Detection of Bottle Marine Debris Using Unmanned Aerial Vehicles and Machine Learning Techniques

Round 1

Reviewer 1 Report

Environmental pollution is an important global problem. Typically, it is caused by plastic or metal pollutants. There are simple solutions for preventing pollutants that require two steps: the first one is sensing and the second one is acting. For broad scanning, it also requires a planning stage. In this study, a sensing system was proposed that covers the first step used for detecting bottles on the marine side with the help of UAVs and AI together.

Thank you for submitting your paper. The work done here draws attention to a significant subject. The article's aim is of interest to the community. The experiments and sample preparation were well-designed and tested. However, the current manuscript has several weaknesses that must be strengthened to obtain a documentary result that is equal to the value of the publication. In conclusion, a revision is needed before considering the publication of the submitted manuscript.

The following are a few minor/major comments that should be resolved:

Point 1. As a suggestion, using a custom drawn Figure will be better to explain FoV, imaging area, and UAV altitude relationship for readers.

Point 2. Please explain, are the captured images captured by the UAV while it is stationary or in the locomotion stage.

Point 3. If possible, please include image filtering steps using additional images in Figure 3.

Point 4. Please explain why the YOLO v2 model is chosen. Please express its importance due to Real-Time object detection.

Point 5. Please explain how the authors intersect the GPS coordinates with imaging Camera positions. Is the GPS module centered with the same center of gravity as the camera? If so, please discuss it.

Point 6. Please mention the imaging angle of the camera. Is there any effect on the shapes in the captured images and their coordinates?

Point 7. Please express the altitude of the UAV while capturing the images.  

Point 8. Is the sensor of the camera CMOS/CCD, is it a rolling shutter camera or a global shutter, please explain this parameter.

Point 9. Please explain if there is any blurring effect on images due to the real-time imaging (while the UAV is moving) and the camera specifications. 

Point 10. Please check the references such as Ref 1.

Point 11. Please check the title "6. Patents”.

 

Author Response

Response to Reviewer#1 Comments of drones-2065826

We sincerely thank Reviewer#1 for his/her valuable suggestions and comments.  We highly appreciate the in-depth review and constructive suggestions.  The manuscript has been revised according to the reviewers’ comments.  The revised contents are highlighted in blue in the manuscript (as attached pdf file).  The point-to-point responses to the reviewer’s comments are given below.

General comments

Environmental pollution is an important global problem. Typically, it is caused by plastic or metal pollutants. There are simple solutions for preventing pollutants that require two steps: the first one is sensing and the second one is acting. For broad scanning, it also requires a planning stage. In this study, a sensing system was proposed that covers the first step used for detecting bottles on the marine side with the help of UAVs and AI together.  Thank you for submitting your paper. The work done here draws attention to a significant subject. The article's aim is of interest to the community. The experiments and sample preparation were well-designed and tested. However, the current manuscript has several weaknesses that must be strengthened to obtain a documentary result that is equal to the value of the publication. In conclusion, a revision is needed before considering the publication of the submitted manuscript.

We are thankful for the reviewers’ evaluation of the manuscript.  The manuscript has been revised according to the reviewers’ comments.  We believe this manuscript is ready for publication.

 

Point 1: As a suggestion, using a custom drawn Figure will be better to explain FoV, imaging area, and UAV altitude relationship for readers.

Response 1:  We have added a new figure (Figure 2) to indicate the relationship between the resolution and the UAV altitude.  The corresponding descriptions for Figure 2 are given below:

The field of view is based on image size and resolution. The resolution is a function of the UAV altitude (FL), sensor width (SW), focal length (FL), and image width (IW)

 (Please see Lines 129- 136)

 

Point 2: Please explain, are the captured images captured by the UAV while it is stationary or in the locomotion stage.

 

Response 2:  The UAV was operated in a hovering mode (approximately stationary) that can minimize the motion of the camera and reduce the blurring effects on the captured images.  However, we noted to readers that the blurring effects might be an important issue for real-time surveys.  (Please see Lines 123 - 125)

 

Point 3: If possible, please include image filtering steps using additional images in Figure 3.

