Seasonal variation in habitat selection for a Neotropical migratory songbird using high‐resolution GPS tracking

Deciphering how environmental heterogeneity affects population dynamics in migratory species is complicated by the redistribution of individuals in time and space across the annual cycle. Approaches that tackle this problem require information about how migratory species respond to ecological factors across time and space, and how they are linked across migratory periods. Using high spatial resolution (10 m) GPS tracking of individual male songbirds, we quantified for the first time (1) localand landscape-scale habitat selection across the annual cycle and (2) patterns indicative of regional habitat selection for individuals within five populations of wood thrush (Hylocichla mustelina) throughout their breeding range. Wood thrush exhibited seasonal variation in localand landscape-scale habitat selection. Within stationary periods, wood thrush occupied forested habitats and proximity to forest edge was an important predictor of habitat selection at the local scale. In contrast, during migratory periods wood thrush exhibited greater behavioral flexibility indicative of a more generalist approach to habitat selection. Landscape habitat selection was only identified during the breeding season (average patch size) and could be a response to the extensive forest fragmentation in the North American breeding grounds. We also identified individual population distribution patterns indicative of regional habitat selection during fall migration and winter period, but not spring migration. Seasonal changes in habitat selection at multiple spatial scales suggest the factors driving habitat selection patterns are aligned with life-history stage and may be dependent on regional differences in landscape composition. These results highlight the importance of a full annual cycle approach to ecological studies that address how migratory species respond to spatial and temporal environmental heterogeneity.

Abstract. Deciphering how environmental heterogeneity affects population dynamics in migratory species is complicated by the redistribution of individuals in time and space across the annual cycle. Approaches that tackle this problem require information about how migratory species respond to ecological factors across time and space, and how they are linked across migratory periods. Using high spatial resolution (10 m) GPS tracking of individual male songbirds, we quantified for the first time (1) local-and landscape-scale habitat selection across the annual cycle and (2) patterns indicative of regional habitat selection for individuals within five populations of wood thrush (Hylocichla mustelina) throughout their breeding range. Wood thrush exhibited seasonal variation in local-and landscape-scale habitat selection. Within stationary periods, wood thrush occupied forested habitats and proximity to forest edge was an important predictor of habitat selection at the local scale. In contrast, during migratory periods wood thrush exhibited greater behavioral flexibility indicative of a more generalist approach to habitat selection. Landscape habitat selection was only identified during the breeding season (average patch size) and could be a response to the extensive forest fragmentation in the North American breeding grounds. We also identified individual population distribution patterns indicative of regional habitat selection during fall migration and winter period, but not spring migration. Seasonal changes in habitat selection at multiple spatial scales suggest the factors driving habitat selection patterns are aligned with life-history stage and may be dependent on regional differences in landscape composition. These results highlight the importance of a full annual cycle approach to ecological studies that address how migratory species respond to spatial and temporal environmental heterogeneity.

