Using artificial intelligence (machine vision) to increase the effectiveness of human-wildlife conflict mitigations could benefit WAW

1. Overview

This report explores using artificial intelligence (AI) to increase the effectiveness of human-wildlife conflict (HWC) mitigations in order to benefit wild animal welfare (WAW). Two concrete examples are providing more funding, research and direct work into reducing fatalities due to 1) collisions between bats and wind turbines, and 2) culling crop-raiding starlings. The report aims merely to raise awareness of this topic and introduce the idea for discussion, but not yet strongly suggest it is a cost-effective intervention on par with other interventions-see uncertaintieslimitations, and potential for harm.

What's the problem profile?

  • HWC is increasing due to human expansions and climate change, (Gross et al., 2021) and is starting to be considered in government strategies and policy. The expected future impact of innovative and effective solutions to HWC could be even larger than currently appreciated.
  • Lethal control or other methods which significantly impact animal welfare are still widely used (such as culling), despite preventative non-lethal strategies growing in more recent wildlife management approaches.
  • Currently deployed AI systems directed towards HWC could be expanded further within the next 10-20 years as they become more reliable, more effective, and cheaper. We should not assume they will prioritize WAW concerns, or be widely used for animals of WAW concern, so this should be embedded before they are potentially rolled out at scale.
  • There are already companies working on AI solutions for specific problems involving endangered species, such as protected areas using AI assisted technology for poacher detection. There is already proof-of-concept of an NGO-backed early warning AI system, ‘WildEyes’, with this type of solution being invested in by a local governmental department in Tamil Nadu, India. Buy-in from a range of stakeholders (especially when it benefits humans and profits too) offers a way in with conservationists and researchers who may not otherwise consider WAW. Research and development (R&D) on AI-assisted HWC mitigations would likely attract researchers who would not otherwise consider or be motivated by WAW concerns.

What should we be doing differently?

A very tentative theory of change: if machine vision-based methods prevent HWC, they could be adopted, even on a small scale → helps drop prices → allowing for systems to be more widely adopted → leads to more support and R&D → continued price drops and adoption → could create space for legislation to ban harmful or lethal methods of animal control → preventing HWC could reduce apathy and antagonism towards “problem species” and make it easier for people to consider the welfare of animals, while also directly reducing negative WAW effects of HWC.

  • This report highlights two examples of HWC where advocates could influence AI-assisted mitigation to directly affect substantial numbers of animals, and spread welfare considerations in software and norms:
    • Wind turbine collisions are a leading anthropogenic cause of bat deaths and cause a significant number of bird deaths (600,000 to 949,000 bats and 140,000 to 679,000 birds annually in North America). We should expect fatalities to increase due to expansions in wind power.
    • Culling of crop-raiding species. In one year, the USDA’s Wildlife Services culled 1,028,642 European starlings responsible for agricultural crop damage, because other mitigations are ineffective. Despite this, starlings still cause extensive damage each year. More effective mitigation measures would hold value and could prevent culls.
  • There are a number of research and advocacy directions to address these:
  • Conduct cost-effectiveness estimates on different HWC mitigations; in some cases, Randomized controlled trials (RCT) and Cost-Effectiveness Analyses (CEAs) may find mitigations other than AI-assisted ones are more effective.
  • Improved research and data collection of how many animals HWC affects. For example, estimating bird and bat turbine fatalities is difficult, (hence wide ranging estimates); utilizing cameras and automated detection could help us collect more accurate data.
  • Research on improving accurate small object detection from a distance, so that automated detection and curtailment systems at wind farms can reliably detect smaller species.
  • Utilizing machine vision techniques in long-wave infrared cameras placed near wind turbines could allow them to detect bats and avoid fatalities.
  • Researchers could apply machine learning techniques, utilizing large amounts of existing data on well-studied animal populations, or use machine vision to gather this data to predict where mitigation will be most effective. E.g., identifying high-risk areas for wildlife where turbines should not be erected.
  • R&D into flock detection systems which could be integrated with automated deterrents to prevent wide-scale crop damage by flocks of starlings - if effective, this could be adopted in favor of wide-scale culling.
  • Extending protected status to a wider number of bats, and utilizing this to get leverage over wind energy companies.
  • Adding more bird species/a species with a large population to protected species acts and/or automated curtailment detection lists covered by already deployed AI mitigation technologies could reduce their fatalities by >60%.
  • Funders and researchers can also engage in institution-building to create a robust interdisciplinary field at the intersection of AI development and wildlife management:
    • Establish an interdisciplinary conference tying experts from engineering, AI, animal welfare, ecology, animal behavior, conservation, and perhaps even from the social sciences to discuss examples and opportunities for AI to increase the effectiveness of HWC mitigations, or even wildlife management methods more generally, that improve animal welfare.
    • Create an accessible applied methods journal for publication of this interdisciplinary work.
  • See section 3 for the uncertainties and section 4 for the limitations of this report.

2. How much would interventions speed up adoption of effective AI mitigations?

I think the development of AI-assisted technology that could help alleviate HWC will continue. Forecasts from respected researchers (Cotra 2022) and community platforms like Metaculus[1], give good reasons to think AI systems might be readily capable of much more WAW work within 10-20 years. To give an example of the speed of adoption of HWC mitigation technology, WildEyes took just over five years to go from idea to development to deployment. However, it does benefit from addressing a well-known HWC (human-elephant conflict) and a charismatic species of public concern. Technological solutions proven to work better, or at least as well, as current mitigations are better candidates for being adopted even without added pressure from the WAW movement. Currently endangered species are both the lowest hanging fruit and the most likely to already be assisted without further effort. However, these are almost by definition affecting small numbers of individuals. I am skeptical that AI-assisted mitigation systems will be widely deployed for the other species WAW focuses on, or that they will be developed with animal welfare considerations at the center, unless there are deliberate interventions. The crossover of WAW and AI in HWC is a minority concern. For example, WILDLABS is a network bringing together a global conservation technology community, but it is not specific to AI technology and only 4% of all members belong to the HWC group. There are also special issues in journals on AI in conservation, e.g., in Sensors, and there are some articles in high impact journals like Nature and Frontiers, however a focus on animal welfare is missing - and there needs to be more opportunities for interdisciplinary, rather than field-focused, work and discussions. The possibility of AI systems being rapidly rolled out in 10-20 years adds urgency to trying to get WAW concern established in AI system use, as opposed to trying to establish this later when the systems are already widely deployed without WAW considerations.

