Managing African Swine Fever: Assessing the Potential of Camera Traps in Monitoring Wild Boar Occupancy Trends in Infected and Noninfected Zones, Using Spatio-temporal Statistical Models

The recent spreading of African swine fever (ASF) over the Eurasian continent has been acknowledged as a serious economic threat for the pork industry. Consequently, an extensive body of research focuses on the epidemiology and control of ASF. Nevertheless, little information is available on the combined effect of ASF and ASF-related control measures on wild boar (Sus scrofa) population abundances. This is crucial information given the role of the remaining wild boar that act as an important reservoir of the disease. Given the high potential of camera traps as a non-invasive method for ungulate trend estimation, we assess the effectiveness of ASF control measures using a camera trap network. In this study, we focus on a major ASF outbreak in 2018-2020 in the South of Belgium. This outbreak elicited a strong management response, both in terms of fencing off a large infected zone as well as an intensive culling regime. We apply a Bayesian multi-season site-occupancy model to wild boar detection-nondetection data. Our results show that (1) occupancy rates at the onset of our monitoring period reflect the ASF infection status; (2) ASF-induced mortality and culling efforts jointly lead to decreased occupancy over time; and (3) the estimated mean total extinction rate ranges between 22.44% and 91.35%, depending on the ASF infection status. Together, these results confirm the effectiveness of ASF-control measures implemented in Wallonia (Belgium), which has regained its disease-free status in December 2020, as well as the usefulness of a camera trap network to monitor these effects.

The recent spreading of African swine fever (ASF) over the Eurasian continent has been acknowledged 17 as a serious economic threat for the pork industry. Consequently, an extensive body of research focuses 18 on the epidemiology and control of ASF. Nevertheless, little information is available on the combined 19 effect of ASF and ASF-related control measures on wild boar (Sus scrofa) population abundances. This 20 is crucial information given the role of the remaining wild boar that act as an important reservoir of the 21 disease. Given the high potential of camera traps as a non-invasive method for ungulate trend 22 estimation, we assess the effectiveness of ASF control measures using a camera trap network. In this 23 study, we focus on a major ASF outbreak in 2018-2020 in the South of Belgium. This outbreak elicited 24 a strong management response, both in terms of fencing off a large infected zone as well as an intensive 25 culling regime. We apply a Bayesian multi-season site-occupancy model to wild boar detection-26 nondetection data. Our results show that (1) occupancy rates at the onset of our monitoring period 27 reflect the ASF infection status; (2) ASF-induced mortality and culling efforts jointly lead to decreased 28 occupancy over time; and (3) the estimated mean total extinction rate ranges between 22.44% and 29 91.35%, depending on the ASF infection status. Together, these results confirm the effectiveness of 30 ASF-control measures implemented in Wallonia (Belgium), which has regained its disease-free status 31 in December 2020, as well as the usefulness of a camera trap network to monitor these effects. 32

