Patterns of species diversity in a network of artificial islands

Artificial island habitats such as human‐made wetlands are emerging novel ecosystems. Understanding the drivers of diversity in such artificial systems is essential for balancing the goals of biodiversity conservation and human socio‐economic needs.


| INTRODUC TI ON
Island systems have long fascinated scientists: Research on island diversity has advanced our understanding of key evolutionary and ecological processes (Darwin, 1859;Wallace, 1869;Warren et al., 2015). The seminal theory of island biogeography explained island species diversity as a dynamic balance between immigration and extinction processes (MacArthur & Wilson, 1967). The predictions of this theory have been broadly confirmed by empirical studies in the decades since its publication (Warren et al., 2015). Most of these empirical studies have focused on natural systems, or remnants of natural systems, but there is incipient interest in humancreated artificial islands as well (Rejmánek & Rejmánková, 2002;Sánchez-Zapata et al., 2005).
The original applications of island biogeography were to species diversity on 'true' islands such as oceanic islands (e.g. Hawaii) or continental-shelf islands (e.g. islands in the Sundaic shelf) where insularity and distance from the mainland strongly regulate the rates of immigration and extinction (Heaney, 2000;MacArthur & Wilson, 1963;Whittaker et al., 1989Whittaker et al., , 2008. Later on, the theory was interpreted to be relevant to natural habitats that resembled true islands. For example, the temperate mountaintops of the Great Basin, the caves of Greenbrier Valley, and the islands in human-made lakes such as Lago Guri, Venezuela and Thousand Island Lake, China, all exist in isolated biogeographical contexts analogous to those of natural islands, and thus, their diversity dynamics can be examined through the lens of classic island biogeography theory (Brown, 1971;Culver, 1970;Hu et al., 2011;Qie et al., 2011;Terborgh et al., 1997).
The theory has also been adapted to inform conservation and management decisions, especially in the context of fragmented terrestrial habitat patches, where the landscape has been modified by humans, but the patches themselves are remnants of the original natural landscape (Diamond, 1975;Harris, 2013;Lomolino & Perault, 2001;Shaffer, 1981;Thomas, 1990).
With the growing demands of human population, a new kind of island system is emerging globally. People are creating artificial island habitats to meet their socio-economic and cultural needs (Kueffer & Kaiser-Bunbury, 2014). Green roof-top gardens on buildings, valued for energy-efficiency, hydrological benefits and aesthetic appeal, are increasingly becoming common in cities (Oberndorfer et al., 2007), and are analogous to islands because of their semi-isolation from one another. Artificial aquaculture ponds provide aquatic habitat in regions where such habitat may be rare or absent (Dai et al., 2021).
A beneficial side effect of artificial islands is their supplemental role in biodiversity conservation. Roof-top gardens and aquaculture ponds provide crucial habitats for amphibian, bird, insect and butterfly species in urban, rural and production landscapes (Bai et al., 2018;Belcher et al., 2018;Kadas, 2006;Kloskowski, 2010;Wang et al., 2017).
Another, less-studied, artificial island system is wetland habitat created and managed by humans in dry biomes. For instance, artificial watering points in the semi-arid regions of Australia boost kangaroo and bird populations and widen geographic ranges of large herbivores and breeding ranges of invertebrates (James et al., 1999).
Artificial wetlands promote connectivity and foraging habitats for bats, birds and frogs (Oertli & Parris, 2019;Sirami et al., 2013;Stahlschmidt et al., 2012). In dry and populous countries such as India, artificial wetlands nestled in sweeping agricultural lands form the lifeline of the farming community by supporting irrigation and small-scale fisheries and providing a suite of cultural and provisioning services (Bassi et al., 2014;Chimalakonda, 2012). They also augment landscape biodiversity and create habitat islands for species, especially resident and migratory birds (Sundar & Kittur, 2013;Thiere et al., 2009). It is crucial to understand the dynamics of the ecological communities of these artificial islands if the goals of reconciling agricultural production with biodiversity conservation are to be achieved (Lindenmayer et al., 2008;Scherr & McNeely, 2008).
Our focus here is on the artificial wetlands of India, which are emerging novel ecosystems and are likely to become more widespread in the future (Hobbs et al., 2006;Oertli, 2018). We ask the basic question of what structures the diversity of water bird communities in these artificial island systems. Our hypothesis is that diversity of wetland birds in this system is primarily determined by an immigration-extinction balance, as in traditional island biogeography and metacommunity theory (Leibold et al., 2004). We recognize that 'extinction' here represents local extinction, that it may be due to emigration of individuals as well as mortality, and that extinction and colonization events may occur over rapid timescales (days).
Under this hypothesis, wetland species composition fluctuates substantially over time, with a high degree of randomness in species composition at any one time. An alternative hypothesis is that diversity is primarily regulated by the local habitat and niche structure of particular wetlands. Under this hypothesis, wetland species composition is more stable over time, with species presence and absence being strongly related to biotic and abiotic wetland characteristics.
The rationale for our hypotheses is that water birds, with their strong dispersal abilities and movement patterns, would fly easily between the spatially proximate wetlands in the landscape. Although these two hypotheses are not mutually exclusive, the extent to which one or the other is more correct has implications for conservation. Under the first hypothesis, wetland area and connectivity are fundamental determinants of diversity and conservation planning must be at the regional scale. Under the second hypothesis, conservation planning can be more focused on individual wetlands. We tested the hypotheses by looking at four classic macroecological patterns and assessing which hypothesis was more consistent with the observed data. We investigated spatial and temporal biodiversity metrics and dynamics via correlations with dissimilarities in wetland size and distance and null models, respectively.
In addition, we examined the species-area relationship (SAR), and species-abundance distributions (SAD) through different models with underlying niche versus dispersal assembly themes to disentangle which of these drives the macroecological patterns. The temporal patterns are more informative because the two hypotheses make more strongly contrasting predictions about what should be observed.  Figure 1; Appendix S1, Tables S1.2 and S1.3). Agriculture is the primary occupation in this region, with rice, cotton and maize being the major crops grown there. The shorelines are mainly characterized by vegetation including Ipomea spp., Typha spp., Prosopis spp. and agricultural fields. The aquatic vegetation is mainly comprised of Hydrilla spp., Najas spp., Potamogeton spp., Nelumbro spp. and Nymphaea spp.
The climate is generally dry and hot except during the southwest monsoon in the months of June-October. Natural water bodies are rare in the landscape and waterbirds would be rare in the landscape were it not for the artificial wetlands. Thus, the wetlands are analogous to islands.
Sampling was performed during the wintering months when migratory birds were present (mornings approximately 7:00-9:30 am and evenings approximately 4-6 pm). The point count locations at each wetland were the same every year; at each point count location, we observed birds using Nikon binoculars and Celestron spotting scope and noted down species identities and abundances.
Since there is variation in wetland area over time due to irrigation use and fluctuations in monsoonal rain, we assessed the wetlands' areas every winter using Google imagery and the software ArcGIS.
For each wetland, we thus obtained six values of species richness and abundance (from each census) and three values of area (for each year). The only exceptions were wetlands 2, 9, 14, 22, 23 and 54, which were censused only five times, and wetland 27, which was censused only four times due to logistical roadblocks.

| Betadiversity patterns
We used the 'betapart' package in R to analyse betadiversity (Baselga & Orme, 2012). We constructed site-species incidence matrices for each census to compute two dissimilarity measures of betadiversity: turnover and nestedness (Baselga, 2010). The beta. multi function gives sor = sim + sne where sne is the nestedness dissimilarity measure, sim is the turnover component, and sor is the overall betadiversity. The values of all three metrics range from zero to one. We assessed both spatial and temporal betadiversity. We also used the function beta.pair to check for correlation between pairwise dissimilarities and other geographical variables such as isolation and area. We used incidence-based instead of abundancebased indices because our aim was to explore species replacement, whether via immigration-extinction dynamics or niche-based processes (Baselga, 2010).
For spatial betadiversity, if immigration-extinction dynamics are structuring the wetland bird communities, we would expect turnover ( sim ) to be high (i.e. account for the majority of betadiversity), F I G U R E 1 Study region containing the artificial wetlands sampled over the three years bordered by pictures of water bird species encountered during the surveys. The administrative centre Warangal is shown near the centre of the map. The inset shows the location of the region in India. because any given species may by chance be present at some wetlands and absent at others at any given time (although there could also be a contribution from nestedness sne if there is large variation in both isolation across wetlands and migration ability across species). In the extreme case that all sites contain different sets of species and do not share any species, then we would have sim = 1.
On the other hand, if local niche processes are dominant, we would expect nestedness ( sne ) to be high (i.e. account for the majority of betadiversity), because small wetlands will typically have a subset of the niches and hence a subset of the species of larger wetlands (note that this is a stronger statement than saying that small wetlands have fewer species than larger wetlands, which we explore below with the species-area relationship). The betadiversity metrics were calculated for each sampling season separately, to allow finer-grained examination of the data. Combining data from all sampling seasons leads to similar results (not shown).
To probe the spatial betadiversity further, we tested whether dissimilarities in species composition were correlated with dissimilarities in the wetland characteristics such as area or proximity (geographic distance from other wetlands) by bootstrapping parameters estimated from distance-decay functions. If immigration is important, we would expect wetlands that are closer in space to be more similar in species composition, whereas if local niche structure is important, we would expect wetlands of similar area to be more similar in species composition.
To assess temporal betadiversity, we computed the measures for the 15 pairwise combinations of time steps arising from six censuses, and hence, for each wetland, there were 15 observed values of temporal turnover metrics. In instances where wetlands were sampled only five times, there were only ten combinations, and for the wetland with only four censuses, there were only six combinations. The averages and the standard deviations of sor, sim and sne for each wetland were calculated. If immigration is important, we would expect to observe high temporal betadiversity driven mostly by temporal turnover ( sim ), whereas if niche structure drives local wetland biodiversity, we would not necessarily expect any substantial change in species composition through time at a given wetland.
Furthermore, we investigated whether species pools in wetlands sampled closely in time differed from wetlands that were sampled farther in time. We assessed whether temporal betadiversity increased over time within a wetland by regressing the metrics sor , sim and sne against the length of the time interval between censuses and comparing the results to a null model in which the order of censuses was randomized (see Appendix S3 for details). The slope of this relationship indicates the rate of change of betadiversity metric with inter-census length (measured in days) over the range of the observed data.

