 Reference Base  Sentinel animals: Enriching artificial intelligence with ... |
|
Eikelboom, J.A.J., 2021. Sentinel animals: Enriching artificial intelligence with wildlife ecology to guard rhinos. Dissertation presented to the University of Wageningen, The Netherlands, pp. 1-209
|
|
|
|
| |
| Location: |
Africa |
| Subject: |
Distribution |
| Species: |
African Rhino Species |
 |
|
The survival of both African rhinoceros species is under threat due to large-scale poaching. The pressure that poaching currently exerts on rhino populations is too large to solely wait for long-term conservation strategies, e.g., demand and corruption reduction campaigns, to take effect. Consequently, protection efforts aimed at the short-term survival of the rhino species seem to be urgently needed. Unfortunately, current rhino protection efforts fail to prevent large rhino population declines as conservation officers often fail to localize poachers before they can kill a rhino. Therefore I aimed to develop a poacher early warning system that provides conservation officers with more situational awareness, which can therefore decrease the risk of shootouts between poachers and conservation officers.
For this task I focused on developing a ''sentinel-based poacher early warning system'', for which I envision nature reserves where abundant prey animals are tracked and where the movement responses of these animals are automatically used to detect the presence and infer the location of poachers. Hence the term: ''sentinel'', as the animals themselves will take the role of game wardens. The benefit of such a system is that it could be working at all times and is not limited solely to rhino poachers. Apart from the obvious wildlife conservation challenge this thesis poses, it also tackles a major scientific challenge: to be able to detect abrupt changes in an environmental variable based on animal movement. In order to solve this challenge, a myriad of environmental and animal movement variables needed to be considered in interaction in a single model. This premise lead me to the use of a non-traditional statistical approach for wildlife ecologists: artificial intelligence.
This thesis brings together a number of coherent papers about wildlife conservation, movement ecology and artificial intelligence, aimed at investigating the necessity, analytics and applicability of a sentinel-based poacher early warning system. In Chapter 2 I critically evaluated whether rhino protection efforts aimed at the short-term survival of the species are actually needed. I examined this by investigating if legal international rhino horn trade could be an ultimate solution for rhino poaching. Through an integrative review of scientific and grey literature about rhino horn trade legalization, I identified four main mechanisms through which a legal rhino horn market would influence the remaining wild rhino populations: 1) financial viability for private rhino owners, 2) rhino horn demand, 3) laundering of rhino horns, and 4) behaviour of rhino horn consumers. Subsequently, I determined through plausible reasoning that only the increased revenue for rhino farmers could potentially benefit rhino conservation. Conversely, the global demand for rhino horn is likely to increase to a level that cannot be met solely by legal supply. Moreover, corruption is omnipresent in countries along the trade routes, which has the potential to negatively affect rhino conservation. Finally, programmes aimed at reducing rhino horn demand will be counteracted through trade legalization by removing the stigma on consuming rhino horn. After combining these insights and comparing them with criteria for sustainable wildlife farming, I concluded that legalizing rhino horn trade will likely negatively impact the remaining wild rhino populations. To preserve rhino species I suggest to combine long- and short-term conservation approaches, by prioritizing the reduction of corruption within rhino horn trade, increasing the rhino population within well-protected 'safe havens' and implementing educational programmes and law enforcement targeted at rhino horn consumers.
In Chapter 3 I investigated how much tropical animal populations in general are impacted by hunting, apart from solely considering African rhinos. I did this by analyzing how much human hunters alter the abundance and spatial distribution of animals in the tropics. Through a systematic review and a mixed effects meta-analysis I estimated that bird abundances declined on average by 58% (95% CI: 25-76%) and mammal abundances by 83% (95% CI: 72-90%) in hunted compared with unhunted areas. Mammal population densities were higher inside than outside protected areas, but hunting pressure reduced mammal abundances even within protected areas. Furthermore, I determined that bird populations were depleted within 7 kilometers and mammal populations within 40 kilometers from roads and settlements, which function as access points for hunters. These results signify that the impact of hunting on both the abundance and distribution of tropical animals is very large. Although these results suggest that the effect of hunting within protected areas is less detrimental than outside reserves, gazettement of protected areas seems insufficient to safeguard wildlife populations if not accompanied with improved reserve management, effective law enforcement and on-ground protection efforts.
