Empowering Wildlife Guardians: An Equitable Digital Stewardship and Reward System for Biodiversity Conservation Using Deep Learning and 3/4G Camera Traps

Round 1

Reviewer 1 Report

In this manuscript, the authors suggest a framework for rewarding wildlife guardians (humans) for successful protection of wildlife. This system is based on a camera trap and machine learning approach in which an automated detection of an individual of a species results in an automated transfer of a (currently arbitrary) amount of money from the “account” of the species to the guardian.

The manuscript is well-written and suggests an interesting and relatively little-discussed framework for wildlife conservation. In general, the manuscript includes a lot of background material on conservation biology and on the ML methods being used – these are interesting and potentially useful to a reader new to this area, although probably not necessary for the main thrust of the manuscript. I would note that the manuscript is a mix of two ideas that fit together loosely but do not need to be presented together – a very thorough (but somewhat standard) description of the fine-tuning of an ML model to recognize five mammal species, and a somewhat cursory description of a potential application of functional ML models and connected cameras to the idea of “interspecies money”.

My specific comments on the manuscript are below:

1.         An important initial question about “interspecies money” is how it would compare to existing conservation funding mechanisms that supply funds to local communities for wildlife protection but do not rely on the relatively “high tech” funding or detection methods described here. These do exist, and some (though not all) have been reported to be successful – what are the pro’s and con’s of this new mechanism compared to the existing? This is not my area of expertise, but many examples are given and discussed in Sodhi and Ehrlich’s textbook, for example.

2.         From a population biology perspective, the idea that each species (rather than each individual organism) gives a payment based on detections raises a number of questions. It’s entirely possible that the population could be shrinking, perhaps by quite a lot, but detections are still positive and the guardians are hence still being paid – should guardians be paid when populations are decreasing by the population is not (yet) extinct/extirpated? It’s possible that the number of detections would be correlated with population size, but this approach would still seem to reward guardians based on population size and not trends. It’s also possible that a population could increase due to no particular action on the part of the guardians. All of this suggest that really what should be monitored and related to payment is abundance, not number of detections. This information is not readily available from camera traps – what would be the downside (compared to the approach proposed) to doing a simple yearly population census using traditional survey methods?

3.         In what ways is this approach better or worse than the original “Interspecies Money” concept based on individual identification? As individual ID from camera trap imagery is undoubtedly a more difficult problem than species ID, this new framing is certainly more practical – but is it in fact better? What does it mean to ask a guardian to protect a population (effectively) rather than an individual? In what ways might the help (or dilute) a guardian’s sense of responsibility and control?

4.         The way in which the evaluation datasets were generated (L206) could use additional explanation – why were 10 separate data sets created and averaged? What exactly is the “margin for error” and “confidence interval” in this context?

5.         Obviously no classifier will (probably ever) be perfect at identification. How will the proposed payment system handle this, and what repercussions might this have for its viability?

6.         Related to the above, do the authors envision a human in the loop reviewing classification? This is potentially feasible to detect false positives, but what about false negatives? It seems very probable that situations will arise where a guardian observes or claims that an animal was present but not detected and requests payment.

7.         Any system like this will be open to the possibility of cheating. In what ways might this occur and how could it be mitigated? L525 for example raises the idea that only one individual of the species needs to be present and repeatedly walk by the camera for successful payment, while all other individuals in the population could be poached.

8.         It is not clear why an object detecting model is needed for this application – is it possible that a model not intending to predict bounding boxes might have higher PR for the image as a whole?

9.         The “results” focused on the BioPay approach are fairly cursory – really there’s nothing more to report other than the number of detections per species during the deployment period.

Author Response

Reviewer #1 Comments:

==================

The manuscript is well-written and suggests an interesting and relatively little-discussed framework for wildlife conservation. In general, the manuscript includes a lot of background material on conservation biology and on the ML methods being used – these are interesting and potentially useful to a reader new to this area, although probably not necessary for the main thrust of the manuscript. I would note that the manuscript is a mix of two ideas that fit together loosely but do not need to be presented together – a very thorough (but somewhat standard) description of the fine-tuning of an ML model to recognize five mammal species, and a somewhat cursory description of a potential application of functional ML models and connected cameras to the idea of “interspecies money”.

