FlexibleNet: A New Lightweight Convolutional Neural Network Model for Estimating Carbon Sequestration Qualitatively Using Remote Sensing
Round 1
Reviewer 1 Report
It is urgent and very important topic to reduce computational time with keeping or improving accuracy. And I am very happy to review your paper and have some comments and suggestion as follow:
(1) It looks this study is only one case and there are so many kind of forest such as tropical, boreal, etc. So, I wonder how robust or generalizing your proposed method is.
(2) Is formulation (5) corrupted? Can you please check it and let me know how to interpret three lines formulation?
(3) Figure 7 is too hard to read it. Can you please revise it?
(4) In line 308, how do you select learning rate 0.001? Do you have any result with different rate case?
Author Response
Dear Colleague,
Thank you for reviewing my research topic. Your comments very constructive and helpful to improve my paper.
Please find attached the my responses to your comments and remarks.
Sincerely
M. Awad
Author Response File: 
Author Response.docx
Reviewer 2 Report
In this paper, author propose a new lightweight CNN network named ElasticNet.
The Elastic Deep Neural Network is a general framework that systematically “elastifies” an arbitrary network to provide novel trade-offs between execution time and accuracy.
This new Elastic Network was tested on estimating sequestered carbon qualitatively in the aboveground forest biomass from Sentinel-2 images.
Public Comment section:
1. The title is not suitable: why the author use a name “ElasticNet” so commonly with “linear regression ElasticNet” in machine learning (for example, see sklearn.linear_model.ElasticNet, https://scikit-learn.org/stable/modules/classes.html#module-sklearn.linear_model). I recommend renaming the ElasticNet network.
2. In “2. Materials and Methods”, I think that author should specify in the text that the area study is located in Lebanon. (We see Lebanon in Figure 1. Area of study, but this isn’t mention in text).
3. To check: in line226, “Then a new activation function is used the Leaky Rectified Linear activation function (LReLU)”, remove new before activation function, because it’s not new function, and it’s known as LeakyReLU in many paper [35] and frameworks LeakyReLU — PyTorch 1.13 documentation, https://pytorch.org/docs/stable/generated/torch.nn.LeakyReLU.html
4. In equation (4), the author suppose the number of rows d equals to the number of columns w (d= w), then estimate n and the filter size. If d ≠ w, what we do?
5. In equation (5), the author suppose the number of filter then, he uses fsize, should use fnum to indicate the number of filter. Then some rules imposed to choose the filter numbers by a new variable m without explain why and how to know the limites of these numbers or variables.
6. The same remarks for equation (7), explain why and how to know the limites of theses layer depths.
7. Figure 7 is not clear; I recommend modifying it by high quality figure with details.
8. To do alignment in Table 2. and equations (8), (9) and (10)
9. Figure 10, 12 and 14 are not clear. I recommend modifying it by high quality figures with details: in the first line “Loss” for all nets ((a) ElasticNet (b) ResNet50 (c) EfficientNetB5 352 (d) MobilNetV3 Large) and “Accuracy” in the second line.
10. To do alignment in Figure 11, images and (a), (b), etc.
11. The question about data in-situ, if the author has measurements in field to validate the results and compare with the nets output ((a) ElasticNet (b) ResNet50 (c) EfficientNetB5 352 (d) MobilNetV3 Large).
12. Resume the main difference between this paper and papers in additional references.
Conclusion: The paper needs improvements AND MAJOR REVISION.
Additional references:
[1]. Y. Zhou, Y. Bai, S. S. Bhattacharyya and H. Huttunen, "Elastic Neural Networks for Classification," 2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), 2019, pp. 251-255, doi: 10.1109/AICAS.2019.8771475.
[2]. Y. Bai, S. S. Bhattacharyya, A. P. Happonen and H. Huttunen, "Elastic Neural Networks: A Scalable Framework for Embedded Computer Vision," 2018 26th European Signal Processing Conference (EUSIPCO), 2018, pp. 1472-1476, doi: 10.23919/EUSIPCO.2018.8553186.
[3]. Yu, D.; Xu, Q.; Guo, H.; Zhao, C.; Lin, Y.; Li, D. An Efficient and Lightweight Convolutional Neural Network for Remote Sensing Image Scene Classification. Sensors 2020, 20, 1999. https://doi.org/10.3390/s20071999
[4]. Chen, Y.; Chen, X.; Lin, J.; Pan, R.; Cao, T.; Cai, J.; Yu, D.; Cernava, T.; Zhang, X. DFCANet: A Novel Lightweight Convolutional Neural Network Model for Corn Disease Identification. Agriculture 2022, 12, 2047. https://doi.org/ 10.3390/agriculture12122047
Author Response
Dear Colleague,
Thank you for reviewing my research topic. Your comments were very constructive and helpful to improve my paper.