 

Response 3: Thanks for the suggestion. We have revised the image by adding sample images in the filtering steps as shown in Figure 4. (Please also see Lines 213-215).

 

Point 4: Please explain why the YOLO v2 model is chosen. Please express its importance due to Real-Time object detection.

 

Response 4: We thank the reviewer’s comment. The YOLO v2 was chosen because this model has been proven a useful tool to identify marine debris with satisfactory accuracy and computing speed.  YOLO v2 has been applied in many studies in recent years [1-4].   In addition, the YOLO v2 gets a score of 78.6 on mean average precision on testing the well-known dataset of Pascal VOC 2007, which is the best one among many other models [5].  Note that other new models (e.g., YOLO v7) have recently been developed; future works on testing how these new models improve the performance of identifying marine debris could contribute to the detection of marine debris.  (Please see lines 141-148)

 

 

Point 5: Please explain how the authors intersect the GPS coordinates with imaging Camera positions. Is the GPS module centered with the same center of gravity as the camera? If so, please discuss it.

 

Response 5:  The UAVs were operated in a hovering mode with a camera heading 90 degrees toward the ground.  Therefore, the GPS module is approximately centered with the same center of gravity as the camera.  (Please see lines 128-129)

 

Point 6: Please mention the imaging angle of the camera. Is there any effect on the shapes in the captured images and their coordinates?

 

Response 6: The oblique-view reconnaissance of the camera causes a significant effect on object coordinates [6] and on the object shapes due to radiometric and geometric deformations [7].  Therefore, the UAV was operated in a hovering mode with a camera heading 90 degrees toward the ground while capturing images [8-10].  (Please see lines 126-129)

 

Point 7: Please express the altitude of the UAV while capturing the images.

 

Response 7: Thanks for the suggestion. We have added information related to UAV altitude in section 2.2 and Figure 2. The description is given as follows:

The two drones captured images every 5 m from the fly heights of 5 m to 60 m. (Please see lines 122-123 and Figure 2).

 

Point 8: Is the sensor of the camera CMOS/CCD, is it a rolling shutter camera or a global shutter, please explain this parameter.

 

Response 8: Those drones are all equipped with a camera sensor of 1-inch complementary metal oxide semiconductor (CMOS) and a 20-megapixel camera fixed on a three-axis gimbal.  The P4P has a lens of 24 mm focal length and a mechanical shutter supporting the capture of 4K at 60 fps. The M2P camera has a lens of 28 mm and a rolling shutter shooting 4K video. (Please see Lines 166 - 120).

 

Point 9: Please explain if there is any blurring effect on images due to the real-time imaging (while the UAV is moving) and the camera specifications.

 

Response 9:  The UAV was operated in a hovering mode that can minimize the motion of the camera and reduce the blurring effects on the captured images. However, we noted to readers that the blurring effects might be an important issue in the case of real-time surveys.  (Please see Lines 123-126)

 

Point 10: Please check the references, such as Ref 1.

Response 10: Thanks.  We have revised reference 1 as below:

“Conservancy, O. Tracking trash 25 years of action for the Ocean. Organisation Report. ICC Report 2011.” (Line 469)

 

Point 11: Please check the title "6. Patents”.

Response 11: Thanks. We have removed it.  (Line 450)

 

