AI-Driven Real-Time Monitoring of Ground-Nesting Birds: A Case Study on Curlew Detection Using YOLOv10

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The research brings an interesting approach to using artificial intelligence in bird monitoring. What is missing is scientific contribution and originality, since similar research on other species has existed for a long time.

Suggestions are incorporated into the manuscript

 

Comments for author File: Comments.pdf

Author Response

Question 1: The research brings an interesting approach to using artificial intelligence in bird monitoring. What is missing is scientific contribution and originality, since similar research on other species has existed for a long time. Suggestions are incorporated into the manuscript.

Response 1: The authors would like to thank the reviewer for taking the time to read the manuscript. We appreciate your feedback and the time taken to edit the manuscript. We have updated the manuscript as suggested:

• Added the scientific name of the Curlew to the keywords as suggested.
• This sentence relates to how predators and prey are being pushed closer together due to habitat loss (often due to development and farming). In this site there is a geriatric population of Curlews as most chicks to not survive to adulthood.
• We agree, we didn’t add the scientific names of the predators in the introduction. We thought it would be best to included them in the class names in the methodology.
• We agree, this sentence was repetitive, and we have removed it. Thank you for highlighting it.
• Corrected AI to Artificial Intelligence as it’s the first mention as suggested.
• We agree with this comment, its important as the methodology would work for all bird species with the correct amount of training data and variance. We have change this to birds to make it applicable to all bird species.
• We have added an example for this comment “How do you intend to protect the nest? Any form of anthropogenic influence in the bird nesting zone is often undesirable”. Such interventions include the removal of eggs and chicks for incubation.
• Curlew eggs and chicks are sometimes removed for incubation. This has been successfully undertaken in Ireland. With the real-time workflow proposed in this paper conservationists can be alerted in time to intervene where required.
• We agree with this comment “If you want to protect nests from natural enemies (foxes, badgers), it would be a good idea to include these species in the monitoring”. These species are included in the model (with the corresponding results). However, during this deployment, these species where not seen hence there are no inference results for these classes.
• Thank you, we agree and have removed the repetition as suggested.
• We would like to leave the picture of the Curlew detection in if possible? We think it helps to show what we are trying to do and the environment where the study was conducted. We are also quite proud of this image. It’s has taken us a long time to get these models out into a study. But if you feel strongly about this, we can remove it.
• We would also like to leave this in if possible as we feel it sets the scene for the for the paper. “The remainder of this paper is structured as follows: Section 2 outlines the meth-odology, including data collection, model training, and evaluation metrics. Section 3 presents the results, including both model performance and real-time inference for the curlew case study, followed by the discussion in Section 4. The conclusions and future directions are presented in Section 5.” But if you feel strongly about this, we can remove it.
• We agree the first paragraph in the material and methods has been removed as it is detailed in the remainder of the methodology.
• We agree with this comment “Scientific names in italics in parentheses after common names”. This has now been corrected thanks for highlight it.
• This is a fair comment “I recommend that authors study the ways of writing the names of plants and animals.” As computer scientists where often abstracted from this but we do have a number of ecologists on the paper so this should have been flagged. Thanks for highlighting it.
• The number of tags on the y axis label relates to the number of total bounding boxes for each species. We have updated the caption to reflect this.
• All images were evaluated manually as well as the model to make sure all species were captured in the results. If a different species was spotted and misclassified, we would have reported the error in the inference results like we have for other species.
• The reference for the model architecture is in the paragraph number [27].
• We would like to keep the technical details of the model architecture. This is a relatively new model which hasn’t been deployed or used in this setting before. How it works helps to justify how it is trained and partially why we get the results presented in the paper.
• We would like to keep the technical details regarding the hardware and hyperparameters of the model for reproducibility. In our opinion too many papers don’t include these. If other researchers want to train their own model they would need this information. We have gone through a long period of testing different hyperparameters to get the results presented in this paper. We feel strongly about sharing these settings so other researchers can build their own models.
• We would like to keep in the details regarding the inferencing hardware. This shows how expensive it is to run the AI model in terms of cost and power. We feel that this is important and something that is often overlooked in papers presenting AI solutions. AI is great but there is an ongoing cost to it which is important to highlight here.
• We completely agree with this comment thanks for raising it “How many cameras were there in one location? how the locations for the cameras were chosen”. We have added in the following: Each nesting site was monitored using a single camera, nine of the cameras were placed one meter away from the nest cup while the remaining two cameras were placed two meters away due to nest protection measures. The cameras were deployed in various habitats including two heathland, eight silage/hay and one rough grassland.
• Thanks for the following comment “Perhaps a photo related to the species being researched would be more appropriate”. This figure has been removed inline with comments from other reviewers.
• The accuracy during inference was checked using the evaluation metrics outlined in the methodology. All the inference results (model prediction) where manually checked by a person and a confusion matrix produced to determine the results.
• The results have been checked multiple times. The results presented in the confusion matrix are correct. Even though pheasants weren’t detected during deployment some of the curlew species where misclassified as pheasants hence why it was included. Another reviewer suggested altering the confusion matrix slight (which we have now done) to make this clearer.
• During the study no other small birds were detected here. However, if a species that’s not in the model was detected it would be misclassified as another species. Hence why we are trying to add in as many species in as possible. It would be great if you can get in touch directly and add suggestions for other species. We have added the following to the paper “It is important to acknowledge that the detection of species not represented in the model's training set can lead to misclassifications. Therefore, expanding the training dataset to include a broader spectrum of species is necessary to improve the model's accuracy and reliability”.

