Towards Automatic Burrow Detection for Sustainable River Levees

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. What is the main question addressed by the research?

The authors were inspired by the fact that animals dig burrows to protect themselves from predators and harsh weather. However, this burrowing also damages the soil, increases erosion, and can cause river banks to weaken or collapse. Based on these well known facts, the authors have developed their main research question i.e. "can an AI-based system using drone-acquired imagery automatically and accurately detect and map animal burrows on river levees?". The main motivation behind this question is that early detection could help identify risks sooner and support proactive maintenance, instead of relying on traditional inspection methods that are time consuming and resource heavy.
I feel that the authors have clearly stated their main research question. Since animal burrows tend to weaken river levees and increase the risk of flooding, the authors have focused on whether AI can use drone images to automatically find and map these burrows. The problem was well explained and it provided linkages to real-world safety and infrastructure issues. However, the main research question could be written more clearly in one short sentence at the end of the introduction to better guide the reader.

2. Do you consider the topic original or relevant to the field? Does it address a specific gap?

I feel that the topic is indeed original and also I feel that it is quite relevant. It is not that burrows detection in river levees is a never explored research topic; however prior researchers have relied mostly on geophysical methods and manual inspection. What is the novelty factor for this research is that it addresses a clear gap by proposing a scalable, automated, and cost effective computer vision approach using drones and machine learning, which has not been widely implemented for this specific application.  Because of this, the research topic is very relevant and timely for sustainability, flood protection, and strong infrastructure. Using AI and UAV images for this specific problem is still not very common in existing studies, which makes this work original and useful. 
However, there is still some room for improvement. The authors could make this section stronger by clearly comparing their approach with one or two similar AI based studies to better show what is new in their work.

3. What does it add compared with other published material?

As mentioned earlier, most past studies depend on geophysical methods or manual checks to find burrows on river levees. In contrast, this research fills a clear gap by proposing a scalable, automated, and economic computer vision approach using drones and machine learning. The authors have presented a new pipeline that combines drone images, pixel level classification using Random Forest, and morphological processing to detect burrows on river levees. Unlike geophysical methods such as GPR and resistivity surveys, this approach allows fast monitoring over large areas. The method is also flexible and can be adapted to different environmental conditions, which makes it more practical to use in real situations.
To improve this further, the authors could add a simple baseline comparison to clearly show how much better their method performs compared to simpler approaches.

4. Specific improvements the authors should consider regarding methodology

The pixel level analysis in this study has used only RGB values and did not consider any spatial or texture features, which I am pretty confident could help to improve the detection accuracy. The authors may consider to add some simple texture features, such as local patterns or texture descriptors, in order to may make the model more robust. The study is also based on images collected at a single time period, which means it does not show how burrow detection may change across different seasons or times. Future work could include time series drone imagery to better understand seasonal changes and long term trends in burrow detection. In addition, the paper does not clearly state the dataset size, as it refers to it as “XXX” in the experiments section. I would suggest that the authors provide the exact number of images and burrow instances to improve the reproducibility. Overall, the methodology has been well organized and clearly explained. The authors have carefully designed the workflow, data collection, labeling process, and iterative learning strategy, and the experimental setup shows strong attention to robustness and parameter tuning.
There are a few methodological details that need more clarification. The paper does not clearly mention the exact number of images and burrow instances used in the experiments. The authors also need to better explain why they chose to use only RGB pixel features and what limitations this choice may have. Furthermore, the paper does not clearly explain how consistency in manual annotations was ensured among experts. The model is trained and tested using data from only one river section, and testing it on different rivers or regions would help show how well it works in other locations. These issues are mainly related to clarity and reporting and do not represent major methodological weaknesses.