Reference

1. Boudjit, K.; Ramzan, N. Human detection based on deep learning YOLO-v2 for real-time UAV applications. Journal of Experimental & Theoretical Artificial Intelligence 2022, 34, 527-544.
2. Han, X.; Chang, J.; Wang, K. Real-time object detection based on YOLO-v2 for tiny vehicle object. Procedia Computer Science 2021, 183, 61-72.
3. Raskar, P.S.; Shah, S.K. Real time object-based video forgery detection using YOLO (V2). Forensic Science International 2021, 327, 110979.
4. Sridhar, P.; Jagadeeswari, M.; Sri, S.H.; Akshaya, N.; Haritha, J. Helmet Violation Detection using YOLO v2 Deep Learning Framework. In Proceedings of the 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), 2022; pp. 1207-1212.
5. Redmon, J.; Farhadi, A. YOLO9000: better, faster, stronger. In Proceedings of the Proceedings of the IEEE conference on computer vision and pattern recognition, 2017; pp. 7263-7271.
6. Chen, Y.; Li, X.; Ge, S.S. Research on the Algorithm of Target Location in Aerial Images under a Large Inclination Angle. In Proceedings of the 2021 6th IEEE International Conference on Advanced Robotics and Mechatronics (ICARM), 2021; pp. 541-546.
7. Jiang, S.; Jiang, C.; Jiang, W. Efficient structure from motion for large-scale UAV images: A review and a comparison of SfM tools. ISPRS Journal of Photogrammetry and Remote Sensing 2020, 167, 230-251.
8. Fallati, L.; Polidori, A.; Salvatore, C.; Saponari, L.; Savini, A.; Galli, P. Anthropogenic Marine Debris assessment with Unmanned Aerial Vehicle imagery and deep learning: A case study along the beaches of the Republic of Maldives. Sci Total Environ 2019, 693, 133581, doi:10.1016/j.scitotenv.2019.133581.
9. Gonçalves, G.; Andriolo, U.; Gonçalves, L.; Sobral, P.; Bessa, F. Quantifying marine macro litter abundance on a sandy beach using unmanned aerial systems and object-oriented machine learning methods. Remote Sensing 2020, 12, 2599.
10. Papakonstantinou, A.; Batsaris, M.; Spondylidis, S.; Topouzelis, K. A Citizen Science Unmanned Aerial System Data Acquisition Protocol and Deep Learning Techniques for the Automatic Detection and Mapping of Marine Litter Concentrations in the Coastal Zone. Drones 2021, 5, 6.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors,

I really like the tackled task and the motivation behind this work. The authors used a object detector (Yolov2) to detect bottle marine debris (BMD) on beaches. For this a data set was recorded. Based on this data set, the evaluation was performed. As a result an augmentation pipeline and the optimal flight height / ground resolution was proposed.

The data generation is well planned and documented. The overall structure of the paper is easy to follow and the approaches are understandable.

But I still have four major concerns:

 - The used object detector, Yolov2, is "quite" old. There is already a Yolov7 out there (https://arxiv.org/abs/2207.02696). And for this task a two-stage-detector, e.g. Faster R-CNN, is maybe even a better fit, because there is no real time requirement.

 - The hyper-parameter "anchor-boxes" was not mentioned. For anchor-based object detectors, like Yolov2 , this defines the possible bounding boxes. This parameter has to be adapted to the ground resolution of the images (exactly to the size of the objects), otherwise it is not a fair comparison.

- For object detection the most common metric is mean Average Precision (mAP), which was not even mentioned in this work. Maybe the author could at least reason why it is not helpful in this application.

- The claim that 0.5 cm / pixels are optimal for this use case is only shown for one object detector. The authors showed that other works found similar values, but still with this small set of experiments, it is risky to propose this kind of general recommendation.

 

Further, there are some minor issues:

Line 56: Keras is a framework not a model

Line 85: Replace * with \times or an x

Line 145 and Fig. 1: Please be consistent with the naming: Designed site 1 or designed site 1 (Captialized or not)

Line 121: Is the word training necessary in this sentence?

Line 153: Inverting a image sounds strange. Maybe flipping is better.

Line 157: The checkpoint path seems unnecessary for the method itself.

Line 159-161: It is hard to follow why you need 57 detectors? Could you maybe clarify this?

Fig. 2: Please check the capitalization in this figure. It is not consistent. And Please add the training to this figure.

Line 197: There is a lot of space left on this page

Line 234-235: This is not fully clear. Why do you name data sets now detectors? Do you mean the model at this point? Please clarify this part.

Tab. 1: The alignment of this table, especially the second column, is not optimal.

Fig. 5 and 7.: The line styles are hard to distinguish. The Precision curvature also seems strange in these plots.

Fig. 10: the a) is hard to read

Line 415: The section Patent is empty. Please remove or fill this section.

Reference 1: Seems very short. Is this reference complete?