INTRODUCTION
The annual cycle for a migratory species is characterized by long-distance movement of individuals between multiple geographically, and often ecologically, disparate locations (Newton 2008, Akesson andHansson 2014). In each location, habitat heterogeneity drives individual settlement decisions and habitat selection processes often have individual fitness consequences (Hutto 1985, Mayor et al. 2009). Migratory birds, despite their small size, travel some of the longest distances of any animal throughout their annual cycle (Alerstam et al. 2003). Although empirical studies of habitat selection have been conducted within breeding (Orians and Wittenberger 1991, Mitchell et al. 2001, Lee et al. 2002, migration (Buler et al. 2007, Mccabe and Olsen 2015, Lafleur et al. 2016, and winter (Chandler andKing 2011, Fraser et al. 2017) periods, few species have been assessed across multiple seasons. Moreover, studies rarely use the same individuals or populations (but see Beatty et al. 2014, Pickens et al. 2017. As such, we know relatively little about the breadth of suitable habitats and how individuals select habitats throughout the annual cycle for migratory birds , McGarigal et al. 2016. Adaptive habitat selection has been theorized as a hierarchical decision-making process by which the factors that drive habitat suitability (based on fitness costs and benefits) and the mechanisms for assessing suitability occur at different spatial and temporal scales across the annual cycle ( Fig. 1; Johnson 1980, Wiens 1989, Mayor et al. 2009). At each spatial or temporal scale, the decision-making process may involve novel and multiple interacting criteria (e.g., patch size, predation rate) and be constrained by the effects of criteria from higher scales (annual vs. diel; Rettie andMessier 2000, Mayor et al. 2009). When considering spatial scales, habitat selection decisions at the broadest scale will shape the species' geographic range and patterns of regional population distribution (hereafter referred to as "regional"; Johnson 1980, Meyer andThuiller 2006). Hutto (1985) refined this idea for migratory birds and suggested nonhabitat factors are expected to drive decisions at the regional scale. For example, an individual's genetically programmed and learned migration route (Hutto 1985), weather Aborn 2000, Buler andMoore 2011), and physiological condition (Studds et al. 2008, Rushing et al. 2015 are factors expected to play a role in regional habitat selection. In contrast, as the spatial scale decreases, Hutto (1985) suggested that habitat selection will increasingly be driven by the costs and benefits associated with the habitat itself. At intermediate scales (hereafter referred to as "landscape"), selection drives the choice of a particular habitat type and home range. Finally, at finer spatial scales, selection of habitats within the home range (hereafter referred to as "local") or microhabitat selection is most closely linked to the ultimate factors driving habitat suitability (e.g., food or nesting site availability, nest predation risk). Equally important to consider is how the decision-making process changes across temporal scales or levels (Cody 1985, Wiens 1989. For example, factors limiting an individual's fitness can vary between diel and seasonal decisions (scales) or equally across different seasons (levels; Mayor et al. 2009). Therefore, habitat selection may be a hierarchical process that operates not only across multiple spatial scales ranging from regional to microhabitat but also across temporal scales.
Over the last thirty years, the multi-scale approach to habitat selection has been incorporated into many ecological studies across a range of vertebrate taxa, largely in terms of spatial scales (Wiens and Milne 1989, Orians and Wittenberger 1991, Rettie and Messier 2000, Grand and Cushman 2003. However, the scope of this research has often been restricted to studies of single seasons on single populations, limiting the inferences that can be drawn (McGarigal et al. 2016). For migratory songbirds, most ecological studies have been conducted during the breeding season . When habitat use across seasons has been documented, many migratory songbirds, even those considered habitat specialists, appear to show shifts in habitat use during the nonbreeding season (Petit 2000, Zuckerberg et al. 2016. This lack of consistency in habitat associations suggests that individuals can maximize their fitness across seasons through behavioral flexibility in their habitat settlement decision making (i.e., nesting vs. refueling sites). However, documenting behavioral flexibility in habitat selection has remained challenging due to the difficulty in tracking small mobile animals across large spatial and temporal scales. Therefore, two large information gaps related to behavioral flexibility in habitat selection need to be addressed: (1) What factors drive habitat selection decisions across seasons for individual birds, and (2) what factors underlying habitat selection vary across the species range? v www.esajournals.org Fig. 1. Conceptual hierarchy of the decision-making process of habitat use by a migratory songbird. At higher spatial scales (A) and (B), the process is expected to be largely inflexible and constrained by extrinsic habitat factors. At lower spatial scales (C) and (D), the process likely involves the assessment of the intrinsic factors of the habitat based on cues and exploratory assessment. Different patterns of migratory connectivity are illustrated (A). Strong connectivity (solid lines) occurs when most individuals from one breeding population move to the same nonbreeding locations. Weak connectivity (dashed lines) occurs when individuals from the same breeding population move to different nonbreeding locations.
Here, we evaluated the spatial and temporal patterns of habitat selection at three spatial scales (local, landscape, and regional) and across all seasons (breeding, fall migration, winter, and spring migration) of the annual cycle for wood thrush (Hylocichla mustelina), a Neotropical migratory songbird using fine spatial resolution archival GPS geolocators deployed across five distinct breeding populations. First, to test the hypothesis that factors driving habitat suitability change across seasons, we explored individual variation in habitat selection at the local and landscape scales. Since wood thrush are characterized as forest species that are tolerant of forest fragmentation (Evans et al. 2011), we predicted that if habitat suitability remains constant across seasons, thrushes will consistently select local and landscape feature characteristics of forest habitats. In contrast, if the factors that drive habitat suitability exhibit seasonal variation, we predicted a shift to more generalist habitat selection during migratory periods as nonhabitat features (e.g., temporal constraints on migration, energetic demands) drive the decision-making process and the mechanisms for accurately assessing unfamiliar habitats become difficult for birds. In addition, we predicted that wood thrush will exhibit local-and landscape-scale habitat selection during stationary (breeding and winter) periods, but selection will shift to landscape-scale features during migration, as they provide a quick visual cue that can be used in flight to assess habitat quality (Buler et al. 2007, Beatty et al. 2014. Second, to test for seasonal changes in regional habitat selection among five breeding populations we assessed the strength of migratory connectivity. Migratory connectivity is most commonly used to describe the strength of regional redistribution of migratory individuals across the annual cycle, but here we used it to identify the presence of population-specific regional habitat selection decisions as individuals move large distances throughout the year. Based on earlier work on wood thrush migratory connectivity with low-resolution tracking devices (Stanley et al. 2015), we predicted the strength of regional habitat selection would vary seasonally. Specifically, we predict regional habitat selection during fall migration driven by population-specific migration routes (Stanley et al. 2015). Wood thrush show moderate connectivity from breeding to winter sites (Stanley et al. 2015), suggesting weaker patterns of regional habitat selection on winter sites. Finally, we predict no regional habitat selection during spring migration due to the convergence of migratory routes at the Gulf of Mexico.

