Exploring Anti-poaching Strategies for Wildlife Crime with a Simple and General Agent-Based Model

The most common threats to plant and animal species worldwide are the destruction of their habitat and the over-exploitation of natural resources due to human activities like wildlife poaching and the illegal trade in wildlife products [1]. Recently, criminologists have also started to study wildlife crime. Understanding the processes behind crime often leads to practical implications for crime prevention strategies to improve its effectiveness. These strategies are based upon opportunity theories, seeking to create interventions that reduce victimization by removing or disrupting opportunities for crime. This includes increasing offender perceptions of risk and effort while minimizing perceptions of reward [2].

A commonly used strategy to prevent wildlife crime inside a protected area is aimed at increasing the poacher’s perception of risk. This is a data-driven approach that includes the analysis and mapping of poaching incidents [3]. Rangers are then deployed at poaching “hotspots” to either prevent or detect illegal activities. The downside of a data-driven approach is that it is heavily biased towards the spatial aspects of how the observation data was collected [4]. Ranger patrols are often not able to cover the whole protected area on a regular basis, leading to an incomplete knowledge of where and when poaching activities occur [5]. This is also referred to as the dark figure of crime [6]. Furthermore, when a new anti-poaching strategy has been put in place, rarely has it been evaluated in a standardized and systematic way.

As an alternative, wildlife crime can be studied as a complex and dynamic system. Its complexity does not only arise from the strategic interdependencies between animals and poachers. Poachers are also being tracked by rangers and seek to avoid interaction. This creates a second layer of strategic interdependence. To tackle the complexity of these systems, formal models allow researchers to study the implications of model assumptions. They help to understand why and under what conditions the models generate unexpected predictions [7]. Formal models allow researchers to explore the effectiveness of anti-poaching strategies even before they are implemented. They also allow to derive testable predictions about the conditions under which they are effective or not. For example, a successful strategy can turn ineffective at one moment, not because rangers failed to implement it, but simply because animals responded to changes in their environment. Subsequently, this motivates different poacher behavior which makes the adopted anti-poaching strategy ineffective.

Agent-based modeling (ABM) is a prominent formal modeling technique and particularly suited to study wildlife crime. In an agent-based model, each individual agent makes autonomous decisions, reacting to its environment and the behavior of other agents. AMB is a rigorous method, but hardly restricts the modeler in the choice of assumptions [8]. ABM allows researchers to tailor models to specific settings, for example by applying their model to a specific protected area, endangered species, or anti-poaching strategy. Such information is not only relevant for practitioners, but also expands the application of ABM into new areas of wildlife crime.

This study is aimed at exploring the dynamic interactions between the agents involved in wildlife crime using agent-based modeling. The objective was to develop a simple and general model that captures these dynamics under different anti-poaching scenarios. As an illustration, the model is demonstrated by applying it to the context of rhino poaching in South Africa.

For most cases of wildlife crime, there are three types of agents: animals, poachers, and rangers. The interactions between these agents can be described as a triple foraging process [9]: “animals search for food, poachers search for animals, and rangers search for poachers”. The model simulates a “world” with a population of animals where poachers go in and out to hunt for this species. The rangers try to disrupt and catch the poachers.

The world is a simplified representation of a protected area, like a national park. This virtual park is divided into grid cells that contain information about the environment, like the amount of resources available to the animal, how long it takes to move through this cell and any signs of animal, poacher, or ranger presence. The advantage of using a simplified model of a park is that dynamics cannot be driven by idiosyncratic characteristics of a specific setting, allowing the modeler to derive general conclusions about the implications of model assumptions. Nevertheless, the model is formulated in such a way that many idiosyncrasies of real parks can be implemented, which allows the study of dynamics also in specific settings.