3. Uncertainties

As this is an initial report with very preliminary/foundational ideas, there are some uncertainties:

  • Key counterfactual: If randomized control trials and cost-effectiveness analyses find non-AI mitigations are significantly more effective than AI assisted mitigations.
  • Number of animals affected: Existing HWC estimates are mostly regional, but better data on national and global numbers of animals could make this a more or less pressing area to work on. However, estimates of bat deaths alone suggest working on this would be impactful.
  • Species-specific problems: It may not be possible to customize system parameters for the species we care about most.
  • Conflict-specific problems: The physical locations where most conflict occurs may not always be amenable to these technologies.
  • Social factors in HWC: To the extent that attitude change is an important outcome of such interventions beyond the direct impact on animal fatalities, there is substantial uncertainty about whether and how it is affected. It has been suggested that antagonism towards animals or species may remain robust over time and there are deeper-seated factors (such as misinformation or cultural beliefs) that may lend to this (Dickman, 2010). Therefore, even if HWC is resolved, perceptions of animals and their treatment may not alter. However, this may only be applicable in cases where there is uncertainty about whether the “problem animal'' still poses a threat. It is often recommended that awareness raising, education, and outreach go hand in hand with mitigating certain HWC, so this would also need to be considered.
  • Harms: Dual-use technologies could be used to make animal lives worse. Due to the great uncertainty about the net suffering of wild animals, AI-assisted HWC mitigations could increase other types of suffering. See section 8.3 for further details.
  • Undesirable outcomes/trophic effects: For example, we do not know what the implications of a surplus ~1 million starlings per year would be if culling was prevented. Would they feed on neighboring, AI-unguarded crops, and would that create undesirable economic disparities or displace HWC fatalities to another location? What ecosystem impacts would we see?

4. Limitations of this report

  • Full scope of technology not addressed: There are AI-assisted HWC mitigations I have not been able to address in this report, such as the use of drones. From initial expert outreach, their use may be limited based on laws and regulations, and their effectiveness in the field may depend on how easy it is for end-users to utilize the technology. Though it could be worthwhile for further exploration and investigation of the use of drones. (Singer & Tse, 2022) also discuss expanding wildlife-vehicle collision prevention AI to smaller animals.
  • Neglecting HWC of most concern in the Global South: When writing this report with WAW in mind, the examples I discuss may largely be applicable to countries in the Global North (and specifically North America because of data and study availability). The stakeholders for the mitigations I discuss may be conservation organizations, farmers, and larger companies. However, people and areas that may need effective solutions the most are protected areas, national parks, and crop farmers in the Global South (which is usually underfunded in comparison). With any research or funding coming from the Global North that contributes to work on HWC in these areas, there is a need to ensure work is being carried out in collaboration with affected communities, and to invest in local capacity-building. This may add value to AI-assisted systems that can be one-off installations and may not require substantial long-term funding.

5. Background on human-wildlife conflict

Broadly, HWC refers to interactions between humans and wild animals with negative consequences for people and their resources and/or for wildlife and their habitats (e.g.,SARPO, WWF, 2005IUCN SSC HWCTF, 2020). Examples of HWC directly negatively affecting humans include wild animals damaging infrastructure, crops and other vegetation, and attacking livestock. People often kill wildlife to mitigate or even retaliate against such threats. Other examples of HWC directly negatively affecting wildlife include destruction of habitats for human developments, hunting, wildlife-vehicle collisions, and wind turbine collisions.

In many regions of the world HWC is increasing as a result of human population growth, encroachment, and transforming land use (e.g., agricultural expansions, urbanization, and infrastructure development). This pushes humans, human structures, and agricultural practices in closer proximity to wildlife, raising the likelihood of interactions and natural resource competition - a key cause of HWC (Gross et al., 2021). HWC has become a global concern, particularly in terms of conservation, sustainable development, and food security. Several countries are now explicitly beginning to include HWC in national policies and strategies, not only for wildlife management, but also in poverty alleviation and development (FAO & UNEP, 2020, p. 17).

Climate change is further intensifying HWC by exacerbating resource scarcity, increasing competition and the need to share spaces even further (examples of this can be found in Abrahms, 2021). Whilst climate change-driven HWC hasn't seen significant research attention, governmental bodies are increasingly recognizing this challenge and developing policies which incorporate climate into the management of certain HWCs (Bhatt et al., 2018). In the same vein, research and development of HWC mitigation strategies and interventions could be framed as part of climate adaptation. Therefore, there may be an opportunity for the WAW community to shape such policies.

Current HWC efforts are largely reserved for charismatic species, or species of key conservation concern, or where human-wildlife conflict leads to economic and social damages to humans - this sidelines the animal welfare implications. For example, Anand and Radhakrishna (2017) reviewed HWC trends in India and found that all of the publications they accessed focused exclusively on the negative impacts of wildlife on humans. Working to emphasize animal welfare in HWC mitigation could help improve WAW right now. The direct impacts to WAW of current HWC practices include the number of animal deaths as a result, which are often stressful and perhaps even drawn out (e.g., animals chased in retaliatory killings or poisoned) and presumably very painful (e.g., train and turbine collisions). HWC may also exacerbate negative perceptions of wildlife among the people affected, which is associated with reduced support for species conservation (Goodrich, 2010). Humans can suffer economic losses, loss of life, and trauma from conflicts, which reduces tolerance levels for wildlife and is associated with increased apathy for the welfare of wild animals (this applies at a community level, even among people who have never personally experienced HWC (Dickman, 2010)). Therefore, preventing HWC could potentially alter the perception of “problem animals.” If preventing HWC improves attitudes toward affected animals, it may make it easier for people to consider the welfare of those animals.

There are many methods implemented to respond to different HWC. In the case of elephants alone, more than 80 non-lethal mitigation strategies have recently been identified and compiled in a human-elephant coexistence toolbox resource. However, for other species, non-lethal methods of prevention are lacking. Many HWC mitigations are reactive and some lead to deaths or significant impacts on animal welfare. Developing preventive methods that do not compromise animal welfare and are more animal-friendly should be prioritized, with a particular need for effective mitigations to avoid more extreme and harmful solutions such as lethal control. When it comes to mitigating HWC, the integration of several strategies is often required for effective reduction in conflict. Solutions preventing HWC from occurring in the first instance would provide the greatest benefit.

6. Background on current use of AI techniques

The utilization of AI in conservation continues to grow[2]. AI techniques and methods can be utilized to establish effective systems in HWC mitigation, and in an ideal world, prevent conflicts from occurring in the first instance. Machine vision has a key role to play in this and has been developed successfully in other conservation areas.

Deep learning tools such as machine vision with artificial neural networks, specifically convolutional neural networks (CNNs). CNN algorithms can learn (i.e., increasingly improve their ability) to execute tasks with examples. This is an example of a supervised ML technique, as a subject expert needs to pinpoint features that will be used to identify and classify objects in images. The CNN architecture employs filters to sequentially identify the most important features of an object and finally classify the image output in binary (e.g. bird/not bird) or multiple classes (e.g., small bird/medium sized bird/large bird). CNNs have been used successfully for automatically identifying, classifying, and monitoring wild animal populations using images and videos primarily from camera trap data (Norouzzadeh et al., 2018). These methods are used relatively widely in conservation and academia - initiatives such as 'Wildbook' help to facilitate this. Wildbook uses AI to identify individual wild animals from different species to create a wildlife image inventory for data collection. It is an open-source, user-friendly software platform to help collaborative projects store and manage wildlife data for scientific studies. It’s used by over 900 wildlife researchers globally to track over 118,000 individual animals and collect data on populations from over 50 species[3].

Due to rapid development in the field, CNNs have now become a more robust option for automated animal detection. Though published material on research and development for its specific use in HWC mitigation exist, it is still in its infancy in comparison with other conservation uses, and is a lesser utilized method of harnessing such techniques. Below, I review a case study of a scalable and tractable AI-assisted system for mitigating HWC, which has been deployed and proven more effective than typical mitigation methods using RCTs, while identifying where and how welfare considerations can be maximized. Secondly, I will discuss a potential theory of change for how increasing the effectiveness of HWC mitigation could have significant benefits to WAW through eliminating the need for lethal methods. The final example of how AI could improve HWC mitigation is informing where/when mitigations should be implemented to be most effective.

7. Using AI-assisted systems for HWC mitigation

1. A case study: preventing bird-turbine collisions

This is an example of a successful AI intervention that has proven effectiveness and reliability in solving a HWC problem better than existing methods, at scale. It is estimated that 600,000 to 949,000 bats and 140,000 to 679,000 birds die from turbine collisions per year in North America alone (Choi et al., 2020). It is likely that wildlife collisions will continue to increase due to wide-scale wind energy development across the globe. The International Energy Agency states wind additions doubled in 2020 and additions through 2026 may be almost 25% higher on average than between 2015 and 2020. Predictions for the number of terawatt-hours of electricity generated from wind power globally in 2030 are also 2.7x higher than published records for 2020. This rapid expansion of the wind energy industry will lead to increases in animal injuries and deaths, which makes it valuable to develop effective methods for preventing this HWC. AI solutions have been intentionally designed to detect certain birds falling under protected species legislation, with successful deployment in several countries. There is potential for: 1) advocates to utilize a similar mechanism of getting government leverage over companies; 2) expanding this to the protection of a wider number of bird species; 3) exploring research and development for how this system could be used to prevent mass bat mortalities.

The system: automated informed curtailment

Generally speaking, these systems are programmed to: identify specific birds/classes of birds → decide if the collision risk is great enough → if so, trigger curtailment upon detection (temporarily shutting off the blades) and/or other deterrents such as sound → thus preventing collision.

Example product: IndentiFlight

The ‘IdentiFlight’ system can be used to detect eagles within 1km of wind turbines (it has been tested for detection of birds as large or larger than American kestrels (Falco sparverius)).

The system utilizes a CNN and is made up of a stereo unit that captures pairs of images eight times per second while following a bird's flight path, and calculates distance to the bird every 125 meters per second. Bird dimensions are estimated from the images using the distance and fixed angular field-of-view of the cameras. IdentiFlight calculates ~200 image attributes (e.g estimates of body length, wingspan, posture, color, etc.) which are used to determine bird classification with a confidence value (how well the calculated attributes match the pre-trained library of eagle attributes in the system). It can quickly discriminate between objects too far away, too large, or too small. The system continuously scans the sky in a 360-degree radius and can follow a bird's flight path continuously (unless it goes out of view and then it gets 'reset' i.e. detected as a different bird). It chooses the classification with the highest confidence for the whole record of the bird flight (i.e., as the bird and its features become more visible, confidence levels increase). If the bird moves into non-ideal perspective again the highest classification is still preserved, even if it is at a low confidence value, as the system is intentionally biased to err towards classifying non-eagles as eagles.

Image of an Identiflight bird detection system.

What it is designed to replace: human observers

Airports use radar systems widely for detection of birds, however they cannot perform classification of birds or distinguish them from flying objects, thus requiring detailed analysis of data, for example via biologist consultation. High prices, power consumption, and government-imposed limitations on beam power and frequency limit the scalability of radar systems to other contexts (Gründinger, 2017). To prevent turbine collisions of protected species, some wind power facilities employ human observers stationed at vantage points who watch for eagles and order certain turbines to be powered down.

Using automated machine vision systems would be quicker and cheaper than observers who also have reduced visibility in comparison. Furthermore, AI solutions allow for otherwise conflicting needs to be met. For ornithologists, preferable solutions to date have been periodic turbine shut-downs. However, this limits wind farm power production and thus affects company profits (Gradolewski et al., 2021). Therefore, wind farm developers and operators would want to avoid unnecessary breaks in power production and minimize the time turbines are shut off, whereas environmental authorities want a high reliability of collision prevention, particularly for rare and large birds.

Why it is tractable: conservation concern + proven efficacy

There is a current conservation concern for big birds in particular, and there has been a challenge in designing systems that meet requirements specified by local environmental authorities, wind farm developers, and turbine manufacturers. In the United States, Duke Energy Renewables was found to violate laws protecting the Migratory Bird Treaty Act by failing to make all reasonable efforts to build their wind energy projects in a way that would avoid risk of avian deaths by collision with turbine blades, despite being warned by United States Fish and Wildlife Service (United States of America v. Duke Energy Renewables, 2013). This required the company to use human observers to prevent eagle-turbine collisions, but after several years they collaborated in the testing and implementation of IdentiFlight.

