Landscape genetics identifies streams and drainage infrastructure as dispersal corridors for an endangered wetland bird

Abstract Anthropogenic alterations to landscape structure and composition can have significant impacts on biodiversity, potentially leading to species extinctions. Population‐level impacts of landscape change are mediated by animal behaviors, in particular dispersal behavior. Little is known about the dispersal habits of rails (Rallidae) due to their cryptic behavior and tendency to occupy densely vegetated habitats. The effects of landscape structure on the movement behavior of waterbirds in general are poorly studied due to their reputation for having high dispersal abilities. We used a landscape genetic approach to test hypotheses of landscape effects on dispersal behavior of the Hawaiian gallinule (Gallinula galeata sandvicensis), an endangered subspecies endemic to the Hawaiian Islands. We created a suite of alternative resistance surfaces representing biologically plausible a priori hypotheses of how gallinules might navigate the landscape matrix and ranked these surfaces by their ability to explain observed patterns in genetic distance among 12 populations on the island of O`ahu. We modeled effective distance among wetland locations on all surfaces using both cumulative least‐cost‐path and resistance‐distance approaches and evaluated relative model performance using Mantel tests, a causal modeling approach, and the mixed‐model maximum‐likelihood population‐effects framework. Across all genetic markers, simulation methods, and model comparison metrics, surfaces that treated linear water features like streams, ditches, and canals as corridors for gallinule movement outperformed all other models. This is the first landscape genetic study on the movement behavior of any waterbird species to our knowledge. Our results indicate that lotic water features, including drainage infrastructure previously thought to be of minimal habitat value, contribute to habitat connectivity in this listed subspecies.


| INTRODUC TI ON
Research on animal movement behavior, in particular how landscape features affect dispersal, is essential for predicting, understanding, and managing the impacts of ongoing changes in climate and landscape structure on animal populations (Hanski, 2001;Holyoak & Heath, 2016;Knowlton & Graham, 2010;van Strien et al., 2014).
Although direct data on animal movement can be time-consuming and expensive to collect, the development of indirect methods using genetic markers to estimate rates of dispersal has greatly increased understanding of population connectivity (Anderson, Kierepka, Swihart, Latch, & Rhodes, 2015;Epps, Wehausen, Bleich, Torres, & Brashares, 2007;Lowe & Allendorf, 2010). These indices can be especially important for studying the movement of behaviorally cryptic species that are difficult to study through other means like mark-resighting (Finnegan et al., 2012), and provide estimates of genetic differentiation due to dispersal and subsequent gene flow (Sexton, Hangartner, & Hoffmann, 2014;Waser & Strobeck, 1998).
The field of landscape genetics provides an analytical framework to assess the potential effects of landscape structure and composition on genetic differentiation in wildlife populations (Manel & Holderegger, 2013;Manel, Schwartz, Luikart, & Taberlet, 2003).
According to the most basic landscape genetic model, isolation by distance, genetic similarity among populations (or individuals) is correlated with geographic or Euclidean distance (Wright, 1943).
The concept of effective distance extends this model by incorporating information on the movement behavior of an organism and by considering distances between points or populations in addition to the permeability of the intervening landscape matrix to movement for that particular organism (Adriaensen et al., 2003;McRae, 2006). Effective distances are quantified using resistance surfaces, spatially explicit models that reflect hypotheses about the degree to which specific landscape features or cover types impede or facilitate movement in a raster format (Spear, Balkenhol, Fortin, McRae, & Scribner, 2010;Storfer et al., 2007). Among a suite of surfaces, those that best explain data on spatial genetic variation are assumed to represent the most likely representation of how a given set of landscape features affects movement in an organism (Cushman & Landguth, 2010;Zeller, McGarigal, & Whiteley, 2012).
