2.1 Study area

West Dongting Lake National Nature Reserve (WDLNNR: 28°48′–29°07′N, 111°57′–112°19′E) is a Ramsar site and important wintering ground for many waterbirds species in the East Asian-Australasian Flyway (EAAF). The lake has subtropical monsoon climate with a distinct wet/dry season where mean annual precipitation ranges from 1,200 to 1,415 mm. The average annual temperature ranges from 16.2 to 17.8°C with 259–277 frost-free days. The elevation of the lake ranges from 27 to 30 m. During wet season (July–September), the majority of the lake is inundated with flood when water level rises from 20.13 m to 30.32 m. During dry season (November–February) as the water level drops, an array of sub-lakes naturally converts to five main types of habitats, namely grassland, wet meadows, mudflats, shallow water and deep water (Lai et al., 2013; Li et al., 2014; Lu, Jia, et al., 2018) (Figure 1).

Since 1970, the Lake has undergone dramatic anthropogenic changes such as alteration of hydrological regimes and reclamation (Lu, Jia, et al., 2018; Zhang et al., 2016), and large area was converted to Poplar (Populus spp.) and common reed (Phragmites australis) plantation. These changes have led to sharp reduction in migratory waterbirds species richness and abundance (Li et al., 2014; Lu, Shi, et al., 2018; Zhang et al., 2016). Therefore in 2015, the Chinese government started a restoration program, which aims to sustain suitable habitats in some sub-lakes through managing water level within the premises of WDLNNR. In the restored lakes (referred as R1 thereafter), controlled water level is managed to allow the development of permanent mudflats, grasslands, shallow water and open deep-water habitats during the wintering period. In the unmanaged sub-lakes (referred as R2 thereafter), the inundation regimes are determined by the hydrology of collected streams and local rainfall and are highly unpredictable and the water depth fluctuate in matter of hours and days.

2.2 Data collection

2.3 Building social-behavioral association network (SBAN)

We constructed a social-behavioral association network for R1 and R2 using the abundance-based activity data. This network includes three parts describing individual species identity, distribution in shared space (i.e., microhabitat), and activity to be performed in a particular position in a shared time. Note that prior to the habitat selection event, there is a site selection event playing its part, which the data do not allow us to explore. Together with site selection, the social-behavioral association network model has four bipartite projections. The first bipartite is a (S_ij, S) projection which aggregates species-site selection in shared environment (Figure 2a). This projection connects sites (S) on the basis of shared species (S_ij). Note that first bipartite projection was not studied. The second bipartite projection is shared habitat use network within same site (S_ij, S H). It connects species (S_ij) that visit the same habitat (H) (Figure 2b) in shared space (i.e., S). The third bipartite projection is an intra-/inter-species interaction through shared activities (H, A S_ij; Figure 2c) where individuals of species (S_ij) interact with one another through presence in the same “aggregation” showing the robustness of habitat (H) for shared activities (A S_ij). The fourth bipartite projection is (S_ij, A_ij B) where species (S_ij) switch between activities (A_ij) in response to a short period behavior (B) (Figure 2d).

2.3.1 Defining nodes and edges in social-behavioral association network

The two primary aspects of networks are a multitude of separate entities (nodes) and the connections between them (edges). Consider the habitat association of two species, S₁ and S₂, denoted X₁ = {H, A|S = S₁} and X₂ = {H, A|S = S₂}, where X₁ and X₂ are the set of activities in space performed by species S₁ and species S₂, respectively. The intersection of these trajectories, X₁ ∩ X₂ contains all points in space where species S₁ and S₂ co-occur with their complete set of direct interactions. Because X₁ and X₂ are conditioned on specific individual species, the network produced by them retains species-specific information in shared space, forming its nodes. Restructuring the conditional term within the habitat association network changes the network projection. For example, if and , then the intersection Y₁ ∩ Y₂ is the set of species who visited both H₁ and H₂. A network based on this condition retains information about specific habitat patch use, which forms its nodes. Edges link the individuals of a species S who visited habitat patch H, but this intersection lacks the information on activity performed by the species. For activity, we reconfigure the network. Given that and then the intersection Z₁ ∩ Z₂ denotes that the species performed activity A₁ and A₂ within the same habitat, forming its nodes.

