Dynamics and predicted distribution of an irrupting ‘sleeper’ population: fallow deer in Tasmania

27 Sleeper populations of non-native species can remain at low abundance for decades before 28 irrupting. For over a century, fallow deer ( Dama dama ) in the island state of Tasmania, Australia, 29 remained at low abundance and close to the region in which they were released. Recently, there are 30 indications the population has increased in abundance and distribution. Here, we spatially quantify 31 the population change using a time series of annual spotlight counts from 1985 to 2019 (total of 32 5,761 counts). Next, we predict the potential for further range expansion, using global occurrences 33 to characterise the species’ climatic niche, and remote-camera surveys (n = 3,225) to model fine- 34 grained habitat suitability. Spotlight counts of fallow deer increased by 11.5% annually, resulting in a 35 40-fold increase from 1985 to 2019. The core distribution increased 2.9-fold during this 35-year 36 period, and now spans c. 27% of Tasmania’s land area. Satellite populations have established in 37 locations where farmed deer have escaped or been released, suggesting that humans have 38 facilitated some of the range expansion via new introduction events. Based on climate and habitat 39 suitability, our models predict that 56% of Tasmania is suitable under the current climate. This 40 suggests range expansion is likely to continue unless the population is actively managed, which could 41 include the eradication of satellite populations and containment of core populations. This case study 42 cautions that despite over a century of slow population growth, sleeper populations of non-native 43 species can abruptly increase. 44


Introduction 46
Sleeper populations of introduced species can persist at low abundance for decades before being 47 Spotlighting counts 110 The Tasmanian State Government has conducted annual spotlight surveys along 172 transects across  111 northern, eastern, and central Tasmania from 1985 to 2019. The surveys were established to 112 monitor harvested herbivore populations, while also recording all sightings of free-ranging mammal 113 species, including fallow deer . The same transects are surveyed each year; 114 however, the number of transects increased from 132 in 1985 to 172 in 1993 (Table S1). The 115 transects do not sample large areas of western Tasmania, largely due to inaccessibility, and also 116 because wildlife are not harvested in those areas Driessen, Mallick 2003). 117 Each transect comprises a 10-km section of road and adjacent land. Transects are surveyed at night 118 with the aid of a spotlight from a vehicle driven at 20 km/hr (for details, see . 119 Transects are surveyed once each year during the summer months. This ensures comparability 120 between years but prevents the use of statistical techniques like occupancy modelling that require 121 repeat surveys within a year. We consider the count of fallow deer per transect as an index of 122 abundance. 123

(ii)
Remote camera surveys 124 We used camera traps to model the current distribution and habitat preferences of fallow deer in 125 Tasmania. We compiled a database of our own camera deployments between 2014 and 2020, 126 totalling 3,225 camera sites, and 315,120 camera-nights (Table S2). For each camera, we recorded 127 the number of unique detection events of fallow deer, the duration of camera deployment, and 128 environmental covariates (described below). We defined a detection event as unique if at least 30 129 min separated consecutive occurrences of fallow deer on the same camera. 130 Global occurrences of fallow deer 131 To characterise climatic conditions suitable for fallow deer, we collated occurrence records from 132 regions where fallow deer are likely to be near to equilibrium. We used citizen-science data from 133 Europe, where fallow deer occur as both a native and long-established introduced species, and the 134 Australian mainland, where fallow deer have been widely introduced across a broader range of 135 environmental conditions than in Tasmania. 136 For Europe, we downloaded 58,381 presence-only occurrences of fallow deer from the Global 137 Biodiversity Information Facility (GBIF 2020). We selected only human-observed records (i.e., 138 excluding fossil records). To reduce the influence of intensively sampled areas, we randomly thinned 139 records so that at least 5 km separated occurrences, leaving 4,726 occurrences ( Fig 1A). Substantial 140 regional biases in survey effort were apparent across Europe, with many fewer mammals reported in 141 eastern than in western Europe (Fig 1A). Ignoring this survey bias would increase the likelihood of 142 misclassifying areas with low survey effort as unsuitable for fallow deer (Phillips et al. 2009). Thus, as 143 recommended (Phillips et al. 2009), we selected pseudo-absences according to the same survey bias 144 that generated the occurrences. We did this using the reported occurrences of any mammal in the 145 GBIF database as a proxy of relative survey effort across Europe. From these mammal occurrences, 146 we randomly selected 4,726 records as pseudo-absences. Distributing pseudo-absences in this way 147 helped to distinguish heavily sampled areas where fallow deer are probably absent (e.g., northern 148 Norway; Fig 1A) from areas where there is low survey effort and that could contain suitable habitat 149 (e.g., eastern Europe; Fig 1A).  unconfirmed and probably erroneous records from the South Australian desert and randomly 174 thinned records so that occurrences were separated by at least 5 km, leaving 490 occurrences (Fig  175   1B). We distributed pseudo-absences across the area we expect fallow deer have had the 176 opportunity to occupy. We also included areas around failed introduction attempts that might 177 indicate unsuitable habitat; most notably the Cobourg Peninsula, Northern Territory, where 52 178 fallow deer were introduced in 1912, but did not establish (Chapman, Chapman 1980  where β1 and β2 denote intercepts, and random field denotes a shared spatial random field. We 220 modelled both data sources using the negative binomial distribution to account for overdispersion in 221 the count data. From this model, we predicted a map of fallow deer distribution. We note that due 222 to imperfect detection, this map inherently represents a conservative estimate of the current 223 distribution. 224 We fitted the spatiotemporal and joint-likelihood models using v2. Because fallow deer have probably not yet dispersed to all suitable habitat, we modelled fine-247 grained habitat suitability using camera data from the core distributional range in Tasmania where 248 they are likely to be closer to equilibrium. We defined the core range using a 90% minimum convex 249 polygon around the locations where fallow deer have been observed on camera (Fig 5). There were 250 674 camera sites within the core-range polygon, of which we withheld 20% as a test dataset for 251 model validation. We projected the model of habitat suitability in the core range across all of 252