You are completely correct “You have detected individuals in your research!!”, we have removed the word species.

Reviewer 2 Report

Comments and Suggestions for Authors

Summary:

This manuscript provides an interesting and useful addition to the rapidly evolving space of biodiversity monitoring by demonstrating a near-real time application of computer vision to monitor bird nests in real landscapes across a variety of locations in Wales. The manuscript communicates clearly. The real-time pipeline from camera to predictions is clearly described. Model development, training, and assessment are thoroughly presented. Overall the study provides a compelling test case for the deployment of real-time wildlife camera processing for species monitoring.

Review:

The manuscript does a good job of communicating the gap in knowledge (real-time image processing) that it is addressing. Successful deployment of an AI model with camera traps and the model and pipeline design are all well communicated. I had only a couple of areas where I thought some additional explanation would be helpful for people using this paper as a guide for thinking about and developing their own versions of this project.

1)      Site information: Very little information is given about the 11 sites in Wales that were used for the study. Were they varying habitat types? All sheep fields? Different color or structure of vegetation? I ask because deployment in fairly similar habitats would mean image backgrounds are relatively uniform and that is a different computer vision challenge than image backgrounds that range from rocks to trees and everything in between. I coupe of sentences about the deployment conditions and perhaps a brief addition to the discussion about possible challenges or limitations of this work given the deployment conditions would be useful to help the reader interpret the results and potential.

2)      Sheep: The discussion about why some curlews were identified as pheasants was well done and useful. The confusion matrix also suggests that there is also some loss of actual curlew detections because they are being detected as sheep, which is perhaps a little more puzzling. Not overly concerning, because the numbers are relatively low, but I’d love to see a sentence speculating on what image features are causing curlews to be confused with sheep – is it image quality or partial captures of the animal? I’ll admit this is slightly me being really curious and this is not a major concern.

3)      Table 2: On this note, I saw that for the rows, Phasianus colchicus was not given a row because there are no actual birds, However, the table description references the diagonal number for true positives and without the pheasant row this is less straight forwards. I’d recommend adding the row of 0s. It takes little space, makes the table more directly interpretable, and drives home the message about the model confusion.

