics. For example, what is the satellite's sensor was not highlighted? According to the platform's open data source, what is the access link? Therefore, I suggest further detailing and incorporating these characteristics into the article's material and methods.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report (New Reviewer)
Comments and Suggestions for AuthorsComments for author File: Comments.pdf
Author Response
Please see attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report (New Reviewer)
Comments and Suggestions for AuthorsDear Authors,
Thank you for taking the main suggestions into account. The article has improved significantly after these revisions.
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
Comments and Suggestions for AuthorsI reviewed the manuscript entitled "Unraveling Forest Diversity: Sentinel-2 Time Series Neural Networks for National-scale Tree Species Mapping with Pure and Mixed Classes, Validated via Probability Sampling" by Schadauer et al. The authors attempted to map tree species in Austria using a dense time series of Sentinel-2 and neural networks. The topic is interesting and the methodology is sound. The manuscript is well-written and almost all necessary explanations are provided. The authors have already investigated their objectives from different perspectives, making the study an exhaustive research in the remote sensing and forestry fields. Accordingly, I recommend a minor revision, and my comments are provided below:
1- The abstract is too long, and I suggest that the authors make it more concise by focusing on the main message of the study.
2- Line 109; please correct the citation of Delwart.
3- Table 2, Please add suitable references for the vegetation indices, and modify the caption to be more suitable.
4- What was the motivation/intention of using ResNet models, while there are many other deep neural network models?
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript titled "Unraveling Forest Diversity: Sentinel-2 Time Series Neural Networks for National-Scale Tree Species Mapping with Pure and Mixed Classes, Validated via Probability Sampling" employs extensive data for the classification of forest tree species. The authors provide detailed descriptions of the experiment's specific content and methods. After a thorough review of the manuscript, it is deemed necessary to make substantial revisions. I would like to provide you with the following feedback on the review:
1. Presently, the introduction lacks an in-depth discussion of research involving deep learning methods in tree species classification. It is suggested to augment this section with additional content.
2. Section 2.4 mentions the utilization of ALS LiDAR point cloud data to generate slope and aspect data but fails to specify the data's source and relevant parameters.
3. In Figure 2, the red markings indicating all Training areas are not prominently displayed and are even difficult to discern. It is recommended to emphasize these markings for better visibility.
4. The placement of the sentence at line 392 is deemed inappropriate.0
5. In Section 2.10, the author discusses the inherent difficulties in labeling mixed data and proposes generating new training data for mixed and sparse category tree species through data synthesis. Can the newly generated training data positively impact tree species classification? This requires quantitative experimental comparison.
6. In Section 2.12, the authors design a neural network, "MLP-ResNet," for tree species classification. The corresponding structure depicted in Figure 5 is rudimentary and fails to effectively convey the network architecture. Integration of Figures 5 and 6 is suggested to highlight improvements in the network and its input-output components.
7. In Section 2.13, the author suggests allocating 5% of the training data as a validation set to prevent overfitting during network training. Is this proportion too small for meaningful evaluation?
8. Section "2.14 Feature Impact - SHAP Analysis" seems misplaced and would be better suited for the third chapter.
9. The "Materials and Methods" section extensively details each data preprocessing step. I think the authors need to condense this section to eliminate unnecessary verbosity.
10. The placement of Line 802 in the text appears inappropriate and requires adjustment.
11. The analysis in the Results section remains superficial, necessitating a deeper examination incorporating accuracy metrics.
Overall, the manuscript requires a refined summary of proposed methods and a comprehensive exploration of innovations in forest tree species classification. Additionally, further substantiation of the designed neural network's effectiveness is necessary, possibly through integration with more advanced techniques. Furthermore, there are instances of unclear or inaccurate language expressions throughout the manuscript.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe paper describes classification of forest stands deploying Sentinel 2 imagery and NFI data. The work and analysis itself is professionally done based on dense phenology series. The NFI data and manually delineated data, were used for validation. In NFI data only 55 percent accuracy which is interesting results and it is worth to publish such results in order to give information to the other researchers.
The spatial separation of data showed higher accuracies in lower and higher elevation levels, but it should be explained or showed the distribution of the species in these parts.
My overall conclusion is that it is nice work, and worth of publishing but some substantial changes should be done to the text itself
- description of NN architecture is unclear, did you develop some modification to the existing model or new model? Please rememeber that no all the readers are familiar with the tiny details and some broader introduction or presentation of what you have done in relation to NN modification would be appreciated.
- NN were trained 9 epochs, 128 batch and LR 0.00025. WHy exactly this settings, sometimes changing the setttings may provide different results. Did you try more settings or only these.
- the utilization of semivariograms is not clear to me at all. Would it be possible to again clearly explain why and what is the purpose?
- NN execution and running itself is not explained in detail. Did you use some existing scripts, did you develop your own application?
- In general the paper is very hard to read, tables and charts are super abundant probably not necessary many of them. Please reformulate the text in a simpler way following one story with clear methods, objectives and results. All of it is there now but in very confusing manner.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThank you my comments were responded and addressed. I do not have any further comments only minor, the new version of paper is not clear what remains what is accepted. If possible send please the final version for final check