The authors would like to thank the reviewer for their kind word support for our work. We really appreciate it. We recognise the paper is a little long, but we hope this will provide a comprehensive set of carefully selected references for interested readers to build on.

1. An important initial question about “interspecies money” is how it would compare to existing conservation funding mechanisms that supply funds to local communities for wildlife protection but do not rely on the relatively “high tech” funding or detection methods described here. These do exist, and some (though not all) have been reported to be successful – what are the pro’s and con’s of this new mechanism compared to the existing? This is not my area of expertise, but many examples are given and discussed in Sodhi and Ehrlich’s textbook, for example.

Thank you for raising this question. Yes this is really interesting. There are examples of micropayments that we reference in the paper but far too little of the global biodiversity budget, if any, is reaching the poorest people on the planet who often live in the most biodiverse regions of the world. COP26 pledged 1.7 billion dollars annually but this is far less that the 143 billion dollar budget. We do not yet know how Interspecies Money will compare as there is no realiable system in place to evaluate it. And this would need to be a longitudinal study if we are to measure impact. But it is plausible. It is a radical shift that will likely empower and improve the lives of poor people and the animals they live with but there will be a number of significant problems to overcome before it is fit for purpose. Even our own monitory systems are not without problems but they have matured and become much more sophisticated over time. BioPay will have the benefit of building on the success of our own financial institutions therefore initial installations will be largely successful in terms of money transactions but there will  be a number of different issues to address in terms of how the animal-guardian mappings will operate. In this sense Interspecies Money is high risk high gain. It is not an incremental technology but foundational  in nature and this makes it very difficult to compare with other approaches as nothing quite like it exists. It is hoped that this paper will spare interest in the area and move research in this area to fully understand the utility of Interspecies Money and  the impact it has on biodiversity management.

2. From a population biology perspective, the idea that each species (rather than each individual organism) gives a payment based on detections raises a number of questions. It’s entirely possible that the population could be shrinking, perhaps by quite a lot, but detections are still positive and the guardians are hence still being paid – should guardians be paid when populations are decreasing by the population is not (yet) extinct/extirpated? It’s possible that the number of detections would be correlated with population size, but this approach would still seem to reward guardians based on population size and not trends. It’s also possible that a population could increase due to no particular action on the part of the guardians. All of this suggest that really what should be monitored and related to payment is abundance, not number of detections. This information is not readily available from camera traps – what would be the downside (compared to the approach proposed) to doing a simple yearly population census using traditional survey methods?

Thank you for these questions. In terms of “It’s entirely possible that the population could be shrinking, perhaps by quite a lot, but detections are still positive and the guardians are hence still being paid – should guardians be paid when populations are decreasing by the population is not (yet) extinct/extirpated?” This is a difficult question. We are working at the moment with population biology scientists to see whether we can calculate population size using Conservation AI. At the moment we do not have answers for this. However, camera trap configurations will play a part in how animals are detected and what the expected detection rate will be. In the paper you will see that Zebras were seen far more than say a cheetah. This is because the population size of Zebras was significantly larger than cheetahs. If for some reason the Zebra population size started to decrease we would expect the number of detections to also decrease. However, this would need a separate study in itself to provide any kind of sensible answers. If a population was decreasing, then maybe this would trigger an intervention/investigation. For example, they may be an imbalance between predators and grazing animals. In this case you may need to redress the balance by boosting underrepresented animals and relocating predators to other areas. This is a problem with lions at the moment along the eastern flank of South Africa where they are overkilling animals such as antelope because of their large numbers. The hope is the local community (guardians) would interject and protect animal populations in order to protect their income. These of course are very difficult questions to answer. In terms of yearly assessments this would not be practical. Remember, guardians are likely to be among the poorest in the world. The whole idea is to propose a monetary system that rewards them immediately for the work they do. Many of these will rely on that money to live and will not necessarily be able wait a year for money to emerge for services they have done throughout that year. If people become desperate then they may resort to poaching or the over hunting of animals to survive. As the literature in the paper shows there is a serious problem with conservation. We are literally sleep waling into a six-mass extinction. Therefore, radical ideas, such as Interspecies Money\BioPay are needed to look at alternative ways to manage biodiversity. It may work, it may not – it is still far too early to tell. But what this paper does is open up the debate and raise interesting questions like the reviewer has done.  

3. In what ways is this approach better or worse than the original “Interspecies Money” concept based on individual identification? As individual ID from camera trap imagery is undoubtedly a more difficult problem than species ID, this new framing is certainly more practical – but is it in fact better? What does it mean to ask a guardian to protect a population (effectively) rather than an individual? In what ways might the help (or dilute) a guardian’s sense of responsibility and control?

Thank you for raising these questions. In terms of which is better “Interspecies Money” and BioPay – there is no comparison at the moment. We mention in the paper that Interspecies Money is primarily based on facial recognition whereas BioPay uses species detection. There is no comprehensive study on the facial recognition aspect of Interspecies Money but we currently do have a number of RAs working in this area. Once a solution has been developed (we are using FaceNet to facially recognise Orangutans) we will do a comparison of results. BioPay maybe a much broader implementation of Interspecies money which might have a broader application in conservation and biodiversity management. Yes we agree, doing facial recognition with camera traps is extremely difficult. We have tried this with lions with mixed results. You really need high resolution images to do this. Therefore, BioPay would be more practical. Whether it is better or not is still too early to say. We need more data from experimental deployments. In terms of your question on “What does it mean to ask a guardian to protect a population (effectively) rather than an individual” is a good question. These subtleties are still difficult to answer. However, there are examples of both methods, i.e., farmers managing cattle, or specialised units looking after and training individual animals like orphaned gorillas. Animals need to be valued from a number of different points of view and models need to be developed accordingly to understand how Interspecies Money and BioPay would work. We think this is one of the nice aspects of this work – it opens up far more questions than the paper answers which is indicative of a good research idea. Hopefully, many of these questions will be picked up by the research community because it is going to take a large number of researchers to address the problems that arise while successfully implementing the Interspecies Money/BioPay concept. In term of your question “In what ways might the help (or dilute) a guardian’s sense of responsibility and control?” the rewards received for caring from animals will be a strong incentive. It is worth noting that guardians are likely to be among the poorest people on the planet who are in need of money for food, shelter, education and healthcare. The idea is that this will be a self-regulatory system where animals and the poorest people among us benefit from Interspecies Money/BioPay. We hope these explanations go some way to answering the questions raised by the reviewer.  

4. The way in which the evaluation datasets were generated (L206) could use additional explanation – why were 10 separate data sets created and averaged? What exactly is the “margin for error” and “confidence interval” in this context?

The authors would like to thank the review for raising this issue. Using multiple folds (10 in this instance) is a common approach used in object detection which provides a more comprehensive understanding of the performance of and object detection model because it:

• Increases diversity: object detection models perform differently on different datasets due to variation in object types, sizes, orientations, and backgrounds. Using multiple datasets increases the diversity of the data which allows the model to be evaluated using a wide range of scenarios.
• Ensures robustness: Models that perform well on a single dataset may not necessarily generalise well to new datasets. Evaluating the model on multiple datasets and averaging the results assesses the robustness and ability to perform consistently across different datasets.
• Provides a fair comparison: Comparing the performance of different object detection models is important to evaluate them on the same datasets to ensure a fair comparison. Using multiple datasets and averaging the results, reduces the impact of dataset bias and allows a more reliable estimate of the model's true performance to be obtained.
• Gives more representative results: object detection evaluation metrics, such as mAP, can be sensitive to small variations in detection results. By averaging the results across multiple datasets, we can obtain more representative and stable evaluation metrics that reflect the model's overall performance.

We hope this addresses the reviewer’s concern.

5. Obviously no classifier will (probably ever) be perfect at identification. How will the proposed payment system handle this, and what repercussions might this have for its viability?

