Next Article in Journal
On-Board Multi-Class Geospatial Object Detection Based on Convolutional Neural Network for High Resolution Remote Sensing Images
Next Article in Special Issue
Rapid Large-Scale Wetland Inventory Update Using Multi-Source Remote Sensing
Previous Article in Journal
Quantitative Precipitation Estimation in the Tianshan Mountains Based on Machine Learning
Previous Article in Special Issue
Integrating UAV-Derived Information and WorldView-3 Imagery for Mapping Wetland Plants in the Old Woman Creek Estuary, USA
 
 
Article
Peer-Review Record

Spatial Analysis of Intra-Annual Reed Ecosystem Dynamics at Lake Neusiedl Using RGB Drone Imagery and Deep Learning

Remote Sens. 2023, 15(16), 3961; https://doi.org/10.3390/rs15163961
by Claudia Buchsteiner 1,*, Pamela Alessandra Baur 1,2 and Stephan Glatzel 1,2
Reviewer 1:
Reviewer 2: Anonymous
Reviewer 3:
Reviewer 4:
Remote Sens. 2023, 15(16), 3961; https://doi.org/10.3390/rs15163961
Submission received: 6 June 2023 / Revised: 5 August 2023 / Accepted: 7 August 2023 / Published: 10 August 2023

Round 1

Reviewer 1 Report

Thank you for the opportunity to review this paper.  It's really well written and covers a worthwhile topic nicely.

The phenocam and drone phenology are offset from each other.  With the phenocam phenology greening up sooner than the drone.  I suspect that this may be due to view angle effects (phenocam is side looking,  whereas drone is nadir looking).  It would be worth commenting on this in the discussion.

As part of future work (not needed for the manuscript) it would be intriguing to see the Sentinel 2 greenness time series for the same location (to see how well the phenology matches up)

The RMSE results could use a bit more explanation.  If I'm reading this correctly, the RMSE for the control sites where you were able to attach markers above the canopy were ~2pixels, whereas the RMSE for the other sites (more evenly distributed, ,but more spatially ambiguous) was a lot higher.  If my interpretation is correct, it'd be worth detailing that more, as useful guidance for people trying to apply similar methods in other locations.

 

 

Author Response

Please see the attachment for the point-by-point responses.

Author Response File: Author Response.pdf

Reviewer 2 Report

The major flaw or to a certain extent fatal problem is the insufficient time coverage for the data they collected and used in this study. Both in the introduction and in the conclusion, the authors explicitly confirm the standpoint of the whole research is the impact of the changing climate on this regional area, while meanwhile only less than a year data is used and from which the conclusions are based. I suggest either changing the tone throughout the manuscript to a short-term, regional study and not claiming on climate change (which could be discussed in the discussion section absolutely), or more data from multiple years would be collected and studied so as to answer the question of the impact of climate change on the Lake Neusiedl. 

However, if the authors revise the manuscript either way, the whole story would be radically different. Therefore, I'm regretting to but have to give a rejection recommendation.  On the other hand, I would love to encourage the authors to resubmit the manuscript whenever they finish revising the current version. 

Author Response

Please see the attachment for the point-by-point responses.

Author Response File: Author Response.pdf

Reviewer 3 Report

The aim of the research is to monitor the reed belt of Lake Neusield using RGB aerial photographs in a selected year / vegetation period. The methodology and processing workflow are well described, the presentation of the results is aduquate, varied and informative. The discussion of the results is well detailed and most of the questions raised by the results are answered and compared with other research.

The methodology used for classifying the ortomosaics into three land cover classes is nowadays "fashionable", and can be implemented in a commercial GIS environment using predefined geoalgorithms. However it was carried out correctly, some questions remain open:

- Why is it neccessary to apply a deep learning model to classify three well separated land cover types? You mentioned a few references [30,31] that justify your choice, but do you have your own experience / comparisions with traditional approaches such as maximum likelihood or simple thresholding based on RGB indices?

- What is the resolution of the othophoto (2019-06-26) used to derive the 6 check points? Did this influence the RMSE values obtained? Same question regarding the linear distribution of GCPs.

- Why did not you use (get) the same GSD values and area covered for all campaigns? Did you create a new mission plan each time?

- You mentioned you had to train the model for each orthomosaic. Did you use a generic model and compare the performance metrics with the single ones? Is it really worth the greater investment of time?

- Did you use the validation data (20% of the chip dataset) to evaluate the training (on the "ground truth area")? If so, how did you validate the classification on the entire orthophoto? Have you quantified the classification errors? E.g. Does Fig.7 display a ground truth area or is it a random subset of the study area?

- You discuss possible causes of missclassification (Line 385-396). Do you find any suggestion in the literature to overcome these problems?