In this model, animals are distributed over the virtual park and individually make decisions on where to move next based on the characteristics of the surrounding grid cells, choosing cells that are most attractive. At the same time, animals also change their environment by consuming resources. The resources recover over time. While the animals move over the landscape, they leave signs that can be detected by poachers, influencing their movement as they search for a target. Poachers can also detect signs of ranger activity and try to avoid those areas. Hence, poachers make decisions based on, among others, signs of recent animal activity and ranger activity. Poachers always start and end their hunt at the border of the virtual park. If a poacher encounters an animal within his observation radius, he kills it and the poached animal is removed from the park. The poacher then returns to the park’s border. When the poacher reaches the border, he successfully escaped and cannot be caught by rangers. Poachers remember the areas where they were successful, areas with high ranger activity and use that information to plan their subsequent incursions.

Rangers carry out patrols and search for signs of poaching. Just as poachers tend to go to areas with the highest animal activity, rangers tend to go to areas with the highest poaching activity. If a ranger detects a poacher within his observation radius, the ranger catches the poacher, removing him from the park. Rangers remember where they found poaching signs and tend to patrol those areas more frequently. Rangers can start either at a base camp inside the virtual park or at one of its borders.

Dynamics are broken down to a sequence of discrete events. At each event, first all animals decide in a random order where to move, followed by the moving decisions of poachers, and rangers. Finally, the simulation program updates the resources available at each cell, taking into account the consumption by animals and resource recovery. The decision-making of all three agents is similar and can be easily adjusted by adding relevant variables for a particular problem. This can also be used to create specific scenarios or conditions.

Rhino poaching in South Africa has surged since 2008, in response to significant increases in black market prices for rhino horn [10]. This has led to discussion among conservationists, law enforcement and governmental organizations about effective anti-poaching strategies. Protected areas often have limited resources available, and this forces ranger commanders to implement patrol strategies that are as efficient as possible. A standard anti-poaching strategy is to deploy ground based patrol teams around rhino locations and searching for poaching activity. Almost all protected areas in South Africa are fenced or partially fenced, and fence patrols play an important role in the early detection of illegal fence crossings.

The model was applied to study the interactions and dynamics of rhino poaching and two different patrol strategies in a virtual fenced park. The free software ‘NetLogo’ [11] was used to program the virtual park and agents. The model, its code, and the used parameters are available online at the following website: http://modelingcommons.org/browse/one_model/5016.

Having the agents in place, several scenarios or strategies can be simulated and compared in terms of how man rhinos have survived at the end of each simulation run. The different anti-poaching strategies were compared with a ‘worst-case scenario’: a virtual park without any ranger teams. For each patrol strategy, the number of ranger teams and the duration of their patrols were varied, in combination with different numbers of poachers. Ranger and poacher numbers were considered a categorical variable with three levels: 1, 2, and 4 poachers or rangers. Patrol duration was also considered as a categorical variable with three levels: 50, 100, and 200 events. The number of camps was set at 1 for all scenarios with rangers. Each simulation run lasts no longer than 10,000 events, but ends earlier if either all rhinos are killed, or all poachers are caught. The outcome variable was the number of surviving rhinos at the end of each run. Each combination of settings was ran 100 times.

The Shapiro-Wilk test was used to test for normality. Further analyses were performed with the Kruskal-Wallis test. A post hoc comparison using Dunn’s test with the Bonferroni adjustment was performed if the Kruskal-Wallis showed significant differences between the groups [12]. These conservative non-parametric methods were applied to reduce the possibility of type I error.

5.1 Standard Patrols

The effect of adding more ranger teams carrying out standard patrols was studied in a virtual park with 1 poacher, 2 poachers, and 4 poachers. The amount of time that the poachers and ranger teams are allowed to spend inside the park was fixed at 50 events. The number of surviving rhinos was significantly different between the number of poachers and number of ranger teams (Kruskal-Wallis test; H = 567.99; d.f. = 8; P < 0.001). The number of surviving rhinos increased with increasing ranger team numbers (Fig. 1a). There was no significant difference between one ranger or two ranger teams on the surviving rhino numbers when the poacher numbers were kept constant. The same applied for two rangers and four ranger teams for the same number of poachers.