Introduction 34
African swine fever (ASF), a viral disease that causes high mortality among domestic pigs (Sus scrofa 35 domesticus) and wild boar (Sus scrofa), originates from East Africa and is regarded as one of the most 36 important threats to the European pig industry. Recently, ASF has been re-introduced to the wild boar 37 populations on the European mainland, presumably due to infected meat spills in the environment (1). 38 Most likely, this spillage mediated the recent spread of ASF through a new epidemiological cycle, 39 designated the wild boar-habitat cycle, which involves both direct and indirect viral transmissions. 40 Direct transmissions occur through contacts among wild boar, whereas indirect cases result from viral 41 reservoirs in the environment, such as ASF-infected carcasses (2). This new role of wild boar in the 42 epidemiology of ASF has led to new management guidelines of wild boar populations in infected areas 43 (3, 4). Management strategies include continuous carcass removal from the infected zone, coupled with 44 intense culling of wild boar within a buffer zone (5). Together, these strategies are expected to 45 effectively reduce ASF transmission by removing viral sources from the environment in the infected 46 zone and by depleting the susceptible wild boar population in the buffer zone. The latter is essential, 47 since the number of individuals remaining in the host population of the buffer zone will determine the 48 probability of the spread to a noninfected area, i.e., host threshold density. In the infected zone, after 49 the epidemic phase, culling of the remaining wild boar will determine the probability for the disease to 50 become endemic, i.e., critical community size (6). To evaluate measures aimed at counteracting ASF, 51 sound information on the joint effect of the disease and culling efforts on population trends of wild 52 boar within the managed areas is crucial (3). 53 Over the last decade, the use of remote cameras, henceforth referred to as camera traps (CTs), has 54 become popular when monitoring trends in medium-size to large-size mammals, including wild boar 55 (7, 8). Photographic captures (i.e., detections) by CTs can be translated into information on the 56 distribution and density of a focal species. However, density estimation by CTs is still hindered by 57 imperfect detection in many cases (i.e., not detecting a focal species, when present) (9). Given the 58 elusiveness and nocturnality of wild boar, it is among the species subjected to severely limited 59 detectability. Moreover, traditional density estimation methodology requires individual identification, 60 hence cannot be applied to many common species that lack natural markings, including wild boar (10, 61 11). Statistical frameworks such as the random encounter model (REM) allow for density estimation 62 of unmarked populations, while accounting for imperfect detection, using CTs (12, 13). However, the 63 need for auxiliary data collection, restricts the use of REM in many cases (14). Occupancy models on 64 the other hand overcome both imperfect detectability and the need for individual recognition of 65 animals, without requiring additional information. They proceed by simultaneously estimating site-66 occupancy and the probability of detecting a focal species, given its presence (15). Extending 67 occupancy models to so-called multi-season site-occupancy (MSO) models, enables estimating rates 68 expressing population changes through time. One of these rates, the extinction rate, is of prime interest 69 when assessing the combined effect of a viral disease, such as ASF, and culling efforts on a host 70 population. 71 In the current study we evaluate the use and effectiveness of a CT network to monitor wild boar 72 population trends throughout the recent ASF epidemic in Wallonia (Belgium  Table  108 2). A posteriori, the strata were superimposed by a hexagonal grid layer (x-spacing of 500 m, area of 109 21.65 ha/ site) ensuring that each camera was assigned a unique grid cell (Figure 1). Throughout the 110 sampling period, camera locations were fixed. All cameras were installed by mounting them on trees 111 approximately 50 centimetres above ground, facing North. We did not use baiting, nor did we select 112 for trails. Monthly check-ups were performed to determine battery levels and to verify camera 113 operability. Each camera trigger was followed by a series of five photographs, without a delay between 114 consecutive triggers. All images were manually annotated, using the Agouti software platform (19). 115 After omitting data from the excluded zone (5 deployments; 5.15%) (Figure 1) is within the ASF-infected zone, 0 otherwise (Figure 1). An overview of these covariates and a priori 127 defined models are given in Table 1. For model-specific predictions (P(Model)) consult 128 Supplementary Table 3. 129