Thus, landscape genetic analyses of threatened and endangered bird taxa are urgently needed for a better understanding of the impacts of continued anthropogenic landscape change. Among avian taxa, rails (family Rallidae) are among the most poorly understood with regard to their movement ecology, due to their cryptic behavior and tendency to inhabit densely vegetated habitats (Ripley, Lansdowne, & Olson, 1977;Taylor, 2010). Rails also exhibit the interesting behavioral-evolutionary tendency to colonize widespread and isolated islands or habitat patches, while appearing to have a natural antipathy to disperse after colonization, often becoming flightless (Livezey, 2003;Steadman, 2006). Coupling these behaviors with the discrete and naturally fragmented nature of many wetland ecosystems, and further isolation by anthropogenic landscape change, wetland-specialist birds like rails are a convenient study system for landscape genetic research. The sensitivity of wetland ecosystems to a diversity of anthropogenic threats (Green et al., 2017;Strayer & Dudgeon, 2010) makes wetland birds a group for which landscape genetic research is likely of substantial importance to conservation.
Our interest is in one member of the Rallidae, the Hawaiian gallinule (Gallinula galeata sandvicensis, Figure 1), which is an endangered subspecies of the common gallinule endemic to freshwater wetlands of the Hawaiian Islands (United States; Bannor & Kiviat, 2002).
Hawaiian gallinules were once found on the five main Hawaiian Islands, but were extirpated from all islands other than O`ahu and Kauai during the late 19th to mid-20th century (Banko, 1987).
Habitat loss from anthropogenic landscape change and exotic, invasive wetland plants, as well as predation from introduced mammalian predators drove severe population declines and reductions in the subspecies' range (Griffin, Shallenberger, Fefer, Sharitz, & Gibbons, 1989;USFWS, 1977). This precipitous decline was eventually halted with legal protection, the establishment of National Wildlife Refuges and state protected areas on these two islands, and the advent of habitat management (predator control and vegetation restoration).
Wetland habitats on O`ahu are distributed with varying degrees of F I G U R E 1 An adult Hawaiian gallinule stands on lilly pads at in a golf course water hazard in Kailua, Hawai`i. Photograph credit Amanda Sandor geographic isolation and are embedded in a complex landscape mosaic of anthropogenic land cover, including highways, recreational areas (e.g., golf courses and resorts), military bases, agricultural land, and residential areas. Connectivity among these fragmented subpopulations is considered an important factor in the conservation and management of Hawaiian gallinules (Reed, DesRochers, VanderWerf, & Scott, 2012a,b;Underwood et al., 2013;US Fish and Wildlife Service, 2011). However, little is known about the movement behavior of Hawaiian gallinules, and their cryptic behavior leads to poor detection rates, limiting the efficacy of mark-resight studies (DesRochers, Gee, & Reed, 2008).
Using information from expert opinion and published and unpublished literature, we identified several landscape characteristics and features that may influence movement and gene flow in Hawaiian gallinules. These included roads, urban land cover, (K. Doyle, Hawaii Division of Forestry and Wildlife, pers. comm.), forested areas, steep slopes, and high elevation terrain (Banko, 1987;Perkins, 1903) as potential barriers. By contrast, we expected that mesic areas and open habitats would promote dispersal and gene flow. We specifically predicted that streams and rivers might facilitate dispersal, given reports from a related species and anecdotal observations from experts in this taxon (Nagata, 1983;Takano & Haig, 2004).
This information made it clear that there are a great many alternative explanations for how Hawaiian gallinules might perceive landscape features during dispersal between wetlands. Consequently, we generated a suite of resistance surfaces of the O`ahu landscape matching our proposed hypotheses about the relative resistance of both natural and anthropogenic landscape features to gallinule movement across the landscape. The models are based on observations and biologically informed speculation about distributions and movements of Hawaiian gallinules across the landscape. Once these models were created, we simulated the movement of gallinules across the landscape according to these hypotheses and compared their fits to data on microgeographic genetic differentiation of Hawaiian gallinules on O`ahu (pairwise F ST using microsatellite markers; van Rees, . Our goal was to determine the relationships among landscape features and observed genetic differentiation to identify landscape features that are important to maintaining connectivity among Hawaiian gallinule populations.