Line 76: There are two headings without text in between. Personally, I looks odd for me, but this is only a esthetic point.

Author Response

Response to Reviewer#2 Comments of drones-2065826

We sincerely thank Reviewer#2 for his/her valuable suggestions and comments.  We highly appreciate the in-depth review and constructive suggestions.  The manuscript has been revised according to the reviewers’ comments.  The revised contents are highlighted in blue in the manuscript (as attached pdf file).  The point-to-point responses to the reviewer’s comments are given below.

 

General comments

Dear authors,

I really like the tackled task and the motivation behind this work. The authors used a object detector (Yolov2) to detect bottle marine debris (BMD) on beaches. For this a data set was recorded. Based on this data set, the evaluation was performed. As a result an augmentation pipeline and the optimal flight height / ground resolution was proposed. The data generation is well planned and documented. The overall structure of the paper is easy to follow and the approaches are understandable.

We are thankful for the reviewers’ evaluation of the manuscript.  The manuscript has been revised according to the reviewers’ comments.  We believe this manuscript is ready for publication.

 

Point 1: The used object detector, Yolov2, is "quite" old. There is already a Yolov7 out there (https://arxiv.org/abs/2207.02696). And for this task a two-stage-detector, e.g. Faster R-CNN, is maybe even a better fit, because there is no real time requirement.

 

Response 1: We thank the reviewer’s comment. The YOLO v2 was chosen because this model has been proven a useful tool to identify marine debris with satisfactory accuracy and computing speed.  YOLO v2 has been applied in many studies in recent years [1-4].   In addition, the YOLO v2 gets a score of 78.6 on mean average precision on testing the well-known dataset of Pascal VOC 2007, which is the best one among many other models [5].  Note that other new models (e.g., YOLO v7) have recently been developed; future works on testing how these new models improve the performance of identifying marine debris could contribute to the detection of marine debris.  (Please see lines 141-148)

 

 

Point 2: The hyper-parameter "anchor-boxes" was not mentioned. For anchor-based object detectors, like Yolov2 , this defines the possible bounding boxes. This parameter has to be adapted to the ground resolution of the images (exactly to the size of the objects), otherwise it is not a fair comparison.

 

Response 2: Thanks for the comments. The definition of hyper-parameter “anchor-boxes” and its used number was included in Section 2.3 as follows:

“YOLO v2 uses anchor boxes to detect objects in an image. Anchor boxes are predefined boxes that best match the given ground truth boxes and are defined by K-mean clustering. The anchor boxes are used to predict bounding boxes [5]. Estimating the number of anchor boxes is an important step in producing high-performance detectors [6].” (Lines 149 – 152)

“Further note, we used seven anchor boxes for training in this study.” (Line 187)

 

Point 3: For object detection, the most common metric is mean Average Precision (mAP), which was not even mentioned in this work. Maybe the author could at least reason why it is not helpful in this application.

 

Response 3: Thanks for the comment.  The mean Average Precision was not used in this study because we believe the other indexes (precision, recall, and F1-score) are enough for the assessment process.  Average Precision is a way to summarize the precision-recall curve into a single value; in other words, it is a precision recalculation based on accumulation recall value. The mean precision is calculated more independently and is eligible to represent the percentage of correctly detecting results. Moreover, F1-score also plays the role of statistical measurement of precision and recall to capture the whole picture of the detecting performance. Therefore, we believe that the mean precision, mean recall, and mean F1-score might be sufficient for evaluating our detecting results, and this is consistent with other studies [7-11].

 

Point 4: The claim that 0.5 cm / pixels are optimal for this use case is only shown for one object detector. The authors showed that other works found similar values, but still with this small set of experiments, it is risky to propose this kind of general recommendation.

