Autumn stopover hotspots and multiscale habitat associations of migratory landbirds in the eastern United States

Significance

Understanding the en route habitat requirements of migratory birds is critical for conservation but difficult to know at a large scale. We mapped stopover density of landbirds during autumn migration for the eastern United States using radar data. At a coarse scale, we found that birds migrate across a relatively broad front, underscoring the importance of widespread, locally based conservation efforts. At finer scales, we identified stopover hotspots that consistently support high densities of migrants. We demonstrate that forests provide the most important habitats for autumn migrants and that deciduous forest fragments in heavily deforested regions support especially high densities of migrants. We also present evidence that the now-agriculture-dominated Midwest constitutes an inland migration barrier for forest birds.

Abstract

Halting the global decline of migratory birds requires a better understanding of migration ecology. Stopover sites are a crucial yet understudied aspect of bird conservation, mostly due to challenges associated with understanding broad-scale patterns of transient habitat use. Here, we use a national network of weather radar stations to identify stopover hotspots and assess multiscale habitat associations of migratory landbirds across the eastern United States during autumn migration. We mapped seasonal bird densities over 5 y (2015 to 2019) from 60 radar stations covering 63.2 million hectares. At a coarse scale, we found that landbirds migrate across a broad front with small differences in migrant density between radar domains. However, relatively more birds concentrate along the Mississippi River and Appalachian Mountains. At a finer scale, we identified radar pixels that consistently harbored high densities of migrants for all 5 y, which we classify as stopover hotspots. Hotspot probability increased with percent cover of all forest types and decreased with percent cover of pasture and cultivated crops. Moreover, we found strong concentrating effects of deciduous forest patches within deforested regions. We also found that the prairie biome in the Midwest (now mostly cropland) is likely a migration barrier, with large concentrations of migrants at the prairie–forest boundary after crossing the agricultural Midwest. Overall, the broad-front migration pattern highlights the importance of locally based conservation efforts to protect stopover habitats. Such efforts should target forests, especially deciduous forests in highly altered landscapes. These findings demonstrate the value of multiscale habitat assessments for the conservation of migratory landbirds.

Billions of birds migrate every year, and in the process of doing so, they provide important ecosystem functions with respect to energy flow, nutrient cycling, and plant propagule dispersal (1). However, populations of many migratory species are in steep decline due to habitat loss, overexploitation, and climate change (2, 3). Long-term surveys reveal a staggering 28.3% decline in migratory bird populations in North America since 1970 (4). Some of the steepest population declines among North American birds involve landbirds, primarily passerines, moving through the eastern half of the continent (4). Such declines underscore the need for a better understanding of the habitat requirements of migratory birds.

However, the conservation of migratory landbirds is especially challenging due to the vast spatial extent of their migrations coupled with a lack of understanding of the factors limiting their populations across the entire annual cycle (5, 6). Historically, research has been biased toward studying birds where they are resident for long stretches of time, such as their summer breeding grounds and, increasingly, their winter nonbreeding grounds. The migration period itself remains poorly studied (7), despite the fact that the greatest annual mortality for migratory birds can occur during migration (8, 9). Given that habitat loss and degradation are likely the major threats to migratory birds (10), identifying and protecting key habitats along migratory pathways are crucial to the conservation of these birds (6, 11, 12).

While many migratory waders and waterfowl use well-defined networks of discrete stopover sites and show a high degree of philopatry (13), most migratory landbirds, especially passerines, are thought to migrate across a broad front, taking advantage of multiple stopover opportunities and exhibiting lower interannual fidelity to particular sites (14, 15, but see refs. 16 and 17 for counterexamples involving particular species). Beyond these local case studies, there is little empirical information on the large-scale migration patterns of migratory landbirds due to a lack of broad-scale surveys that match the highly dynamic spatial–temporal patterns of bird migrations. The resulting lack of information regarding migration patterns and habitat requirements en route has made migratory landbirds difficult to protect.