Author Response

Question 1: Site information: Very little information is given about the 11 sites in Wales that were used for the study. Were they varying habitat types? All sheep fields? Different color or structure of vegetation? I ask because deployment in fairly similar habitats would mean image backgrounds are relatively uniform and that is a different computer vision challenge than image backgrounds that range from rocks to trees and everything in between. I coupe of sentences about the deployment conditions and perhaps a brief addition to the discussion about possible challenges or limitations of this work given the deployment conditions would be useful to help the reader interpret the results and potential.

Response 1: The authors would like to thank the reviewer for taking the time to read the manuscript. We appreciate your feedback and have addressed the comment above in the following way:

• Added more details regarding the camera setup and deployment: “Each nesting site was monitored using a single camera, nine of the cameras were placed one meter away from the nest cup while the remaining two cameras were placed two meters away due to nest protection measures. The cameras were deployed in various habitats including two heathland, eight silage/hay and one rough grassland”.
• We have also added this to the discussion: The cameras were deployed in slightly varying habitats although the overall colour and structure of the vegetation didn’t vary significantly between each site. However, the density of vegetation especially in the hay and silage fields varied depending on the growth. Model performance would likely degrade in environments that differ significantly from the habits used in this study. It would therefore be important to add additional training data to increase environmental variance.

Question 2: Sheep: The discussion about why some curlews were identified as pheasants was well done and useful. The confusion matrix also suggests that there is also some loss of actual curlew detections because they are being detected as sheep, which is perhaps a little more puzzling. Not overly concerning, because the numbers are relatively low, but I’d love to see a sentence speculating on what image features are causing curlews to be confused with sheep – is it image quality or partial captures of the animal? I’ll admit this is slightly me being really curious and this is not a major concern.

Response 2: Thank you for your comment regarding the misclassification of curlews as sheep. We have added a discussion addressing potential reasons for this confusion, including overlapping visual features and image quality factors. “In addition to the misclassification of curlews as pheasants, there were instances where curlews were incorrectly identified as sheep. We speculate that this error may arise from overlapping visual features, such as similar coloration under certain lighting conditions or partial captures of the birds that obscure distinctive characteristics. Furthermore, environmental factors like rain or condensation on the camera lens can impair image clarity, making it challenging for the model to accurately discern specific features of curlews”.

Question 3: Table 2: On this note, I saw that for the rows, Phasianus colchicus was not given a row because there are no actual birds, However, the table description references the diagonal number for true positives and without the pheasant row this is less straight forwards. I’d recommend adding the row of 0s. It takes little space, makes the table more directly interpretable, and drives home the message about the model confusion.

Response 3: Thank you for your suggestion regarding the confusion matrix. We have updated the table by adding a row of zeros for Phasianus colchicus to reflect the absence of actual pheasants in the dataset.

Reviewer 3 Report

Comments and Suggestions for Authors

Using AI-driven intelligent algorithms to improve wildlife protection is an important research direction, and I believe this article has made a useful attempt in this field. However, this article only proposes a bird monitoring solution and does not improve the intelligent monitoring algorithm. Therefore, I think this article is not within the scope of the Remote Sensing journal, and it is recommended to submit it to zoology journals such as Animals.

In addition, I still have some suggestions:

1. Lines 65-71 are too limited in saying that the application of trap cameras is limited to filtering blank applications, because there are currently a large number of applications based on trap cameras. New literature research needs to be supplemented.

2. The meaning of the other gray lines on the P-R curve lacks legends.

Author Response

Thank you for your thoughtful feedback and acknowledgment of the importance of using AI-driven intelligent algorithms for wildlife protection. While we appreciate your recommendation regarding zoology journals, we respectfully believe that our work aligns well with the scope of Sensors, particularly because our research emphasises the development and deployment of a custom pipeline for real-time monitoring of ground-nesting birds using AI, accelerated compute and sensor technology.