 

Response 4: Thanks for the comment.  We have softened the tone for the statement in the discussion and conclusion:  In the planning of the study, we have tried to add complexities that would match to real-world situations.  For example, different colors and sizes (from 5 to 50 cm) of marine bottles were used in the training process.  Fifty-seven detectors were used to increase the randomness in the auto-detection process.  The detection algorithms and analysis procedures were also tested and studied in both the designed and testing(real-world) sites.  The results indicate that resolution is an important factor, which definitely affects the performance of detection.  We also found that the image resolution used in other studies ranges from 0.11 to 0.82 cm/pixel (Table 3), and many studies chose a resolution of around  0.5 cm/pixel in their surveys.  As a result, a resolution of 0.5 cm/pixel might be a considerable choice that has the potential to apply in large-scale surveys.  (Please see lines 420-429)

 

Point 5: Line 56: Keras is a framework, not a model.

 

Response 5: We thank you for figuring out this point. The content was recorrected in Line 56 – 57 as below:

“Kako et al. (2020) employed a deep learning model based on Keras framework to estimate plastic debris volumes.”

 

Point 6: Line 85: Replace * with \times or an x

 

Response 6:   Corrected. Thanks. (Please see lines 90)

 

Point 7: Line 145 and Fig. 1: Please be consistent with the naming: Designed site one or designed site 1 (Capitalized or not).

 

Response 7: Thank you. We have revised the manuscript to make it consistent throughout the manuscript.

 

Point 8: Line 121: Is the word training necessary in this sentence?

 

Response 8: The word “training” was removed.  (Please see lines 139-140)

 

Point 9: Line 153: Inverting an image sounds strange. Maybe flipping is better.

 

Response 9: Thanks for the suggestion; the word has been changed in Line 182.

 

Point 10: Line 157: The checkpoint path seems unnecessary for the method itself.

 

Response 10: Thanks for the suggestion, the description of training options has been revised in lines 185 – 186.

 

 

Point 11: Line 159-161: It is hard to follow why you need 57 detectors? Could you maybe clarify this?

 

Response 11: Thanks for the comment. We have added more descriptions in Lines 279-283:

Because the objects of the marine debris might be complicatedly distributed in real-world situations, a small number of detectors might not have good performance in various conditions.  Therefore, 57 detectors with different initial settings of parameters were trained in different runs to increase the randomness.  The results from the 57 detectors were then compared to evaluate the performance of these initial settings.

 

Point 12: Fig. 2: Please check the capitalization in this figure. It is not consistent. And Please add the training to this figure.

 

Response 12: The Figure has been revised with descriptions. (Please see the Lines 191 - 193, Figure 3).

 

Point 13: Line 197: There is a lot of space left on this page.

 

Response 13: Thanks for your comment. The figures’ sizes have been changed to reduce the space left on pages.

 

Point 14: Line 234-235: This is not fully clear. Why do you name data sets now detectors? Do you mean the model at this point? Please clarify this part.

 

Response 14: Thanks for the advice. The sentence has been revised. (Please see Lines 277 – 278)

 

Point 15: Tab. 1: The alignment of this table, especially the second column, is not optimal.

 

Response 15: Thanks for your comment. The table alignment has been adjusted. (Please see Table 2, Lines 275-276)

 

Point 16: Fig. 5 and 7.: The line styles are hard to distinguish. The Precision curvature also seems strange in these plots.

 

Response 16: Thanks for your comment. Line styles and colors have been modified (Line 298, Figure 7; Line 326, Figure 9).  

 

Point 17: Fig. 10: the a) is hard to read.

 

Response 17: Thanks.  Corrected. ( Line 380.)

 

Point 18: Line 415: The section Patent is empty. Please remove or fill this section.

 

Response 18: Thanks. We have removed it.  (Line 450).

 

Point 19: Reference 1: Seems very short. Is this reference complete?.

 

Response 19: Thanks.  We have revised reference one as below:

“Conservancy, O. Tracking trash 25 years of action for the Ocean. Organisation Report. ICC Report 2011.” (Line 469)

 

Point 20: Line 76: There are two headings without text in between. Personally, I looks odd for me, but this is only an aesthetic point.

 

Response 20: Thanks for the suggestion.  We think this is an aesthetic point, and we keep the original format.