Recent advances in radar ornithology provide novel opportunities to study bird migration and stopover ecology at unprecedented spatiotemporal scales (18, 19). Unlike traditional field surveys conducted by people, radar “observes” birds through reflected electromagnetic radiation, thus comprehensively measuring the aggregate bird biomass in the atmosphere at relatively fine scales across broad extents and with less observer bias. Although individual species cannot be identified from radar signals, the comprehensive spatial and temporal coverage of radar provides density estimates of all migratory birds collectively, thereby generating results that contribute to the conservation of entire migratory bird populations (20, 21).

Radar images have been used to map the stopover density of migratory landbirds near major geographic barriers at coastal regions, including the northeastern United States (18, 20), the Great Lakes (22, 23), and the Gulf of Mexico (24, 25). These regional studies highlight local stopover hotspots with high densities of migrants and reveal positive associations between migrant density and deciduous forest cover. However, broad-scale stopover habitat mapping across the vast interior of the continent away from large open-water barriers has not been done. Although landbirds may have more stopover opportunities within the large tracks of vegetation of the interior, they are not expected to be equally abundant in all locations (26). For instance, forest habitat is limited in the agricultural Midwest due to large-scale deforestation accompanying Euro-American settlement, which could constrain bird migrants throughout this region (27, 28).

Here, we used data from 60 weather surveillance radar stations (NEXRAD) in the eastern United States covering 8.9 million radar pixels (pixel area: 0.4 to 21.8 ha) to determine whether and where there are continental stopover hotspots for migratory landbirds. We defined hotspots as radar pixels showing consistently high seasonal densities of migrants during autumn migration. In addition, we evaluated the habitat and land cover characteristics important to migratory landbirds across multiple spatial scales. In order to assess multiscale habitat associations, we calculated proportional land cover at local (within radar pixel), landscape (5-km buffer), and regional (within radar domain) scales. We also divided our study area into three avifaunal biomes (eastern forest, northern forest, and prairie) to explore biome-specific habitat requirements. We tested whether, as hypothesized in the literature (e.g., ref. 29), autumn landbird migration moves along a broad front when analyzed at the scale of the entire eastern United States. We predicted that, even if landbirds migrate in a broad front across the continent, there would be areas measured at a finer scale that consistently support high densities of migratory landbirds (hotspots), as has been found in previous regional studies (e.g., ref. 20). We also predicted that such hotspots would be tied to intrinsic habitat features at multiple spatial scales (30) and might be clustered in forest patches in largely deforested landscapes (18). Last, we explored whether the habitat associations of migrants differ among the three different biomes included in our study area.

Results

Coarse-Scale Stopover Patterns across Radar Domains.

We mapped seasonal stopover densities across 60 radar domains (sampling areas of ~80 km radius around each radar station) from 2015 to 2019, covering an area of 63.2 million hectares, effectively sampling 24% of the land area of the entire eastern United States. Across radar domains (coarse scale), we found an overall homogenized pattern of average autumn stopover density (Fig. 1; yearly maps from 2015 to 2019 are provided in SI Appendix, Fig. S1), with the seasonal cumulative density of migrants falling below 200 cm²/ha/season for the majority of radar domains (SI Appendix, Fig. S2); this represents fewer than 180 birds/ha stopping over throughout the autumn migration (the conversion is provided in SI Appendix, Methods). The relatively few radar domains supporting higher seasonal migrant densities are along the Appalachian Mountains and the Mississippi River (Fig. 1). We found the highest migrant density (665 cm²/ha/season) at the KHTX radar site in Alabama, which also has restricted spatial coverage due to topography (Figs. 2 and 3).

Fig. 1.

Coarse-scale cumulative stopover density of landbirds (square centimeters per hectare) during autumn migration averaged across 2015 to 2019 for each radar domain.

Fig. 2.