Specifically:

• First-of-its-Kind Deployment: This study represents the world’s first deployment of real-time AI-based monitoring specifically for curlews, a vulnerable ground-nesting bird species. The novelty lies not only in the biological application but also in the integration of AI, camera trap sensors, and real-time data transmission pipelines to address urgent conservation needs. This end-to-end technological solution directly pertains to the journal's focus on sensor-based systems.
• Custom AI Pipeline and Model: While our primary goal was to develop a solution for monitoring curlews, we built a custom pipeline integrating real-time 3/4G-enabled camera traps with a YOLOv10x model specifically trained for curlew detection. This system leverages cutting-edge sensor and AI technology, enabling immediate insights for conservation interventions, which would not be possible with traditional wildlife monitoring approaches.
• Focus on Sensor Integration: The research highlights the technical challenges and solutions involved in integrating sensor data with AI for dynamic and proactive monitoring. This includes designing a real-time data processing pipeline, optimising hardware (camera sensors, servers, and communication technologies), and ensuring seamless inferencing through AI algorithms. These aspects make the research relevant to readers of Sensors, as they extend beyond zoological observations to technological innovation.

Question 1: Lines 65-71 are too limited in saying that the application of trap cameras is limited to filtering blank applications, because there are currently a large number of applications based on trap cameras. New literature research needs to be supplemented.

Response 1: Thank you for your feedback regarding the scope of the literature review. We have expanded the discussion to include additional foundational work and have revised the text to emphasise that many existing solutions rely on models designed to classify a large number of species across diverse taxa, rather than tailoring their application to specific geographical regions or ecosystems. This approach often overlooks the unique requirements of monitoring vulnerable species, such as ground-nesting birds, where accurate species-specific classification is critical for conservation efforts. We have updated the paragraph to reflect this distinction and clarify the limitations of these existing solutions. As you can appreciate this is a large field and we have focused out background on platforms that offer similar functionality to what we have proposed in the paper.

Question 2: The meaning of the other gray lines on the P-R curve lacks legends.

Response 2: Thank you for your observation regarding the gray lines on the Precision-Recall (P-R) curve. These lines correspond to the individual class-wise Precision-Recall curves for the 26 classes included in the model. Unfortunately, due to a limitation in the visualization functionality of the Ultralytics framework, it is not possible to include legends for all individual classes when the number of classes is large. However, the blue line in the plot represents the mean performance across all classes, with a mean Average Precision (mAP) of 0.976 at an IoU threshold of 0.5, as indicated in the legend. We recognise the importance of clarity in the visual representation of results and will explore alternative visualisation approaches in future iterations to improve the interpretability of multi-class Precision-Recall curves.

Reviewer 4 Report

Comments and Suggestions for Authors

The present work represents an interesting case study concerning AI-Driven Real-Time Monitoring of Ground-Nesting Birds. The paper is well structured and written and the quality of english allows a correct comprehension of the topic.

Anyway, some little changes have to be performed in order to go ongoing with the publication process. If Authors will folllow the suggestion given, I will certainly recommand the manuscript for the publication. 

Introduction. 

Line 60-64. The contribution of camera traps for monitoring wildlife is quite interesting, expecially associated to ground-nesting birds. Please provide more details about this topic, also if it is not the main aim of the work. I think that this part may contribute to enrich the paper. 

Line 105-109. I think that this part may be removed, it seems too excessive.

Materials and methods

Line 130. Please about this affirmation insert a reference or report past studies using this type of input resolution. 

Line 196. In my opinion Authors may resume the key points of the process, without reporting point by point the several passages. This could help the reader in a easy comprehension of the process. 

Figure 7. It is not necessary in this way. 

Conclusions.

Line 574-587. This part is already in part mentioned in the results section. Authors could skip it or more summaryze it. 

At the end, Authors speak about the importance of the study for empowring new knowledges about bird conservation and ecology. Authors mentioned also the contribution of citizen science in research of this type, but in the rest of the paper this information is not deeply discussed. So please provide a better comprehension for the reader of the role of the citizen science and future challanges that could be achieved with this approach. 

 

Author Response

The authors would like to thank the reviewer for taking the time to review and provide comments on the manuscript.

Question 1: Line 60-64. The contribution of camera traps for monitoring wildlife is quite interesting, expecially associated to ground-nesting birds. Please provide more details about this topic, also if it is not the main aim of the work. I think that this part may contribute to enrich the paper.