Open image in new window

The effect of patrol duration was studied in a virtual park with 1 poacher, 2 poachers, and 4 poachers. The duration of a poacher’s hunt was fixed to 50 events. The number of rangers was set to one. The number of surviving rhinos was significantly different between the number of poachers and patrol duration (Kruskal-Wallis test; H = 680.58; d.f. = 8; P < 0.001). While the surviving rhino numbers decreases with increasing number of poachers, the results show no significant difference between the different patrol durations (Fig. 1b).

5.3 Comparison of Anti-poaching Strategies

The two anti-poaching strategies were compared with a ‘worst-case scenario’: a virtual park without any ranger teams. Two parks were created, one with one poacher, and one with four poachers. The duration of a poacher’s hunt was fixed to 50 events. The comparison provides insight in how effective each patrol strategy is compared to a park without any patrols. The number of surviving rhinos was significantly different between the different patrol types and number of rangers in the park with one poacher (Kruskal-Wallis test; H = 271.56; d.f. = 6; P < 0.001; Fig. 2a) and in the park with four poachers (Kruskal-Wallis test; H = 451.35; d.f. = 6; P < 0.001; Fig. 2b). As one might expect, the numbers of surviving rhinos was the lowest when four poachers were present. Still, when one poacher was present approximately half of the initial 70 rhinos survived. The average number of rhinos surviving was significantly higher when rangers were present and patrolling. In the virtual park with one poacher the difference between the two types of patrols decreases slightly with increasing rangers. Interestingly, the number of surviving rhinos was not significantly different between one fence patrol team and the two and four standard patrol teams in the one-poacher park. In the park with four poachers the differences between the patrol types are increasing with increasing ranger teams. The number of surviving rhinos was not significantly different between one fence patrol team and two and four standard patrol teams.

Open image in new window

This study introduces an agent-based model to study the dynamic interactions between animals, poachers, and rangers. The model provides a general framework that can easily be applied to a specific setting or context. In this study the model was applied to an abstract, virtually fenced park to explore the effect of an increase in ranger teams and an increase in patrol duration for standard patrols and fence patrols on the numbers of surviving rhinos.

The results show that the more rangers are being deployed, the less rhinos were poached. From a situational crime prevention perspective [2], providing more ‘boots on the ground’ is a way to increase the risks for poachers. The increase in formal surveillance leads to a higher chance of getting detected and reduces the chances of success. The increase in ranger teams also means that a greater area can be covered or more frequently covered. This improves the knowledge on the spatial and temporal distribution of illegal activities, and hence uncovering the dark figure of crime [6].

The general approach in building this model leads to the assumption that all ranger teams are equal: there are no differences in detection rate and in performance among the ranger agents, regardless of where or when it was deployed in the protected area. However, the effectiveness of one team and their ability to respond to a poaching event is heavily influenced by the amount of training they have received, and the equipment they carry on their patrol [13]. The effect and efficiency of having more patrol teams is only possible when all rangers went through proper training and have adequate supplies of good anti-poaching equipment. The strength of a general model is that it can be easily adjusted. For example, one can introduce variation in the ranger’s ability to detect poacher signs, or different levels of experiences. This can then lead into an analysis of the costs and returns for investing more in ranger teams, equipment, or perhaps in technology, like drones or GPS-transmitters for tracking animals. It also allows to test the proposed rule-of-thumb of 1 ranger per 20 km² by Bell and Clarke [14]. This was not possible in the current study because the applied context was still a general approach with units that do not map directly to real world measurements.