The network of these nodes connected through edges forms a “real set of interactions” offering wintering waterbirds community dynamics in a shared environment. We created five different SBAN networks (Figure 3a), one for each habitat; with five separate nodes (n = 5) for each species (N = 14: A–N; A_1, A₂, A₃, A₄, A₅ …. N₁, N₂, N₃, N₄, N₅), that represent two activities (foraging and roosting) and three behaviors (competition, aggression and courtship). Each node possesses attributes of the species identity, foraging guild and activity or behavior (Figure 3b). Pairwise observations were converted into edges to populate the adjacency matrix for interactions that represent how two nodes relate to each other and can be used to describe how they associate or interact for social relationships (Castles et al., 2014) in SBAN network.

2.4 Assigning node importance and edge weight

Assigning importance to any given species in mutualistic networks is a key task when evaluating species rank over habitat types. Using the corSparse function from package “qlcMatrix” (Cysouw) in R (R Core Team, 2016), we computed the Pearson correlation (−1 to 1) of activity-based abundance data and used its absolute values as importance between the columns of sparse matrix to make undirected networking plots utilizing PageRank™ (Csardi & Nepusz, 2006; Heath, 2018). Species network constructed using package “ggnetwork” (Briatte, 2016), where nodes were arranged by their relative order of importance, thereby resulted in a highly packed matrix network. We assumed negative Pearson correlation values as zero to simplify the network plots because negative correlation was just seen in one habitat type (BL-R1). For each habitat type, we made use of hub centrality to rank the species importance (keystones) in the network system.

For interactions (edges) we used PageRank™ algorithm (Bryan & Leise, 2006) to rank the important species (S) over habitat (H) for activities (A) to be performed. For details of the algorithm, please refer to (Bryan & Leise, 2006). In brief, the measure for node importance is the maximum number of edges it connects and range between 0 and 1, where 0 indicates no connection between nodes (less important), and 1 indicates maximum pairing in their respective network system (Croft et al., 2008, 2009; Farine & Whitehead, 2015). Species may move from one habitat type to another at a site to perform or switch between activities. This movement between habitats results in the mutual interactions between species. The problem is similar to that of ranking web pages. Thus, we define a species as important if it interacts (directly or indirectly) with other species that are in turn important. Since species can randomly switch between habitats in the given site, this habitat selection becomes complex and primitive.

2.5 SBAN networks comparison using two-way analysis of variance (ANOVA)

We used analysis of variance (ANOVA) to test the difference of the SBAN characteristics (i.e., number of nodes, number of edges, network density, and SIPS and BIPS) between R1 and R2. Following a significant ANOVA, we used LSD (least significant difference) for post hoc pairwise comparisons. LSD is found very effective and well suited for detecting true difference in the means after a significant ANOVA at 5% significance level (Brown & Behrmann, 2017; Fraiman & Fraiman, 2018; Mason et al., 2003).

3.1 Social-behavioral association networks (SBAN) in lakes with distinct hydrology

Figure 3a shows the social-behavioral association network built for the wintering migratory waterbirds community in the managed (R1) and unmanaged lakes (R2). The networks formed in sites with different hydrological regimes displayed great difference in all measured attributes including number of nodes and edges, edge density, and SIPS and BIPS (Figure 3a, Table 2). Surprisingly, compared with unmanaged sites, the sites with controlled hydrology had more species, more active nodes and edges, higher edge density, and BIPS and SIPS but differences were not statistically significantly different at treatment level (p = .296; Table 3).

For different habitats within a site, mudflats had far more complex network followed by shallow waters and grasslands while bare grounds and deep waters had the simplest networks, no matter what the hydrological regime was (Figure 3a, Table 2). For example, we recorded 14 species in mudflat at R1 site, which formed a network of 40 nodes connected through 780 edges, resulting in the highest density score of 32.3 (Table 2). On the other end, we only recorded 3 species in bare ground at R2 site without hydrological control, where the waterbirds formed the simplest social-behavioral network with six nodes linked by 15 edges, with lowest edge density of 0.6 (Table 2). Differences in these network attributes tested by two-way analysis of variance (ANOVA) at 0.05 significance level were highly significant (p = .000; Table 3).