McClure et al. (2018) examined its efficacy compared to human observers in a proof-of-concept study. IdentiFlight was able to detect 96% of bird flights detected by human observers and 562% more than observers did (due to their difficulties with identifying birds from distance). The automated eagle detection had a false negative rate of 6% and false positive rate of 28% (it is “programmed to have a low false negative rate at the cost of sometimes misclassifying a ‘non-eagle,’” as it is intentionally designed to protect eagles (McClure et al., 2018, p. 30)). The median time from detection to classification was 0.4 seconds. The system missed 53 birds - however, only four of these were detected by observers - but detected 5,958 birds that the observers missed, making IdentiFlight comparatively more effective. McClure et al. (2021) tested the system’s real-world effectiveness in addressing the problem using a before-after-control-impact study. They found a 75%-89% reduction in eagle fatality rates at a test site where IdentiFlight was deployed relative to the control site without the curtailment system. The number of fatalities within the treatment site decreased by >60% after the mitigation was implemented. The system is currently deployed in five countries around the world: the USA, France, Germany, the Netherlands, and Spain.

Considerations:

  • The time taken for blades to stop depends on the turbine model (anything from 20s to >1 min) (McClure et al., 2021).
  • Initial costs may be high, however the costs of employing long-term observers are likely even higher and have been proven less effective.
  • Some systems have been designed to combine deterrents with curtailment, such as strobing lights (Gradolewski et al., 2021) and auditory stimuli (such as ‘DTBird’ which is also commercially available, though I did not choose this as a case study as recent publications of efficacy tests, including compared to other mitigations, were not as readily available). Curtailment is largely reserved for rare or big birds, and long-distance light/sound deterrents for other birds. In some cases the additional deterrent methods are required to encourage birds to change their flight path and avoid turbine stopping zones altogether; in other cases, where possible, wind energy companies might opt for only deterrents to completely avoid the economic impacts of curtailment (Gradolewski et al., 2021).
  • In Gradolewski et al. (2021), the authors also focused on system efficacy for identifying large birds, however their CNN was trained to classify birds in three different categories: small, medium and large; and was able to detect small birds reliably within 150m. Further research and development is needed for the improvement of accurate detection of smaller objects from a distance.
  • Painting turbine blades may also help prevent bird deaths, however more robust testing is being conducted - this could be an alternative to AI, or something that could be used alongside AI systems.

Potential for further impact:

There is promise that these kinds of systems will continue to be rolled out - in several locations protected species acts and legal requirements will encourage their uptake. Due to regulations, several wind power companies have a legal incentive to implement bird collision mitigation measures for protected species. This is a highly valuable mechanism advocates can use to gain leverage over companies. IdentiFlight currently only specifically intends to protect golden eagles (aquila chrysaetos), bald eagles (Haliaeetus leucocephalus), and red kites (Milvus milvus). Other bird species that were common within the testing site included turkey vultures (Cathartes aura), red-tailed hawks (Buteo jamaicensis), and common ravens (Corvus corax), and ferruginous hawks (Buteo regalis). So while the system is designed to protect a limited number of species, it can have a wider positive impact and could be programmed to detect other species. From a welfare perspective, it would be ideal for systems to protect more species of birds that are at risk. However, the conflicting priorities of involved stakeholders makes this difficult. Therefore, adding a bird from a relatively numerous species to a protected species list could be beneficial. It would be worthwhile to identify whether there is a species of bird at risk, in large numbers, not currently protected that could be added to capture a lot of the value of protecting all birds by just adding that one species to the detection list. This would be worthwhile if it would be relatively quick or easy to do this.

However, wind turbines are by no means the leading cause of bird deaths. Despite many studies being conducted on the impacts of turbines on birds, resources may be better spent deterring birds from colliding with buildings, which causes between 365 and 988 million deaths annually (Loss et al., 2014). While there are clear merits to this technology for the conservation of large birds of prey and other protected species, and it can ultimately prevent painful injury and death, there would potentially be greater value in further research on how it can be used to protect bats, given that turbine collisions kill 600,000 to 949,000 bats annually in North America alone. It is one of the leading causes of bat mortality; a global review showed wind turbine collisions were “the most frequently observed cause of multiple mortality events” in bats worldwide (O’Shea et al., 2016). Taking the data from this study, the average number of bats that die at once in a multiple mortality event is approximately 40. Not only are they affected in higher numbers than birds, bats can also suffer barotrauma (internal hemorrhaging) due to the sudden drop in air pressure around turbine blades. Studies have typically focused on automated detection of bats from their acoustic cues, and deployed systems, such as ‘DTBat’, that use one to three automated ultrasonic detectors to trigger curtailment in response to bat presence. However, recent evidence from Voigt et al. (2021) suggests that when monitoring bat activity near wind turbines, the limited sensitivity of ultrasonic microphones and attenuation of ultrasonic echolocation calls poses limitations, such as severely constrained detection distances which do not cover the total area of risk. It would perhaps be more beneficial to explore how the detection of bats near turbines and prevention of collisions could be improved to save a higher number of bats. If this were to be explored using machine vision techniques, long-wave infrared cameras would have to be utilized, as bats are most affected between dusk and dawn. Thermal imaging is currently used to monitor bat behavior and can have a range of up to 110m. There has been success with automated recognition of thermal signatures of larger species, however, the small size of many bat species may make this difficult, even for typical object detection algorithms that rely on morphological properties to be used. Preliminary studies have found it may be possible to detect and classify bats from flight track patterns (Cullinan et al., 2015) which could be a way to overcome this challenge. More species-specific research on this would be needed, which highlights the fact that specific solutions are often needed for specific problems. However, research and development in this area could be a way to significantly improve WAW through improving systems to mitigate fatal HWC on a great scale.

In some places, like North America, the majority of bats are not legally protected, so the incentive to consider mitigations is lacking. Therefore, in tandem with the improvement of mitigations and proving their efficacy, advocates could push for the protection of a wider range of bat species, or those particularly at risk of turbine collisions, and the enforcement of protection at wind farms. Ways to go about this might include highlighting their conservation status and that turbine collisions are a leading cause of mass mortality events in bats, which could affect whole populations (O’Shea et al., 2016)[4]. This could speed up the adoption of effective systems that prevent bat deaths at wind turbines. There may be other ways to raise the status of a specific species, but this type of intervention was not the focus of this report.

