Rangers and modellers collaborate to build and evaluate spatial models of African elephant poaching

Abstract

Globally, tens of thousands of wildlife rangers patrol wide areas and record evidence of poaching activity such as elephant carcasses and snares. Such data have significant potential to inform conservation, but patrols are non-random in space and time, so conclusions from raw patrol data may be biased. Here we model spatial patterns of elephant poaching based on detections of carcasses by ranger patrols in the Zambezi Valley, Zimbabwe (201 carcasses, 2000–2017), using different methodological scenarios to correct for patrol bias. We follow a participatory modelling framework, using interviews with practitioners (rangers and managers) to help build and evaluate these models. We found that poaching patterns in the bias-corrected scenarios differed among themselves and from the uncorrected scenario. Practitioners interrogated the credibility of the predictions in each scenario and thus helped discern true poaching patterns from those explained by patrol bias. We uncovered proximity to water as the strongest driver of poaching, likely reflecting both poacher and elephant behaviour. Our results show that it is essential to account for observer bias before developing management actions (such as ranger patrol strategies) from raw observational data. We further demonstrate the value of combining multiple lines of evidence (statistical models and interview responses) for more robust inference in the face of uncertainty.

Introduction

Monitoring trends within socio-ecological systems (species populations, illegal harvest rates, etc.) is essential for adaptive management, helping managers understand and manage change (Nichols and Williams, 2006). Evaluating anti-poaching strategies, for example, requires reliable measurement of real poaching trends. Data on biodiversity and threats are however difficult to gather at relevant scales, and are often biased and imprecise (Field et al., 2007). Time and resource constraints often mean that monitoring data are collected by people doing other jobs, such as wildlife rangers detecting snares while on patrol or fisherman providing records of bycatch species landed. Such opportunistic data present unique challenges to interpretation (Keane et al., 2011). A drop in the detection of poachers' snares, for example, may reflect a shift in patrolling to a ‘non-hotspot’ area, rather than an actual change in poaching levels.

Another challenge to interpreting observational data is the complexity of the underlying mechanisms generating the data. The behaviours of data generators (e.g. poachers), data collectors (e.g. rangers) and species of concern (e.g. elephants) are likely to interact in complex ways and their relative influence is difficult to disentangle. Dobson et al., (2018), for example, show how deterrence of poachers by rangers can confound inferred trends on the prevalence of illegal activity. Imperfect detectability of illegal activity (like bushmeat snares in thick forest; O'Kelly et al., 2018), and patrol observations that are biased towards certain areas (Critchlow et al., 2015), may similarly confound true patterns.

Participatory modelling is a promising way to design quantitative models that are robust to uncertainty arising from the bias and complexity discussed above (Voinov and Bousquet, 2010). Bringing together people familiar with the system of interest provides essential qualitative context to modelling (Milner-Gulland and Shea, 2017). These may be fisherman, wildlife rangers, or protected area managers that have a grounded understanding of how a system works (e.g. where elephant poaching happens) and how data are collected (e.g. what affects ranger movements). Participatory or collaborative modelling involves using the qualitative insights of on-the-ground practitioners and stakeholders in both the design and validation stages of statistical/mathematical modelling (Voinov and Bousquet, 2010). Quantitative models are vulnerable to the data and assumptions used to build them, while qualitative insights are often subjective or incomplete. Combining multiple lines of evidence (statistical outputs and interview responses) is a useful way of addressing this uncertainty. Engaging practitioners in modelling may also create a sense of ownership that amplifies its real-world relevance (Basco-Carrera et al., 2017).

Globally, tens of thousands of park rangers spend significant amounts of time on patrol, encountering plants, animals, and illegal activities. Such data are becoming an increasingly important source of information for both science and conservation (Gray and Kalpers, 2005; Moore et al., 2018). The MIKE programme (Monitoring of the Illegal Killing of Elephants), is a high-profile example of the use of data collected by ranger patrols to inform local and international conservation policy (CITES Secretariat, 2019). MIKE covers 60 sites across Africa, within which >19,000 elephant carcasses have been detected by rangers to date. The data have been used in high profile global and continental analyses (Wittemyer et al., 2014; Hauenstein et al., 2019), but less so at the local site level. In this paper, we investigate spatial patterns in poached elephant carcasses detected by rangers at a MIKE site in the Zambezi Valley, Zimbabwe. We combine quantitative models with interviews with widlife rangers and their supervisors to address the following research questions:

• What spatial patterns are evident in poached elephant mortalities at the case study site?

• How are these patterns influenced by monitoring bias?