Shrub cover homogenizes small mammals’ activity and perceived predation risk

Altered disturbance regimes, increasing atmospheric CO2, and other processes have increased woody cover and homogenized vegetation in savannas across the planet. African savannas with extensive versus minimal woody cover often have vastly different animal communities. However, we lack a clear mechanistic understanding of why animal communities are changing with vegetation structure. Our goal for this study was to understand how vegetation structure in an African savanna shaped the perceived predation risk of small mammals, hence affecting their activity. Using a reciprocal measure of standard giving-up-densities, amount of food eaten, we found sharp declines in rodents’ perceived predation risk and increased rodent activity underneath shrub cover. This response was consistent across species; however, species showed subtle differences in their responses to grassy vegetation. Our findings suggest that areas of minimal or extensive shrub cover (shrub encroachment) may be homogenizing rodents’ perceptions of predation risk and thus shaping their use of space.


Materials and Methods
When feeding, animals will remain in a resource patch until the nutritional benefits of feeding in that patch no longer outweigh the costs of feeding in the patch [46][47][48] . To tease apart the influence of the different costs associated with feeding in a patch (i.e. metabolic, predation, and missed opportunity costs), Brown 47 introduced the "Giving-up Density" (GUD) methodology. By placing identical artifical foraging patches in a landscape, the opportunity costs of foraging are equalized across the environment, and hence the relative predation costs will be proportional to the harvesting rate at each foraging patch. Less food remaining in a foraging patch suggests a "safer" location where the animal perceives lower costs to continuing foraging in order to maximize nutritional intake. By comparing the GUDs obtained in artificial food patches in different habitats or microhabitats, we can better understand why animals use the landscape in the way that they do 27 .
We used GUDs 47 to assess small mammals' perceived predation risk across a fine-scale gradient of shrub and grass cover in the savannas of Eswatini. Additionally, by using camera traps to monitor foraging patches, we quantified the activity levels of individual species and the small mammal community as a whole.
Study sites. We conducted GUD experiments in Mbuluzi Game Reserve (30 km²) in Eswatini (26.1564°S, 31.9824°E, Fig. 1). Located in the lowveld region adjacent to the Lubombo Mountains on basaltic soils, this protected area is currently managed for wildlife conservation and ecotourism with minimal resource extraction (e.g. wood and grass harvest). This area has a subtropical climate and a unimodal rainfall pattern with the wet season during October-March (when 75% of the annual rainfall is received) and the dry season during April-September (25% of the annual rainfall) 49 . Yearly precipitation in the region typically ranges from 500-750 mm 50 . Senegalia nigrescens and Sclerocarya birrea were the dominant tree species in the savannas. The dominant grass species included Themeda triandra and Panicum maximum, while Dichrostachys cinerea was the dominant shrub. This region, and our site, has seen a drastic increase in shrub cover over the last the 70 years, from 2% to >40% at the time of the study 51,52 . The most common large mammals on the study sites included impala (Aepyceros melampus), warthog (Phacochoerus africanus), giraffe (Giraffa camelopardalis), zebra (Equus quagga), blue wildebeest (Connochaetes taurinus), kudu (Tragelaphus strepsiceros), and nyala (Tragelaphus angasii) 13 . Elephants (Loxodonta africana), white rhinoceros (Ceratotherium simum), and buffalo (Syncerus caffer) were extirpated from the region around 1920 53,54 . Small mammal predators on the site included jackals (Canis mesomelas, C. adustus), serval (Leptailurus serval), and large-spotted genet (Genetta maculata), as well as snakes (e.g. black mamba [Dendroaspis polylepis]) and predatory birds such as the spotted eagle owl (Bubo africanus), and the lizard buzzard (Kaupifalco monogrammicus).