Study system
Wood thrush are medium-sized, long-distance Neotropical migratory songbirds that breed in deciduous and mixed forests of eastern North America and winter in the broadleaf and palm forests ranging from southern Mexico to northern Panama (Evans et al. 2011). They are a multibrooded, omnivorous species that primarily forages on the ground. Wood thrush maintain territories during the winter and breeding seasons; however, they show extensive flexibility in territorial behavior. Tracking studies have found they engage in regular off-territory forays and can make large-scale relocations (permanent long-distant movements, 1-25 km), throughout stationary periods (Rappole et al. 1989, Lang et al. 2002. Their global population size has declined by 60% since 1966 based on breeding ground surveys (Sauer et al. 2013).

Field methods and GPS telemetry
To determine spatiotemporal variation in habitat selection across the annual cycle, we studied wood thrush from 2014 to 2015. Archival GPS tags (Model PinPoint-50, 1.8 g, 50 fixes, Lotek Wireless) were deployed on 137 breeding adult male wood thrush over the summers of 2014 and 2015 across five populations in Delaware, Indiana, North Carolina, New York, and Minnesota (Appendix S1: Table S1). The tags provide location estimates for individual birds with an estimated accuracy of 10 m. Tags were recovered by returning to sites the following summer and recapturing individuals. We recovered 23 tags and retrieved data from 21 tags (Appendix S1: Table S1). Two of the 23 tags were unresponsive (no data retrieved); nine of the 23 lost antennas sometime after release (partial data recovery), and 12 of the 23 were retrieved with antennas intact. We obtained an average of 26 out of a maximum of 50 points per tag (range 1-46; v www.esajournals.org Appendix S1). Most missing fixes, excluding those involving lost antennas, occurred during migratory periods (Appendix S1: Table S2). Due to low success of location fixes during migratory periods, annual cycle stages (breeding, fall migration, winter, spring migration) were defined based on the first or last location recorded at stationary stages (e.g., spring migration ends the day before first recorded point on breeding grounds). Migratory periods were identified as northward or southward movements of at least 20 km over consecutive location fixes (>2 points, 6-8 d).

Habitat features
To quantify the structural characteristics of habitats occupied by wood thrush across the annual cycle, we paired location fixes to geospatial habitat data. Location estimates spanned eastern North America and Central America; therefore, we restricted geospatial data to those sources that covered the entire study area (detailed description of satellite imagery in Appendix S1: Table S3). At each location fix, we examined structural characteristics of the habitat used and available to wood thrush at two spatial scales (local and landscape) using the extract function in the R package raster (Hijmans 2016). Local habitat features represented remotely sensed data interpolated at each point location and included percent tree cover, Enhanced Vegetation Index (EVI, index of primary productivity), patch size (ha, connected pixels of >30% tree cover), and proximity to non-forest (m, defined as <30% tree cover; detailed description of interpolation methods in Appendix S1: Table S3). Landscape characteristics represented remotely sensed data interpolated from a buffer of 1 km radius around each point location and included average percent tree cover and average patch size (ha). A 1-km buffer was used because off-territory forays of wood thrush typically range from 150 m to 1 km (Rappole et al. 1989, Lang et al. 2002. To assess landscape composition, we interpolated data from classified land cover layers using a 5-km buffer because classified land cover layers suitable for a 1-km buffer were unavailable across our study area (Appendix S1: Table S3). Landscape composition was therefore analyzed separately from the other landscape characteristics.