Statistical model 130
We analyse the CT data using a longitudinal multi-season occupancy model (MSO), defined as a state-131 space model (20), to make inference on wild boar's site-occupancy. The sampling grids used are 132 smaller than wild boar's home range, hence occupancy should be interpreted as habitat use (21). 133 Detection histories were constructed using the R package CamtrapR (22). For site = 1,2, … , , at 134 survey day = 1,2, … , , in observation month = 1,2, … , , the detection history is 1 when wild boar 135 were observed during a 24-h period ( = 1) or 0 when no boars are caught on camera that day 136 ( = 0). These are assumed to follow a Bernoulli distribution, such that, 137 where is the probability of detecting the focal species and is the latent occupancy status 138 (unoccupied = 0; occupied = 1) at site during observation month . Note that we do not use 139 survey day-specific, nor site-specific covariates to model , hence the detection probability 140 simplifies to . The occupancy status is modelled as, 141 Where is the occupancy probability, from now on simply referred to as 'occupancy', at site during 142 observation month . Unlike dynamic MSO, we do not take probabilities of wild boar surviving or 143 colonizing a site from observation month to + 1 into account, as the high degree of zero-inflation 144 in our data complicates joint inference on all these processes. We define = { , }, which 145 collects all processes that will be modelled as a function of covariates and random effects, using a logit 146 link. A general model formulation for , = 1, 2, can be defined as 147 where are intercepts, are vectors of process-specific slope parameters with their corresponding 148 covariate matrix . The term models spatially unstructured overdispersion as a normally distributed 149 random effect, ,1 is a smooth function modelling temporal variation for each observation month and 150 ,2 is an isotropic two-dimensional smooth function modelling spatial variation in occupancy patterns, 151 for the longitude and latitude of each site's centroid. Both ,1 and ,2 are modelled using 152 Gaussian processes (GP), with an exponentiated quadratic covariance function. While ,1 uses an exact 153 GP, we model ,2 by means of the Hilbert space reduced-rank Gaussian process (HSGP) approach as 154 the number of sites in our study area is large (23, 24). This approach yields substantial speed gains 155 when dealing with large number of sites through approximate series expansions of the GP's covariance 156 function. 157 Model fitting was performed using Stan (via the R package rstan), a probabilistic programming 158 language that enables Bayesian estimation through a dynamic Hamiltonian Monte Carlo (HMC) 159 sampler (25). For each MCMC iteration, we also derive site-specific growth rates = ( −1) , 160 average monthly growth rates ̅ = most prior evidence is placed on scales that can be estimated from the data (i.e., larger than the smallest 166 difference between any pair of CT locations and smaller than the largest difference between any of 167 these pairs). 168 The full model that would contain two random effects terms for each of these processes, in addition to 169 covariates, was computationally infeasible to fit and, furthermore, does not necessarily reflect a 170 sensible data-generating process. Hence, we consider a set of sensible reduced models based on 171 ecologically plausible considerations, preventing multicollinearity, and computational feasibility 172 ( Table 1). Multicollinearity was avoided by including one of two covariates, when their Spearman rho 173 correlation estimate | | < 0.6. Subsequently, we select the most appropriate model by means of a 174 model selection procedure. 175 Model selection through approximate leave-one-out cross-validation was performed using the R 176 package loo (27). Following the authors' recommendations, leave-one-out (LOO) expected log-177 predictive densities were used to rank our a priori selected candidate models ( Table 2). Our ranking 178 procedure consists of a two-step approach. First, the top-ranked occupancy model is retained by 179 comparing LOO for selected combinations of fixed and random effects at the occupancy ( ) level, 180 while keeping detectability ( ) constant (Table 2, step 1). Secondly, the detection process is modelled 181 using fixed effects only, while adopting the top-ranked occupancy model from step 1 ( Table 2, step  182 2). 183 All models were fitted using four parallel MCMC chains with 4000 iterations, which included 2000 184 iterations that were discarded as burn-in iterations for all candidate models; this always resulted in 185 satisfactory convergence (Supplementary Figure 1) Table 2 presents the model selection process, which yielded a final model consisting of an occupancy 193 process and detection process that will be detailed in the following subsections. 194

Detectability 195
This is a provisional file, not the final typeset article The detection model according to LOO ( can be observed (Supplementary Figures 2-3). 203

Occupancy 204
All of the tested covariate combinations performed better than the intercept model ( 1) (Figure 2), reveal an overall decline in occupancy, both in the ASF-infected and noninfected zone. 216 Finally, prediction maps for the estimated occupancy from March 2019 until May 2020, are displayed 217 in Figure 3 (see Supplementary Figures 4-5 for 2.5 th and 95 th percentile maps). 218

Occupancy growth rate 219
Posterior means of occupancy growth rates, i.e., , are lower than one regardless of the site and season 220 (Supplementary Figure 6). For , , total growth rates (in fact, extinction rates, due to their negative 221 trend) in occupancy, posterior means range between 0.0865 and 0.7756 (0.9135 and 0.2244), while 222 those for average monthly growth rates ̅ lie between 0.8049 and 0.9757 (0.1951 and 0.1243). Finally, 223 posterior mean and 95% HPDI for , and ̅ averaged over the ASF management zones (designated 224 , and ̅ ) are given in Table 3.