5. Are the conclusions consistent with the evidence? Do they address the main question?

I do feel that the conclusions drafted by the authors indeed align with the experimental results. Hence, I can say that this study has fairly demonstrated that the proposed method has the potential to achieve high level of precision and F1-scores under optimal parameter configurations to ensure automated burrow detection. The discussion clearly explains the balance between precision and recall, which supports a practical approach for real world use. Importantly, the authors do not make exaggerated claims and clearly acknowledge the balance between precision and recall.
The conclusions could be made a bit stronger by including some clear numbers or performance ranges from the results, instead of mostly relying on general or descriptive explanations. Referring to key values such as precision or F1-score ranges would help readers better understand how well the method actually performs.
General Comments:-
The paper is clear and well organized, and it is easy to read and understand. The structure makes sense, and the figures and tables help to explain the noteworthy work done by the authors. The practical focus of the study indeed makes it useful for real world users and decision makers. To improve the paper further, I would suggest that the authors do some minor English editing, especially in the results section where a few sentences are long and not very clear. A careful proofreading should be done so as to help fix small spelling mistakes, for example changing “drone-adquired imagery” to “drone-acquired imagery.”
In addition, the authors should clarify the dataset size and clearly explain how images were manually annotated. The discussion section could be improved from a clearer explanation of limitations, such as the effect of dense vegetation on burrow detection and the need to validate the method using time series data across different seasons. The authors may also consider discussing whether adding other sensor data, such as multispectral or thermal images, could improve detection under difficult lighting or vegetation conditions. It would be also useful to expand the discussion on how this system could be used in real levee management practices and whether it could reduce costs compared to traditional inspection methods. Finally, all placeholder text (for example “XXX” in the experiments section) should be replaced with actual values before publication.
These are mostly clarity and improvement suggestions and do not affect the overall quality or contribution of the paper.

Later supplementary comments:

1. Summary of the manuscript and key contributions

 

Authors of this research paper have handled an important sustainability related issue, that of detecting animal burrows on river levees. Rapid burrowing of soil on river levees could cause major problems such as structural weakness and soil erosion, which may further escalate to catastrophic events such as floods.  The authors have done a rigorous review of existing literature and they have studied methods used by similar researches. Instead of focusing on manual, human centric methods, the authors have opted for AI-based solution using emerging technologies such as drones and image processing. Their main objective was to support early risk identification of map burrows and help in proactive levee maintenance.  The critical contribution of their paper is to share the design of a scalable computer vision based system that utilizes multi domain technologies such as UAV image processing, supervised ML algorithm Random Forest and image processing. This approach is distinctively different from the other traditional methods that require many number of human members on the field (river levees) as well as some methods requiring heavy equipment's. Traditional methods and human intensive and may suffer from related issues such as slowness, possibility of human induced errors and possibility of injuries. The method proposed by the authors is indeed a cost effective method  and is more scalable than the traditional methods. It introduces a human in the loop aspect that provides their system a great ability of adaptability to different environmental conditions. For the reasons aforementioned herein, I would indeed suggest that this paper is relevant to the sustainability journal.

 

2. Evaluation of methodology, analyses, and conclusions

 

Authors have explained their work in a structured step by stem manner, which enhances both readability as well as understandability. In simple terms, their system mainly uses an aerial device UAV (drone) that takes pictures of the river levees. Their system then performs pixel-level classification of the images using a Random Forest algorithm to identify pixels that may belong to burrows. Their system learns from these examples to find similar burrows in new images. They have applied morphological and geometric filtering to reduce noise and help in identifying authentic burrow like shapes. Their system focusses on surface visible patterns consistent with burrow openings. Similar approaches have been used in related works such as litter detection on sea beaches or river banks etc. This system is likely to work well where the burrows are not hidden i.e. they are distinctly visible. This also means that the system cannot work for hidden voids, for which best approach would be the traditional geophysics one that uses magnetic and seismic measures. As stated in previous section that this system is not human centric i.e. doesn't need active involvement of human on the field, however, they have provisioned for the human in the loop aspect for evaluating and reviewing system outputs and results and also human involvement would help to improve the ML model over time. The way the authors have tested their method is careful and well planned. They have evaluated how often the system finds burrows correctly and how often it makes mistakes. They have also used different settings to see what works best. This helps to show that their method is stable and works reasonably well in practice.

 

Authors have documented their works quite professionally, however, some details could possibly be explained more clearly. The exact dataset size (number of images and burrow instances) is not explicitly stated and is referred to as “XXX” in the experiments section. Hence, the paper does not clearly say how many images were used, and the tests were done on only one river area and during one time period. This means it is not yet clear how well the method would work in other places or during different seasons. That means the reproducibility of this research has negative consequences. Another possible limitation is that this method uses only color information  (RGB values) from images, so it may struggle in areas with heavy vegetation or strong shadows.

 

Overall, the conclusions match the results shown in the paper. The good thing authors do not claim that their system is perfect. Instead, they have explained that the system is meant to help inspectors by quickly pointing out risky areas, not to fully replace human experts. Overall, the results support their main idea that this approach can help find burrows earlier and make levee monitoring easier and faster. Therefore, the conclusions are consistent with the results and directly address the main research question, but they could be further strengthened by referencing specific performance ranges rather than relying mainly on qualitative interpretation.