2. Exploring a theory of change: preventing HWC with ‘pest’ or ‘nuisance’ species

If machine vision-based methods can make HWC prevention highly effective, they could be adopted, even on a small scale → this helps to drop prices → allowing for systems to be more widely adopted → which would lead to more support and R&D → leading to continued price drops and adoption → this could create space for legislation to ban harmful or lethal methods of animal control, such as culling, which wouldn’t be needed if the problem is solved (or more realistically, nearly solved) → it may also reduce apathy and antagonism towards species[5]. This is a very tentative possible theory of change. A HWC that this could apply to is ‘pest’ species in agriculture, such as the culling of European starlings (Sturnus vulgaris) in the USA. Starlings can potentially ruin entire harvests of crops, particularly fruits, and also consume livestock feed, potentially contaminating it with their droppings. Earlier this year, Wildlife Services, a department of the United States Department of Agriculture (USDA), came under fire when it was revealed that they had killed over 1.75 million animals across the country, 1,028,642 of which were European starlings (Milman, 2022). The USDA states that they deploy lethal methods for bird control when non-lethal methods are unavailable or deemed impractical or ineffective. Below is an overview of a proposed AI-assisted system that could be developed to detect and deter flocks of starlings. If proven more effective than alternative methods, which have been found ineffective at solving the problem or are highly costly, it could prevent wide-scale crop damage and thus the need for culling. There is potential for: 1) funding/influencing funding for AI-assisted solutions with high potential impact; 2) advocates to raise awareness among stakeholders of effective non-lethal mitigations; 3) investigating if a similar flock detection system could be used to prevent mass bat deaths.

The non-lethal deterrent typically used is scaring

Mechanical: Nets can be wrapped around individual plants or stretched over entire orchards or vineyards. This incurs high purchase costs as it is a short-lived method which needs replacing every one to five years, and costs range from $7,000 to $30,000 USD per acre, depending on the netting system used. Furthermore, installation and removal is labor and time intensive, and the netting has negative ecological impacts. Hawk-kites and scarecrows are another option, but show mixed results, so unmanned aerial vehicles/drones/’robobirds’ have been developed in recent years to be flown above affected areas and deter flocks. However, there are limits to flight times, legislation restrictions, and constraints from control and battery charging requirements.
Acoustic: Hunters may shoot blanks to scare birds without death or injury, however, guards can only watch over part of the total area and this ideally requires someone working at different times throughout the day. Employment costs also need to be considered. Sonic bird repellers which emit raptor sounds, starling calls that signal danger avoidance, or sounds at frequencies unpleasant to birds can be deployed. Directional ultrasonic repellers are often effective but only perform optimally in small enclosed areas. Another method is using firing cannons during ripening periods. This method is relatively cheap, but disturbs other nearby animals and people. Furthermore, acoustic methods gradually become ineffective as animals habituate to constant disturbances (Kingshay, 2013).

Optical: Distributing shiny objects disturbs all types of birds, but smaller birds particularly are scared by the reflections, which resemble raptor eyes. However, starlings are intelligent and again, over time, this method may become less effective. Holographic strips and lasers (agrilaser systems) are alternative methods.

Natural enemies: some farmers may hire a falconer to predate upon ‘pest’ species. This method is time-intensive and costly, and it also causes suffering for the targeted species. From the perspective of conserving animal welfare, methods which cause the least suffering yet remain effective would be ideal. Harris Hawks (Parabuteo unicinctus) may be trained and flown above the area without chasing or catching starlings to deter them; though this is effective, it is still costly. I could not find costs in the US, but comparative costs in the UK are as follows: £3045 GBP for ‘setting up’, which includes expenses such as equipment, shelter, and food + £3045 GBP for one to two hours of staff time each day the hawk is flown.

A proposed solution

Marcoň et al. (2021) propose a vision detection system to trigger a scaring element only when a flock of starlings is detected, so that the process is not continuous (preventing habituation to deterrent stimuli), and also eliminates undesired sonic disturbance. The system is an optical detector, utilizing a camera system and ML algorithm to track items in a predefined area. Images of different flying objects taken within the study area were annotated and categorized, for example based on if they were images of birds, insects, helicopters, or flocks. They were then classified into more specific subsets (e.g., bees vs. flies), and for birds, specifically based on their individual flight characteristics (e.g., soaring birds vs. birds with retracted wings). Classified images were used in training, validation, and testing in the ML phase. The system was set so that it would trigger a scaring task if the optical unit detected a flock with more than 30% confidence. Testing of the classifier showed a 100% precision for flocks (i.e., it detected 100% of the true positives) with only a 5.7% error rate (the proportion between all misclassified objects and the total number of detected objects). It evaluated flocks correctly up to 300m - any distance further than that causes an increase in false detections. Large flocks and even small formations of several birds were highly detectable. The idea is for a detection event to trigger a deterrent stimulus.

There is a deployed bird deterrent system, ‘AVIX’, which also uses AI in a similar way. This product can distinguish birds from other objects and uses laser beams to deter them - however, there is no trigger system for an entire flock, just a single bird. Since flocks raiding ripening fruit would cause the most widespread damage, the proposed system has a significant advantage over systems which are not programmed to detect flocks, and could be more valuable for effectively preventing raids and thus the need to resort to lethal methods. The authors emphasize that their system also allows incorporation of multiple scaring methods, not just one. By only triggering the deterrent when needed, this addresses limitations of alternative methods by reducing the adaptability of the flock to the scare sounds, compared to current commercial devices which emit continuous sound. For low consumption, the system includes a time delay switch to activate the setup one hour before sunrise and shut off one hour after sunset; this combined with the trigger system reduces the electricity costs and battery charging requirements compared to other technological solutions.

Image of the complete set up at a vineyard in Bořetice, Moravia, the Czech Republic - taken from Marcoň et al. (2021), p. 6.