226
Discussion 227 To assess the effectiveness of CTs to monitor wild boar population trends throughout the recent ASF 228 epidemic in Wallonia (Belgium), we have built a spatio-temporal MSO model. This was done 229 according to a two-step approach, selecting the best detection covariates and subsequently occupancy 230 covariates with respect to the LOO (27) statistics from a set of a priori defined models. 231

Detectability 232
For all model comparisons (relative to the top-ranked detection model) the standard errors for Δ LOO 233 values are smaller than two times |Δ LOO|, hence a certain degree of uncertainty as to which model 234 provides the best fit to the data remains after our selection procedure (Table 2, step 2). Nevertheless, 235 we believe that using a GP to model monthly temporal variation in wild boar detection probability is a 236 sensible choice, given the ability of GPs to balance ecological realism with model flexibility (29). 237 When using CTs, detection probabilities are known to be affected by, among others, vegetation 238 denseness, background surface temperature and weather conditions (30), all of which depend on the 239 seasonal variation to some extent. Hence, a certain degree of seasonality in detection probabilities is 240 not uncommon. Morelle et al. (17) report higher probabilities of detecting wild boar in summer months 241 compared to fall (4.90E+04), winter (4.34E-03) and spring (4.90E+04). In this study, posterior mean 242 detection probability for wild boar is low, although some additional heterogeneity attributed to the 243 observation month was observed. In 2019, summer months display a somewhat higher probability of 244 detecting wild boar as compared to winter months, yet there is insufficient evidence that a periodic 245 trend exists. We rather suggest that the main effect at play, is a density-dependent effect (31, 32), more 246 specifically a decline in detection probability that coincides with a decreasing wild boar density. In 247 addition, the intensive culling regime adopted throughout the ASF epidemic possibly led to an 248 increased risk perception by wild boar, incentivizing them to restrict their movements and seek hiding 249 places. Lower activity levels negatively relate with photographic rates (33). Similarly, low probabilities 250 of detecting wild boar could reflect decreased movement. 251

Occupancy 252
The top-ranked occupancy model consists of a multiplicative effect the ASF infection status ( ) of 253 the zone, the observation month ( ) and the proportion of deciduous forest land-use class ( ), which 254 is followed by a fully additive model of these covariates (  Table 1). This is not surprisingly given 260 that the zone was already infected with ASF for several months (i.e., since September 2019) before the 261 study's onset. Although it is uncertain whether the low initial occupancy in the infected zone is driven 262 by ASF alone, mortality rates approaching 100% have been reported (34). Interestingly, the inclusion 263 of a HSGP that accounts for extra-variability due to spatial correlation did not prevent the existence of 264 a strong boundary between the ASF-infected and noninfected zones with respect to their occupancy 265 patterns (Figure 3). This boundary effect remained, even after omitting the information about ASF 266 infection status from our model (results not shown). In that case, the variation in occupancy previously 267 accounted for by , was taken up by the spatial GP ( 2 ). Hence, we are confident that this boundary 268 effect is not an artefact. Instead, we argue that a sudden rise in occupancy at the infected-noninfected 269 boundary results from having fenced off the ASF-infected zone (Figure 1). The latter effectively 270 creates two separate subpopulations with each of them subjected to distinct population dynamics, 271 leading to an unprecedented extinction rate in the infected zone. 272 Despite the strong impact of ASF, the infected zone seems to have had quite some refugee sites that 273 display higher occupancies compared to surrounding areas as of March 2019 (Figure 3). As the 274 epidemic progressed, occupancy dropped in most of these subareas, with only one patch in the South 275 displaying markedly higher occupancy towards May 2020. Given the remoteness of this patch, it could 276 be that wild boar in this area were shielded from ASF to some extent. A more likely explanation is that 277 these refugee sites reflect the area's suitability for remaining wild boar in terms of habitat quality and 278 food availability. Similarly, we argue that latent ecological preferences drive the heterogeneity in 279 occupancy observed within the noninfected zone, where higher occupancies were observed in the 280 central axis (horizontally) throughout the study period (Figure 3). Indeed, looking at the spatial random 281 effect alone, both the Southern patch of the ASF-infected zone and central axis of the noninfected zone 282 are associated with some of the highest values (Supplementary Figure 7). A number of potential 283 This is a provisional file, not the final typeset article ecological drivers of wild boar occupancy were observed and subsequently modelled; we have 284 considered the proportion of broad-leaved tree land cover class, which is known to positively affect its 285 occupancy (35-37), as a fixed effect in our final model. We find this covariate to be not significant at 286 the 5% significance level, but significant at the 10% level (results not shown), which suggests an effect 287 that needs further investigation in future studies. 288 Finally, we obtain an overall declining trend (posterior mean OR( ) of 0.76; Supplementary Table  289 4) in wild boar occupancy for both ASF-infected an noninfected zones (Figure 2). This indicates that 290 occupancy probabilities continue to drop as a consequence of the joint effect of ASF-induced mortality 291 and culling efforts in the infected zone. While the decline in occupancy in the buffer zone results from 292 culling alone. Together, these findings support the management strategies adopted in Wallonia 293 (Belgium). In addition, we find that occupancy declines have different rates between the zones, with a 294 more moderate decline seen in the ASF-infected zone (posterior mean OR( ⋅ ) of 1.13; 295 Supplementary Table 4). Possibly, differences in hunting pressure could explain this variation in rates 296 of occupancy decline. Although wild boars have been culled in both ASF-infected and noninfected 297 zones, the latter was more densely populated throughout the entire study period, which likely reduces 298 the search effort by hunters and leads to increased hunting success (Supplementary Table 6). In 299 addition, we note that between the epidemic-onset and its peak (September 2019 -February 2019), an 300 occupancy decline was likely much higher in the ASF-infected zone. Importantly, our model reveals 301 that the ⋅ interaction term is insignificant when looking at 95% HPDI. Uncertainty about the 302 existence of zone-specific occupancy rates, is also reflected in the small Δ LOO between a model with 303 and one without the interaction term ( Table 2, step 1). 304