Tasmania. 253
We compiled five explanatory variables that we expect would influence fallow deer abundance and 254 distribution: i) the percentage of grassy vegetation types (%grassyVeg); ii) the percentage of forest 255 cover (%forest); iii) the percentage of alpine treeless vegetation (excluding alpine grasslands); iv) 256 terrain ruggedness index and v) year of survey. For covariate details, see Table S3. We expected that We expected that climate would influence the broad-scale distribution of fallow deer, and that deer 286 would discriminate suitable habitat mostly from within climatically suitable areas. Nevertheless, it is 287 possible that deer could use habitat that our models suggested was in climatically unsuitable areas. 288 We therefore used a weighted product model (Tofallis 2014) to combine the maps of climate and 289 habitat suitability. We first scaled the maps of climate and habitat suitability by dividing pixel values 290 by the maximum value of each map. This constrained suitability to the range 0-1 but did not force 291 any values to zero, which is advantageous because it preserves the proportionality of the data 292 (Tofallis 2014). Next, we combined these scaled maps of climate (C) and habitat suitability (H) using a 293 weighted product model where each map is raised to the power of its weight and then the maps are 294 multiplied together. We determined the weights for each map by selecting the combination of 295 weights, constrained to sum to one, that maximised the AUC statistic for the 20% of camera data 296 that was withheld from model fitting. This resulted in a combined map that was calculated as C 0.78 × 297 H 0.22 (Fig 6C). Finally, we divided the resulting map by the maximum value, constraining relative 298 suitability values to a maximum of one. 299

Results 301
Population trends 302 The annual spotlight surveys revealed a steady and large increase in the spotlight counts of fallow 303 deer from 1985-2019 (Fig 2A). An exponential model fitted the data well (coefficients ± SE: Intercept 304 = -218.18 ± 23.6, βyear = 0.109 ± 0.012; adjusted R 2 = 0.71). The exponential model suggests the 305 spotlight counts have increased at an approximately constant rate of 11.5% (SE: 1.2%) per year over 306 the last 35 years, representing a 40-fold increase from 1985 to 2019 (Fig 2A). The exponential model, 307 however, did not explain a large increase followed by a decline in deer detections in the mid-2000s 308 (Fig 2A). This pattern coincided with a period of above-average rainfall followed by a period of well-309 below-average rainfall across much of the species' Tasmanian range ( Fig S2). The number of deer 310 reported as being killed under crop protection and hunting permits has increased substantially in the 311 last five years (Fig 2C), but this offtake has evidently not prevented population growth. 312 The spatiotemporal model of the spotlight surveys shows that fallow deer distribution has steadily 313 expanded (Fig 2D), with a 2.9-fold increase in the model-estimated area of 'core' distribution since 314 1985-1989 (Fig 2B). The joint likelihood model, which simultaneously modelled the spotlight and 315 camera datasets, conservatively estimates that fallow deer now occupy c. 27% of Tasmania's land 316 area (Fig 3). Much of the expansion in deer distribution has occurred in areas where there are 317 anecdotal reports of farm escapes or intentional releases (Fig 2D). The climate suitability models revealed that large areas of Tasmania that are currently unoccupied 344 by fallow deer have climates that could support the species (Fig, 4E). Maximum temperature of the 345 warmest month and minimum temperature of the coldest month had the highest relative 346 importance ( Fig 4C). The concave response curves for the temperature variables and average annual 347 precipitation suggest upper and lower constraints of climate suitability. For instance, maximum 348 temperatures of less than ~18°C in the warmest month, and minimum temperatures less than ~-7°C 349 in the coolest month, resulted in an abrupt reduction in climate suitability (Fig 4D). Highly rugged 350 terrain had a negative effect on fallow deer occurrences (Fig 4D). All models performed better than 351 random, with out-of-sample AUC values of 0.89, 0.75 and 0.76 for the random forest, GAM and 352 boosted regression tree, respectively. Based on an ensemble of these models, weighted according to 353 out-of-sample AUC (Marmion et al. 2009), much of Tasmania has a suitable climate for fallow deer, 354 with highest suitability in the east and the north and the two largest off-shore islands (Fig 4E). Due to 355 its high annual rainfall the south-west of Tasmania

