Gradient Boosting Machine and Object-Based CNN for Land Cover Classification

Round 1

Reviewer 1 Report

Accept in present form.

Author Response

Dear Reviewer 1,

We sincerely appreciate your positive comments and valuable suggestions for the improvement of the manuscript.

Many thanks to the Editor and the reviewers for your time and effort!

Wish you all the best,

Kind regards,

Van

Author Response File: Author Response.docx

Reviewer 2 Report

The authors significantly improved the quality of the paper by addressing most of the comments.

Your introduction section was a little incomplete where the motivation of your methodology related to object-based CNN for classification was not well established. In the introduction section, can you add one more paragraph for the High-Resolution Remote Sensing Images application using the recently established state of the arts object-based CNN deep learning technique where they utilized optimal band combinations (e.g. three-band combinations) and exhibited significant accuracy? They successfully developed an automatic extraction framework for remote sensing applications from high spatial resolution optical images using CNN architecture in a large-scale application based on multispectral band combinations. I provided the latest advanced research work that highlighted the importance of the band combinations in the use of multispectral datasets on model classification accuracy for remote sensing applications. Please go through those studies carefully. Please introduce these latest advanced research works, and their potential impact, and their limitation in terms of algorithms. Then you introduce your work with superiority along with proper justification. It would help the manuscript tremendously if the state-of-the-art was more streamlined in the introduction section. The authors should explain this aspect in the introduction section.

 Below studies that will be fit in your review paper:

Li, Yuqi, et al. "Optimized multi-spectral filter array based imaging of natural scenes." Sensors 18.4 (2018): 1172.

Ehsan et al. 2020 “Understanding the Effects of Optimal Combination of Spectral Bands on Deep Learning Model Predictions: A Case Study Based on Permafrost Tundra Landform Mapping Using High Resolution Multispectral Satellite Imagery. J. Imaging 2020, 6, 97.”

Abdalla, Alwaseela, et al. "Color Calibration of Proximal Sensing RGB Images of Oilseed Rape Canopy via Deep Learning Combined with K-Means Algorithm." Remote Sensing 11.24 (2019): 3001.

Park, Ji Hyun, et al. "RGB Image Prioritization Using Convolutional Neural Network on a Microprocessor for Nanosatellites." Remote Sensing 12.23 (2020): 3941.

Author Response

Dear Editors and Reviewers,

We sincerely appreciate your positive comments and valuable suggestions for the improvement of the manuscript. We have studied the comments carefully and made corrections according to these helpful suggestions and revised the manuscript accordingly. Below are point-to-point responses to address the comments/suggestions from the reviewers.

I hope that all the responses could meet the reviewer's requirements.

Many thanks to the Editor and the reviewers for your time and effort!

 

Reviewer 1:

Many thanks for your interest in our work.

 

Reviewer 2:

The authors significantly improved the quality of the paper by addressing most of the comments.

Your introduction section was a little incomplete where the motivation of your methodology related to object-based CNN for classification was not well established. In the introduction section, can you add one more paragraph for the High-Resolution Remote Sensing Images application using the recently established state of the arts object-based CNN deep learning technique where they utilized optimal band combinations (e.g. three-band combinations) and exhibited significant accuracy? They successfully developed an automatic extraction framework for remote sensing applications from high spatial resolution optical images using CNN architecture in a large-scale application based on multispectral band combinations. I provided the latest advanced research work that highlighted the importance of the band combinations in the use of multispectral datasets on model classification accuracy for remote sensing applications. Please go through those studies carefully. Please introduce these latest advanced research works, and their potential impact, and their limitation in terms of algorithms. Then you introduce your work with superiority along with proper justification. It would help the manuscript tremendously if the state-of-the-art was more streamlined in the introduction section. The authors should explain this aspect in the introduction section.

Below studies that will be fit in your review paper:

Li, Yuqi, et al. "Optimized multi-spectral filter array based imaging of natural scenes." Sensors 18.4 (2018): 1172.