Equipment needed for one system:

  • 2 video cameras (1750 USD)
  • NVIDIA Jetson Nano single board-computer for fast processing (connected to the cameras) (174 USD)
  • 1x Raspberry Pi HQ (64.44 USD) equipped with Sony IMX477R sensor (80.16 USD) + 16mm f/1.4 PT3611614M10MP lens (80.43 USD)
  • 1x Arducam 8 Mpx USB webcam CCTV 5–50 mm (68.31 USD)
  • A module for temperature and humidity measurements (6.96 USD)
  • 1x SIM card to transmit data over an 4G network every 30mins (1 USD)
  • Cooling fan (5.80 USD) and heat sink (2.32 USD)
  • Reflective foil (8.04 USD)
  • WiFi (ESP8266 module) (3.48 USD)

Total cost: approx $2244.94 USD

Considerations:

  • This is just an initial proof-of-concept prototype, which still needs integration of scaring methods such as lasers and auditory stimuli.
  • RCTs are needed to prove its effectiveness over time compared to other mitigation methods and compared to a control site with no mitigations.
  • I found it difficult to find out just how widely current mitigation methods are used - I think this could be a beneficial area of research, in order to assess how easily affected farmers could switch to AI-assisted systems.
  • In the case that affected agricultural areas, such as vineyards, do not have a connection to an electricity grid, an alternative is required. In this study, a small island network was built, which additionally required a solar charger, a photovoltaic panel and a voltage regulator to supply electricity to the microcomputer and cameras.
  • The number and placement of cameras needed will be dependent on the camera parameters as well as the area and altitude profile of the area being observed. For the study area, the authors in this study calculated a radius of 300m could be covered for a full 360 degree range by at least eight sensing modules. This poses a limitation in that each system has a reduced field of view, so several modules are needed.
  • Initial costs may be high, but the equipment could be made cheaper and more scalable with product development. Most agricultural areas are also owned by large multinational companies - scalability may depend on whether they or farmers bear the brunt of the cost, or whether government organizations could be persuaded to subsidize costs. It also benefits from being a system with a one-time installation which could continue to deter birds, avoiding crop damage, and thus prevent their deaths from poisoning.

Potential for impact:

With the deployment and adoption of any mitigation, there will be a need to effectively demonstrate how the system works better than current methods to alleviate the problems end-users are facing. If the expected effectiveness of this system for deterring starlings is high in comparison to (or in combination with) other mitigations, and it successfully prevents birds from damaging crops and contaminating feed, it could potentially eliminate the need for lethal control. Culling also does not necessarily stop a problem recurring, it just controls the population of a species. So if effectiveness can be proven, perhaps it would be likely for conservation and government organizations to invest funds into R&D, or for agricultural companies to buy the end product. This would perhaps require some advocacy work to convince stakeholders that this method is preferable to poisoning, or a funder to subsidize its free distribution and training costs for a period to incentivise quick awareness and adoption. Pimentel et al. (2000) estimated that starlings damaged 800 million USD worth of agriculture crops in the United States annually. Preventing this extent of financial damage to the agricultural sector alone could be a significant motivator for investing in HWC mitigation systems that are proven highly effective and consequently eliminate the need for culling. This flock detection system could also potentially be used for detecting bats near turbines, especially as turbine collisions are the leading cause of mass mortality events in bats. This may help to overcome issues of poor visual detection and poor acoustic detection described above.

The tentative theory of change I explore could apply to several conflicts involving numerous species that are seen as crop-raiding animals. Another example is wild boars, who can cause extensive and costly damage to agricultural land. AI-assisted systems could potentially be used to detect and deter such animals. Farmers in India have been demanding that wild boars be formally categorized as ‘vermin’ to facilitate mass culling. While these demands have not been met, governments in Goa and Kerala have authorized permissions to hunt wild boars, suggesting other states where this conflict is high may follow suit. As wild boars are one of the widest-ranging mammals across the world, several other regions face a similar problem. In the USA there are now over 6 million feral pigs; they are widely seen as an invasive pest and are poisoned, trapped, or shot.

3. Using ML predictions of where/when HWC will occur to redirect mitigations

Improving the effectiveness of existing HWC mitigation systems could have a high pay-off on its own. An additional way AI could be used to make HWC mitigations most effective, and feed into this potential theory of change, is by predicting when and where HWC might occur in the first place. This could ensure that mitigation systems are placed in the most optimal locations. Using ML methods to understand, identify, and predict the conditions in which HWC may occur would allow proactive action to prevent HWC. ML algorithms can handle vast amounts of data, AI can learn and rationalize – allowing modeling with higher accuracy and speed than humans, without having to explicitly provide instructions. Thus, animal behavior, animal movement, migration patterns, economic, psychosocial, and spatial-temporal data on HWC could be synthesized and analyzed to facilitate predictions.

Currently, Wildlife Conservation Trust (WCT) is collaborating with Singapore Management University and Google Research India in its Social Good Programme to design AI models to predict HWC in Maharashtra, India. These models are currently being tested. WCT has already been conducting extensive research in this area, so plentiful data is available. Similarly, there are other locations and animal populations that are well-studied, with large amounts of data from camera traps and GPS remote sensing methods to analyze movement behavior, and community surveys to collect conflict-related data, which could be used in this context too. In other words, for some locations, there would not even be a need to collect this vast and fine-scale data. Other sources of data can be utilized too. WCT utilized the Forest Department’s compensation records, government data on human and livestock populations, and Google satellite data.

Along with their primary uses, the AI-assisted systems I have described above could also potentially be used to contribute to gathering and storing large amounts of data on individual species counts and behavioral data that could be used to predict where HWC might occur. The turbine collision case study provides a clear example:

Due to the difficulty and differences in regulatory requirements for collecting this data, reliable estimates for mortality rates as a result of turbine collisions are difficult to acquire, which is why there is such a large range in figures. Here, machine vision without the need for a curtailment system can be utilized to automatically identify species and collision incidents to inform interventions. McClure et al. (2021) also suggested the best case scenario would be avoiding the erection of turbines in high-risk areas in the first instance, but it can be difficult to predict impacts before construction. While curtailment systems will still hold high value and will still be needed in the meantime, being able to predict where these high-risk areas are would be ideal in deciding where to erect turbines. For example, machine vision systems could also be used to monitor bat activity and inform ML predictions to identify migration pathways and high risk areas more quickly, before turbines are constructed. We may not be able to stop the development of wind farms in all high-risk areas, so when avoidance isn’t preferred an effective automated curtailment system should still be implemented.

8. Broader considerations, obstacles, and opportunities

A global survey of 248 conservation technology developers, users, and academics by Speaker et al. (2022) suggests coordination, capacity-building, and funding issues are hampering the development and implementation of conservation technologies and need to be addressed. Some of these wider considerations could be applied specifically to AI-assisted methods for mitigating HWC and improving WAW, and the potential for harm should be considered too.