Data collection: estimating perceived predation risk. We conducted GUD experiments in the austral winter from May-August 2016 when the rodents are most abundant in this region 61 . We placed five artificial foraging patches comprising a circular plastic tray (30 cm diameter, 2 cm high) in a transect line as follows: 1 m within a shrub (Shrub); at the shrub/grass interface (Edge; approximately 50% shrub coverage); and 50 cm, 1 m, and 3 m into an open grassy area (Grass: 50 cm, 1 m, 3 m; Online Resource 3). We defined a "shrub" to be a connected cluster of woody plants with a volume of at least 2 m³, and 1-3 m in height. Several shrub species were present in our study area, but where possible we selected Dichrostachys cinerea bushes for our experiments (≈70% of transects) as this was the most common encroaching woody species at the site. The last patch in the grassy area, which marked the end of the transect, was placed 3 m from the focal shrub and a minimum of 3 m from other shrub clusters in all directions. We placed transects >50 m apart and sampled a total of 15 transects. We did not place the foraging patches near large (>5 m tall) trees to avoid the confounding effects of canopy cover and reduce the influence of avian seed consumers.
Before collecting GUD data, we habituated the rodents to the artificial patches by putting them out with millet seeds for at least 3 days. Once we saw evidence that rodents were using the patches (removed food, droppings, seed fragments, or shells), we mixed 600 g of sand with 40 ml (33 g) of hulled millet seeds per foraging patch. We determined the appropriate amount of seed and sand to use during a trial period to ensure that the animals were not consuming all the seeds in a tray 27,28 . Mixing the food with sand mimics a natural foraging scenario where there are associated costs with time spent foraging 47 . To ensure that the rodents were familiar with the sand and seed mixture, we put these artificial foraging patches out for a minimum of 2 days before data collection began. We captured the foraging of nocturnal and diurnal small mammals by sifting out the remaining seeds once every 24 hours at sundown. To determine the GUD, we measured the volume of seeds remaining, to the nearest ml, in each foraging patch each day. We then refilled the patches with 600 g of sand and 40 ml (33 g) of hulled millet seeds. We monitored the foraging patches over the course of 5-7 days (24-hour periods) and generated count data by recording the number of mls removed to the nearest whole ml (0-40 ml, Food Eaten).
To remove the confounding influences of non-target species 21,27,62 and estimate the time of visits by the entire community and individual species, we attached a Reconyx Hyperfire Infrared Camera (programmed to 40 cm short-range focus and set to high sensitivity; Reconyx, Holmen, WI, USA) to a stake with a crossbar 50 cm above each tray. While the placement of these 10 × 15 cm cameras above the artificial patches might have reduced the perceived predation risk of small mammals, it was unlikely to be significant as the total cover provided by these cameras 50 cm above the patches was minimal. Moreover, by focusing a camera above the patch and placing a ruler on each tray, we ensured species identification 63 that would not have been possible with a side mounted camera.
We set cameras to take a series of three photos in a row (1 second apart) and then pause for 1-minute before the camera was triggered again. At each foraging patch we recorded the number of three-picture series taken every 24 hours and the maximum number of each species within each series. To generate a community level metric of activity (Minute-Visit) we summed the maximum number of rodents detected in each series. To generate species-specific indices (Minute-Visit) we summed the maximum number of individuals of each species in the series. Some foraging patches were visited by avian species. We used the rodents' percentage of total Minute-Visits to estimate the amount of food that rodents consumed, eliminating the impacts of avian consumers. We assumed that the correlation between Minute-Visits and Food Eaten was the same for both rodents and birds, because it was not possible to approximate the correlation for each group separately due to the unknown rate of food depletion at the patches over each 24-hour period. All applicable institutional and/or national guidelines for the care and use of animals were followed and the project was approved by the University of Florida IAUCC (protocol 201609284).