Please find attached my responses to your comments and remarks.
Sincerely
M. Awad
Author Response File: Author Response.docx
Reviewer 3 Report
The author proposed a ElasticNet model for carbon sequestration estimation, I have some questions:
1. The author should explain more details about why the proposed ElasticNet can achieve better performance than the other state-of-the-art in a lightweight architecture.
2. I suggest the author should list contributions of this paper in the introduction.
3. From line 358 to 362, the author placed constraints on running the proposed model because of lacking a large dataset. Please cite references to support this claim.
4. Please describe more about the importance of carbon sequestration estimation.
Author Response
Dear Colleague,
Thank you for reviewing my research topic. Your comments were very constructive and helpful to improve my paper.
Please find attached my responses to your comments and remarks.
Sincerely
M. Awad
Author Response File: Author Response.docx
Reviewer 4 Report
To overcome the time and computer resource consumption of CNN, the authors present a new Lightweight model called (ElasticNet). In my opinion, the manuscript’s contributions are that ElasticNet was better or comparable to the other lightweight or heavy CNN models concerning to the number of parameters and time requirements. However, some weaknesses should be addressed, especially the introduction and experiment.
Major issues:
1) The introduction is not detailed enough. It is unable to highlight the specific significance of this study. For example, the popular Lightweight model in recent years have not been given in detail. In addition, the authors should also introduce some traditional machine learning methods. I suggest that the authors add some literature analysis, such as
[1] Super-Resolution Mapping Based on Spatial-Spectral Correlation for Spectral Imagery [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(3): 2256-2268.
[2] Pansharpening by convolutional neural networks,” Remote Sens., vol. 8, pp. 594, Jul. 2016.
2) The paper structure is suggested to be modified. Data and methods should be presented separately in two sections. In addition, the current data introduction is simple. The authors should elaborate on why this area was chosen for research.
3) The experimental part needs further improvement. First, it is suggested that the authors compare the SOAT deep learning networks. In addition, the selected experimental data are few, and at least one real data should be added for verification.
Minor issues:
1) There are some grammatical errors in the article, which need further careful proofreading. It is suggested that the authors invite professional personnel to proofread.
2) It is suggested that the authors analyze the advantages and disadvantages of the proposed model and its prospects in the conclusion part of this paper.
Author Response
Dear Colleague,
Thank you for reviewing my research topic. Your comments were very constructive and helpful to improve my paper.
Please find attached my responses to your comments and remarks.
Sincerely
M. Awad
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
I understand there are no difference for height estimation among forest / tree types. However, there are very big difference about spatial density of tree based on forest type. It maybe impacted to height estimation beacuse sentinel-2 spatial resolution is only 10m. How do you tackle this tpic?
Author Response
I would like to thank the reviewer for the helpful remark that made my research work better.
Point 1: I understand there are no difference for height estimation among forest / tree types. However, there are very big difference about spatial density of tree based on forest type. It maybe impacted to height estimation beacuse sentinel-2 spatial resolution is only 10m. How do you tackle this tpic?
Response 1:
Dear Reviewer,
Thank you very much for raising essential ideas and questions that are helpful in improving my current research and future ones. I agree that the spatial density of trees based on forest type is a more complex issue to handle using moderate spatial resolution RS images such as Sentinel-2. However, in this research, and since the forests in the area of study are less complex than other forest areas like boreal or tropical forests, we created the forest canopy density using three vegetation indices Advanced Vegetation Index (AVI), Bare Soil Index (BI), Canopy shadow index (CSI), and Elevation layer. Further investigation in the future may include time series NDVI to separate deciduous forest trees from that evergreen which can further enhance the research work. As an example, Tropical deciduous forests are less dense compared to tropical evergreen forests. Tropical Evergreen Forests are found in regions with more than 200 cm of rainfall. Many other research alternatives can be applied in the future.
Reviewer 2 Report
Thanks for your modification.
No comments .
Author Response
Thank you
Reviewer 3 Report
The author have addressed all my questions, and I have no more other extended question.
Author Response
Thank you
Reviewer 4 Report
Thank the author for his reply. I have no other questions.
Author Response
Thank you