Local and landscape habitat selection
To identify variation in spatiotemporal patterns of habitat selection at the individual level, we explored three measures. First, at the landscape scale, we examined the consistency of land cover associations for individual birds across seasons. Second, we quantified the repeatability of use of habitat features by individuals at both the landscape and local scales across the stationary (breeding and winter) and migratory (fall and spring migration) periods. Finally, to disentangle whether seasonal differences in habitat usage were driven by changes in habitat availability across regions or shifts in factors driving habitat selection, we employed a multi-scale approach to determine habitat selection within a use-availability design.
First, to characterize consistency in land cover associations we utilized the Shannon diversity index (Shannon 1948) to estimate seasonal diversity in landscape composition for each individual. The index was assessed each season based on landscape composition (the number of classified land cover classes) at each point location used by an individual with the R package vegan (Oksanen et al. 2017). High landscape diversity indicates that individuals showed plasticity in land cover associations and suggests more generalist habitat selection within a season.
Second, for each habitat feature we calculated the repeatability of use among individual birds between stationary and migratory periods. Repeatability quantifies the proportion of between-individual variation from observations relative to the total variation in the population for a repeated measure (Lessells and Boag 1987). To calculate repeatability, we used the r package rpt (Stoffel et al. 2017) based on the structure of the best-fit model describing each habitat feature (see Appendix S2).
Third, we constructed discrete choice models for each season to determine seasonal habitat selection preferences of wood thrush. These models examine the probability of an individual choosing a location based on a choice set of alternative available locations (Cooper and Millspaugh 1999). If habitat features are used disproportionate to their availability, it is assumed they confer a fitness advantage (Manly et al. 2004). If these habitat features change across seasons, we conclude this represents v www.esajournals.org behavioral flexibility in habitat preferences. GPS location fixes were considered as used resource units. Available resource units were generated by creating a buffer around each used resource unit and generating up to 20 random points (detailed methods in Appendices S1, S5). Therefore, each used resource unit had its own set of available units. For all used and available points, local and landscape habitat features were extracted as defined in the habitat features section.
A multi-scale approach was used to identify habitat selection patterns that would be ecologically relevant to wood thrush (Wiens 1989, Buler et al. 2007). Since relevant spatial extents were unknown for wood thrush (size of relevant study area), we delineated four spatial extents encompassing distances moved within an individual's territory (0.5 km), during off-territory forays (5 km), and during relocation events (15 and 25 km) based on prior information on wood thrush spatial ecology (Rappole et al. 1989, Lang et al. 2002. Off-territory forays have typically been recorded as covering <1 km; however, more recent tracking has found individuals can travel longer distances (Stanley 2019); therefore, we chose a larger spatial extent to account for these movements. The spatial extent defined the size of the buffer created around each used resource point and served to generate the available points for each used point. The Bayesian mixed conditional logistic regressions were used to model the probability of a wood thrush choosing a used resource unit out of the available choice set. The mixed conditional logistic regression was adapted from the methods of Beatty et al. (2014), and a Bayesian framework was employed because comparable mixed conditional logistic regressions were not available in existing maximum-likelihood R packages. We developed a set of candidate models to test alternative hypotheses about the habitat features influencing habitat selection (Appendix S1: Table S4). All models, with the exception of the null model, included individual bird as a random variable to account for variation in selection patterns among birds. At the 0.5-and 5-km spatial extent, we evaluated two models (local and null); and at the relocation level (15 and 25 km), we evaluated four models (local, landscape, full, and null). These models were run for each season for a total of 48 models (12 models 9 4 seasons).