 

Reference

1. Boudjit, K.; Ramzan, N. Human detection based on deep learning YOLO-v2 for real-time UAV applications. Journal of Experimental & Theoretical Artificial Intelligence 2022, 34, 527-544.
2. Han, X.; Chang, J.; Wang, K. Real-time object detection based on YOLO-v2 for tiny vehicle object. Procedia Computer Science 2021, 183, 61-72.
3. Raskar, P.S.; Shah, S.K. Real time object-based video forgery detection using YOLO (V2). Forensic Science International 2021, 327, 110979.
4. Sridhar, P.; Jagadeeswari, M.; Sri, S.H.; Akshaya, N.; Haritha, J. Helmet Violation Detection using YOLO v2 Deep Learning Framework. In Proceedings of the 2022 6th International Conference on Trends in Electronics and Informatics (ICOEI), 2022; pp. 1207-1212.
5. Redmon, J.; Farhadi, A. YOLO9000: better, faster, stronger. In Proceedings of the Proceedings of the IEEE conference on computer vision and pattern recognition, 2017; pp. 7263-7271.
6. Loey, M.; Manogaran, G.; Taha, M.H.N.; Khalifa, N.E.M. Fighting against COVID-19: A novel deep learning model based on YOLO-v2 with ResNet-50 for medical face mask detection. Sustainable cities and society 2021, 65, 102600.
7. Fallati, L.; Polidori, A.; Salvatore, C.; Saponari, L.; Savini, A.; Galli, P. Anthropogenic Marine Debris assessment with Unmanned Aerial Vehicle imagery and deep learning: A case study along the beaches of the Republic of Maldives. Sci Total Environ 2019, 693, 133581, doi:10.1016/j.scitotenv.2019.133581.
8. Gall, S.C.; Thompson, R.C. The impact of debris on marine life. Marine pollution bulletin 2015, 92, 170-179.
9. Gonçalves, G.; Andriolo, U.; Gonçalves, L.; Sobral, P.; Bessa, F. Quantifying marine macro litter abundance on a sandy beach using unmanned aerial systems and object-oriented machine learning methods. Remote Sensing 2020, 12, 2599.
10. Papakonstantinou, A.; Batsaris, M.; Spondylidis, S.; Topouzelis, K. A Citizen Science Unmanned Aerial System Data Acquisition Protocol and Deep Learning Techniques for the Automatic Detection and Mapping of Marine Litter Concentrations in the Coastal Zone. Drones 2021, 5, 6.
11. Takaya, K.; Shibata, A.; Mizuno, Y.; Ise, T. Unmanned aerial vehicles and deep learning for assessment of anthropogenic marine debris on beaches on an island in a semi-enclosed sea in Japan. Environmental Research Communications 2022, 4, 015003.

 

Author Response File: Author Response.pdf

Reviewer 3 Report

The following revisions are required.

1. In literature review, add 3 to five more relevant and latest techniques.
2. Add Comparison table at the end of section 2 or table 2 and compare with at least 3 to 5 Techniques with appropriate parameters.
3. Please make sure your paper has necessary language proof-reading.
4. What is the possible limitations of the proposal in this paper if any? More detailed discussions are necessary to be added.
5. The Conclusion section should be compressed further to summarize the whole paper and make it concise enough.

Author Response

Response to Reviewer#3 Comments of drones-2065826

We sincerely thank Reviewer#3 for his/her valuable suggestions and comments.  We highly appreciate the in-depth review and constructive suggestions.  The manuscript has been revised according to the reviewers’ comments.  The revised contents are highlighted in blue in the manuscript (as attached pdf file).  The point-to-point responses to the reviewer’s comments are given below.

 

Point 1: In literature review, add 3 to five more relevant and latest techniques.

 

Response 1: Thanks for the comment. Three relevant references have been added in Section 1. (Please see Lines 57-61, Lines 66 - 67)

 

Point 2: Add Comparison table at the end of section 2 or table 2 and compare with at least 3 to 5 Techniques with appropriate parameters.

 

Response 2: Thanks for your comment. We have added the relevant information in the table (Table 3). (Please see Lines 419, Table 3)

 

 

Point 3: Please make sure your paper has necessary language proof-reading.