| Species-abundance distributions
The SAD is a frequency distribution of species abundances in a community. The SAD is a classic community pattern that cannot on its own diagnose community assembly mechanisms (Chisholm et al., 2014;Laurance et al., 2009), but does contain useful information. For example, monotonically decreasing SADs (log-series or geometric) are characteristic of communities with very high or very low immigration, whereas hump-shaped SADs (e.g. lognormal) are characteristic of communities with intermediate immigration (Mouquet & Loreau, 2003). We fit four different SAD models to each wetland bird community. Fitting was done using the 'sads' package in R (R Core Development Team, 2013). We fit the model separately for each wetland and for every census. We did not use average abundances over time as we wanted to examine temporal trends in SADs and see whether the best-fit SAD model remained similar in shape over time. Next, for each census, we fitted a pooled SAD to the data from all 57 wetlands, which we conceptualized as one large metacommunity.
The four SAD models we fitted were the broken-stick, lognormal, log-series and Volkov models. The broken-stick distribution traditionally comes from a niche apportionment model where niche space in a community is broken simultaneously among the species of the community (MacArthur, 1960). The log-normal model, in which the logarithms of species abundances follow a normal distribution, was originally phenomenological but can be derived from mechanistic population models in special cases (Engen & Lande, 1996). The log-series was also long used as a purely phenomenological distribution but has more recently been shown to arise mechanistically from neutral metacommunity models (Hubbell, 2001;Volkov et al., 2007).
The Volkov model represents a local community under neutral drift with immigration from a metacommunity (Volkov et al., 2003). The best-fit model was selected based on Akaike's information criterion (AIC). The model with the lowest AIC was considered the bestfit model and any model with AIC difference with the best model less than two was deemed statistically indistinguishable from it (Burnham & Anderson, 2003). We used Preston plots, that is, frequency distributions of log 2 abundance classes, to represent SADs in the analyses (Preston, 1962). In the context of our hypothesis that diversity of the wetlands is determined by immigration-extinction dynamics (our first hypothesis) rather than local niche processes (our second hypothesis), we expected the best fit to the data to be given by the Volkov model with a high value of the immigration parameter, or potentially the log-series model, which is a limiting case of the Volkov model under high immigration.