Vegetation assessment. At each foraging patch along the gradient, we measured the height of the shrubs and the grass (Max Ht). We also included a binary categorical variable (Binary Ht) to indicate long grass (>40 cm) or short grass (<40 cm), because savanna rodents appear to respond to grass height at this specific height 26 . In an effort to avoid collinearity we did not use (Max Ht) and (Binary Ht) in the same models. We estimated the horizontal visual obstruction (HVO) of the vegetation surrounding each tray using a Robel pole, averaging the horizontal coverage from measurements taken from 4 m away in each cardinal direction 64 . To quantify fine-scale vegetative cover we estimated the percentage of ground covered (Ground Cov) as the combined coverage of shrubs, grass, and forbs looking down onto a 1 m² circular plot from 1.5 m. We estimated the percentage of cover in each quadrant of the circle and then averaged these estimates into a total percentage. We used HVO to evaluate www.nature.com/scientificreports www.nature.com/scientificreports/ cover from visual-based terrestrial predators and Ground Cov to quantify cover from avian predators 42 . Finally, to control for the effect of using shrub clusters, we included a binary categorical variable representing Shrub Size (Small = 2-4 m³; Large >4 m³).

Data analysis.
To aid in interpretation of our analysis we standardized the directionality of our response variables. To do this, we used the inverse of the GUD, the amount of food eaten (Food Eaten) and Minute-Visit, both of which decrease with increased perception of risk. To understand rodent communities' perceptions of fear within and beyond shrubs, we first combined all species' responses together and evaluated their models to explain variation in the foraging patches. We compared sets of six shrub proximity models (see below) for the response variables Food Eaten and Minute-Visit. This allowed us to understand if rodents had punctuated responses to shrub cover that fell into broad categories (i.e. shrub vs. grassy areas), or if they were continually increasing, decreasing, or varied at each location. We compared models with distance from the shrub as a continuous variable, a categorical variable for each foraging patch (5 levels), two binary categorizations (Shrub/Edge vs. Grass and Shrub vs. Edge/Grass), and three vegetation categories (Shrub vs. Edge vs. Grass). We evaluated each model with generalized linear mixed models, with transect as a random effect, and our count data fitted to negative binomial distributions with a log link function in the package glmmTMB 65 in Program R (R Version 3.3.3, www.r-project. org, accessed 19 Sep 2017).
Further, we investigated the response (Food Eaten and Minute-Visit) of all rodents combined to different measures of vegetation structure in grassy areas (i.e Shrub and Edge excluded). We developed a suite of 15 a priori grassy areas models with single variables and additive models using the variables Max Ht, Binary Ht, Shrub Size, HVO, and Ground Cov. We also modeled a squared term for the HVO measurement (HVO²) as horizontal visual obstruction may cease benefiting rodents at elevated vegetation densities 66 . We created a model with both Ground Cov and HVO, as both of these metrics may influence prey perceptions of risk 42 . We also built models with Max Ht added to Ground Cov and HVO, as grass height may also influence rodent foraging and risk perception 28,40 . We standardized (z-score) the vegetation covariates so that their means fell at zero.
We also evaluated the suites of six shrub proximity models and 15 grassy areas models for the activity levels (Minute-Visit) of each species individually. For each species we only included activity data from transects where it was detected at least once. We included data from all foraging patches at each of these transects, even if species were not detected at all foraging trays, because we assumed that all foraging patches were available to all foragers. We analyzed these data using negative binomial distributions in glmmTMB 65 in Program R.
To evaluate all of our suites of models, we ranked them based on their Akaike Information Criterion adjusted for small sample size (AICc 67 ), prioritizing the most parsimonious models. We considered models that were <2 AICc units of the best model to be competing models 68 . We evaluated the parameters in these competing models, considering model parameters with β estimates and 95% Confidence Intervals (CI) that did not include 0 to be relevant predictors. For categorical data, we examined the 95% CIs of each category for overlap between them and considered overlapping categories to be redundant. Additionally, we only considered quadratic terms to be important predictors if the 95% CIs of both parameters did not include 0.