Results from this study will provide important information for the subspecies' recovery and to predicting the potential vulnerability of Hawaiian gallinules to future modifications to the landscape as attributable to land use and climate change.

| Study area
We studied Hawaiian gallinules on the island of O`ahu (Hawai`i,USA,21.3156 N,, one of two islands that make up the subspecies' entire range. We collected genetic samples at 12 locations on the island, which represent all known major wetland habitats for the subspecies on O`ahu ( Figure 2). Sampled sites were distributed across the island's low-elevation coastal plain and included sites in

| Genetic data
We obtained multilocus genotypes for 152 Hawaiian gallinules at 12 wetlands from a previous study (van mtDNA sequences are accessioned in GenBank, MF673902-MF673904).
We defined wetlands as complexes of spatially proximate and hydrologically linked water bodies. The sampled wetlands included all major breeding areas for the Hawaiian gallinule on the island, and our sample accounts for at a minimum 30% of the known population of the island (Reed et al., 2011;US Fish and Wildlife Service, 2011).
We captured gallinules using walk-in cage traps baited with attractive food items (fresh fruit and cracked corn) and extracted DNA from 4 to 6 body feathers collected from each captured bird. We obtained estimates of interpopulation genetic variance (F ST ) among the 12 wetland sites from van Rees, ; these estimates were based on microsatellite genotype data collected from 12 autosomal loci and 520 base pairs (bp) of the NADH dehydrogenase 2 (ND2) region of mitochondrial DNA (mtDNA). All microsatellite loci

| Landscape variables
We represented landscape variables and movement capacities using resistance surfaces, in which landscapes are modeled as a raster grid, assigning different resistance values to landscape cover classes or features according to the hypothesized difficulty of passing through such features  Figure 3). We analyzed 20 resistance surfaces that addressed 10 hypotheses pertaining to the movement ecology of Hawaiian gallinules (Table 1). These hypotheses were derived from expert opinion and literature on this and related taxa. We named these surfaces according to the datasets from which they were derived; the named groups are Elevation, Topographic Wetness Index (TWI), Land Use (LU), Roads, and Proximity-to-Water.
To avoid issues with scaling across predictor variables, all surfaces were resampled to 30 m resolution and resistance was scaled from 1 to 100, where 1 is minimal resistance and 100 is maximum possible resistance. We assigned values to different landscape features based on expert opinion and available field evidence, with the objective of defining relative degrees of resistance (e.g., roads have higher resistance than agricultural fields), rather than specific numerical values (e.g., roads have a resistance value of 70, rather than 40) (Figure 3a-d) (see Spear et al., 2010;Zeller et al., 2012).
Elevation datasets were derived from 30 m resolution digital elevation models from the Hawaii Department of Commerce et al. (2007). Three types of surfaces were created using digital elevation models, with two versions each (based on the values assigned to open water, see below), for a total of six resistance surfaces (Table 1).
The first of the elevation-based hypotheses is binary models, in which we assigned a low resistance to all pixels below an empirically derived elevation threshold (resistance value = 10), and assigned a high value (80) to all pixels above that value (Hypothesis 1). The  van Rees and Reed, unpubl. data). Birds could move through high elevations, but would move more readily (by a factor of 8) through low elevations. For linear resistance surfaces, we assumed a direct linear relationship between elevation and landscape resistance, with minimal resistance (resistance value = 1) at coastal elevations and maximum resistance (resistance value = 100) at maximum elevation for the island (Hypothesis 2). Finally, for slope surfaces, we assumed a linear relationship between degree of slope and resistance, where the maximum slope was given a value of 100, and flat ground was given a resistance value of 1 (Hypothesis 3).