Regional habitat selection
To identify regional habitat selection, we estimated migratory connectivity between (1) breeding and wintering grounds and (2) breeding and northern Gulf of Mexico stopover sites. We employed the MC metric from the MigConnectivity package (Cohen et al. 2018) using orthodromic distances between breeding and winter or stopover sites (detailed methods in Appendix S1). We also described populationspecific patterns of local-and landscape-scale habitat features (characteristics of used habitats) to determine whether patterns of regional habitat selection corresponded to population-specific patterns of habitat use (see Appendix S4). To examine the consistency of annual schedules between breeding populations, we determined the timing of annual events for all individuals (see Appendix S3).

Statistical analysis
All analyses were conducted in R 3.4.3 (R Core Team 2017). The Bayesian mixed conditional logistic regression models were run with the R package jagsUI (Kellner 2017) using the software JAGS 4.3.0 (Appendix S1, S5; Plummer 2003). Generalized mixed-effects models were run in the R package nlme (Pinheiro et al. 2017) with restricted maximum-likelihood estimation. Generalized linear models were run in the R package stats (R Core Team 2017). For full model specifications and fit assessment, see Appendix S1.

Patterns of movement
We obtained breeding ground location estimates from 21 birds, fall migration locations from 20 birds, winter locations from 19 birds (1 partial), and spring migration from 18 birds for a total of 554 GPS location fixes across 21 birds ( Fig. 2; Appendix S1: Table S2). During the winter period, of the 18 birds tracked across the whole period, eight individuals engaged in intra-winter movements occupying two or more territories an average of 59 AE 21 km (AESE, range 1.6-180 km) apart (Fig. 2C). The remaining 10 individuals remained at a single winter territory, moving an average of 110 AE 26 m (AESE) between locations from November to April.

Local and landscape habitat selection
The diversity of land cover types occupied by individual wood thrush varied by season (v 2 = 45.9, df = 3, P < 0.001; Fig. 3). We detected no difference in land cover diversity within stationary or migratory periods (breeding-winter Z = À0.26, P = 1.0), but we did detect a difference in land cover diversity between the migratory and stationary periods (breeding-fall Z = 5.8, P < 0.001; breeding-spring Z = À3.5, P = 0.003; winter-fall Z = 5.5, P < 0.001, winterspring Z = 3.2, P = 0.08). Within the breeding season, the major land cover type used was a matrix of mixed-use agriculture and natural vegetation, while the dominant land cover type on the winter grounds was forest (Fig. 3). The high diversity of land cover types selected by wood thrush during fall and spring migration included forest, agriculture, agricultural mosaics, wetlands, and wooded savannahs.
The repeatability of habitat features as birds moved within the migratory periods was very low, close to zero. Within the stationary periods, repeatability of local and landscape tree cover, proximity to non-forest, and average patch size was moderate (R > 0.3; range, R = 0.33-0.48, Table 1). EVI and patch size had low repeatability (R < 0.3) during stationary periods. v www.esajournals.org Discrete choice models indicated that the habitat features of the resource units (locations) wood thrush selected (compared with what was available) varied across seasons and spatial extent. The top discrete choice model of habitat selection for most spatial extents during the breeding, winter, and fall migratory period was the local habitat feature model ( Fig. 4; Appendix S1: Table S5). The exception was during the breeding season at the 25-km spatial extent, in which the full model (local and landscape habitat features) was the top model. During spring migration, the null model was the top model across all scales, indicating no habitat features measured were important for habitat selection.
On the breeding grounds at all spatial extents, birds selected resource units with high percent tree cover (Fig. 4). At the smallest spatial extent (0.5 km), birds chose resource units with low EVI values. At the two largest spatial extents (15 and 25 km), wood thrush selected resource units proximate to non-forest. Additionally, on the breeding grounds, at the 25-km extent birds selected resource units that were found within  landscapes with high average patch size (1-km buffer), but at the local scale, resource units in small forest patches were selected. During fall migration, birds selected resource units with large patch size at the 0.5-, 5-, and 25-km spatial extent and resource units with high EVI values at the larger spatial extents (15 and 25 km). On wintering grounds, at all spatial extents, birds chose resource units with large patch sizes; at the 5-, 15-, and 25-km spatial extents, birds chose resource units proximate to non-forest; and at the 25-km extent, they chose resource units with high tree cover.