Response 1: Thank you for yourcomment. We agree that elaborating on the role of camera traps, particularly for monitoring ground-nesting birds, would enrich the paper. In response, we have expanded the relevant section to provide a more detailed discussion on the utility of camera traps. Specifically, we have highlighted their ability to facilitate non-invasive monitoring, collect critical data on nesting behavior, breeding success, and predator activity, and their importance in addressing conservation challenges associated with habitat loss and predation. This additional information aligns with the broader context of the study and reinforces the significance of employing such technologies in biodiversity monitoring.

Question 2: Line 105-109. I think that this part may be removed, it seems too excessive.

Response 2: We would also like to leave this in if possible as we feel it sets the scene for the for the paper. “The remainder of this paper is structured as follows: Section 2 outlines the meth-odology, including data collection, model training, and evaluation metrics. Section 3 presents the results, including both model performance and real-time inference for the curlew case study, followed by the discussion in Section 4. The conclusions and future directions are presented in Section 5.” But if you feel strongly about this, we can remove it.

Question 3: Line 130. Please about this affirmation insert a reference or report past studies using this type of input resolution.

Response 3: Thank you for your comment, this is an important point. The reported average resolution of 972 x 769 pixels corresponds to the characteristics of the provided training dataset used in this study. As this value directly reflects the properties of the dataset rather than being derived from external sources or past studies, we removed the claim that it aligns with a typical input resolution. To clarify this point, we have revised the text to explicitly state that this resolution is based on the dataset characteristics and not drawn from prior studies.

Question 4:  Line 196. In my opinion Authors may resume the key points of the process, without reporting point by point the several passages. This could help the reader in a easy comprehension of the process.

Response 4: Thank you for your feedback. While we understand the importance of conciseness for improving readability, we believe that retaining the detailed descriptions of the model architecture, hardware, hyperparameters, and inferencing process is critical for the following reasons:

1. Novelty of the Model: The YOLOv10x architecture used in this study is relatively new and has not been deployed in this specific setting before. Including technical details of the model's design and functionality helps justify its application, training process, and the results presented.
2. Reproducibility: We aim to provide sufficient technical details about the hardware and hyperparameters to ensure reproducibility. Many studies omit these details, which can hinder other researchers from replicating or building upon the work. By sharing the specific settings and processes, we contribute to greater transparency and facilitate future research in this area.
3. Cost and Resource Considerations: Details about the inferencing hardware highlight the financial and power-related costs of deploying the AI model, an aspect often overlooked in AI-focused papers. By providing this information, we aim to bring attention to the practical implications of AI deployment, ensuring a balanced understanding of both its capabilities and resource requirements.

Given these points, we respectfully propose keeping the detailed descriptions while ensuring that the information is presented in a structured and coherent manner to enhance accessibility for readers. If this is something you feel strongly about, we can take a further look at it.

Question 5: Figure 7. It is not necessary in this way.

Response 5: Thank you for your comment. We completely agree that the images don’t really add anything to the paper, and we have removed it.

Question 6: Line 574-587. This part is already in part mentioned in the results section. Authors could skip it or more summaryze it.

Response 6: Thank you for your feedback. We have revised the section to summarise the AI-driven classification system more concisely, minimising overlap with the results section.

Question 7: At the end, Authors speak about the importance of the study for empowring new knowledges about bird conservation and ecology. Authors mentioned also the contribution of citizen science in research of this type, but in the rest of the paper this information is not deeply discussed. So please provide a better comprehension for the reader of the role of the citizen science and future challanges that could be achieved with this approach.

Response 7: We appreciate the reviewer’s feedback regarding the role of citizen science in our study. In response, we have expanded the discussion to include not only the contributions of citizen scientists in data collection and model development but also the challenges associated with this approach, such as the absence of expert annotation and data quality issues. These additions provide a more comprehensive understanding of the role and implications of citizen science in AI-driven conservation efforts.