3.2 Keystone species and their dominances over habitat types

Species changes their role at different habitats (Table 2). Moreover, besides deep waters and shallow waters in both regimes, other habitats were far dissimilar in keystone species composition and their dominances (Table 2; Figure 4a). In deep water and shallow water habitats, C. columbianus was the keystone species, however, its dominancy was challenged by A. alba in deep waters of R2. Ardea alba, Ardea cinerea and Ciconia nigra were keystone in mudflats of R1 while Platalea leucorodea and Anser fabalis were keystone species at R2 mudflat habitats. Anser fabalis and Ardea alba were keystones at grasslands of R2 while Anser albifrons was keystone at R1 grasslands. Anser cygnoides in R2 bare ground and Anser albifrons, Anser fabalis and Ciconia nigra at R1 bare grounds were keystones.

3.3 Species interaction preference scores and behavior interaction preference scores

Except bare grounds at R2, all other networks had inter-species interactions, indicating the prevalent competitions for available resources in the wintering grounds. In the managed sites, mudflats had the highest value of inter-SIPS (734), followed by shallow water (460), grassland (233), bare grounds (73), and deep waters (54). On the other hand, mudflats (46), followed by shallow water (36), grassland (20), deep water (12), and bare grounds (5) had the lowest value of intra-SIPS. Intra-SIPS values in the uncontrolled sites were also highest for mudflats (490) followed by shallow water (348), grassland (115), deep water (114) and bare ground (11), while intra-SIPS values for mudflats was 38 followed by shallow waters (30), deep water (22), grassland (21) and bare grounds (4) (Figure 4b, Table 2).

For BIPS, the indicator of activity synchrony, nearly all habitats have behavioral interaction (intra- & inter-) in both regimes (Table 2), suggesting there was activity synchrony between species. In managed habitats (R1), highest activity synchrony (inter-BIPS) was in mudflats, followed by shallow water, grassland, deep waters and bare grounds with 580, 369, 176, 52 and 40 respectively. The lowest value for this activity synchrony was seen in bare ground (12) of unmanaged sites (R2), followed by deep water, grassland, shallow water and mudflats having 108, 108, 283 and 293 scores respectively (Figure 4c, Table 2).

3.4 Species SBAN networks were far dissimilar in different habitat types

Some of the habitats (both in R1 and R2) were significantly different since the P-value was found to be <0.05. The mean comparison between the habitats revealed that BG and DW were not different from each other but significantly different from SW and MF, GL was significantly different from MF at 5% level of significance while MF and SW were not different from each other at 5% level of significance, supported our results (Appendix S2). These assertions were further supported by dramatic differences in keystone species composition at different habitat types, as previously described.

4.1 Species' dominance in different habitat types

Stability and robustness assessment in complex ecosystems is a basic problem in conservation ecology (Dunne et al., 2002, 2004). The loss of an individual “keystone or hub” species may induce effects on other important species, resulting in re-configuration of the entire network structure. Thus, the relative “importance” of a given species within a network can be viewed as a function of the integrity and robustness within the network.

In shallow and deep waters, C. columbianus was hub of the network in both regimes and dominated over other species such as A. cygnoides, A. alba, G. leucogeranus, A. cinerea, A. albifron, and Larus spp, among which C. columbianus is the heaviest (averaged at 6.4 kg, (Wilman et al., 2014). Our findings provide support for a key prediction that large body size helps species acquire high dominance ranks over occupied habitats. This dominance pattern was found across a range of animals ((Haley et al., 1994; Neumann et al., 2011; Pelletier & Festa-Bianchet, 2006; Pusey et al., 2005).

The single species dominance was not observed in other habitat types. In mudflats, species with the same foraging guild and similar body size, such as herons (A. alba, A. cinerea) and storks (C. nigra), were the keystone. Such distribution patterns in gregarious species, led to the exhibition of aggression or competition behaviors between dominant species, especially when body size allow some species a greater access to limited resources. At these habitats, the aggregation size predicts within and between species contact preferences (Fretwell, 1969; Manlove et al., 2018; Smith & Metcalfe, 1997). These contacts induce pressure on other species to synchronize different activity to avoid potential resource competition thus segregating the temporal niche and minimizing the risks of competition in shared space (Marra, 2000), supported by our findings.

