Resource selection of a nomadic ungulate in a dynamic landscape

Nomadic movements are often a consequence of unpredictable resource dynamics. However, how nomadic ungulates select dynamic resources is still understudied. Here we examined resource selection of nomadic Mongolian gazelles (Procapra gutturosa) in the Eastern Steppe of Mongolia. We used daily GPS locations of 33 gazelles tracked up to 3.5 years. We examined selection for forage during the growing season using the Normalized Difference Vegetation Index (NDVI). In winter we examined selection for snow cover which mediates access to forage and drinking water. We studied selection at the population level using resource selection functions (RSFs) as well as on the individual level using step-selection functions (SSFs) at varying spatio-temporal scales from 1 to 10 days. Results from the population and the individual level analyses differed. At the population level we found selection for higher than average NDVI during the growing season. This may indicate selection for areas with more forage cover within the arid steppe landscape. In winter, gazelles selected for intermediate snow cover, which may indicate preference for areas which offer some snow for hydration but not so much as to hinder movement. At the individual level, in both seasons and across scales, we were not able to detect selection in the majority of individuals, but selection was similar to that seen in the RSFs for those individuals showing selection. Difficulty in finding selection with SSFs may indicate that Mongolian gazelles are using a random search strategy to find forage in a landscape with large, homogeneous areas of vegetation. The combination of random searches and landscape characteristics could therefore obscure results at the fine scale of SSFs. The significant results on the broader scale used for the population level RSF highlight that, although individuals show uncoordinated movement trajectories, they ultimately select for similar vegetation and snow cover.

Yes -all data are fully available without restriction maximum, or minimum). In addition to testing two models, we ran all models with an increasing 268 number of pseudo-absence points. Coefficients did not change much after using 20 times the 269 number of pseudo-absence points as presence points so we present results from these models (S1 in the data. We therefore re-ran the models without year as a random intercept. Models were still 272 singular, because individual also explained little variance, and so we also ran the models without 273 random effects using the glm function in the stats package [62]. In winter, year was also singular 274 in the linear models, so we re-ran the models without a random intercept for year.

275
Step selection functions 276 To examine selection of individual Mongolian gazelles for NDVI we used step-selection of 2015) for all three spatial scales (1, 5, and 10day step-lengths) for the winter models because 305 mixed-effects models such as those proposed by [73] usually did not converge for our data set.

306
Yet for the growing season models the 5 and 10 day step scales had too little data to split it by 307 year, we therefor ran these models by individual and ran the 1 day models by both individual-

308
year and individual as comparison. year. For the models by individual, we chose mean NDVI during the 2015-2017 growing seasons.

331
For the winter models we used mean snow cover for a given year because we also hypothesized

347
On the population level, in both seasons, quadratic models offered a better fit than linear 348 models (   Few individuals showed significant selection at any scale (Fig 2). If sample size were a problem, 383 we would expect to find many more individuals selecting at the 1 day scale then at the 5 and 10 384 day scale, but this was not the case (see S4-S6 Tables for sample sizes). For the growing season, 385 we did not have sufficient data to run individual-year models for the 5 and 10 day scales, so we 386 report on results from models run by individual at all scales (but see S2 and S3 Figs and S7 Table   387 for the 1 day SSFs run by individual-year). For those individuals who did show selection during 388 the growing season the results were similar to the RSFs, with individuals selecting for above 389 average NDVI, often at the upper end of available NDVI, at all scales (Fig 3; S4 and S5 Figs). In 390 the SSFs, though, this was the result of selection taking a quadratic or linear shape (Fig 3; S7 391 Table). In winter, we had enough data to look at selection by individuals within a year. Gazelles  Table). This meant that individuals selected snow cover 394 both above and below the mean snow cover during a given year. This mix of selection types and 395 a lack of consistency were seen at all scales (S7 and S8 Figs). We also examined if gazelles that 396 had more than one year of data showed the same selection across years. While most gazelles 397 showed consistent selection, up to a quarter did switch selection strategies (1 day scale = 6 398 individuals out of 19 with more than one year of data, 5 day scale = 5 out of 17, 10 day scale = 4 399 out of 7, Fig 5). Switching selection could also include switching between selecting and not 400 selecting.      might be choosing areas with higher NDVI because these have denser vegetation cover and thus 451 more forage. Therefore with NDVI we might not be able to detect for selection for forage quality 452 according to the FMH, but the observed selection suggests that gazelles find areas with higher 453 forage cover. Yet at some point preference for higher NDVI begins to decrease again, likely 454 because these areas reflect tall, dense vegetation that is difficult to digest and serves as a hiding 455 place for predators such as wolves.

456
In winter the population selected for intermediate snow cover. This supports our 457 hypothesis that in winter gazelles avoid areas of high snowfall, but seek out areas that provide 458 enough snow for hydration without restricting movement or access to forage.  Table) were a problem we would have expected to see many of what search strategies can be used when a landscape has large, homogeneous patches. If 520 gazelles were to use perception, this would require perception of resources beyond their sight.

521
The Eastern Steppe is a flat, treeless landscape. In similar landscapes in the Serengeti wildebeest 522 (Connochaetes taurinus) were observed moving toward a thunderstorm they heard 24 kilometers 523 away or toward rainstorms that darkened the sky as far as 80 kilometers away [82]. In addition,

524
Mongolian gazelles which have found good resources could potentially call to other gazelles 525 several kilometers away to share that information [83]. Therefore gazelles might be able to find 526 good resource patches, even if they are far away.

527
Alternatively, given the large patch sizes of vegetation and their unpredictable dynamics 528 using perception might not, or only partially, be feasible, and gazelles may in many cases use 529 random search strategies to find new resources. This is because in unpredictable resource 530 landscapes, memory and gradient following are not expected to be useful in finding resources. Whether gazelles use random searches or use perception to choose between large, 535 homogeneous habitat patchesselection for foraging patches is likely to occur only rarely. If this 536 is the case then SSFs might be too fine-scaled to pick up on these rare events.
Step-lengths can 537 only be increased to a certain point before sample size becomes limiting or the area examined 538 becomes similar to an RSF approach. Using SSFs with nomadic animals may therefore require 539 using integrated step-selection functions to associate changes in movement (e.g., speed, turn   552 An additional factor that may complicate the SSF approach is the possibility that resource 553 configuration in the Mongolian steppe may change over time.   we would expect additional biotic and abiotic factors to weaken but not eliminate our ability to 595 detect selection for vegetation or snow cover. It might have also been easier to detect selection if 596 data on the digestibility or composition of the vegetation was available for the growing season 597 and data on snow depth was available for winter. NDVI and snow cover are the best available 598 metrics that are available across time and at broad scales, but are only rough proxies for habitat 599 quality. SSFs that yet need to be studied and we hope our study inspires future research into the how