 

3. Constructive feedback and suggestions for improvement

 

This research paper has been written professionally and if suggested enhancements are implemented, then it could be considered as a strong research paper:-

 

1. Authors should clearly explain about the dataset used (size, number of images, burrow instances used for training the model etc.). All placeholders for e.g. "XXX" should be replaced with actual values.
2. The authors should more clearly describe the manual annotation process used in the study. This includes explaining how the experts have labeled burrow and non-burrow areas in the images, how these labels were used to train the model, and how the human reviewers checked and corrected the model’s predictions. It would also be useful to explain how the corrected labels were added back into the system to help to further improve the model. Providing these details would also enable readers to better understand the human involvement in the approach and thus enhance transparency and reproducibility.
3. The paper would be easier to understand if the authors clearly explain how their system could be used in current levee management work, and how it compares with traditional inspection methods in terms of cost and daily operations.
4. The paper has a few small spelling and typing mistakes, such as “drone-adquired” instead of “drone-acquired,” so the authors should carefully proofread the manuscript.
5. Overall, this is a useful applied research study that fits well within the scope of the Sustainability Journal.

Comments on the Quality of English Language

To improve the paper further, I would suggest that the authors do some minor English editing, especially in the results section where a few sentences are long and not very clear. A careful proofreading should be done so as to help fix small spelling mistakes, for example changing “drone-adquired imagery” to “drone-acquired imagery.”

Author Response

1) The main research question could be written more clearly in one short sentence at the end of the introduction to better guide the reader.

We added a clearer sentence now.

2) However, there is still some room for improvement. The authors could make this section stronger by clearly comparing their approach with one or two similar AI based studies to better show what is new in their work.

Unfortunately, we were unable to find or use any system that could be, even loosely, compared to the one we produced. This, we feel, is in fact one of the strengths of our contribution: providing the public with a tool that fills in a real and evident gap. 

3) To improve this further, the authors could add a simple baseline comparison to clearly show how much better their method performs compared to simpler approaches.

Again, there simply is no baseline; currently, these operations are performed manually, with a relevant material effort, both time-consuming and error-prone.

4) The pixel level analysis in this study has used only RGB values and did not consider any spatial or texture features, which I am pretty confident could help to improve the detection accuracy. The authors may consider to add some simple texture features, such as local patterns or texture descriptors, in order to may make the model more robust. 

This is a very good suggestion. We are, in fact, considering this possibility; however, this is a major technical change that requires a relevant effort, and we are currently considering this possibility for a second, future version.

5) The study is also based on images collected at a single time period, which means it does not show how burrow detection may change across different seasons or times. Future work could include time series drone imagery to better understand seasonal changes and long term trends in burrow detection. 

Unfortunately, image acquisition for test purposes is very time consuming. Currently, the system is being used in test phase; we plan to report potential season-dependent issues in the future. 

6) In addition, the paper does not clearly state the dataset size, as it refers to it as “XXX” in the experiments section. I would suggest that the authors provide the exact number of images and burrow instances to improve the reproducibility. 

We did so now, alongside a long series of improvements over the first submission. 

7) There are a few methodological details that need more clarification. The paper does not clearly mention the exact number of images and burrow instances used in the experiments. The authors also need to better explain why they chose to use only RGB pixel features and what limitations this choice may have. Furthermore, the paper does not clearly explain how consistency in manual annotations was ensured among experts. The model is trained and tested using data from only one river section, and testing it on different rivers or regions would help show how well it works in other locations. These issues are mainly related to clarity and reporting and do not represent major methodological weaknesses.

We briefly discussed these issues in the current version. 

8) The conclusions could be made a bit stronger by including some clear numbers or performance ranges from the results, instead of mostly relying on general or descriptive explanations. Referring to key values such as precision or F1-score ranges would help readers better understand how well the method actually performs.

Thank you for this suggestion; we did include these considerations now. 

9) A careful proofreading should be done so as to help fix small spelling mistakes, for example changing “drone-adquired imagery” to “drone-acquired imagery.”

Done.

10) In addition, the authors should clarify the dataset size and clearly explain how images were manually annotated. 

Done.

Reviewer 2 Report

Comments and Suggestions for Authors

Overall, a good application paper, well presented. A detailed review has been provided. 

Comments for author File: Comments.pdf

Author Response

1) Proper reasoning was added for Random Forest over deep learning method selection; however, clarification of the dataset size (currently marked as “XXX” in the experiments section) is needed.

Done.

2) A short discussion on generalization limits, seasonal variability, and transferability to other river systems is required.

Done.

3) Please report Recall explicitly alongside Precision and F1-score for better interpretability.

As it is known, having both Precision and F1 allows one to extract Recall, if needed. Moreover, the numerical results are already a bit hard to read; we feel that adding a new parameter could worsen the situation.