Additional morphological characteristics and intransitivity for coexistence that explain shifts in dominance, beyond body mass toward the root of the phylogeny (e.g., families) deviate from the overall mass-to-dominance relationships (Allesina & Levine, 2011; Laird & Schamp, 2006, 2008; Miller et al., 2017). This has been previously demonstrated for plant competition in laboratory (Keddy & Shipley, 1989; Kerr et al., 2002; Soliveres et al., 2015). However, tests of this idea in animals are rare (Levine et al., 2017). For instance, A. albifrons and P. leucorodia, even though these two species have different feeding guild, both species were dominant in mudflats of R2, this apparently mutual dominance between these species may simply be a function of resource competition, same as in case with grassland at R2, where A. fabalis and A. alba were dominant.

4.2 Inter- and intra-species interaction and activity synchrony

Measuring important habitat features (food and cover) in meaningful ways for management is often difficult, labor intensive and costly (Henschel & Ray, 2003). Consequently, wildlife managers often rely on surrogate variables, such as frequency of habitat selection (as described in previous section) and interaction profiles, which are relatively easy to measure in the field or can be interpreted from databases (Kays et al., 2015; Stonehouse et al., 2015; Wilmers et al., 2015).

With the realized social-behavioral association network, species interaction preference scores (SIPS) for interaction (resource competition between species) and intra-species interaction (resource competition between individuals of a species) can be easily calculated. The SIPS can also be used to assess habitat quality. For example, deep waters at managed lakes were the most suitable habitat for C. columbianus while bare grounds at unmanaged lakes were best for A. cygnoides (Table 2; Figure 4a). Grasslands at unmanaged lakes were more suitable with lesser amount of intra-species interaction than at managed lakes while shallow water at managed lakes were better habitats with reduced intra-species interactions. High (absolute) interaction scores in mudflats and shallow waters at both R1 and R2 sites suggested that while many water bird species preferred these habitats, spatial competition (especially between hub species) was also strong in these habitat types.

Numerical measures of synchrony are complicated to define, identify and deal with. In general, it increases when many species in subnetworks exhibit broad dominance hierarchy (Shizuka & McDonald, 2012, 2015). Linking these dominance hierarchies to behavioral interactions between species gives insights for activity synchrony in subnetworks (habitat patches) among co-occurring species (Mukherjee et al., 2009; Rasool et al., 2015; Suselbeek et al., 2014), which in our case are defined as BIPS (behavioral interaction preference scores).

At both managed (R1) and unmanaged lakes (R2), BIPS were linked to resource competition and aggression behavior. Inter-species behavioral interactions of C. columbianus in deep waters and shallow waters resulted in avoidance of these habitats by Swan geese as foraging grounds (but they used the habitats for roosting). As described earlier, this might be due to the function of body size and aggression. Grassland habitats gave insight to the increased BIP score possibly because of competition between species, therefore driving other species to avoid or switch their activities to minimize the potential competition. Based on frequency of site selection events, it can be said that the species (especially G. Leucogeranus, G. monacha, C. boyciana, C. nigra and A. albifrons) somehow avoided unmanaged lakes (R2) while the species selecting these lakes were thereby friendly to share the habitats for given activities.

4.3 Influence of hydrological regimes on habitat selection, interactions and network structure

Reliable information about the quality of habitat features is critical to predict the performance of occupying animals for their conservation (Beck et al., 2014). Studies on animal behaviors and interactions among individual species can provide answers to many questions in wildlife ecology and conservation. Using social-behavioral association network approach, we found that there were more inter- and intra-species interactions in managed lakes than unmanaged lakes (Figure 4a). This finding was further supported by the total number of species that selected managed lakes over unmanaged lakes (Table 2). Generally, for each habitat type, more species were recorded in managed lakes than in unmanaged ones. Out of the total 14 species that we focused in this study, all of them were observed using the mudflats and 13 were recorded in shallow waters at the managed lakes. However, in unmanaged lakes, we found only eleven species using these habitats. Moreover, ten species selected grassland habitat patches at managed lakes while just six species selected grassland patches of unmanaged lakes. In bare grounds, eight species in managed and three species in unmanaged lakes were found. These results provided a clear indication of better habitat quality at managed lakes. As animals select habitat at a variety of spatial and temporal scales, this information becomes relevant to the scale of management and reflects the functional value of the habitat for population persistence (Godvik et al., 2009; Silveira et al., 2003; Wegge et al., 2004).