Two additional hypotheses, with a total of four resistance surfaces, were based on the TWI (Beven & Kirkby, 1979). The TWI is a simple hydrological model that uses a digital elevation model (a spatial representation of elevation across a landscape) to approximate the likelihood that water would accumulate at any single point under uniform rainfall conditions. The TWI is a unitless measure, with higher values in areas that are likely to support standing water or mesic conditions, which are strongly associated with the occurrence of common gallinules and gallinule habitats (Bannor & Kiviat, 2002).
We calculated TWI using the Geomorphology and Topology Toolbox (Evans & Oakleaf, 2011) in ArcGIS 10.2. In binary TWI models, we divided the landscape between low-resistance pixels (resistance value = 1) at or above a threshold TWI value, and high-resistance pixels (resistance value = 100) below that value (Hypothesis 4). We used a threshold value (TWI value = 11.5) for binary TWI surfaces based on van Rees and Reed (2014) (Ripley et al., 1977). Although genetic (Miller, Mullins, Haig, Takano, & Garcia, 2015) and observational (Takano & Haig, 2004;Worthington, 1998)  shoreline wetlands, and estuarine marshes from this dataset because they are not used by Hawaiian gallinules (Banko, 1987), and we retained rivers, streams, freshwater low-elevation wetlands, and other water features (drainage ditches and irrigation infrastructure). We then used the Euclidean distance tool to generate a raster dataset where each pixel was assigned a value based on its proximity to the nearest water feature. This category of resistance surfaces is based on anecdotal accounts that Hawaiian gallinules tend to travel along river margins, observations by the authors that the birds appear behaviorally inhibited from moving far from water, and evidence from related taxa that movement occurs along riparian corridors (Nagata, 1983;Takano & Haig, 2004;Hypothesis 10

| Effective and euclidean distances
Because little is known about the movement behavior of Hawaiian gallinules, we calculated effective distances among all pairwise combinations of occupied (and sampled) habitat patches using both cumulative least-cost paths (Coulon et al., 2004;Michels et al., 2001) and resistance distances (McRae, 2006), which differ in their assumptions of an organism's movement behavior and knowledge of the surrounding landscape (Coulon et al., 2004). Circuit-theory approaches simultaneously consider all potential movement pathways when calculating effective distances, while least-cost path simulations consider only a single optimal pathway, thereby assuming that the animal has complete knowledge of the landscape. We also calculated topographically adjusted Euclidean distances between population pairs using the near-to-table tool in ArcGIS.

| Landscape genetic analyses
There is disagreement in the recent literature on which statistical methods are most appropriate for assessing the relationships between landscape features and genetic differentiation (Shirk, Landguth, & Cushman, 2017;Zeller et al., 2016 Akaike, 1973;Hurvich & Tsai, 1989) and by calculating the R 2 β statistic (Edwards, Muller, Wolfinger, Qaqish, & Schabenberger, 2008 that other methods may exhibit toward highly complex models. We fitted mixed effects models with REML estimation using the lmer function in the package lme4 (Bates, Maechler, & Bolker, 2011) in R, and calculated AIC values and generated AIC tables using AICtab function in the package AICcmodavg (Mazerolle, 2017). We calculated R 2 β using the package pbkrtest (Halekoh & Højsgaard, 2014). We implemented a post hoc analysis to test three additional models (see Supporting information Table S2) designed to account for three potential confounding factors that might have biased model selection toward Proximity-to-Water models. These were (a) that any resistance surface consisting of low-resistance, linear features (corridors) outperforms all others, (b) that surrounding wetland habitat at source and destination nodes was driving patterns of simulated effective distance, and (c) that the spatial arrangement of features in the Proximity-to-Water resistance surfaces, and not their resistance values, was driving their ability to describe observed genetic structure. Scenario 1 was a special concern, considering that the Proximity-to-Water resistance surfaces were the only ones that contained linear features that could act as corridors. To test Scenario 1, we repeated our methods using an inverse version of the roads map, in which O`ahu's roads had low resistance, acting as corridors.