The reviewers would like to thank the reviewer for raising this interesting issue. We do mention this is the paper “The BioPay service performed as expected and the results show the successful transfer of funds between animals and the associated guardian. The detection results during inference show that overall detection success was high with a small number of miss-classifications. This would mean that money between species accounts would be made when in fact that animal was not actually seen. This will always be a difficult challenge to address, but a small margin of error in this case is negligible. Obviously, this may not be the case when species are appropriately valued where highly prized animals could transfer large amounts of money when they are misclassified. This will need to be considered in future studies.” How you address this is difficult which is a problem we face in our own financial offerings. For  example, in the case of childcare benefits in the UK. If you earn above a certain threshold you need to inform child benefit so that the payments can be stopped. However, not everyone does this and they continue to get paid regardless. The same happens with unemployment benefits where people claim benefits but continue work for cash in hand. As with this case a robust BioPay system would have to have in places to detect such occurrences possibly through auditing, and moneys would have to be re-distributed to the rightful places. We would like to thank the reviewer for pointing this out as these subtle questions will greatly help develop BioPay moving forward. We hope this answers the reviewer’s questions.   

6. Related to the above, do the authors envision a human in the loop reviewing classification? This is potentially feasible to detect false positives, but what about false negatives? It seems very probable that situations will arise where a guardian observes or claims that an animal was present but not detected and requests payment.

The authors would like to thank the reviewer for raising these questions. Yes, there would need to be a human in the loop. A BioPay system like any other financial would need to be carefully audited. In BioPay the classifications would need to be mapped to payments. In terms of false negatives yes this may be a problem and guardians would have to have the right to raise concerns and appeal decisions. Again, this problem occurs in any financial system you have. Whether it is buying goods, getting building works done or investing money. Where money is involved, there will always be disagreements. We have added this concern  to the paper. We hope this satisfies the reviewer’s concern.  

7. Any system like this will be open to the possibility of cheating. In what ways might this occur and how could it be mitigated? L525 for example raises the idea that only one individual of the species needs to be present and repeatedly walk by the camera for successful payment, while all other individuals in the population could be poached.

The authors would like to thank the reviewer for raising this interesting point. The short answer is you could cheat the system in theory. When dealing with money there needs to be some very serious safeguards in place in much the same way we have with people. We do not confess to knowing all requirements for BioPay and we are currently working on a much bigger project to connect BioPay to real financial systems, but it will be some time before we fully understand the strengths and weaknesses of the system. In practice, such as the study in South Africa, it would be highly unlikely for this to happen as every model we train has person and car in it. Therefore, we would likely see misbehaviours in the images (not always, but it would be difficult to continually fool). You would notice animal behaviours that are different, such as herding. Maybe you could use anomaly detection to notify appropriate regulators when things are different – much like we get with the grading in fast food restaurants and takeaways. Any financial system would need to be regulated and if necessary, the uses of that system may need to be penalised in some way. For example, we have this with personal credit scores – if you are a good payee, you get good deals, if you are not, you get more expensive deals or none at all. This would be the same for BioPay, if people misuse it then they will not be able us and thus will not benefit from it. In this case you may find it becomes a self-regulatory system because local communities as a whole will not risk losing that financial revenue stream. These are questions for future work. The main idea was to get out the idea of Interspecies Money and show how it could be achieved using a very small prototype system. We hope that this answers the reviewer’s questions.

8. It is not clear why an object detecting model is needed for this application – is it possible that a model not intending to predict bounding boxes might have higher PR for the image as a whole?

Object detection is a crucial component in the approach. You would not necessarily need to show the binding boxes, but it helps to show the reader what objects have been detected and what it thinks the class is (important for transmitting funds from one account to another). You probably would get a higher PR for image classification but what happens when there are multiple species in the same image, for example a zebra and a rhino (if it classifies an image as a zebra then the funds will be transmitted from the zebra account to guardian and the rhino will be ignored – and vice versa). By using binding boxes, you identify the individual species and transfer the funds from both accordingly. We hope this addresses the reviewer’s concern.  