| Species-area relationships
The SAR is simply the relationship of species richness to area. Empirical SARs are generally increasing (Lomolino, 2000), but they can be nearly flat when niche-assembly dominates (Chisholm et al., 2016). We fitted four different SAR models to the dataset. The first was the classic power-law SAR, S = cA z , where S is species richness, A is area, and c and z are fitted parameters. The power-law SAR has a long history of application to habitat areas in mainland landscapes. The second model was a semi-logarithmic model that has been broadly applied in island biogeography: S = a + blogA, where a and b are constants. The third was a breakpoint regression that is similar to the semi-logarithmic model but allows two separate slopes and estimates a threshold area below which one slope applies and above which the other applies (Chen et al., 2020;Wang et al., 2018). This model has been used to capture the 'small-island effect' (Lomolino & Weiser, 2001;Niering, 1963), a flat phase in the SAR often observed at small areas in archipelago data. The fourth was an island diversity model that incorporates the effects of niches and the effects of immigration and, again consistent with the small-island effect, predicts a flat SAR at small scales where niches dominate and an increasing SAR at large scales where immigration dominates (Chisholm et al., 2016). This fourth model could be particularly informative because it estimates a critical wetland area above which the effects of immigration dominate and below which the effects of niches dominate. We fit the power law and the Chisholm model using Mathematica (student edition 11.2) NonLinearModelFit function to fit S against logA and the semi-logarithmic model using the LinearModelFit function. We fit the breakpoint model in R using the segmented.lm() function in the 'segmented' package. Direct comparison of goodness-of-fit between the three models was done using root mean squared error and AIC. We carried out the SAR analyses for all species combined as well as for resident species alone.

| RE SULTS
A total of 76 species of water birds and 116,070 individuals were counted across the six censuses (Appendix S1, Table S1.1). Migratory birds (~25,000 individuals, 28 species) contributed to 23% of the overall abundance of water birds in the wetlands. The most abundant and widely occurring resident bird species was the Eurasian Coot (Fulica

| Betadiversity analyses
The spatial betadiversity among the wetlands was consistently high and was driven primarily by spatial turnover rather than nestedness (Appendix S3, Table S3.1). The average multisite sor across the six censuses was 0.93 (standard deviation 0.04). The two components sim (turnover) and sne (nestedness) had average values of 0.88 and 0.05 (standard deviations 0.08 and 0.01), respectively. Relationships between the pairwise compositional dissimilarity matrices and the dissimilarities in area and geographic distance, for each census as well as all censuses combined, were not statistically significant ( Figure 2; Appendix S3, Tables S3.3 and S3.4).
The temporal betadiversity patterns were consistent with random assembly. There was little noticeable increase in temporal betadiversity over time within wetlands: Of 57 wetlands, 39 showed a positive relationship between sor and time interval between censuses, but in only three of these cases was the relationship statistically significant at the = 0.05 threshold. This lack of correlation suggests either that (a) species exhibited strong site fidelity over the study period or that (b) the species are frequently randomly shuffled across wetlands with a strong component of randomness. The temporal randomization tests were used to test (a) versus (b), because if hypothesis (a) is true then randomly switching up species between wetlands should produce patterns that are very different from the data. The observed sor , sim and sne values fell within 95% CI of randomized values through null models for 54 out of 57 of the wetlands (e.g. Figure 3 for wetland 44) thus providing support for hypothesis (b). This result coupled with the high intercept value of sor and the low value of sne for each wetland suggests random assembly of birds in the wetlands.
The mean inter-census turnover, averaged across all the wetlands, was 0.31. This means that across two censuses at any given wetland, in the census with lowest species richness about 31% of species recorded were not recorded in the other census . The corresponding values for individual wetlands ranged from 0.14 to 0.53. Temporal betadiversity was driven mainly by turnover rather than nestedness (Appendix S3, Figure S3.1 and Table S3.2), as for spatial betadiversity ( Figure 2). Temporal betadiversity was higher than spatial betadiversity, that is, the difference in composition between two censuses at a single wetland (mean sor = 0.42) was on average greater than between two wetlands across space (mean sor = 0.24).