When resources in protected areas are limited, a different strategy is to go on longer patrols to increase patrol area coverage. However, the results of this study showed that longer patrols were not more effective in protecting rhinos than shorter patrols. Rangers need special training to be able to survive under harsh conditions and in rough environments for an extended period of time. Hence, law enforcement commanders might prefer to send out their teams on shorter patrols, rather than longer ones if indeed they are not more effective. Another study by Nyirenda and Chomba [15] found that shorter patrols of 2 to 8 days were more suitable for the Kafue National Park in Zambia, but stressed that this finding might not be applicable to other protected areas with different environments or in a different social-cultural context. Their statement also applies to the current model. In this study, only three levels of patrol duration were used in a virtual park where poachers only hunt for rhinos. Furthermore, the model also assumed that there are no changes in the ranger’s ability to detect signs throughout their patrol. In other words, the model ignores the possible effect of fatigue on the ranger’s ability. The longer waiting times between long patrols tries to take that into account, but also results in a lower frequency of patrols. This would also explain why no significant differences between patrol durations were found. The model can be further improved by creating a new variable that represents energy or fatigue of the ranger team and how it influences their detection rate.

The results showed that fence patrols were more effective in protecting rhinos than the standard patrols. More specifically, one fence patrol team was equally effective as two or four standard patrol teams. Fence patrols have a higher likelihood of picking up poacher signs, because poachers always start and end at the borders. In addition to that, while patrolling, the ranger agents also leave signs that influence the poacher’s decision-making and likely deflects them. As stated by Eck and Weisburd [16] “offenders avoid targets with evidence of high guardianship”. Hence, offenders seek out new areas or time periods with low guardianship. This is referred to as crime displacement [16]. In the case of wildlife crime, poachers might be spatially displaced to other areas with low security, or temporally displaced by operating at different hours. Especially temporal displacement can be of concern when deploying fence patrols as those are much more linear and predictable compared to standard patrols. While spatial displacement can be observed from the current model, it does not explicitly measure it. Hence, it is still unknown if the observed spatial displacement in the model is a good representation of actual displacement. The model can be easily adjusted to study spatial and temporal displacement. For example, evidence suggests that poachers are more active around a full moon period [17], probably because the light allows them to move through the bush more easily or faster. If more teams are being deployed around these times, poachers eventually are displaced to other moon phases or perhaps even to times during the day. It would be interesting to use the model to study how ranger patrols displace poachers in time and space.

The suggestions mentioned above are just some possibilities to further improve the current model. However, before any of these suggestions are worked into the model, it is important to stress a few limitations. First, no sensitivity analysis was performed to see how each parameter influences the model outcomes. This is especially important when one is interested in studying the conditions under which systems can become critical. ten Broeke et al. [18] suggested the ‘one-factor-at-a-time’ as a good starting point for a sensitivity analysis of an ABM. Furthermore, the current model was not calibrated to a specific protected area, but built around the general context of rhino poaching in South Africa. A true representation of the actual system requires more data on the agent’s behavior. In most cases, data on animal behavior is widely available, however data on ranger and especially poacher behavior is more difficult to collect. Such data usually comes from various sources, each with its own standards. This makes it challenging to create accurate rules for the ranger and poacher agents. However, the model can also be used to test different potential poacher strategies to see which one best reflects the observed behavior. Once these limitations are accounted for, a next step can involve applying the model to a specific context or problem.

This study presented a general model to study the dynamic interactions between the three agents that are involved in wildlife crime. The general, abstract approach was done intentionally to keep the model from getting too complex, yet with rules that result in realistic behavior of the animals, poachers, and rangers. The general framework of the model can easily be expanded to include more levels of complexity. When applied to rhino poaching under different anti-poaching strategies, the model provides some general insight in how the different strategies influence the behavior of rhino poachers. The results show that fence patrols are more effective in preventing wildlife crime than standard patrols. Strikingly, even deploying more ranger teams does not increase the effectiveness of standard patrols compared to fence patrols. The model presented here should be regarded as a first step to understand the complexity of wildlife crime and only benefits from further improvements and extensions.