**1. Introduction**

Habitat loss and fragmentation are among the most relevant threats to arthropod biodiversity [1]. Agricultural expansion, a fforestation with exotic tree species, and urbanization are the primary drivers of loss of natural or seminatural habitats and their insect communities [2], leading to small habitat fragments and decreased connectivity between them [3]. Classical island biogeography theory attempted to explain the e ffect of island size and distance from mainland sources on the diversity of species [4]. This concept was applied for terrestrial habitat fragments and the di fferences between oceanic islands, and isolated habitat

fragments are now well-recognized [5,6]. The predictive power of habitat area was also demonstrated for arthropods of terrestrial islands [7,8].

The effect of decreasing connectivity on arthropods is highly taxon-specific. Habitat generalists and highly mobile species may cover large distances in a strongly modified landscape matrix [3]. The spatial proximity of suitable habitat fragments is more important for arthropods that are habitat specialists and have low mobility; thus, they may form isolated populations [9]. Furthermore, the conversion of a continuous habitat into disjunct habitat fragments usually increases the length of the edges between fragments and the surrounding matrix, which may significantly change the characteristics of edges, and the plant and animal diversity of communities [10,11].

Spillover is the movement of organisms across habitat edges [12]. Its effect is more pronounced near edges than in the central part of the habitat [13]. Most of the studies focused on how the influx of predators from seminatural habitats relates to the pest control services in agricultural fields [14–16]. Only a few studies found spillover from natural habitats [17,18]. For example, Madeira et al. [19] argue that spillover from adjacent crop habitats shapes carabid, rove beetle, and spider assemblages in fragmented seminatural grasslands.

Small habitat fragments are important biodiversity refuges [20] and may harbour a large proportion of the regional species pool in arable landscapes [21]. Species richness and density of arthropods in small fragments can be as high as in large ones [22,23]. However, there are some species that are disadvantaged in small habitats [24]. Changes in species richness and community composition can lead to alterations of ecosystem functioning and stability [25,26]; consequently, habitat fragmentation may broadly affect species interactions [27–29]. Furthermore, the effect of fragmentation on different ecosystem functions depends on the specific function and species identity [30]. Species of certain functional groups, such as larger body size or higher trophic level, may be more vulnerable to habitat loss, and this may have an effect on ecosystem functioning, resulting in a weaker top down effect in food webs [31]. However, the net effect of fragmentation remains controversial [32]. Large variation exists in how plant and animal species and species interactions respond to fragmentation. For example, Tong et al. [33] found that seed predation of acorn weevils (*Curculio glandium* Marsham) was high in large, less isolated fragments. In contrast, Elzinga et al. [34] found higher rates of seed predation on white campion (*Silene latifolia* Poir.) by the specialist moth lychnis (*Hadena bicruris* Hufnagel) in small fragments.

Insect and seed predation are important ecological functions because of the associated community-structuring effects [35,36]. Measuring species interactions such as insect and seed predation is challenging. Instead of measuring the function itself, studies often use densities of predators as a proxy [37], which can be misleading [38,39]. Here, we aimed to study the effects of fragmentation (i.e., increasing isolation, decreasing fragment size, and edge effect) directly on predation in two grassland ecosystems.

We chose forest-steppes and kurgans due to their intense exposure to fragmentation and their special role in nature conservation in the steppe zone [40]. Both types of steppe fragments have the potential to preserve the natural flora, fauna, and act as local biodiversity hotspots [40,41]. Forest-steppes are mosaics of grassland and forest fragments at the contact zone between closed-canopy temperate forests and steppe grasslands. They are among the most complex ecosystems in Eurasia, and their elements play a key role in landscape dynamics [41]. Kurgans (burial mounds) are artificial formations and were developed for burial purposes by steppic people (mainly in the range of IV–I millennia BC) by piling soil on the grave of an important person. The height of the kurgans ranges between half and a few meters, with the diameter between a few meters and 100 m [42]. These relatively small landscape elements represent important refuges for Eurasian steppe wildlife [43]. Both ecosystems are of high natural conservation value, harbouring numerous rare and protected plant and animal species. The fragment size and landscape structure of the two ecosystems are in different scales: small-scale landscape structure and relatively large fragment size in the case of forest-steppes, and large-scale landscape and small fragment size for kurgans.