| Species-abundance distributions
For the 333 empirical SADs from each wetland-census combination, the log-series model was in most cases the best fit. In many cases, the AIC of the best-fit models for each wetland differed from the AIC of other SAD models by less than two points, indicating that multiple models gave reasonable fits. The log-series and the Volkov model were among the top competing models (difference of AIC less than two with the best fit model) for 251 and 256 wetland SADs, respectively. We also assessed how the best-fit model for SADs changed at each wetland across the six censuses. None of the wetlands had the same best-fit SAD model across all the censuses (Figure 4). The closeness of goodness-of-fit statistics meant that the best-fit SAD model was often not a clear choice.
For the pooled community data (all 57 wetlands combined), the AIC values of each model for all the censuses are in Table 1

| Species-area relationships
The water bird densities ranged from 0.3 individuals per ha to 252.5 individuals per ha in an individual wetland and across all censuses.
The mean density over all wetlands in censuses 1-6 was, respectively, 7.9, 11.3, 8.  Table 1; more detail on the fits of the Chisholm model are given in Appendix S2, Table S2.4). We found similar patterns of no small-island effect and fitted SARs falling in the immigration-structured regimes when we conducted SAR analyses on all species as well as resident birds only (Appendix S2, Figure S2.3, Table S2.5 and Figure 5).