Regional habitat selection
The migratory connectivity estimate between breeding and wintering grounds was weak-tomoderate (MC Index = 0.41), indicating regional habitat selection by breeding populations on the wintering grounds. Regional habitat selection was also present at fall stopover locations prior to the Gulf of Mexico crossing based on a high migratory connectivity estimate (MC Index = 0.69) due to breeding populations utilizing region-specific departure points from North America (Fig. 2B). By contrast, during spring migration, connectivity was lower (MC Index = 0.28) at stopovers after the Gulf of Mexico crossing, suggesting no regional habitat selection by breeding populations. As predicted based on our analysis of regional habitat selection, we found evidence of populationspecific differences in habitat use patterns during the stationary periods and fall migration, but not spring migration (Appendix S4). Finally, we found evidence that timing of both spring and fall migration was influenced by breeding location, but not winter location (Appendix S3).

DISCUSSION
For the first time in a migratory songbird, we show direct evidence, from individual birds with high-resolution tracking devices, of seasonal shifts in habitat selection across the annual cycle. Most notably, individual birds relaxed the degree to which they selected habitat features during migratory periods, suggesting greater behavioral plasticity during migration. We also show evidence of regional habitat selection by using Fig. 4. Parameter estimates and associated 95% credible intervals for the top discrete choice model examining habitat selection patterns of individually tagged wood thrush at multiple spatial scale during (A) breeding, (B) winter, (C) fall migration, and (d) spring migration. For each period, discrete choice models were run at multiple extents: 0.5 km (closed square), 5 km (closed circle), 15 km (open square), and 25 km (closed square). Local habitat features include percent tree cover, Enhanced Vegetation Index (EVI), patch size (km 2 ), and proximity to non-forest. Landscape habitat features include average percent tree cover and average patch size (km 2 ). Not all habitat features were present in all top models.
indices of migratory connectivity to assess nonrandom assortment of five populations across regional geographic scales. As migratory species move through different heterogeneous landscapes, they will need to find suitable habitats that meet their needs during different life-history stages. Understanding the factors necessary for persistence will require identifying the breadth and quality of habitats that individuals occupy as they redistribute geographically across the annual cycle (Webster and Marra 2005).