1. Coordination issues

  • It seems many conservation organizations have little knowledge of technology from an operational perspective. Even where there are technology or AI experts, there is a huge gap between theory and applications in the field, and thus deciding what to actually bring to the field to implement. It’ll take interdisciplinary work and collaboration between engineers, AI experts, ecologists, animal behavior experts, conservationists, and even social scientists to fully harness any potential that lies in this area.
    • An interdisciplinary conference on the intersection between AI and HWC or wildlife management, with funding to tie experts from these fields together, could be established. Conservation technology networks like WILDLABS might be a great place to start.
  • Greater consideration should also be given to where research is published. For example, a lot of the research on developing AI-assisted systems for conservation was published in technological journals, and as a result, some developments may not end up on the radar of conservation or animal welfare organizations.
    • An applied methods journal written by technical experts but designed to be accessible for audiences with a wider range of expertise would be beneficial, and could increase the likelihood of citations.
  • There also seems to be a gap between what companies publicize, and the actual end result and efficiency once an AI-assisted, or any other technological, system is deployed in the field - there tends to be a boom of publications and articles before or at the beginning of the deployment phase, but then little after that. This could indicate that field results are worse than expected (highlighting the need for appropriate collaborations), or it could be that once funding is secured and a system is implemented, nobody asks for these studies of efficacy. Like the significance bias that exists in academic research, it does not seem highly likely that we would easily find out if a system had been developed or deployed but then failed. Although it would be highly valuable for this information to be easily accessible in order to have a better understanding of what does and does not work. Through expert outreach, I found conservation technology is sometimes funded and purchased, but end users are not trained or supported to effectively use it long term in the field (this being especially prominent in the case of drone technology).

2. Funding

  • A lot of technologies in conservation need long-term support to gradually implement them. If the funding for that is not considered, then it is not likely to succeed.
    • Therefore, there should be higher expected value placed on systems that do not require long-term funding support. This should especially be paid attention to in areas with historically limited comparative funding and resources - e.g., East Africa and South Asia. In Speaker et al. (2022), upfront and maintenance costs were a key constraint for respondents in developing countries.
  • In Speaker et al. (2022): “respondents identifying as conservation technology developers and testers reported that financing was also a significant barrier for them. They rated securing funding throughout the development cycle (67%) and securing seed funding for projects (62%) as their top 2 constraints” (p. 7).
    • Perhaps involvement of corporate technology companies in conservation technology (e.g. Microsoft and Google) could shift this financing dynamic. If appropriate, there may be a potential opportunity here for EA to fund R&D, work to convince others (such as companies/governments/NGOs/philanthropists) to fund and use effective AI-assisted mitigation systems, or inject a small amount of initial funding to test and gather feedback on interventions.
  • Time and expertise needed: More generally, supervised learning techniques require large amounts of human-labeled data and human specialists to train models. This often comes at a cost. However, this technology is continually improving, and different methods, such as transfer learning, can be utilized for slightly smaller datasets and can also reduce computational requirements.

3. Potential for harm

There could potentially be some negative or harmful effects of the development of AI for wild animal welfare, or perhaps even for preventing HWC more generally.

  • Negative intentions: The intention of end-users with access to AI-assisted technology should be considered, for example, autonomous drones are already used for targeted poisoning operations and increasing pest control efficiency. Individuals or organizations could potentially utilize or hack this kind of technology to locate species of interest (this would be particularly harmful in the hands of poachers or trophy hunters, for example). However, perhaps the potential benefits to improving WAW on a large scale outweighs the possibilities of this occurring. To ensure this, bans/moratoriums should be placed for these uses with negative impacts, while technology with positive effects should be supported. In any case, developers should keep such considerations in mind when designing, testing, and deploying any systems - i.e., making the software robust to hacking and manipulation, and controlling the distribution of the technology.
  • WAW arguments:
    • HWC deaths may be less painful than counterfactual deaths;
    • HWC may reduce populations of species with net negative lives.
      • Currently, where the concept of WAW is at an early stage, it may be difficult to determine to what extent these two points would be true. If they are proven true, extending the lives of these animals, by preventing HWC, may not improve their welfare.
    • HWC could reduce species populations which are overall harmful to other animals, such as in the case of predators.
      • Although, there are counterarguments to this counterargument, such as alternative deaths (like disease) being worse.

I am inclined to believe the overall benefits to improving certain HWC mitigations and decreasing the number of animals negatively affected by HWC, such as in the cases I have discussed of birds and bats, would be worthwhile. This could be an opportunity for WAW to have a direct impact; it may currently be an easier or a good first step to work on avoiding harm and suffering. One can get this technology deployed by bringing together a range of stakeholders with different motivations, and a concerted effort by people interested in WAW could ensure it achieves enduring positive impacts on WAW. Additionally, intervening in HWC is seen as acceptable; having WAW-aligned people working in this space would be beneficial, as it is relatively difficult to find an area that already has such a high level of buy-in, and where WAW could influence positive impact that benefits animal welfare at scale.

Acknowledgments

This post is a project of Rethink Priorities–a think tank dedicated to informing decisions made by high-impact organizations and funders across various cause areas. It was written by Tapinder Sidhu. Thanks to Ben Stevenson, Hannah McKay, Holly Elmore, Michael St Jules, Neil Dullaghan, William McAuliffe for helpful feedback and Eric, Fai, Nir for helpful discussions. If you are interested in RP’s work, please visit our research database and subscribe to our newsletter.

Bibliography

Abrahms, B. (2021). Human-wildlife conflict under climate change. Science373(6554), 484–485. https://doi.org/10.1126/science.abj4216.

Anand, S., & Radhakrishna, S. (2017). Investigating trends in human-wildlife conflict: Is conflict escalation real or imagined? Journal of Asia-Pacific Biodiversity10(2), 154–161.https://doi.org/10.1016/j.japb.2017.02.003

Bhatt, J., Das, A., & Shanker, K. (2018). Biodiversity and Climate Change: An Indian Perspective. Ministry of Environment, Forest and Climate Change, Government of India.

Choi, D. Y., Wittig, T. W., & Kluever, B. M. (2020). An evaluation of bird and bat mortality at wind turbines in the Northeastern United States. PLOS ONE15(8), e0238034. https://doi.org/10.1371/journal.pone.0238034.