9. The “results” focused on the BioPay approach are fairly cursory – really there’s nothing more to report other than the number of detections per species during the deployment period.

The authors would like to thank the reviewer for their comment. We appreciate that BioPay is in the early stages of development and in the future work we do say that there is a great deal of work needed to actualise this. In this paper we just wanted to demonstrate the idea. We are currently working on a project to implement BioPay which we will hopefully be able to publish next year. We have extended the section a little based on the comments from reviewer 3 to include a system diagram of how it works – but this is only abstract.

Reviewer 2 Report

The paper provides an interesting model tying conservation to generating funds. Faster R-CNN is well-known architecture, and the details are well known and do not add anything to the major argument presented in the paper and can be omitted.

Table 5 is not necessary but can be replaced by means, medians, standard deviation and range.

More details on system architecture showing how the image from the camera is transmitted, inferred, and how the system is integrated with the pay system is required.

One key issue is that of “ghost” images.  They call these “blanks” and these are typically not easy to detect.  How are these detected?

Author Response

Reviewer #2 Comments:

==================

The paper provides an interesting model tying conservation to generating funds. Faster R-CNN is well-known architecture, and the details are well known and do not add anything to the major argument presented in the paper and can be omitted.

The authors would like to thank the reviewer for raising this issue. We have submitted several papers on object detection and reviewers have mixed opinions about adding the details of the Faster RCNN. Many conservationists we work with are interested in understanding the mechanisms used in a Faster RCNN when detecting different species. We personally think that this is fundamentally important as it informs camera trap layout and species tagging. We agree that many may not find the details interesting but there will be a number of readers who would like to understand the details. We believe that we provide a much clearer description of the Faster RCNN than you will find in other publications – particularly IEEE Transactions where it most often appears – which are mathematically heavy and very difficult to read. Removing it would push conservationists to think of it as a black box, which we do not think is appropriate as more advanced technology becomes more prevalent in conservation studies.

Table 5 is not necessary but can be replaced by means, medians, standard deviation and range.

The authors would like to thank the reviewer for raising this issue. We have had conflicting views when adding tables like this is – some say add them and some reviewers say remove them. The other two reviewers have not requested this table be removed so concerned this may upset them we would like to keep this table in. We personally think is provides value and demonstrates that the variability between random sampling is minimal and that the sample strategy is fit for purpose in terms of model evaluation. We hope the reviewer understands our rationale.

More details on system architecture showing how the image from the camera is transmitted, inferred, and how the system is integrated with the pay system is required.

The authors would like to thank the reviewer for raising this issue. We have now added a diagram in the BioPay section to show how the end-to-end system works with the integrated pay system. We hope this satisfies the reviewer’s concern.

One key issue is that of “ghost” images.  They call these “blanks” and these are typically not easy to detect.  How are these detected?

The authors would like to thank the reviewer for raising this issue. We simply defined blank images as images that do not have any of the trained objects in the model, i.e., if there are no Equus quagga, Giraffa camelopardalis, Canis mesomelas, Crocuta crocuta, Tragelaphus oryx, Connochaetes taurinus, Acinonyx jubatus, Loxodonta africana, Hystrix cristata, Papio sp, Panthera leo and Rhinocerotidae objects detected in the image then we call this a blank/ghost image. We have made this clearer in the paper. We hope this satisfies the reviewer’s concern.

Reviewer 3 Report

The manuscript is well written and well structured but some adjustments are necessary before going ahead in the publication phase. If the authors will follow the suggestion given I will certainly recommend this manuscript to be published on Remote Sensing.

 Majors adjustments:

1. I suggest better discussing in the introduction section the role of remote sensing in detecting animals also using drones and monitoring their habitat and land cover changes thanks to satellites well as the disease risk. To do so consider including the following references:

- https://doi.org/10.3390/rs15092348

- https://doi.org/10.3390/rs15010178

- https://doi.org/10.3390/app13010390

Also, consider including section 2 as a subsection of the introduction and depicting at the end the aim of your study at the end of this section.