Ehsan et al. 2020 “Understanding the Effects of Optimal Combination of Spectral Bands on Deep Learning Model Predictions: A Case Study Based on Permafrost Tundra Landform Mapping Using High Resolution Multispectral Satellite Imagery. J. Imaging 2020, 6, 97.”

Abdalla, Alwaseela, et al. "Color Calibration of Proximal Sensing RGB Images of Oilseed Rape Canopy via Deep Learning Combined with K-Means Algorithm." Remote Sensing 11.24 (2019): 3001.

Park, Ji Hyun, et al. "RGB Image Prioritization Using Convolutional Neural Network on a Microprocessor for Nanosatellites." Remote Sensing 12.23 (2020): 3941.

• Responses to reviewer: Many thanks for your comment. We have revised the manuscript accordingly.

 

Author Response File: Author Response.docx

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

Some comments formulated during my review are presented below. These are as follows:

-Unfortunately, the entire paper is inconsistent and difficult to follow in its present shape, e.g. what is the structure of the paper?
-What is the novelty presented in the paper? 
-Unfortunately, the reviewer see very little scientific novelty in this paper. 
-What are the parameters and their justification for the filters? 
-Sensitivity to parameters is not discussed.
-English should be carefully checked

The defects found in the article and presented above represent only some of its shortcomings. The paper has much morer. 

Reviewer 2 Report

The classification problem of a land image to 6 region types is examined in this paper.  Gradient boosting algorithms are examined and a neural network has been trained in Tensorflow format. An overall accuracy close to 90% has been achieved.

Although this paper is written in a comprehensive way my concerns are the following:

a) No theoretical background is presented justifying why the gradient boosting is better than other classification algorithms

b) although numerous references are listed, none of them has been used for comparison. The comparison between gradient boosting approaches (as well as SVM, RF) is useful but what about other approaches presented in the literature?

c) the authors have examined the NN architecture with gradient boosting they propose in the tensorflow environment but they have not implemented an integrated tool recognizing and displaying the results.

All the above, make the scientific contribution of this paper, weak. It would be fine for a conference paper but a scientific journal would need a extended development background and description.

I am not sure if Fig. 3 has the appropriate way to display the variation differences

Table in page 8 has wrong numbering

Please also correct:

line 97 (...)

line 107, do you mean 54,234 ??? Please use "," for thousands and "." for decimal points. In the Results section all the numbers should change to this format

 

Reviewer 3 Report

Please see the attached document.

Comments for author File: Comments.pdf

Reviewer 4 Report

The authors present ML/DL-based approach for land cover classification. I have a couple of concerns with the presented introduction/ methodology that should be addressed.  These concerns are a result of a general imbalance between introduction/context information and methodological/analysis information. My major comments and questions are as follows:

• The introduction section is not clear.  Recently the state of the arts deep learning technique MASK RCNN was established in remote sensing mapping and land cover applications. It would help the manuscript tremendously if the state-of-the-art was more streamlined in the introduction section. Please introduce MASK RCNN based semantic segmentation example using high-resolution satellite imagery application. This would make it much easier for the reader to understand, how this study fits into the research context. What is the uniqueness of the proposed algorithm and its potential impacts, over other recently established states of the art Resnet based Mask RCNN semantic segmentation methods for Remote Sensing application (Bhuiyan, et al.2020; Li, et al. 2021, Mahmoud, et al 2020, Zhao et al 2018, and so on)? They successfully developed an automatic extraction framework for remote sensing applications from high spatial resolution optical images using CNN architecture in a large-scale application. Please introduce these latest advanced research works, and their potential impact, and their limitation in terms of algorithms. Then introduce your proposed methodology and the novelty of your proposed methodology. The authors should explain this aspect in the introduction section.