However, the landscape matrix between fragments was relatively homogeneous and highly modified for both ecosystems. Our aim was to compare the two systems, and we expected di fferent responses to the local and landscape factors.

We expected all studied fragmentation e ffects to be important determinants of insect and seed predation; however, the magnitude and relative importance of these e ffects, as well as their interaction, is not known. We tested the following hypotheses: (1) Predation rates are higher when connectivity decreases in the landscape, because isolation can enhance the spillover of generalist predators from the matrix. (2) Predation rates are higher in the edges than in the centres of a fragment, as a consequence of the edge e ffect. (3) Predation rates are lower in small than in large fragments, as functional groups of higher trophic levels are expected to be more sensitive to area loss. We aimed to reveal the similarities and di fferences of these questions in the two investigated fragmented grassland ecosystems of the same region using standardized methods.

#### **2. Materials and Methods**

#### *2.1. Study Region and Sampling Design*

We conducted our study on 60 natural grassland fragments in two di fferent regions of the Hungarian Great Plain. We sampled 30 forest-steppe fragments in the central part of the Kiskunság region and 30 kurgans in southern Hungary. The investigated fragments were scattered around four settlements (Dévaványa, Kunágota, Makó, and Szentes) in the case of kurgans, and around three villages (Pirtó, Bócsa, and Kunfehértó) in the case of forest-steppes (Supplementary Material Figure S1). We established two transects of sentinel prey, and two trays of seeds in each centre and edge of every fragment (Figure 1B). Both areas are characterized by a continental climate with 500 to 550 mm mean annual precipitation, and 9.5 and 10 ◦C mean temperature, respectively [40,44]. Forest-steppes comprise extensive dry grasslands dominated by *Festuca vaginata* Waldst. and Kit ex Willd., *Stipa borysthenica* Klokov ex Prokudin, and relatively small forest fragments of poplar (*Populus alba* L.) and hawthorn (*Crataegus monogyna* Jacq.) [41]. Our study focused on dry steppic grasslands. The potential vegetation of kurgans consists of pannonic loess steppic grasslands [40] dominated by crested wheatgrass (*Agropyron cristatum* (L.) Gaertn.) and forage kochia *(Kochia prostrata* (L.) Schrad.) [45].

**Figure 1.** (**A**) Location of study regions in Hungary, Europe. (**B**) Sampling design. Light green represents the area of grassland fragment. Transects of sentinel preys and seed predation trays were minimum of 10 m away from each other, even in same transect position. (**C**) Sentinel prey. (**D**) Seed predation tray.

We selected the study sites on the basis of the size of the fragments and along a landscape configuration gradient by performing preliminary field visits and GIS calculations. We calculated Hanski's connectivity index [46] and hostile matrix percentage to quantify landscape configuration and composition using Google aerial photographs (captured in 2019), the basic ecosystem map of Hungary, and Quantum GIS 3.6.1 software [47]. Since kurgans and forest-steppes had two different spatial resolutions (i.e., kurgans were situated in large-scale agricultural landscapes and forest-steppe fragments were in a matrix of relatively small-scale forest plantations), we performed GIS calculations within a 1000 m radius buffer around the kurgans, and within a 500 m radius buffer around the forest-steppes. For connectivity calculations, we considered all habitat fragments (other forest-steppe fragments and open-sand grasslands for forest-steppes, closed and alkali grasslands for kurgans) that were located around the focal fragment. As we applied the connectivity index to entire predator communities containing many taxa, scaling parameters α and β were set to the value of 0.5 [48]. For hostile matrix calculations, we considered all nonhabitat fragments (coniferous and deciduous plantations, clear-cut areas, young afforestation for forest-steppes, and arable lands for kurgans) and calculated their pooled percentage cover in a buffer around each site. As we found significant correlations between hostile matrix percentage and connectivity in both habitat regions (forest-steppes: Pearson *r* = −0.64, *p* < 0.001; kurgans: Pearson *r* = −0.95, *p* < 0.001; i.e., proportion of hostile matrix significantly decreased with increasing connectivity), we used only connectivity as landscape-level variable in further analyses. Lastly, we selected 15 small (0.16–0.48 ha for forest-steppe; 0.01–0.10 ha for kurgan) and 15 large (0.93–6.88 ha for forest-steppe; 0.20–0.44 ha for kurgan) grassland fragments. Connectivity values of the selected fragments ranged from 0 (isolated) to 2637 (connected) for kurgans (mean = 689) and 24 to 811 for forest-steppes (mean = 394).