Occupancy growth rate 305
We will not discuss growth rates in depth, since they carry on the same messages as the occupancy 306 probabilities discussed earlier, but see Supplementary Figure 6 for a graphical representation. 307 However, it is worthwhile to briefly focus on total occupancy growth rates, as they provide a summary 308 statistic for net change in occupancy.

Limitations 317
The data used in this study do not cover the full ASF episode as it occurred in Wallonia (Belgium). As 318 such, we are unable to report on the full course of the epidemic. Secondly, it has been reported by (26) 319 that sample sizes smaller than 40 lead to insufficient power to detect declines in occupancy under most 320 circumstances. Here, we deploy 69 cameras in the ASF-infected and only 23 cameras in the noninfected 321 zone. Nonetheless, we were able to capture significant declines in occupancy for both zones throughout 322 the study period. Importantly, both ASF management zones had sampling intensities higher than the 323 best scenario (2% of sites sampled) considered by Banner et al. (26). From a modeller's perspective, 324 we did not attempt a full spatio-temporal analysis. However, we believe that it is reasonable to assume 325 that temporal dynamics in site-occupancy are spatially independent, given the relatively small surface 326 area (ASF-infected: 162.826 km 2 , noninfected: 48.229 km 2 ) of both ASF management zones. Finally, 327 we did not include structured and unstructured random effects for both the detection and occupancy 328 process, due to unidentifiability. 329

Conclusion 330
Based on our results, we conclude that ASF infection status was the main driver of wild boar occupancy 331 at the beginning of the monitoring period, which results in a clear difference in occupancy between 332 ASF-infected and noninfected zones. Moreover, we find that fencing off the infected zone has helped 333 to maintain this sharp contrast in occupancy probability throughout the study period. Starting from 334 March 2019, our model strongly supports an overall decline in occupancy until May 2020. 335 Additionally, we attribute a steeper decline in the noninfected zone to the management strategies, 336 adopted to counteract the progression of ASF. Together, these results confirm (1) the effectiveness of 337 ASF control measures implemented in Wallonia (Belgium), and (2) the potential of using a CT-network 338 to monitor the impact of both the disease and the management actions on wild boar populations during 339 an ASF outbreak. 340

Conflict of Interest 341
The authors declare that the research was conducted in the absence of any commercial or financial 342 relationships that could be construed as a potential conflict of interest. 343

Author Contributions 344
Each author's contribution is described using the CRediT roles.   Table 1: A priori defined occupancy (step 1) and detection (step 2) models. Top-ranked models for 502 each step are indicated in bold. Abbreviations: ASF infection status ( ), z-scored proportion of 503 broad-leaved tree land cover class ( ), observation month ( ), biannual (springsummer, autumn 504 winter) seasons ( ), quarterly (spring, summer, autumn, winter) seasons ( ), smooth function 505 for temporal variation 1 ( ), smooth function for spatial variation 2 ( ). Intercepts and slope 506 parameters are given by and respectively. 507  This is a provisional file, not the final typeset article