 

Response 3: Thank you. The manuscript has been revised substantially and thoroughly based on the reviewers’ comments.  We carefully checked the problems in language and wording. We believe it is acceptable for publication in the present version.

 

Point 4: What is the possible limitations of the proposal in this paper if any? More detailed discussions are necessary to be added.

 

Response 4: Thanks for the suggestion.  The detector we used was YOLO v2; note that other new models (e.g., YOLO v7) have recently been developed; future works on testing how these new models improve the performance of identifying marine debris could contribute to the detection of marine debris.  (Please see Lines 430-433)

 

 

Point 5: The Conclusion section should be compressed further to summarize the whole paper and make it concise enough.

 

Response 5: Thanks for the suggestion. The Conclusion was revised. (Please see Lines 438-452.)

 

Author Response File: Author Response.pdf

Reviewer 4 Report

This work considers one of the most important environmental topic for the marine and beach litter. Here, the detection of bottle marine debris was assessed through unmanned aerial vehicle and machine learning techniques. The importance of this work concerns the use of remote sensing integrated with convolutional neural network, in order to detect the bottle marine debris from drone images. Considering the importance of this theme, mostly to plan some management strategies in front of plastic pollution, I strongly suggest the pubblication of this work.

I have only few comments to improve the manuscript.

I suggest to insert the hyperparameter values, for example inserting a table with the values of minibatch size, max epochs, etc., in the section 2.2.

In the result section, it is important also to show the accuracy and loss function connected to the training of the convolutional neural network.

Other comments are inserted in the attached pdf file.

Many thanks and kind regards.

 

 

Comments for author File: Comments.pdf

Author Response

Response to Reviewer#4 Comments of drones-2065826

We sincerely thank Reviewer#4 for his/her valuable suggestions and comments.  We highly appreciate the in-depth review and constructive suggestions.  The manuscript has been revised according to the reviewers’ comments.  The revised contents are highlighted in blue in the manuscript (as attached pdf file).  The point-to-point responses to the reviewer’s comments are given below:

 

 General comments

This work considers one of the most important environmental topic for the marine and beach litter. Here, the detection of bottle marine debris was assessed through unmanned aerial vehicle and machine learning techniques. The importance of this work concerns the use of remote sensing integrated with convolutional neural network, in order to detect the bottle marine debris from drone images. Considering the importance of this theme, mostly to plan some management strategies in front of plastic pollution, I strongly suggest the pubblication of this work.

I have only few comments to improve the manuscript.

We are thankful for the reviewers’ evaluation of the manuscript.  The manuscript has been revised according to the reviewers’ comments.  We believe this manuscript is ready for publication.

 

Point 1: I suggest to insert the hyperparameter values, for example inserting a table with the values of minibatch size, max epochs, etc., in the section 2.2.

 

Response 1: Thanks for the suggestion. The hyperparameter values have been listed in the Table 1 in Line 194 as follows:

Table 1. Training hyperparameter values

Parameter	Value
Training option	Sdgm (stochastic gradient descent with momentum)
Mini-batch size	8, 16
Number of epochs	50, 100, 150, …, 2000
Initial learning rate	10-3, 10-4, 10-5
Learning rate drop factor	0.8
Learning rate drop period	80

 

Point 2: In the result section, it is important also to show the accuracy and loss function connected to the training of the convolutional neural network.

 

Response 2: Thanks for the suggestion. We have added the loss value in the initial of Section 3.1 as follows:

“The datasets before and after background removal were trained separately. Figure 5 shows an example of a loss reduction curve. In the three values of the initial learning rate (10^-3, 10^-4 , and 10^-5), we figured out that the model was unstable during the training process when the initial learning rate was  10^-4. When the initial learning rate was 10^-3 and 10^-5, the loss curve was steady with a small fluctuation. The training loss value decreases when the number of epoch increase, and final aver-age loss varied from 0.28 to 1.02.” (Lines 251 – 256)

We further inserted a figure showing an example of the loss reduction curve during the training process. (Please see Figure 6, Line 257)

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revision is quiet and acceptable. The authors have addressed my comments well, as seen from the Revision information letter and the revised manuscript. There is no further question or problem on my side.