For Scenario 2, we created a resistance surface identical to the Proximity-to-Water 100-m buffer layer, but using a new dataset that only featured streams and drainage infrastructure, and from which all wetland areas had been removed. Finally, for Scenario 3, we tested the explanatory value of an inverse version of the Proximityto-Water 100 m buffer resistance surface. These extra resistance surfaces were tested against our microsatellite genetic dataset.

| Landscape genetic analysis
Models of the Proximity-to-Water group generally explained a higher amount of observed variation in pairwise population differentiation than any other group of models (Table 2) Table 2, mitochondrial DNA in Supporting information). All observed statistically significant p-values from simple Mantel tests were restricted to models from the Proximity-to-Water group. R 2 β values of linear mixed models using the MLPE parameterization were also highest for Proximity-to-Water models, although the difference was less pronounced, and R 2 β values were generally low. The Euclidean distance model performed poorly across all methods of comparison, and models from all groups except for Proximity-to-Water varied widely in performance across methods of estimating effective distance, but always performed more poorly than Proximity-to-Water models. We observed no clear pattern of model support between models with high resistance assigned to ocean water (A models; see Table 1) and models with low resistance assigned to ocean water (B models).
Among Proximity-to-Water models, r values, RS values, and R 2 β values were consistently higher using effective distances calculated with cumulative least-cost paths than those calculated in Circuitscape. While all Proximity-to-Water models using least-costpath had statistically significant p-values, only one, the binary model, had a significant p-value among those with effective distances measured in Circuitscape. The best overall models differed according to both measure of effective distance and method of statistical analysis, with the two-class, linear to 100 m distance, linear to 200 m distance, and negative binomial distance functions scoring highest for at least one statistic and effective distance measure.
Our AIC C analysis of MLPE-parameterized linear mixed models showed clear support for Proximity-to-Water models, with nine of the top 10 models being based on Proximity-to-Water surfaces (Table 3). The least-cost-path simulated versions of models performed better in our AIC analysis as well, with the top five models coming from our least-cost-path datasets, and only three of the top ten models coming from effective distances measured in TA B L E 2 Test statistics from Mantel (r) and partial Mantel tests, as well as mean relative support (RS) and R 2 β values for all landscape resistance models evaluated using data on genetic differentiation (F ST among 12 microsatellite loci) among 12 populations of Hawaiian gallinules on O`ahu Notes. For each model, statistics are given separately for effective distances calculated using cumulative least-cost path (LCP) and resistance distances in Circuitscape (CS). The Euclidean distance model did not include effective distance, so only one value is presented for each statistic, with the exception of partial mantel RS, where mantel r values were compared to those from models run with effective distances calculated using both methods. Asterisks (*) indicate statistically significant p-values at the α = 0.05 level. TWI and LU stand for Topographic Wetness Index and Landscape Use, respectively. a Only one column per statistical method, because Euclidean distance cannot be simulated.
Circuitscape. The second-ranked model had an ΔAIC C > 4 from the top model, indicating a substantial difference in support from the top model (Burnham & Anderson, 2002). The β (coefficient or slope) estimates for the fixed effect of the top five univariate models were 0.51, 0.35, 0.30, 0.19, and 0.15.
Among our post hoc tests, the roads-as-corridors' resistance surface performed very poorly overall, with a low RS and R 2 β value, although the mantel p-value for effective distances created in Circuitscape was near significance (Mantel r = 0.182, p = 0.07). The poor performance of this resistance surface using our least-costpath algorithm implied that linear features alone cannot explain the high performance of Proximity-to-Water models using least-costpath effective distances. The streams and drainage surface with wetlands removed performed comparably to other Proximity-to-Water models, with a high Mantel r value (0.474 for LCP, 0.327 for CS), and lower Mantel p (0.021 for LCP, 0.058 for CS), and high RS.
Although this model did perform more poorly than other Proximityto-Water resistance surfaces, its sustained high performance compared to other resistance surfaces supports our initial interpretation that Proximity-to-Water surfaces are explaining gene flow on a landscape context, and not simply because they include local habitat features at population nodes. Finally, our inverse "water as barrier" resistance surface performed extremely poorly, indicating that the resistance values assigned to the original Proximity-to-Water surfaces are indeed responsible for high model performance.