Results
From the small mammal community we detected 7 species of rodent foraging at the foraging trays, of which we had sufficient visits to model 5 of them (Online Resource 4). Omnivorous species were detected at a majority of the 15 transects (Lemniscomys rosalia: detected on 14 transects, Mus minutoides: 11, Mastomys natalensis: 8) while other species were less common on the transects (Dendromus mystacalis: detected on 5 transects, Steatomys pratensis: 3, Saccostomus campestris: 2, Aethomys ineptus: 1). Almost all transects (14 out of 15) had at least two species of rodents visiting the foraging patches over the course of 5-7 nights. Shrub proximity. The best model, and only competing model, to explain differences in both community Food Eaten and community Minute-Visits was the model that separated the foraging patches into three categories (Shrub, Edge, Grass) where the three grassy area foraging patches were statistically similar ( Table 1). The 95% CI of the categories did not overlap zero and there were no other competing models. The model showed that the predicted amount of food that rodents removed from foraging patches was higher under shrub cover than on the edge or in the grassy foraging patches ( Fig. 2 . We found no indication that the foraging patches away from the shrubs were different from each other or that there was a gradient of response with distance away from the shrub (Table 1).

Grassy areas.
Examining grassy area foraging patches only, Food Eaten and Minute-Visits were both best explained by a model with Ground Cov (Tables 2, 3). Rodents showed the strongest response to ground cover, with the predicted amount of food eaten and activity increasing from 1 to 5 ml and 3 to 24 minutes respectively (Fig. 3). A curvilinear response of rodents to HVO was also a relevant predictor of Minute-Visits (Table 3) with activity peaking around 3.5 decimeters before declining. There were no other relevant variables in competing models (Tables 2, 3). individual species. Species responses to shrub proximity were similar to the community response, with individuals spending more time at Shrub foraging patches than Edge foraging patches and rarely visiting Grass foraging patches ( Fig. 4; Online Resource 5). The visiting patterns of L. rosalia and M. minutoides were best described by a model with three categories (Shrub, Edge, Grass) and no competing models. The best models explaining the response of M. natalensis included a model with each individual patch, a model with distance between patches, and a model with three categories (Shrub, Edge, Grass)    www.nature.com/scientificreports www.nature.com/scientificreports/ categories estimates in the individual patch model, we considered the three categories and distance models to be better representations of the data. The best models for explaining the visits of D. mystacalis and S. pratensis had two categories. The two-category model best explaining D. mystacalis activity was Shrub vs. Edge/Grass, while the combined category for S. pratensis was Shrub/Edge vs. Grass. Adding more categories to these models did not improve parsimony.
Each species responded differently to variation in vegetation structure in the grassy areas. Examining the relevant parameters in competing models (Online Resource 6), one species, L. rosalia, consistently increased Minute-Visits with increasing ground cover (β 1.86 [1.11-2.60]). One species, M. minutoides, showed a curvilinear

Discussion
The rodent species in this study showed sharp declines in their perceived predation risks underneath and on the edge of shrub cover. This perception did not slowly increase with distance from the shrub, rather it was punctuated. The perception of risk increased at the edge of the shrub and again in open areas only 50 cm away from the edge. Despite considerable variation in the body mass, ecology, activity periods (i.e. nocturnal and crepuscular), and habitat use of the rodents in this study 57 , their perception of risk with regards to shrubs was consistent. Variation in species responses, although subtle, was more pronounced in grassy areas. Our findings suggest that the shrub encroachment of >40% coverage (doubling after 10 years 52 ) found on our site 12 and throughout many African savannas 3 may be altering and homogenizing rodents' perceptions of predation risk and shaping their use of space 27,33 .
The rodents' consistent response to shrub cover allows us to consider the community-wide responses to anthropogenically altered savannas. When woody cover has been drastically reduced through firewood harvesting or extensive browsing, we would expect sharp increases in rodents' perceived risks of predation 69 . This enhanced fear may help explain the pattern of depauperate rodent communities in savannas with reduced cover, but where ample food resources persist 12,22,44 . These communities are often dominated by one generalist species such as Mastomys natalensis 36 . In our study, Mastomys natalensis increased activity with decreasing grass height and has been shown to increase activity in open and disturbed environments where food is plentiful 70 .