| DISCUSS ION
The aim of this study was to assess key patterns of water bird species diversity in a network of artificial wetland habitats and to make F I G U R E 2 Relationship of sor (overall spatial betadiversity) , sim (spatial turnover) and sne (spatial nestedness) to difference in area between two wetlands (a)-(c), and geographic distance between two wetlands (d)-(f). Each point on each panel represents a pair of wetlands (the total number of wetlands is 57). The results are for all censuses combined. Panels (a) and (d) show sor , panels (b) and (e) show sim , and panels (c) and (f) show sne . inferences about the processes structuring the diversity. We examined four key community-level macroecological patterns. We found that the patterns were broadly consistent with the hypothesis that bird diversity in our wetlands is structured by immigration-extinction processes, as in traditional island biogeography and metacommunity theory, rather than strong niche processes, although we acknowledge that some of the patterns are also consistent with niche processes and that more research targeted at niche processes is needed to arrive at more definitive conclusions.
Our most compelling evidence for the role of immigrationextinction dynamics came from temporal and spatial betadiversity, which were both high and mainly driven by turnover, rather than nestedness. Nestedness in ecological communities is suggestive of nonrandom patterns of community assembly processes (Gaston & Blackburn, 2008;Ulrich et al., 2009). Larger wetlands can potentially contain unique species, either because they have microhabitats or niches not present in smaller wetlands, which would favour specialist species, or simply because of greater resource availability, which would favour species with larger body sizes and higher food intake requirements (Ulrich et al., 2009). In such cases, we would expect to see a high level of nestedness, with the species of smaller wetlands being a subset of those of larger wetlands. Turnover, on the other hand, can arise from stochastic local extinction or dispersal among wetlands. Thus, the most parsimonious explanation for the observed patterns is immigration-extinction dynamics. For instance, at a given wetland, on average 31% of species recorded in census with the most species were not recorded in the other census, reflecting a large number of local extinctions and colonizations. We attribute these results primarily to the strong flight capability of most wetland birds relative to the average spatial separation of our wetlands. Another possible contributing factor was that wetlands were sampled in winter, when food and water availability is fairly homogenous throughout the landscape, resulting in reduced niche differences across wetlands.
Our wetland bird communities are changing on very rapid timescales. In general, one might expect temporal betadiversity to increase with time (Korhonen et al., 2010), but in our wetlands temporal betadiversity measured over one month is similar to that over two years. One explanation would be that the dynamics operate on timescales longer than the two-year period of our observation, but we discounted this possibility with randomization tests across wetlands that showed minimal site fidelity (Figure 3). An explanation more consistent with the data is that the immigrationextinction dynamics operate on timescales of less than a month, that is, the temporal autocorrelation period is less than one month  Figure S3.2) suggest not only that birds disperse long distances, but also that they disperse largely randomly without   Note: The Volkov model (in bold) was consistently the best-fit model across the six censuses. All three SAR models fit the data well with ΔAIC ≤2. regard to wetland characteristics, although it is possible that dispersal is regulated by some unmeasured biotic or abiotic wetland characteristics.
The overall higher betadiversity signals the value of this network of wetlands for supporting water bird diversity. Our mean values of spatial betadiversity ( sor = 0.93) and spatial turnover ( sim = 0.88 ) are higher than those of previously studied bird communities on terrestrial islands situated in an artificial lake ( sor = 0.77, sim = 0.62 ) and bird communities in woodland lots ( sor = 0.48, sim = 0.38 ) (Jankowski et al., 2015;Si et al., 2015). Bird assemblages in sites with lower species richness are not simply subsets of assemblages at sites with higher species richness. Additionally, there is no single wetland in the landscape with high alpha diversity and high temporal betadiversity that holds the entire suite of water bird species in the landscape. For these reasons, the current dominant conservation strategy of managing single large wetlands in the country (e.g. Ramsar sites and Important Bird Areas), while crucial, is likely inadequate on its own for conserving bird assemblages across large spatial and temporal scales. A landscape-centric approach to wetland conservation, which also involves managing many smaller agrarian wetlands, could significantly benefit regional bird diversity (Sundar & Kittur, 2013).
We identified future avenues of research in these wetland systems. First, there could be seasonal differences in dominant forces structuring water bird diversity (Murgui, 2010). It is possible that summer sampling would yield higher nestedness due to greater variability in abiotic conditions across wetlands. Second, the relationships between betadiversity patterns and abiotic variables (such as wetland area) or life-history traits (body size) could be investigated (Almeida-Neto et al., 2008). Third, we acknowledge that unmeasured niche-related variables, such as food resources, could be driving some of the observed patterns, at least for some subgroups of species: We recommend that future studies in this system focus on collecting data on such variables.
Other evidence for the importance of immigration-extinction processes came from the SADs and SARs, although we acknowledge that caution must be exercised when drawing inferences about mechanisms from such patterns (Fisher et al., 2010;Martiny et al., 2011;McGill, 2003). The SADs were consistent with a model in which immigration to individual wetlands is very high, leading to large numbers of locally rare species, but immigration to the system as a whole is somewhat restricted, leading to fewer rare species at the landscape scale. The best fit SAD model at the individual wetland scale was consistently the classic log-series SAD. This may seem curious because the log-series typically describes SADs at much larger scales; local SADs more often exhibit a strong interior mode characteristic of a log-normal distribution. We attribute the good fits of the log-series to local SAD patterns here to very high immigration rates, which leads to large numbers of locally rare species and a shift of the interior mode to the left (Hubbell, 2001;Mouquet & Loreau, 2003).
Indeed, the estimated local immigration rates to each wetland from the Volkov model, expressed as the fraction of individuals who are immigrants from other wetlands, are high with a mean m of 0.58 (Appendix S2, Table S2.1) and at some wetlands reached the maximum value of unity. Most immigration was attributable to movement between wetlands within our metacommunity, rather than immigration from outside: The Volkov model again performed well for the pooled data at the metacommunity scale and the average estimate of immigration from outside the metacommunity was only 0.008.
The final part of the study, through the examination of SARs, showed no evidence of the small-island effect characteristic of many oceanic archipelagos (Lomolino & Weiser, 2001) and that recent theoretical work has attributed to the action of niche processes where immigration is very weak, for example, on very isolated islands F I G U R E 5 The best fit model of the Chisholm et al. (2016) (blue), power-law (purple) and the semi-logarithmic (brown) models to the species-area datasets for water birds from the first census (the breakpoint model is not shown because its AIC was more than two points greater than the best model). Note the horizontal axis (Area) is on a logarithmic scale and the vertical axis (Species Richness) is on a linear scale. Each point corresponds to species richness at one wetland. The fits of the three models shown were very similar (root mean squared error did not differ by more than 1.5% across models in any given census). Refer to Appendix S2, Figure S2.2 for SARs of all censuses. (Chisholm et al., 2016), thereby again pointing to predominantly immigration-extinction-structured water bird communities. We found similar patterns of no small-island effect when analysing all species together as well as resident species separately. Most species (48 of 76) are resident species that are quite motile at the scale of the landscape studied. Our results do not exclude the role of niches in sustaining species diversity, but strongly indicate that immigration controls the species richness of water birds in the wetlands despite whatever niche processes are operating. Waterbirds can fly thousands of kilometres and would be one of the first groups of species to colonize novel ecosystems such as these wetlands. Given the birds' excellent dispersal abilities and the spatial proximity of these wetlands, our results suggest that it is primarily through ongoing dispersal of water birds between the wetlands, balancing ongoing local disappearances from wetlands, that the community at each wetland is assembled.
Wetland area was positively associated with species richness (explaining ~20% of the variation), consistent with the standard assumptions of island biogeography. We expect the remaining variance to be attributable mainly to other spatial factors such as the position of each wetland in the network and its isolation from other wetlands, which we did not assess (except via the crude measure of pairwise distances between wetlands), and to stochasticity. A network or spatial metacommunity approach to modelling this system may be a fruitful future research direction. Additionally, future research could examine the influence of local factors, such as resource availability, prey availability and disturbances, on bird diversity, but we expect these factors to explain a comparatively small fraction of the variation in wetland bird diversity.