Fine-grained habitat suitability 366
The fine-grained habitat suitability model based on camera trap detections of fallow deer revealed 367 large areas of potentially suitable but as-yet unoccupied habitat in northern, eastern and southern 368 Tasmania ( Fig 5B). All top-performing models included a positive effect (Fig 5C) of %grassyVeg within 369 1000 m of a camera (variable importance = 1; Table S4). There was strong support for topographic 370 ruggedness negatively influencing (Fig 5C) deer detections, with similar support for this variable at 371 spatial scales of 500 m and 1000 m (cumulative importance of the different spatial scales of 372 ruggedness = 0.93; Table S4). Survey year had a positive effect on deer detections (Fig 5C). Given the 373 strong positive effect of %grassyVeg, most predicted suitable habitat (Fig 5B) is in the agricultural 374 regions of Tasmania or areas with native grasslands and eucalypt woodlands with grassy 375 understories. Importantly, though, habitats without a major grass component still seem able to 376 support low deer densities (i.e., positive y-intercept; Fig 5) We combined the maps of climate and habitat suitability using weights that maximised the AUC of 392 withheld test data, indicating a weight of 0.78 for habitat suitability and 0.22 for climate suitability 393 ( Fig 6C). The combined map of climate-habitat suitability (Fig 6D) suggests that 56% of Tasmania is 394 suitable habitat (Fig 6E, categorized based on a threshold of 0 (Fig 6E), this would represent a doubling of 396 the current range extent (Fig 3). introduced grasses, which are possibly an even more palatable food source for fallow deer (Nugent 471 1990). 472 Some patches of high montane grasslands and shrublands in the northern and north-eastern 473 TWWHA appear to be suitable habitat for fallow deer (Fig 5B). maximised out-of-sample AUC, which we expect will yields a more realistic potential distribution of 494 fallow deer in Tasmania. 495 Given the deer population is likely to continue to expand into unoccupied, suitable habitat, the next 496 important research step is to develop a spatially explicit demographic-dispersal model. This model 497 could be parametrised using a pattern-oriented modelling approach, which provides a systematic, 498 data-oriented way of tuning complex simulation models by matching the simulation results with 499 independent real-world data, known as 'targets' (Grimm, Railsback 2012;Grimm et al. 2005). In our 500 case, Approximate Bayesian Computation (Csilléry et al. 2010) could be used to fine-tune the model 501 parameters based only on the simulation parameterisations that produce results that match the 502 'targets' for site occupancy, distribution and abundance (e.g., Fig 2). Once parameterised, the model 503 could forecast how deer distribution is likely to change under a suite of different management 504 scenarios and compare counterfactual scenarios that investigate the long-term effects of deer 505 escapes and unlawful releases. This approach could also test hypotheses relating to the long lag-506 times in range expansion, inflexion points and triggers of population growth. 507

Management suggestions 508
In light of our findings, we make several management suggestions to reduce the threats posed by Next, we suggest that deer be prevented from expanding into protected areas, such as the 521 Tasmanian Wilderness World Heritage Area. Habitats and pathways of likely spread could be 522 identified and prioritised for monitoring and management. Deer sub-populations most likely to 523 contribute to further spread could be identified and prioritised for management. We suggest that 524 deer be contained to the 'traditional range' at a population size that allows landholders to achieve 525 the outcomes desired from their land. Demographic models are needed to develop strategies to 526 achieve pre-determined outcomes. 527 Once management strategies are implemented, the population should be monitored to determine 528 the effectiveness of these strategies. Our habitat suitability maps could inform the establishment of 529 a systematic surveillance network of camera traps in areas with and without deer. Coupled with the 530 long-term spotlighting dataset analysed in this paper, a systematic camera network would be useful 531 for monitoring the deer population, as well as detecting incursions into the TWWHA, escapes from 532 farms, and unlawful releases. 533

Conclusion 534
After more than 100 years of relatively low and stable abundance and restricted distribution, the 535 Tasmanian population of fallow deer is now increasing substantially in distribution and abundance. 536 Our case study highlights that ungulate species, and invasive species more generally, can have 537 'sleeper populations' with long lag-times before the population irrupts (Caughley 1970