- How representative is the selected study area for the lake and reed belt as a whole? Can the findings be generalized to the entire area?

 

Some extra comments:

- Are there any research from the Hungarian side of the area? The proportion of reed belt is even higher there compared to the open water areas.

- The research covers a vegetation period with drought. Can the observed temporal and spatial trends also be valid in normal or wet years?

- Line 86. FertÅ‘ (instead of Fertö)

Author Response

Please see the attachment for the point-by-point responses.

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper “Spatio-Temporal Analysis in the Reed Ecosystem of Lake Neusiedl using RGB Drone Imagery and Deep Learning” presents a study on the spatio-temporal variations within the reed ecosystem of Lake Neusiedl in Austria using high-resolution RGB drone imagery and deep learning techniques. The research focuses on the effects of intensive drought conditions on the reed belt, which covers half the size of the lake. The study analyzes and detects changes in land cover and vegetation development over the growing season of 2021. The authors developed a workflow for land cover classification using deep learning image segmentation and compared the results with phenological analysis from aerial imagery and a Phenocam. The study reveals significant dynamics in the reed ecosystem of Lake Neusiedl, including changes in water surface area, sediment area, and vegetation cover.

Overall, while the paper presents interesting findings and demonstrates the potential of using high-resolution drone imagery and deep learning techniques for analyzing reed ecosystems, it is important to acknowledge these weaknesses and consider them when interpreting the results. Future research could address these limitations to further improve the accuracy and applicability of the methodology.

-The study focuses on a specific reed ecosystem in Lake Neusiedl, Austria, under intensive drought conditions. The findings may not be directly applicable to other ecosystems or regions with different environmental factors. The authors should justify this Limited generalizability.

-The paper provides detailed results but lacks a thorough discussion of the implications and potential applications of the findings. Further insights into the ecological consequences of the observed changes and their significance would enhance the overall impact of the study.

- The study area was subject to access restrictions due to its location within a protected area, which affected the distribution of ground control points (GCPs). Ideally, GCPs should be evenly distributed throughout the study area, but this was not possible in this case. As a result, additional control points had to be derived from an orthophoto, leading to higher uncertainty values in the georeferencing process.

- The authors initially attempted to train a single model using training data from all flight campaigns to classify the orthomosaics. However, due to the significant variations in surface conditions and lighting conditions between flights, this approach was unsuccessful. Instead, individual models were trained for each orthomosaic, which achieved satisfactory results but required additional effort.

-The classification results showed challenges in accurately distinguishing between water and sediment, especially when water levels were shallow. The exact boundary between these classes was difficult to predict, leading to potential misclassifications in areas with varying water depth.

-The model performed poorly in areas with shading and sparse reed vegetation. Shaded areas were often reported as misclassified, despite potentially being more accurate when compared directly to the imagery. This suggests that more care is needed in the preparation of fully labeled ground truth data. Similarly, sparse vegetation areas posed challenges in generating training data, which affected the model's performance in these regions.

- While the paper provides valuable insights into the spatio-temporal changes in the reed ecosystem of Lake Neusiedl, it does not extensively compare its findings with existing studies. This limits the ability to contextualize the results within the broader scientific knowledge and understand how they contribute to the existing body of research on reed ecosystems.

-While the discussion section acknowledges the limitations of the methodology, it could delve deeper into the factors that may have contributed to misclassifications. Exploring the challenges and potential sources of error would provide insights for future research and help improve the accuracy of the classification model.

-The discussion section would benefit from a more comprehensive comparison of the findings with previous studies on reed ecosystems. This would help position the research within the broader scientific knowledge and highlight the novelty or consistency of the results.

 

-The conclusion could be strengthened by providing more explicit recommendations for the practical applications of the research findings in reed ecosystem management and conservation. Identifying potential management strategies or suggesting ways in which the research can contribute to policy decisions would enhance the relevance and impact of the study.

Author Response

Please see the attachment for the point-by-point responses.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The authors have made the suggested modifications. I accept their answers to my questions.

Author Response

Dear reviewer,

thank you very much for your reply. As a further minor revision, on the editor's advice, we have changed the title as follows: "Spatial Analysis of Intra-annual Reed Ecosystem Dynamics at Lake Neusiedl using RGB Drone Imagery and Deep Learning"

Kind regards,
the authors

Reviewer 4 Report

Accept in present form

Author Response

Dear reviewer,

thank you very much for your reply. As a further minor revision, on the editor's advice, we have changed the title as follows: "Spatial Analysis of Intra-annual Reed Ecosystem Dynamics at Lake Neusiedl using RGB Drone Imagery and Deep Learning"

Kind regards,
the authors

Back to TopTop