We would expect most rodent species to thrive in the relatively safe environment of thick shrub cover 22,36,44 presumably created by altered disturbance regimes and increasing atmospheric CO 2 . However, there is clearly a threshold where the benefits of safety from shrub cover are outweighed by the loss of foraging opportunities. Most savanna rodents rely heavily on grass or grass seeds for food 61 and increasing shrubs can reduce the amount of grass and forage for herbivores 13,71 . This might explain the pattern common in other studies where rodent diversity increased with increasing shrub cover until a threshold was reached where diversity declined 12,38 .
The consistent and predictable response of rodents also suggests that woody vegetation, by altering animals' perception of fear, may act as an ecological filter, selecting against certain species or traits such as body size 72 . The reduced activity and diversity of large mammals in savannas with extensive shrub cover cannot be explained by reductions in food alone 16 and increasingly appears to be a function of perceived predation risk [18][19][20]34 from reduced sight lines and escape routes. Alternatively, small-bodied rodents appear to increase activity and diversity around shrubs 12 , potentially due to reductions in their perception of risk 36 , as demonstrated in our study. Additionally, the known reduction of large mammals from areas with extensive shrub cover might enhance rodent populations because they are released from competition for resources 73 . Overall, there is a clear pattern where the perception of risk from woody vegetation favors smaller mammalian herbivores over larger ones. The potential consequence of this is a trophic shift, or a replacement of large consumers with smaller ones, which is likely to have a marked influence on plant community dynamics and composition [73][74][75] . The loss and replacement of large animals is a global conservation pattern, but rarely has it been associated with a specific mechanism [73][74][75] .
Away from shrubs in grassy areas, the rodents' perceived risk of predation changed with ground cover, although metrics of safety (Minute-Visits, Food Eaten) were <20% of those under shrubs. We expected that reductions in ground cover, grass height and horizontal structure would increase rodents' perception of risks 28,39,40 . While perceived predation risk for Steatomys pratensis increased with a reduction in grass height, Mastomys natalensis showed the reverse relationship. Mus minutoides showed a heightened perception of risk with reduced horizontal cover, though its response was curvilinear, with increased perceptions of risk with extensive horizontal cover as well. In general, species' responses to vegetation varied with habitat structure once they were away from shrubs. Accordingly, the heterogeneity of structure in areas away from shrubs will likely lead to variation in different species' perceptions of fear, potentially fostering species coexistence and diversity 25 . However, the influence of structure in grassy areas on rodents' perception of fear was minimal when compared to the influence of shrub cover.
In our study, the patterns of rodent behavior appear consistent with broader patterns of rodent communities in African savannas. However, there are likely a host of other processes shaping rodents' behaviors and community dynamics. For example, we were not able to separate out the influence of rodent species interactions from predation risk, with aggression and attraction potentially influencing variation in species foraging and activity 76 . Nevertheless, because we saw consistency between community and species-level responses, it appears that the influence of species interactions on our response metrics was minimal. Additionally, there is considerable evidence that rodents can be influenced by the broader patch and landscape level variation in woody cover 77 . At these scales, woody cover may alter rodent movement patterns 78 , food resources 13,71 , predator communities 79 , and overall perception of risk 77 . Replicating this work on sites with varying levels of broad-scale encroachment might allow us to parse the relative influence of the broader vs. fine scale vegetation structure on rodents' behaviors and community dynamics.
Due to its consistent and powerful influence on rodent fine-scale perceptions of fear, shrub cover is likely to have a strong influence on the species and communities that rely on it for safety. In turn, the greatly reduced perceived risk of predation in and around shrubs might provide a mechanistic explanation for the patterns of reduced rodent diversity seen in open savannas 36,80,81 . Species' perceptions of fear can have a strong influence on their reproductive fitness, foraging, movements, and physiology 69,[82][83][84] . With shrubs clearly shaping rodents' perception of fear, there can be little doubt that the anthropogenic forces shaping woody vegetation will have a powerful influence on the rodent species and communities in African savannas.