Cullinan, V. I., Matzner, S., & Duberstein, C. A. (2015). Classification of birds and bats using flight tracks. Ecological Informatics27, 55–63. https://doi.org/10.1016/j.ecoinf.2015.03.004.

Dickman, A. J. (2010). Complexities of conflict: The importance of considering social factors for effectively resolving human-wildlife conflict: Social factors affecting human-wildlife conflict resolution. Animal Conservation13(5), 458–466. https://doi.org/10.1111/j.1469-1795.2010.00368.x.

FAO & UNEP. (2020). The State of the World’s Forests 2020. Forests, biodiversity and people – In brief. https://doi.org/10.4060/ca8985en.

Goodrich, J. M. (2010). Human-tiger conflict: A review and call for comprehensive plans. Integrative Zoology5(4), 300–312. https://doi.org/10.1111/j.1749-4877.2010.00218.x.

Gradolewski, D., Dziak, D., Martynow, M., Kaniecki, D., Szurlej-Kielanska, A., Jaworski, A., & Kulesza, W. J. (2021). Comprehensive Bird Preservation at Wind Farms. Sensors21(1), 267. https://doi.org/10.3390/s21010267.

Gross, E., Jayasinghe, N., Brooks, A., Polet, G., Wadhwa, R., & Hilderink-Koopmans, F. (2021). A Future for All: The Need for Human-Wildlife Coexistence (WWF). https://www.unep.org/resources/report/future-all-need-human-wildlife-coexistence

Gründinger, W. (2017). The renewable energy sources act (EEG). In Drivers of Energy Transition (pp. 257–419). Springer.

IUCN SSC HWCTF. (2020). What is human-wildlife conflict? Briefing Paper by the IUCN SSC Human Wildlife Conflict Task Force.

Kingshay. (2013). Practical Control Strategies for Starling Infestations. DairyCo.

Loss, S. R., Will, T., Loss, S. S., & Marra, P. P. (2014). Bird–building collisions in the United States: Estimates of annual mortality and species vulnerability. The Condor116(1), 8–23. https://doi.org/10.1650/CONDOR-13-090.1.

Marcoň, P., Janoušek, J., Pokorný, J., Novotný, J., Hutová, E. V., Širůčková, A., Čáp, M., Lázničková, J., Kadlec, R., Raichl, P., Dohnal, P., Steinbauer, M., & Gescheidtová, E. (2021). A System Using Artificial Intelligence to Detect and Scare Bird Flocks in the Protection of Ripening Fruit. Sensors21(12), 4244. https://doi.org/10.3390/s21124244.

McClure, C. J. W., Martinson, L., & Allison, T. D. (2018). Automated monitoring for birds in flight: Proof of concept with eagles at a wind power facility. Biological Conservation224, 26–33. https://doi.org/10.1016/j.biocon.2018.04.041.

McClure, C. J. W., Rolek, B. W., Dunn, L., McCabe, J. D., Martinson, L., & Katzner, T. (2021). Eagle fatalities are reduced by automated curtailment of wind turbines. Journal of Applied Ecology58(3), 446–452. https://doi.org/10.1111/1365-2664.13831.

Milman, O. (2022, March 25). ‘A barbaric federal program’: US killed 1.75m animals last year – or 200 per hour. The Guardian. https://www.theguardian.com/world/2022/mar/25/us-government-wildlife-services-animals-deaths.

Norouzzadeh, M. S., Nguyen, A., Kosmala, M., Swanson, A., Palmer, M. S., Packer, C., & Clune, J. (2018). Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning. Proceedings of the National Academy of Sciences115(25). https://doi.org/10.1073/pnas.1719367115.

O’Shea, T. J., Cryan, P. M., Hayman, D. T. S., Plowright, R. K., & Streicker, D. G. (2016). Multiple mortality events in bats: A global review. Mammal Review46(3), 175–190. https://doi.org/10.1111/mam.12064.

Pimentel, D., Lach, L., Zuniga, R., & Morrison, D. (2000). Environmental and Economic Costs of Nonindigenous Species in the United States. BioScience50(1), 53. https://doi.org/10.1641/0006-3568(2000)050[0053:EAECON]2.3.CO;2.

SARPO, WWF. (2005). Human wildlife conflict manual. WWF Southern African Regional Programme Office (SARPO).

Singer, P., & Tse, Y. F. (2022). AI ethics: The case for including animals. AI and Ethics, 1-13. https://doi.org/10.1007/s43681-022-00187-z.

Speaker, T., O’Donnell, S., Wittemyer, G., Bruyere, B., Loucks, C., Dancer, A., Carter, M., Fegraus, F., Palmer, J., Warren, F., & Solomon, J. (2022). A global community-sourced assessment of the state of conservation technology. Conservation Biology36, e13871. https://doi.org/10.1111/cobi.13871.

Voigt, C. C., Russo, D., Runkel, V., & Goerlitz, H. R. (2021). Limitations of acoustic monitoring at wind turbines to evaluate fatality risk of bats. Mammal Review51(4), 559–570. https://doi.org/10.1111/mam.12248.

Notes


Date of Artificial General Intelligence predictions and Human/Machine Intelligence Parity by 2040 predictions ↩︎

For example, if you click “Date Range” of this Semantic Scholar search for "Conservation" and "Artificial intelligence", you can see a steady uptick in this proxy measure. Whilst the number of papers published is unclearly correlated with utilization, there is a sense across the conservation sector that more advanced technologies are beginning to be used. ↩︎

Research platforms can be found here: https://www.wildme.org/#/platforms ↩︎

See section ‘Bat population dynamics, MMEs, and implications for the future’ in O’Shea et al. (2016) ↩︎

Alternatively, the threat of a legal ban first could accelerate technological development, as companies and affected farmers may want to get ahead to find alternative and effective solutions. ↩︎

Tapinder Sidhu

Tapinder has a background in Psychology and a Master's with Distinction in Animal Behaviour, with fieldwork experience. She has completed large-scale projects, particularly on African elephant behaviour. Before joining RP as a Research Fellow, Tapinder held a Research Officer role, conducting social and political research for non-profits in the UK. She is now preparing to start a self-proposed PhD on assessing wild elephant welfare through emotion detection.

Previous
Previous

An Introduction to the Moral Weight Project

Next
Next

Does the trajectory of pain matter?