1. Li, Y.; Xu, W.; Chen, H.; Jiang, J.; Li, X. A Novel Framework Based on Mask R-CNN and Histogram Thresholding for Scalable Segmentation of New and Old Rural Buildings. Remote Sens. 2021, 13, 1070.
2. Bhuiyan et al. 2020 “Use of Very High Spatial Resolution Commercial Satellite Imagery and Deep Learning to Automatically Map Ice-Wedge Polygons across Tundra Vegetation Types.” J. Imaging 2020, 6, 137.
3. Mahmoud, A., et al. "Object Detection Using Adaptive Mask RCNN in Optical Remote Sensing Images." Int. J. Intell. Eng. Syst 13 (2020): 65-76.
4. Zhao, Kang, et al. "Building extraction from satellite images using mask R-CNN with building boundary regularization." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2018.

• Can you please provide a table for validation/training/testing samples along with other information such as sample number, patch, sample source, training and validation sites, repository, etc?

• Can you provide a high impactful schematic diagram to understand the proposed research framework where the big impact of the results can be presented?

• You utilized multispectral bands (eg RGB band) in our algorithm. But how did you utilize the optimal band combinations (eg RGB band) to explore the geographic transferability? Also, Data fusion, the process of combining multispectral (MS) and high-resolution panchromatic (PAN) images with complementary characteristics often serves as an integral component of remote sensing mapping workflows. The fusion process generates spectral and spatial artifacts that affect the classification accuracies of subsequent automated image analysis algorithms, such as deep learning (DL) convolutional neural nets (CNN). Recent studies suggested that the DL-based fusion algorithms that preserve the spatial character of original PAN imagery favor the DLCNN model performances in order to enable an accurate mapping effort in remote sensing applications. You did not discuss anything in detail about the fusion algorithm. You should explain these issues?

1. Li, Yuqi, et al. "Optimized multi-spectral filter array based imaging of natural scenes." Sensors 18.4 (2018): 1172.
2. Ehsan et al. 2020 “Understanding the Effects of Optimal Combination of Spectral Bands on Deep Learning Model Predictions: A Case Study Based on Permafrost Tundra Landform Mapping Using High Resolution Multispectral Satellite Imagery. J. Imaging 2020, 6, 97.”
3. Abdalla, Alwaseela, et al. "Color Calibration of Proximal Sensing RGB Images of Oilseed Rape Canopy via Deep Learning Combined with K-Means Algorithm." Remote Sensing 11.24 (2019): 3001.
4. Park, Ji Hyun, et al. "RGB Image Prioritization Using Convolutional Neural Network on a Microprocessor for Nanosatellites." Remote Sensing 12.23 (2020): 3941.
5. Yang, Ronglu, et al. "Representative band selection for hyperspectral image classification." Journal of Visual Communication and Image Representation 48 (2017): 396-403

• There is only relatively little information about the neural network architecture and, more importantly, about the choice of relevant settings that are missing. The architecture should include sub-networks such as Region Proposal Network (RPN), a region of interest (ROI), Conv, class box, etc. Please explain in detail.

 

• How did you utilize the transfer learning strategy? Did you use trained weight? It is unclear what independent evaluation is meant here. You should present the transferability of the Deep Learning Model in your application.
• Can you provide few feature maps during optimization?
• How was the ground reference built? Only on a computer screen, or with the help of some field data and field knowledge? Please explain.
• How do the authors come up with the current optimized DL structure? Can you show some results for model optimization? How did you create DL optimized model without showing any fundamental results? For DL optimization you need to optimize your model for different additional losses such as Smooth-L1 loss; bounding box loss; classifier loss; binary cross-entropy loss; RPN bounding box loss; RPN classifier loss.`
• How did you choose essential tuned hyperparameters (e.g. number of hidden nodes, learning rate, etc.)? You should provide a table for hyperparameters settings. Can you explain more about overfitting issues?
• In your result section, where are the classification accuracy results in terms of mean Average Precision?

• In the discussion section, you should discuss more your results vs previous research-based semantic segmentation.