| D ISCUSS I ON
To our knowledge, this is the first landscape genetic analysis for a terrestrial waterbird species. As such, it represents a step toward overcoming one bias in the growing literature of landscape genetics (Kozakiewicz et al., 2017;Zeller et al., 2012). We found consistent support for resistance surfaces that were based on Proximity-to-Water, while all other resistance surfaces showed low explanatory value and statistical significance. The higher explanatory power and statistical significance of Proximity-to-Water surfaces was robust across four model selection metrics (Mantels r and p-value, RS, R 2 β , and AIC C ), two simulation frameworks (least-cost paths and resistance distance), and two genetic marker types, suggesting that the presence of water features explains 10.7%-63.7% of variation in observed genetic structure among Hawaiian gallinule populations inhabiting wetlands on O`ahu (Table 2). Although the results of simple Mantel tests should be interpreted cautiously (Balkenhol et al., 2009;Zeller et al., 2016), we see congruent patterns in several more robust metrics. Zeller et al. (2016) found that simple Mantel's r and RS performed best when comparing resistance surfaces where landscapes were highly fragmented, and we believe our study system fits this condition well. We also analyzed our results using linear mixed effects models fit with MLPE, which is currently considered the best performing method for performing regressions on matrix data (Shirk et al., 2017), and had similar results using two different methods of model selection. The consistency across model selection metrics and sharp contrast in support compared to all other models of landscape resistance provide evidence that the presence of small wetlands, drainage canals, and streams enhances genetic connectivity in this endangered subspecies. Our findings support suggestions by other authors that Hawaiian gallinules may move along river systems or other linear water features (Nagata, 1983;van Rees & Reed, 2015). The potential use of linear water features as dispersal corridors by Hawaiian gallinules coincides with observations in other tropical birds (e.g., Gillies & St. Clair, 2008;Sekercioglu, 2009;Takano & Haig, 2004). The results of our post hoc tests lend additional credence to our findings, but a biological, mechanistic explanation for the phenomenon is also important to consider.
One potential mechanism may be a "landscape of fear" (Laundré, Hernández, & Ripple, 2010); in this case where Hawaiian gallinules perceive lower predation risk near water features and accordingly are more willing to travel along them. Reduced antipredator behavior near water features has been documented in other rail taxa (Dear, Guay, Robinson, & Weston, 2015), and rails and other waterbirds tend to flee toward water as part of their normal predator escape behavior (Lima, 1993). Perceiving lower risk due to ease of escape, Hawaiian gallinules may accordingly experience fewer behavioral barriers to movement when closer to water features (sensu Harris & Reed, 2002).
The predictions of our Proximity-to-Water models make intuitive sense based on expert opinion and limited observations of Hawaiian gallinule dispersal behavior. Visual inspection of least-cost paths developed using our 100-m corridors, Proximity-to-Water surface, and the least-cost-path function in ArcGIS (Figures 4 and 5 (Figure 4) involves traveling along the coastline and parallel to the Ko`olau mountains, rather than over the mountains, and the least-cost-path between Kawainui marsh and the Olomana golf links makes use of extensive drainage infrastructure and nearby streams, avoiding direct passage through urban areas ( Figure 5). Thus, the Proximity-to-Water layers implicitly feature other aspects of gallinule movement ecology observed anecdotally, specifically no observations at high elevation and a high susceptibility to road mortality in urban environments.
Because model rankings are mixed between different Proximityto-Water models across different criteria of model selection, we cautiously refrain from selecting one of those models as being the best supported overall. Given the limited genetic variation exhibited by the subspecies (van Rees, , and impacts of a recent population bottleneck on genetic diversity within Hawaiian gallinules (Sonsthagen, Wilson, & Underwood, 2017), it could be that our current sample is insufficient to distinguish between functions relating Proximity-to-Water to resistance values. Additional factors affecting dispersal in other taxa (e.g., conspecific attraction ;Smith & Peacock, 1990;Serrano & Tella, 2003) may also influence dispersal rates in Hawaiian gallinules, but were not explored in this study.