#### *2.2. Sentinel Prey*

We assessed the predatory activity of carnivorous insects with dummy green caterpillars of moths made of plasticine, exposed for seven days. This method of sentinel prey is easy to use and appropriate to assess in situ predation pressure [39,49]. Dummy caterpillars were 25 mm long and 5 mm in diameter, and made from light green nontoxic modelling plasticine (Fimo Soft ®, Staedtler Mars GmbH & Co. KG, Nuremberg, Germany). All caterpillars were covered by PlastiDip ® (PlastiDip International, Blaine, MN, USA) silicon spray to avoid drying and eliminate the smell of plasticine [50]. We fixed all caterpillars to 5 cm long wooden sticks with superglue for easier handling.

We attached them to the ground by pushing the end of the stick into the soil. We placed dummy caterpillars in transects, 1 m distance from each other. We used 2400 sentinel preys altogether (2 regions × 30 study sites × 2 transect positions × 2 transects × 10 caterpillars; Figure 1). The transects of sentinel preys were at a minimum of 10 m away from each other even in the same transect position. We installed dummy caterpillars on 21–27 June and collected them from 28 June to 4 July 2019. Potential predators were identified by the attack marks that they left on dummy caterpillars. We inspected the marks by using magnifying glasses and microscopes in the laboratory, following the methods described by Low et al. [51]. Multiple attack marks by the same predator group were assumed to originate from the same predator. Signs by di fferent predator types were considered independent attacks.

#### *2.3. Seed Predation*

We exposed seeds in transparent, plastic trays to assess seed predation. Placing the seeds in shallow containers in the ground is a simple and established way for assessing seed predation [52,53]. We placed 10 seeds of *Triticum spelta* L. as large, and 10 seeds of *Festuca rubra* L. as small seeds in each tray. We used the di fferent sizes to increase attractiveness for a wider range of seed predator arthropods. The trays were round plastic containers, 10 cm in diameter (Figure 1D). We fixed the container to the ground by attaching a plastic stick to the container and dug it into the soil. We excluded birds and rodents by closing the containers with transparent lids and creating 1 × 1 cm openings on their sides (only for arthropods). Altogether, we had 2 regions × 30 study sites × 2 transect positions × 2 trays, resulting in a total of 240 seed predation trays (Figure 1). The containers were a minimum of 10 m away from each other. We installed trays from 31 May to 6 June and collected them from 7 to 13 June 2019. Thus, all trays were exposed for 7 days. Seed predators were assumed to be responsible for missing seeds. We counted the remaining seeds in each tray and inspected them for further predation marks in the laboratory. We considered multiple attack marks on the same seed as one predation event. Several oligo- and monophagous specialist seed predator insects were present on our study sites, but their seed-predation e ffect was not included in our data.

#### *2.4. Statistical Analysis*

Insect predation rates were calculated as the number of sentinel prey items showing signs of predation per total number exposed per transect. Seed predation rates were calculated as the number of missing seeds and remaining seeds with predation marks per total seed number exposed per transect. To test whether connectivity, fragment size, transect position, and their second-order interactions (fixed factors) had a significant e ffect on insect and seed predation rates, we used generalized linear mixed-e ffects models with the model averaging method. Models were fitted with binomial distribution. Connectivity ranged between 0 and 1. We used lmer (lme4) [54] models with fragment ID within village as a nested random-e ffect term. We used seed size as an o ffset variable in models of seed predation rates. We calculated Akaike's information criteria corrected for small sample sizes (AICc) to rank candidate models. The models with <6

ΔAICc of the best model (i.e., the model with the lowest AICc) were used for model averaging [55,56] with the R package MuMIn [57].