Seasonal variation in habitat selection
Fine-scale tracking of individual wood thrush revealed variation in habitat selection across periods of the annual cycle. By employing a novel multi-scale approach, we determined wood thrush were selective at the local scale across the annual cycle and only selective at the landscape scale during the breeding season. The factors that drove habitat selection varied across seasons and spatial extent (size of study area). Habitat selection by migrating wood thrush was less selective, individuals occupied a greater diversity of land cover types than during the stationary periods, and selection focused on local habitat features.
During the winter and breeding periods, we found individuals almost exclusively in forested and mixed forest-agriculture landscape. Alternatively, while migrating, wood thrush selected novel land cover types (e.g., agricultural, wood savannah) 36-46% of the time. Furthermore, during migration the low repeatability in habitat use indicated that individual birds did not consistently select locations with similar structural attributes. Banding and eBird studies have also suggested greater flexibility in habitat selection for songbirds during the migratory period and changes in foraging behavior, particularly for fruit-eating species such as wood thrush (Yong andMoore 2005, Zuckerberg et al. 2016). Analyses were not conducted on the same individuals or population; however, this study confirms that changes in habitat use during the migration period were due to changes in individual habitat selection across periods. Specifically, unlike the stationary periods, we found fewer predictors of habitat selection during fall migration and no response to habitat features during spring migration, which suggests birds were selecting habitats at random with regard to the habitat features we measured during spring migration.
This shift to be more of a generalist when selecting habitat may be favored due to the strong selection pressure to minimize time on migration (Hutto 1985, Alerstam and Lindstr€ om 1990, G omez et al. 2017. This is expected to be particularly relevant during spring migration due to intense selective pressure for early arrival on breeding grounds to secure a territory and begin reproduction (Kokko 1999). Migration schedules should therefore ensure optimal arrival dates on breeding grounds, which is consistent with our finding of faster spring migrations and population-specific schedules (Appendix S3: Fig. S1, Table S2). Increased behavioral flexibility during migration may also be an adaptive response to lower migrants' threshold of acceptable habitat to reduce search time for high-quality stopover sites (Moore and Aborn 2000). Therefore, if individuals cannot quickly locate high-quality habitat they may settle in poorer quality habitats. It has been suggested that compensatory behaviors such as risk-prone foraging strategies (e.g., increased maneuvers or feeding intensity) or shorter stopover duration may allow birds to maintain migration schedules even when foraging in poorer quality habitats (Yong andMoore 2005, Nilsson et al. 2013). Therefore, during spring migration, at fine spatial and temporal scales food may still be limiting individuals, but at larger spatial scales the timing of migration may be the more limiting factor.
During fall migration, wood thrush selected larger forest fragments and, when examined across large spatial extents (15, 25 km), patches with higher EVI values. In contrast to spring migration, birds responded to local habitat features, but to a lesser extent compared with the stationary periods. We found fall migration schedules were more variable across individuals and populations and longer in duration compared with spring migration (Appendix S3: Fig. S1, Table S2), which suggests individuals may fine-tune fall migration to adjust to local conditions en route (Balbont ın et al. 2009, Stanley et al. 2012. Settlement patterns during migrations are thought to be driven largely by food to meet the energetic demands of migration, and this is consistent with the response of birds to EVI, an indicator of primary productivity (Hutto v www.esajournals.org 1985, Buler et al. 2007). Behavioral plasticity between migratory periods could be driven by seasonal difference in environmental conditions linked to food availability. For example, poor environmental conditions at fall stopover sites or poor physical condition carrying over from the breeding/molting period could drive increased food limitations in the fall and lead to higher selective pressure on habitat suitability (Nilsson et al. 2013). In contrast to spring migration, the higher variability in fall migration schedules also suggests timing of arrival on wintering grounds may not be associated with fitness advantages, which could drive differences in behavioral plasticity between seasons.
Landscape-scale habitat selection was only identified during the breeding season and only across the largest spatial extent. The majority of previous studies have also found that local, and not landscape, habitat features better predicted wood thrush occupancy during both stationary periods (winter, Graham and Blake [2001]; breeding, Lee et al. 2002, Valente andBetts [2018] but see Fauth et al. [2000]). Therefore, the relevance of landscape variables during only the breeding season could be driven by several factors. It could be that critical resources unique to the breeding season (e.g., reproduction-related) are influenced at the landscape scale (Hutto 1985). For example, previous research has found increased nesting success in landscape with higher core forest area and increased rates of nest parasitism and predation in landscapes with higher proportion of developed land cover (Driscoll et al. 2005, Lloyd et al. 2005. Alternatively, it has been suggested that the impacts of fragmentation are primarily visible when habitat loss across the landscape is high or moderate (fragmentation threshold; Andr en 1994, Villard and Metzger 2014). The dominant land cover during the breeding season was mixed forest-agriculture, and there was low forest cover across the landscape (Fig. 4, Appendix S2: Table S1). In comparison, during the winter period, forest was the dominant land cover and forest cover at the landscape scale was high. Therefore, the higher fragmentation in locations occupied by wood thrush during the breeding season could be driving the greater importance of landscape configuration (average patch size) in our models.
Overall, three local habitat features-tree cover, patch size, and proximity to non-forestwere top predictors of habitat use during both stationary periods; however, their relevance varied by spatial extent and season. Wood thrush showed a preference for high tree cover and edge habitats during both stationary periods, which has previously been reported for wood thrush (breeding, Vega Rivera et al. [1998], Kaiser and Lindell [2007]; winter, Graham and Blake [2001], Roberts [2011]). Selection of patch size showed contrasting patterns between stationary periods; wood thrush were more likely to settle in smaller forest patches during the breeding season and larger forest patches during the winter season. It is well documented that wood thrush inhabit small forest patches across most of their breeding range (Robinson et al. 1995, Burke and Nol 2000, Lee et al. 2002, but occupancy of small forest fragments is usually (but see Weinberg andRoth 1998, Fauth 2001) associated with decreased reproductive success for wood thrush (Donovan et al. 1995, Burke andNol 2000). Together, these results indicate that factors related to forest structure play a role in habitat selection during the stationary periods. The observed differences in patch size across stationary periods could indicate shifts in the importance of this habitat feature. It could also represent trade-offs in preferred habitat features imposed by the hierarchical process of habitat selection in regions with different landscape composition and configuration (Hutto 1985).