While genetic structure was detected at small spatial scales, addition of whole genomic or reduced representation genomic data may provide greater spatial resolution and increase our ability to detect F I G U R E 4 Approximation of least-cost pathway between James Campbell National Wildlife Refuge and Keawawa Wetland, calculated using the 100-m corridor distance-to-water resistance surface and the least-cost path tool in ArcGIS F I G U R E 5 Approximation of least-cost pathway between Kawainui Marsh and Olomana Golf Links, calculated using the 100-m corridor distance-to-water resistance surface and the least-cost path tool in ArcGIS. For illustrative purposes, the path has been projected over a modified version of the NOAA C-CAP 2011 map of O`ahu, showing urban areas in white and undeveloped areas in dark gray, with water features in medium gray and open water in light gray landscape features that are influencing gallinule movement patterns (Kozakiewicz et al., 2017;Szulkin, Gagnaire, Bierne, & Charmantier, 2016 both contained water during the kills." Also of interest is that kills occurred at night, supporting other observations that common gallinules and common moorhen (G. chloropus) perform higher-altitude flights at night (Roselaar, 1980;Taylor, 2010), which may explain why long flights are rarely observed on O`ahu.
Although our analysis yielded a strong and consistent signal that water features decrease landscape resistance to movement for Hawaiian gallinules on O`ahu, several important limitations to our study are worth noting. First, although genetic data are a useful descriptor of overall gene flow as a result of dispersal, they ultimately represent only a portion of animal movements , specifically movements that lead to interwetland dispersal and successful breeding (Cushman, Lewis, & Landguth, 2014). Accordingly, although they match limited observations in related taxa (Takano & Haig, 2004), and anecdotal observations in this subspecies (Nagata, 1983), results from this study do not necessarily reflect the actual

| Conservation Implications and Priorities for Future Work
Empirical studies on the movement behavior of Hawaiian gallinules will be important to validate the results of our landscape genetic analysis and investigate the fine-scale behavioral decisions that lead to the observed population-level patterns of gene flow. More consistent mark-resighting efforts (Dibben-Young, 2010) and study methods with higher spatiotemporal resolution will be important steps for improving our knowledge of movement behavior in this subspecies. GPS dataloggers and transmitters have been used to great effect in tracking a number of bird species (Gagliardo, Ioalè, Savini, Lipp, & Dell'Omo, 2007;Rodríguez et al., 2012), and translocation studies like that of Gillies and St. Clair (2008) allow for experimental manipulation of dispersal direction and matrix type.
An important implication of this study is that habitats formerly hypothesized to have little value to Hawaiian gallinules (e.g., drainage ditches and canals, forested and vegetated streams, and roadside swales) may actually affect their population persistence by increasing population connectivity. The management dependency of Hawaiian gallinules (Reed et al., 2012a,b) has led to the assumption that unmanaged wetlands and riparian systems are of little to no value for them. This study suggests that such unmanaged water features act as corridors, which may increase population persistence in fragmented landscapes by alleviating problematic consequences of isolation (Gilbert-Norton, Wilson, Stevens, & Beard, 2010). As van Rees and Reed (2015) speculated, shifts in water management toward a greater emphasis on green stormwater infrastructure might simultaneously provide conservation benefits by creating such corridors for waterbirds like the Hawaiian gallinule. Such landscape changes would represent a gain for both the management of imperiled water resources (Giambelluca, 1986;Ridgley & Giambelluca, 1991) and threatened wildlife on O`ahu.

ACK N OWLED G M ENTS
Funding for this project was provided by the Tufts Institute of the

DATA ACCE SS I B I LIT Y
Data from this study (matrices of genetic distances, resistance surfaces in analysis, and effective distance matrices) available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.p90b87p.