| CON CLUS IONS
Multiple lines of evidence point to immigration as the dominant mechanism structuring bird diversity in this artificial wetland system, although we emphasize that some of our observed patterns are also consistent with niche drivers and we cannot yet rule out a strong subsidiary role for such processes, and we encourage future work measuring variables associated with such drivers (e.g. food resource availability). We expect many novel ecosystems will exhibit similar structure, simply because by definition any species thriving in such systems must have arrived there fairly recently and therefore must have good dispersal capability. This has broad implications for the management of novel ecosystems, in particular those comprising networks of smaller wetlands (Hill et al., 2018;Sebastián-González & Green, 2016;Sundar & Kittur, 2013). Although adequate abiotic (e.g. water quality) and biotic (e.g. vegetation) conditions are a prerequisite for a thriving bird community at any given site, there should also be a strong management focus on landscape-scale processes and connectivity. For our wetland system in Telangana, India, we encourage future work fitting mechanistic metapopulation models to our data (Hanski, 1997;Holmes et al., 2020;Reigada et al., 2015), which can be used to project the effects of potential future changes to the wetland network (e.g. draining or drying of existing wetlands, or creation of new wetlands) on bird diversity, thus informing conservation management.

ACK N O WLE D G E M ENTS
We thank Saniya Chaplod and Paloma Noronha for their help with fieldwork and Tak Fung for his discussions.

CO N FLI C T O F I NTE R E S T S TATE M E NT
The authors declare no conflict of interest.

PEER R E V I E W
The peer review history for this article is available at https:// www.webof scien ce.com/api/gatew ay/wos/peer-revie w/10.1111/ ddi.13715.