2. In material and methods consider to include a section study area providing the area in which you perform your study and the location of the cameras. Then better discuss the software adopted to process these images the workflow followed and in an appendix the main characteristics of each device FOV etc...

3. In the discussion consider discussing limits and future perspectives

4. The conclusion is too long try to sum up all in max 10-15 rows. You may move the other parts in the discussion.

Minor adjustments.

1. provide the settings in the appendix of the model adopted and the code to permit scalability.

Author Response

Reviewer #3 Comments:

==================

The manuscript is well written and well structured but some adjustments are necessary before going ahead in the publication phase. If the authors will follow the suggestion given I will certainly recommend this manuscript to be published on Remote Sensing.

The authors would like to thank the reviewer for their kind words and support, it really means a lot to us.

1. I suggest better discussing in the introduction section the role of remote sensing in detecting animals also using drones and monitoring their habitat and land cover changes thanks to satellites well as the disease risk. To do so consider including the following references:

- https://doi.org/10.3390/rs15092348

- https://doi.org/10.3390/rs15010178

- https://doi.org/10.3390/app13010390

The authors would like to thank the reviewer for their comment. We have read the reviewers papers and feel that they do not fully relate to the idea proposed in the paper, but we do recognise that it is an alternative sensor therefore we have added them to the paper with some text highlighting the different sensor technologies that exist. We do a lot of drone work with Conservation AI, and we have published several papers in journals that talk about their use in conservation. In the context of Interspecies Money, drones would be useful as an alternative to species detection but there are several technical challenges that need to be addressed. Particularly when you would like to include smaller animals, such as Pangolins, which are very difficult to detect using consumer drones (we have worked in several areas Africa and many places do not have the money for Matrice drone with a Z30 camera would be needed to detect small animals at high altitudes. We are talking about rewarding extremely poor people who ultimately would not have access to expensive drones or satellite technology.

Also, consider including section 2 as a subsection of the introduction and depicting at the end the aim of your study at the end of this section.

The authors would like to thank the reviewer for there comment. We understand the rationale but feel the message would be diluted if it was changed to a sub header of the introduction. There is an important message in section two that is fundamentally important for conveying the Interspecies Money concept. We would prefer to leave the section as is and avoid upsetting the other two reviewers who did not share the same sentiments.

2. In material and methods consider to include a section study area providing the area in which you perform your study and the location of the cameras. Then better discuss the software adopted to process these images the workflow followed and in an appendix the main characteristics of each device FOV etc...

The authors would like to thank the reviewers for their comments. We have mentioned in the paper in several places that the study was conducted in Welgevonden Game Reserve in Limpopo Province in South Africa, but we have made this explicit at the beginning of the methodology. We are unable to provide a detailed make of the area with specific camera locations due to the sensitivity and the very high-risk animals we were monitoring (i.e., Black and White Rhino) which are highly prized and constantly under threat of being poached. Revealing the locations will allow poachers to remove and secure areas before  poaching commences. We hope the reviewers understand the sensitivities and understand why we cannot reveal the requested information.

3. In the discussion consider discussing limits and future perspectives

The authors would like to thank the reviewer for raising this issue. We have now added some further details on the limitations of the approach that need to be addressed. We have also put forwards some future perspectives and highlighted how the system might progress.

4. The conclusion is too long try to sum up all in max 10-15 rows. You may move the other parts in the discussion.

The authors would like to thank the reviewer for raising this issue. We have now tried to shorten the conclusions. We hope that this satisfies the reviewer’s concern.

Minor adjustments.

1. provide the settings in the appendix of the model adopted and the code to permit scalability.

Thank you for your comment. We have included the hyperparameter settings in the paper which shows how we configured the Faster -RCNN – these parameters are set in the pipeline.config file that comes with the model  checkpoint. Access to the models is free via the Conservation AI platform (http://www.conservationai.co.uk). Upon registration we are happy to discuss Rest-API access. We have now made this clear in the paper. We hope this satisfies the reviewer’s concern.  

Round 2

Reviewer 3 Report

The authors well followed the suggestion given and the manuscript is now suitable for pubblication.