Interannual consistency of stopover hotspot pixels in three avifaunal biomes. Orange line highlights the prairie–forest boundary. Pixels are colored based on the number of years that cumulative seasonal bird density fell within the top 10% (N = 8,866,713) of pixels. Pixels with 5 y of high density are considered hotspot pixels. Variation in sampling coverage across radar sites, shown by the noncircular shapes of some radar domains, is due to topographical beam blockage in different locations. Zoom-in panel for the KPAH radar near Paducah, Kentucky, highlights large forested areas around the confluence of the Ohio and Mississippi rivers as important stopover hotspots for landbirds.

Fig. 3.

Fine-scale cumulative stopover density (square centimeters per hectare) of landbirds during autumn migration averaged across 2015 to 2019 for each radar pixel. Zoom-in panel for the KPAH radar near Paducah, Kentucky, highlights large areas around the confluence of the Ohio and Mississippi rivers that harbor high densities of migrants.

Fine-Scale Stopover Patterns.

We identified 2.3% of the 8,866,713 pixels (0.4 to 21.8 ha) across all 60 radar domains as stopover hotspot pixels because they fell within the top 10% of autumn cumulative migrant density for all 5 y (Fig. 2). Covering an area of 1.53 million hectares, these are distinctive hotspots that consistently supported the highest densities of migratory birds (Fig. 3; yearly maps from 2015 to 2019 are provided in SI Appendix, Figs. S3–S7). In some cases, these hotspot pixels were clustered into larger units that served as concentration spots, such as the large expanses of forests along the Mississippi River and the forest stretches along the prairie–forest boundary in Indiana and Ohio (Fig. 2). The 90% quantile for determining hotspot pixels ranged from 318 to 364 cm²/ha across 5 y (344 ± 19 cm²/ha/season, N = 5), representing more than 300 birds/ha stopping over throughout the autumn migration (refer to SI Appendix, Methods for the conversion). The mean cumulative bird density of hotspot pixels (756 ± 336 cm²/ha/season, N = 202,176) was five times higher than that of the rest of the pixels (159 ± 113 cm²/ha/season, N = 8,664,537).

Multiscale Habitat Associations of Hotspot Pixels.

The single best logistic GAM (Fig. 4; N = 8,866,713, deviance explained = 29.0%, and diagnostic plots in SI Appendix, Fig. S8) identified a robust, multiscale dependency of deciduous forest cover in determining hotspot pixels (Fig. 4 A and B). The probability of a pixel being a hotspot increased with local-scale deciduous forest cover (Fig. 4B). Moreover, the surrounding landscape and regional deciduous forest cover percentages strongly affected that probability: Birds were concentrated in landscapes showing high percentages of deciduous forest cover in otherwise heavily deforested regions (Fig. 4A). We also found a weaker peak in areas with high regional but intermediate landscape forest cover, although with a relatively smaller sample size (SI Appendix, Fig. S9). In addition, we found moderately positive impacts of other forest types (evergreen and mixed forests) and woody wetland cover percentages (Fig. 4 C–E), as well as negative impacts of pasture and cultivated crop cover percentages (Fig. 4 F and G), on the probability of a pixel being a hotspot. When modeling habitat associations of seasonal cumulative stopover density (Fig. 5; N = 8,866,713, deviance explained = 29.9%, and diagnostic plots in SI Appendix, Fig. S10), we discovered a similar scale dependency of deciduous forest cover impacts (Fig. 5A) and various associations with other land cover features (Fig. 5 B–G). Interestingly, low-intensity urban development, characterized by a mixture of impervious surface (accounting for less than 50% of total cover) and vegetation, did not show unidirectional impacts on hotspot probability (Fig. 4H), but it did contribute positively to cumulative migrant density (Fig. 5H).

Fig. 4.

Logistic GAM output of contributing factors to hotspot probability. N = 8,866,713. Lines and dashed lines each represent the overall means and standard errors of the smooths. The scale of the plots has been shifted and transformed to show the predicted hotspot probability when setting other covariates to mean values. (A) Interaction between percent landscape- and regional-scale deciduous forest cover. (B–H) Impacts of other land cover variables at the local scale on hotspot probability. (I and J) Remaining range bias and coastal effects. (K) Geographical pattern captured by latitude and longitude interaction.

Fig. 5.

Gaussian GAM output of contributing factors on seasonal cumulative stopover density (square centimeters per hectare) of landbirds averaged across 2015 to 2019. N = 8,866,713. Lines and dashed lines each represent the overall means and standard errors of the smooths. The scale of the plots has been shifted and transformed to show the predicted bird density (square centimeters per hectare) when setting other covariates to mean values. (A) Interaction between percent landscape- and regional-scale deciduous forest cover. (B–H) Impacts of other land cover variables at the local scale on seasonal cumulative stopover density. (I and J) Remaining range bias and coastal effects. (K) Geographical pattern captured by latitude and longitude interaction.

In models of both hotspot probability and cumulative stopover density, we found some fairly distinctive geographical patterns: Migratory landbirds are relatively concentrated along the Mississippi River at midlatitudes and, with weaker trends, along the Appalachian Mountains (Figs. 4K and 5K).

Habitat Associations within Avifaunal Biomes.

The eastern forest is the most extensively forested avifaunal biome in our study area (SI Appendix, Fig. S11) consisting of both deciduous and evergreen forests (SI Appendix, Figs. S12 and S13). The habitat associations of hotspot pixels within the eastern forest avifaunal biome were generally similar to what we found in the global model that combined all three biomes (SI Appendix, Fig. S14; N = 5,813,069, deviance explained = 29.6%, and diagnostic plots in SI Appendix, Fig. S15). However, an important exception is that hotspot pixels within this biome were concentrated at the immediate boundary between the prairie and eastern forest biomes, and hotspot probability decreased with increasing distance to the prairie edge (Fig. 6A). We found similar evidence of large concentrations of migrants at the prairie–forest edge within a roughly 25-km band in the cumulative density model for the same biome (Fig. 6B; full model in SI Appendix, Fig. S16, N = 5,813,069, deviance explained = 33.3%, and diagnostic plots in SI Appendix, Fig. S17). The absence of more hotspots in the southern portion of the eastern forest biome, where values for regional deciduous forest cover are the lowest (SI Appendix, Fig. S11), may be due to the forest composition of this region. This region is dominated by evergreen (pine) forests (SI Appendix, Fig. S13), with lesser amounts of deciduous forest (SI Appendix, Fig. S12). Thus, in some places, deciduous forest cover is low, while overall forest cover including evergreens is high, providing an opportunity for migrants to spread out.

Fig. 6.

Concentrating effects of the prairie–forest boundary in the eastern forest avifaunal biome. (A) Decreasing hotspot probability away from the prairie edge. (B) Decreasing migrant density away from the prairie edge. Full model details are provided in SI Appendix, Figs. S18 and S20.

Most radar domains in the prairie biome have a low percentage of regional-scale (within radar domain) deciduous forest cover (SI Appendix, Fig. S11), and we found strong positive impacts of the remaining deciduous forest fragments in concentrating birds to the hotspot pixels (SI Appendix, Fig. S18; N = 2,305,050, deviance explained = 46.5%, and diagnostic plots in SI Appendix, Fig. S19). Local (within-pixel) deciduous forest cover alone explained 12.1% of the deviance in this biome.

The northern forest biome had the smallest sample size of pixels (Fig. 2). The geographic coordinates (longitude and latitude) explained most of the deviance for hotspot probability in this biome, with negligible influence from land cover variables (SI Appendix, Fig. S20; N = 733,232, deviance explained = 54%, and diagnostic plots in SI Appendix, Fig. S21).

Sensitivity Analyses.

In sensitivity analyses that varied quantiles and consistency thresholds for determining hotspot pixels, we found similar habitat associations, indicating that our results are robust to variation in such thresholds (SI Appendix, Figs. S22–S29). We also ran spatial sensitivity analyses through a leave-one-radar-out cross-validation (SI Appendix, Figs. S30–S89), confirming that the model results are not dependent on specific radar stations. One notable exception is that the weaker peak of hotspot probability at high regional but intermediate landscape forest cover (Fig. 4A) is mostly driven by the KCCX radar (SI Appendix, Fig. S39) in central Pennsylvania (SI Appendix, Fig. S90).

Discussion

Geographic Pattern of Stopover Hotspots.

We identified important autumn stopover sites (i.e., areas that consistently contain high densities of migratory landbirds) using radar observations covering one-fourth of the entire land area of the eastern United States, a region where migratory birds have been declining for decades (4). At a coarse scale, we found small differences in seasonal bird density between most radar domains, except for one site in northern Alabama with restricted spatial coverage. The relatively small differences in migrant densities across most radar domains are consistent with the hypothesis that landbirds migrate across a broad front as opposed to the behavior of many migratory shorebirds that concentrate at a few key sites (13, 29). Nonetheless, landbirds are not equally abundant everywhere (26): Radar domains along the Mississippi River and the Appalachian Mountains support relatively higher densities of migrants than do other regions. This stopover pattern is in general consistent with the migration pattern noted by Dokter et al. (31), suggesting that there may be two major autumn migration pathways in the eastern United States that converge in Tennessee and Alabama before the birds cross the Gulf of Mexico. The slight difference between the migration traffic pattern by Dokter et al. (31) and the stopover pattern revealed in this study could be attributed to broad-scale variability in the proportion of passage migrants that stop over across radar domains (23, 25).

The concentrating effects of stopover habitats are more apparent at the finer scale of individual radar pixels. We identified hotspot pixels that, on average, support five times the mean density of birds across all non-hotspot pixels. Similar to the geographical pattern revealed at the radar domain scale, we discovered clusters of hotspot pixels near large geological features, e.g., along major rivers such as the confluence of the Ohio and Mississippi rivers (Shawnee National Forest), as well as along the Appalachian Mountains. Overall, we found more hotspot pixels in the central and western parts of our study area, a pattern that partially overlaps priority stopover sites identified in studies using eBird records (32, 33). The major discrepancy between these studies and ours is that the Atlantic Coast, highlighted as a major area for migratory landbirds via eBird records (32, 33), did not stand out in our radar analyses (except for along the Delaware Bay). This is likely due to differences in how these two approaches characterize important stopover areas. Studies using eBird prioritize stopover sites based on the overlapping core ranges of individual species (i.e., areas that target protection of 30% of the abundance for each of over 400 migratory species, 33), whereas we identify stopover hotspots based on aggregate biomass (i.e., total bird density). It could be the case that species passing through inland areas are much more abundant than species migrating along the coast.

Multiscale Habitat Associations of Hotspot Pixels.

We found strong positive associations between deciduous forest cover and stopover hotspot probability. Such effects persist across avifaunal biomes, suggesting that deciduous forests support high densities of migratory landbirds during autumn migration, regardless of geographical region (18, 20, 25, 34). Dense deciduous forest cover is believed to provide high-quality habitat for migratory landbirds due to its greater fruit and arthropod abundance relative to other cover types, which is consistent with the positive association of hotspot probability with local deciduous forest cover percentages (30, 35). The overall positive impacts of deciduous forests on migratory birds at the landscape scale corroborate the hypothesis that landscape features serve as an attractant for landbird migrants approaching the end of their nightly migratory flight (30, 36).

The significant interaction between the percentages of regional and landscape deciduous forest cover emphasizes the importance of forest fragments in heavily deforested regions to migrants (35, 37, 38). Such concentrating effects are also supported by our avifaunal biome–specific results for the prairie biome, in which deciduous forest cover alone explains more than 10% of deviance. Within the eastern forest biome, both hotspot probability and stopover density tend to be higher at dense regional but intermediate landscape forest cover (mostly within the KCCX radar domain; SI Appendix, Figs. S39 and S90). This could be attributed to birds’ preference for edge habitats, especially in densely forested regions (39).

In addition to deciduous forest cover, we also found positive effects of evergreen and mixed forests, as well as woody wetlands, in determining hotspot probability. That a wide range of forest types can serve as hotspots is consistent with the high variability in land cover associations reported for 43 eastern forest birds during autumn migration using eBird data (40). Zuckerberg et al. (40) also showed higher occurrence rates of seed-eating birds in landscapes dominated by agriculture, a finding that differs from the negative correlation with pasture and cultivated crops that we saw in this study and that Buler and Dawson (18) also observed. We suggest that agricultural landscapes may be beneficial to some dietary guilds of birds, but, in terms of biomass, more migratory birds are stopping over in forested habitats compared with agricultural habitats during autumn migration.

Last, we found that low-intensity urban development hosts high densities of migratory birds, although not with interannual consistency. High densities of migrants in urban areas with low-intensity development may be due to the shelter and food that green spaces offer to migrating landbirds (40, 41). On the other hand, it could also be due at least in part to nocturnal migrants being attracted to urban light (20, 42), which would not necessarily draw them to good quality habitat and might even lure them to places with higher mortality rates due to collisions with buildings (43, 44). We encourage future research to investigate the distinction between light attraction and habitat attraction with respect to urban greenspaces.

Midwest Farmlands as a Migration Barrier.

We found strong evidence of birds being concentrated at the prairie–forest boundary after crossing the agricultural Midwest during autumn migration. The 25-km range of concentration at the edge is similar to that seen at shorelines after birds have just crossed major water bodies such as the Gulf of Mexico (25, 34) or long stretches of the Atlantic Ocean (20). This distinct concentration, followed by a sharp decline of bird density and hotspot probability as one moves away from the prairie–forest boundary, suggests that birds are avoiding stopping over in the heavily deforested Midwest until they reach regions with more forest cover. The mean proportion of autumn passage migrants stopping over within the domains of seven radar stations primarily within the prairie biome around the Great Lakes (26% ± 2%; 23) is nearly half the stopover-to-passage ratio for autumn migrants at 12 radar stations within the eastern forest biome along the northern Gulf of Mexico coast (44% ± 4%; 25), supporting our suggestion. However, a more comprehensive assessment of the stopover-to-passage ratio for all the radar stations within these two biomes is needed. Moreover, compared with the more extreme situation facing landbirds that are crossing major water bodies or deserts (45), the largely agricultural Midwest is not completely inhospitable, given the presence of remnant forests that support high densities of migrants (35; this study). This creates both challenges and opportunities to address the decline of migratory bird populations in the eastern United States, as noted below.

Conservation Implications.

Our results help to fill some important knowledge gaps relating to the understudied migration period, which, in turn, could benefit conservation efforts directed toward migratory landbirds (11, 26, 46). Our yearly maps of stopover density confirm a broadly diffuse migration pattern of landbirds across the eastern United States. A broad front of migrating birds requires a broad front of stopover habitat. Thus, creating a small set of discrete reserves at key staging areas, as might be feasible for some shorebirds (13), would fail to protect most migrating landbirds in eastern North America. A well-distributed network of protected areas at multiple scales across the eastern United States is essential to maintaining healthy populations of these species, a point also highlighted in the strategic goals of the international Partners in Flight (PIF) conservation initiative (6). Incentive programs to encourage the protection of forests (see next paragraph), analogous to the existing federal Conservation Reserve Program (47), or local ordinances that prevent forest conversion could be developed to achieve this goal. Such efforts are especially important for protecting the sites that we identified as consistently supporting high densities of migrants at the pixel scale, especially in areas where those pixels are clustered into continuous stretches of hotspot habitat. Some existing public/private partnership programs, such as the PIF and the Migratory Bird Joint Ventures, are already engaged in efforts to protect key habitats for migrants (6).