Regional habitat selection
Due to the hierarchical nature of habitat selection by birds, environmental heterogeneity across a species range will influence which habitat features are limiting populations based on their regional context. Using migratory connectivity, we identified nonrandom assortment of populations at regional scales between breeding and nonbreeding locations, which we interpreted as evidence of regional habitat selection. As predicted based on previous low-resolution tracking studies (Stanley et al. 2015), wood thrush exhibited regional habitat selection on their winter (MC = 0.41) and fall Gulf of Mexico stopover sites (MC = 0.69), but not on their spring Gulf of Mexico stopover sites (MC = 0.28). Weak and v www.esajournals.org moderate levels of migratory connectivity appear to be common in forest-dwelling migratory songbirds (migration, Kole cek et al. [2016], Knight et al. [2018]; stationary, Gilroy et al. [2016], Finch et al. [2017]), which suggests regional habitat selection occurs across many species. Using migratory connectivity to identify patterns of regional habitat selection does not elucidate the factors driving habitat selection at this scale. Hutto (1985) suggested they are largely factors unrelated to the habitat itself, for example, physiological condition influencing regional settlement patterns on the ground (Studds et al. 2008, Rushing et al. 2015. Ultimately identifying what factors drive patterns of regional habitat selection and how flexible they are will have important implications for understanding how populations will be able to respond to changing landscapes. If regional habitat selection does occur, we could expect to see population-specific patterns of habitat selection at finer spatial scales. Due to limited sample size, we were unable to perform these analyses. However, we did describe population-specific trends in the use of local-and landscape-scale habitat features (see Appendix S4). As predicted based on our analysis of regional habitat selection, we identified differences in habitat use patterns during the stationary periods and fall migration, but not spring migration. Due to limited sample size, this analysis does not determine whether the observed patterns were driven by populations differentially selecting locations based on these structural aspects (disproportionate to their ability) or by differences in the distribution of environmental features across the species range. However, one structural feature, distance to edge, had the same regional pattern across both stationary periods; mid-Western populations selected locations further from forest edge. In addition, this structural characteristic had high individual repeatability across both stationary periods. These results hint at the possibility that mid-Western populations may have a preference for more interior forest areas. If populations develop unique habitat selection preferences, either learned or innate, ignoring information about regional habitat selection across the annual cycle could obstruct interpretation of habitat selection patterns at lower spatial scale and our understanding of the conservation value of different habitat features across populations.

Conservation implications
Environmental degradation is causing alterations of the landscapes where most species live. For migratory species that are composed of populations occupying ranges that span continents, deciphering when and where alterations are most impacting species requires considering their full annual cycle ecology (Wilcove andWikelski 2008, Marra et al. 2015). Here, we found different habitat preferences across periods of the annual cycle, suggesting wood thrush populations will respond differently to changes in habitat composition and configuration during different periods of the annual cycle. For example, relaxation of habitat preferences during migration may be an adaptive trait to facilitate a fast migration, but across rapidly changing landscapes, it could lead to birds occupying unsuitable stopover habitats. Recent work has found high use of human-dominated landscapes by migrating songbirds, but whether these represent suitable habitat or ecological traps is not known (La Sorte et al. 2014, Zuckerberg et al. 2016. For example, anthropogenic light pollution has been found to attract nocturnally migrating birds to urban areas (La Sorte et al. 2017, McLaren et al. 2018) and individual tracking data could provide important insight into how artificial light affects the stopover process (Bowlin et al. 2015). Developing a better understanding of the vulnerability of migrating wood thrush to landscape changes may be particularly important at spring Gulf of Mexico stopover sites where the entire population of wood thrush appear to spend time during spring migration. Unfortunately, our knowledge of how species respond to environmental heterogeneity in the nonbreeding season is lacking, even for well-studied species such as wood thrush (Faaborg et al. 2010). As technologies improve, the increased temporal and spatial resolution of GPS tags will allow researchers to investigate not only how ecological responses of populations and individuals change across the annual cycle but also the factors driving these variations within and across populations and individuals.

ACKNOWLEDGMENTS
The funding for this study came from the Strategic Environmental Research & Development Program v www.esajournals.org