In Chapter 4 I studied the link between individual movement rules and emergent collective movement properties, which can both provide information about changes in the perceived environment of animals. For this I used an agent-based simulation model to investigate the indirect effects of fear and resources on animal group structures. In this model only the individual movement rules were directly affected by fear and resources, but through self-organization the effects of fear and resources also became apparent in the size of the formed groups. I specifically focused on the inherent variability in sizes of groups that were generated from identical self-organizing processes. I found that the coefficient of variation of group size generally lied between 50 and 150% in these simulations, depending on both animal density and the resource scarcity/predation trade-off. Given that the variations of group size are already this large in homogeneous and deterministic scenarios, I consider group size an imprecise collective movement proxy for environmental conditions. Considering this imprecision of group size as a proxy and its time lag with changes in environmental conditions, group size can likely only be informative for slowly-evolving environmental conditions and will require information about the recent history of the animal group in order to be informative.
In Chapter 5 I predicted the environment of animals based solely on their movement data. I specifically investigated how much of the variation in different environmental variables influenced animal movement in its multivariate entirety. I did this by linking high-resolution sensor data from cows in a controlled environment to various environmental variables through extensive feature engineering and machine learning to predict the environment from animal movement sensor data. Using this data-driven framework I demonstrated that it is possible to quantify environmental influence on animal movement with the performance metrics of machine learning regression algorithms. Depending on the chosen time window of feature engineering, the influence of environmental variables on different time scales can be studied. Furthermore, different types of animal movement features (e.g., individual- and collective-based, or GPS- and accelerometer-based) can be included separately or in combination in the framework. Even though the aim of this framework is to quantify the exact contribution of separate environmental variables on the total variation in animal movement, the core of this framework can be used to accurately predict environmental variation from animal movement as well.
In Chapter 6 I developed a sentinel-based poacher early warning system in Welgevonden Game Reserve (South Africa). Using sensor data from 138 savanna ungulates combined with experimentally staged human intrusions, I algorithmically detected and localized poachers using animal movement data. I used a three-step analytical process to achieve this, namely: 1) animal behaviour classification, 2) poacher detection, and 3) poacher localization. In the first step I demonstrated the importance of interpreting animal movement as deviations from expectations given recent movement history and similar environmental conditions, given the complex relationship between the animals' heterogeneous environment and movement. I achieved an average precision of 46% to classify animal movement responses to humans versus all other movement. Even though this performance is quite an achievement (given the large class imbalance between normal and response behaviour, the inherent variability in animal movement, and environmental heterogeneity in the study area), it still leads to a substantial amount of misclassification. However, in the next two steps I considered the classified responses of all animals collectively in a spatiotemporal context, which allowed me to drastically improve upon this performance in the detection and localization of 'poachers'. Periods with humans present in the area could be distinguished from periods without humans with 86% accuracy in a balanced validation design, and these humans were localized with less than 500m error in 54.2% of the experimentally staged poaching intrusions. This chapter thus demonstrates the feasibility of the main theme of this thesis, namely to use a sentinel-based poacher early warning system to detect and localize poachers.
In Chapter 7 I investigated the performance of an automated animal detection algorithm for aerial imagery with the intention to gauge the potential of aerial imagery to supplement or replace animal-born sensors to track animals en masse in the near future. Using a deep learning approach I automatically identified large savanna herbivores inside images from an aerial wildlife survey in Kenya, after which I also classified the animal species using the same model. With this approach I managed to detect 90-95% of the number of individual animals that were found by four layers of human annotation, of which I correctly detected 2.8-4.0% extra animals that were missed by all humans. The model did result in 1.6-5.0 false positives per true positive, which emphasizes the importance of manual verification of automatic animal counts from aerial images. In this chapter I specifically demonstrated the potential of semi-automatic aerial animal counts to improve the precision and accuracy of animal population estimates. Furthermore, the results indicated that automated animal detections from aerial images have the potential to find more animals than humans can, especially when the algorithm is supplied with images taken at a fixed rate. Considering the aforementioned, I acknowledge the potential of aerial imagery to supplement en masse tracking with sensor tags. However, given that the detection chance of animals in images decreases substantially with horizontal distance to the camera, I expect animal tracking with cameras to be only suitable for relatively small areas.
Finally, in Chapter 8 I synthesized my combined research in light of both wildlife conservation and wildlife ecology. I argued that the applicability of my developed sentinel-based poacher early warning system lies mainly in the aid it can provide to short-term wildlife protection efforts during the Anthropocene, which can concurrently reduce some of the negative effects associated with 'militarized conservation' (e.g., human rights violations). I plead for collaboration between conservationists working on short- and long-term conservation strategies, to maximize the efficacy of conservation by considering the occasional trade-off between conservation success in the Anthropocene and the development of a society that is in harmony with nature. Furthermore, I forecasted a large role for artificial intelligence in wildlife ecology research, which may drastically change the way scientific understanding is acquired in the near future. Exciting developments related to explainability and causality within artificial intelligence are currently being undertaken by computer scientists, but these scientists do require the input of ecologists to make these developments truly insightful and applicable to the real world.
|
|
|
File available
details
Original text on this topic: