Inferring wildlife poaching in Southeast Asia with multispecies dynamic occupancy models

Determining the ‘space race’ between sympatric species is crucial to understand the effects of interspecific interactions on the extinction risk of species threatened by poachers, predators, pathogens, and invasive competitors. Dynamic two-species occupancy models provide a flexible framework to decompose complex species interaction patterns while accounting for imperfect detection. In particular, these models can describe poachers-wildlife interactions by considering the occupancy, the extinction and colonisation probabilities of wildlife conditional on the presence or absence of poachers and vice versa. We apply our model to a case study on wildlife poaching in the Eastern plains of Cambodia. We used co-occurrence data extracted from the database of the SMART partnership to study the distribution dynamics between poachers and six ungulate species regarded as main prey of tigers. We used 4 years of survey data reporting the locations of ranger patrols on the detection of snares with visual detections or presence signs of the ungulates. Our results showed that a substantial proportion of the sites occupied by ungulate species went extinct over the years of the study while the proportion of sites colonised by poachers increased. We also showed, for the first time, that spatio-temporal heterogeneity in the patrolling effort explains a great deal of the variation in the detection of poachers and ungulates. Our approach provides practitioners with a flexible and robust tool to assess conservation status of species and extinction risk of wildlife populations. It can assist managers in better evaluating, learning and adapting the patrolling strategies of rangers.


Introduction
given site from one primary occasion to another (Fidino et al. 2018). To build the model, we first defined whether a site was initially occupied or not, and in 1 9 0 which state, which was conveniently captured by the initial state probabilities of HMM 1 9 1 (Gimenez et al. 2014). We proceed in two steps for clarity. The first step represents whether a 1 9 2 site is occupied or not by wildlife regardless of the presence or absence of poachers. The 1 9 3 following vector shows the probability of being in an unoccupied site U or a site occupied by W: absence of wildlife, occupy a site: the occupancy probability by poachers, conditional on the presence of wildlife and  Conditional on the initial occupancy state, we describe how the state at a site changes over 2 0 6 time assuming a Markovian process. We define the transition matrix T for a given site between 2 0 7 the states U, OP, OW, WP from primary occasion t to the next primary occasion t + 1 as follows: Each element of the matrix T is the transition probability from one of the occupancy states during 2 1 3 a given year to another or the same occupancy state in the next year. The matrix describes four 2 1 4 main processes: colonisation, extinction, replacement, and fidelity (MacKenzie 2018) that are 2 1 5 detailed in Table 1 in appendix 1.

1 6
These transitions occur at the beginning of each year (January). Within each primary 2 1 7 occasion, the transition matrix T is simply a diagonal matrix of 1's because the status of a site 2 1 8 does not change from one secondary occasion to the other according to the closure assumption. Here, we defined the secondary occasions on a monthly-based time interval (from February to 2 2 0 December). The last step of the HMM consists of linking the observation process to the partially (WD), only poachers were detected (PD), or both were detected (WPD). For the sake of 2 2 8 simplicity, we described detection probabilities of a species (poachers or wildlife) not conditionally on the presence/absence of the other (Fidino et al. 2018). We therefore defined the 2 3 0 observation matrix as: with ‫‬ the probability of detecting poachers in a site given that only poaching-related threats probability of detecting poachers and wildlife conditional on the presence of both. We accounted for the patrolling effort defined as the number of surveys carried out on patrolled on a given occasion, to 1 where the most GPS waypoints were recorded on a sampling 2 4 5 occasion (Appendix B). We also considered the nearest distance to a ranger station as another 2 4 6 site-specific covariate reflecting spatial variation in the patrolling effort. We defined the shortest 2 4 7 distance of each site to a road to account for effects of road access on the monitoring effort of To validate the performance of our model, we assessed the bias and precision in parameter 2 9 8 estimates using two simulation studies focusing on 1) variation in the occupancy design and 2) 2 9 9 variation in the sparseness of the data owed to differences in species detectability and occupancy 3 0 0 (Appendix C). A previous simulation study on occupancy modelling for a single species revealed 3 0 1 that the optimal design for monitoring a rare species was to sample more sites with fewer 3 0 2 surveys, and the one for monitoring a common species was to sample few sites with more with different ecology: one with restricted range (wildlife) and the other with a more widespread 3 0 6 distribution (poachers). In the second simulation study, we asked which level of sparseness in the 3 0 7 data may generate bias in parameter estimation when monitoring species, more or less difficult to 3 0 8 detected and with a more or less extended distribution. Over the study period, we collected 322 presence signs of our six key wildlife species B, Table B1).
The model best supported the "poacher winning" hypothesis and showed the lowest AIC.

2 4
We found that detection probabilities were related to the spatial and temporal variations of 3 2 5 patrolling effort ( and unrelated to rangers' patrolling effort (AIC = 3105.28, Table 3). We also found that 3 2 7 poachers' transition probabilities were not affected by wildlife, meaning that the dynamics in  following "wildlife winning" predictions and had less empirical support compared to the 3 3 5 "poacher winning" model from the first set of hypotheses (Table 3). The initial occupancy probability of poachers (0.06 ± 0 . 1 1 ) was lower than the initial poachers vs 0.20 (Fig. 2). The probability that poachers colonise sites was 3 4 7 independent of wildlife present and was very low (ߛ

4 8
As predicted by "poachers winning" hypothesis, the estimated probability of sites being 3 systems is the presence of rangers. For example, rangers change the distribution of snares (i.e., as "shields" to avoid poachers, similar to prey species sometimes using areas close to human case study, wildlife species were probably unable to differentiate between poachers and rangers, to which predator-prey theories advance our understanding of poacher-wildlife systems is a ultimately improve the effectiveness of anti-poaching interventions.

1 0
Our results also suggest that the patrolling effort in the PPWS and SWS study areas of The strength of some of our inferences is limited because, unfortunately, we could not the other. Unfortunately, our data was insufficient and too sparse to estimate such parameters.

3 3
Therefore, we opted for a simpler representation, prioritizing the tests on the effects of patrolling 4 3 4 effort on the detection of both wildlife and poachers. Also, we acknowledged that pooling six 4 3 5 species of ungulates into one guild defined as "tiger prey" could also generate bias in estimates, 4 3 6 as each species has its own abundance, occupancy dynamics and detection probabilities. To violated. We assumed the population to be closed during the 12-months' primary occasion.  We did not account for spatial autocorrelation among detections and occupancy states, Finally, assessing the quality of fit of occupancy models accounting for imperfect be computed in occupancy models under imperfect detection, but they no longer assess how well To our knowledge, this study is one of the few applications of a two-species occupancy 4 6 6 model extended to a multi-season version (see also Fidino et al., 2018, Yackulick et al. 2014. It poachers between t and t+1 -----Replacement ----from OW to OP ߱ ௐ probability that a site occupied by wildlife only is replaced by poachers from OP to OW ߱ ௐ probability a site occupied by poachers only is replaced by wildlife   Figure 1: Map of the study areas in Cambodia, the Phnom Prich Wildlife Sanctuary (PPWS) and Serepok Wildlife Sanctuary (SWS) within the green polygons. Sampling sites were defined as 10x10 km cells. We show the occurrence patterns of poachers (red crosses) and wildlife species (black dots). Brown lines represent the main patrolling path network (roads), blue lines describe the river streams, black triangles represent the ranger stations as well as temporary ranger's campsites. Note that only 81 cells were used for the analysis as we removed 17 sites that were never sampled (cells shaded in grey).

Figure 2:
Estimated ecological parameters with associated 95% confidence intervals. We provide the estimated occupancy probability for wildlife ( W ) regardless of the presence/absence of the poachers, occupancy probability for poachers which was independent of wildlife occupancy so we denoted it as ( P ) the extinction probability (ε) of a species (wildlife denoted by W or poachers noted by P) given the presence (without an upper bar) or absence (denoted by an upper bar) of the other (W or P) , the colonisation probability of a species (γ) conditional on the presence/absence of the other, the replacement probability of one species by another (ω). Estimates were obtained from the best model displayed in bold character in Table 3 with no effects of wildlife on the transition parameters of the poachers nor on their initial occupancy and spatiotemporal effect of the patrolling efforts on the detection probability of poachers and wildlife.

Figure 3:
Temporal changes in the probability that a site is occupied by poachers only (OP, in red) and occupied by wildlife only (OW, green), or occupied by both, poacher and wildlife (WP, blue), estimated over the study period.