Deciduous forest appears to be the habitat type that harbors the greatest biomass of migratory landbirds in the eastern United States. Forest fragments in largely deforested regions are especially important to migrants. Such sites may represent the “convenience stores” identified in the framework for stopover sites proposed by Mehlman et al. (11); these are places where birds can briefly rest and replenish their energy stores while moving through a generally inhospitable landscape. Mehlman et al. (11) called for more attention to these remnant forest patches because they are often neglected in conservation planning due to their relatively small size and the low nesting success of breeding populations of landbirds within them (48). This is especially relevant to the agricultural Midwest, which we identified as a human-created migration barrier with a long history of land use change and habitat loss (28, 49). Perhaps for this reason, the Midwest has experienced the greatest cumulative loss in migrant biomass over the past decade (4). We therefore urge the protection of remaining forest habitats throughout the Midwest, as well as attention to anthropogenic woodland patches that also provide suitable stopover habitats (35, 50, 51). Protecting and creating more such stepping stones might enable more migratory birds to successfully traverse this human-created migration barrier and thus maintain healthy populations.

Our study also underscores the importance of measuring both cumulative density and consistency of use in order to identify stopover hotspots. Doing so also makes the most effective use of limited conservation resources because the loss of areas that receive heavy and consistent bird use across years is more likely to directly affect the abundance of migratory bird populations (11). Conservation planning based on both metrics is more likely to achieve the mission of “keeping the common birds common” and “helping species at risk” (6). Note, however, that the broad-front migration pattern we document means that conservation measures cannot be limited to only a few places if migrant populations across the eastern half of the continent are to be sustained.

Limitations and Future Directions.

The archiving of NEXRAD data, coupled with recent advances in radar-processing algorithms, has made possible multiyear continental monitoring of migration (19, 21, 52). However, we had to limit the time span of our analyses to 5 y due to the computation power required for this analysis and the extensive manual quality control we felt was necessary to process the radar data. A consequence of our manual quality control is that we only sampled birds aloft on clear nights, which may introduce weather biases. However, while migrating birds may be forced into less favorable stopover sites during extreme weather events, they tend to depart via a “landscape-scale” nocturnal flight rather than by flying during the daytime to better habitats (53), which can also be detected in radar signals. Therefore, our takeoff-based sampling method captures habitat usage in both good and unfavorable weather conditions. Furthermore, most birds migrate under favorable weather conditions (54); hence, our clear-night sampling should be representative of the stopover distributions for the majority of migrating landbirds. Having used strict density (top 10%) and consistency (all 5 y) thresholds for hotspot pixels, we believe our results are both conservative enough and robust enough to provide helpful guidance for conservation planning.

With growing interest in radar ornithology and new developments in automating radar processing, future studies should be able to generate more years of data faster, which should result in additions to and amendments of our results. Despite the unprecedented spatial coverage we achieved, weather radar stations sample only a quarter of the land area of the eastern United States with sufficient accuracy for bird monitoring. It would be helpful to be able to extrapolate migrant densities to areas beyond the domains of radar coverage by using both field surveys and machine learning algorithms; doing so is likely to reveal additional stopover hotspots. Moreover, only a few radar domains cover the forest–prairie boundary in the Midwest, the data from which we used to hypothesize that this large region of cropland constitutes an important migration barrier. We therefore encourage more fieldwork in that region to assess the degree to which large-scale agricultural development poses a barrier to migratory birds.

Another limitation of our departure-based hotspot classification method is that we could not incorporate the metric of stopover duration, which also determines the conservation importance of a stopover site by directly affecting the total food demands placed on that site during migration (e.g., a site supporting 10 birds for 10 d may be as valuable to migrants as another site supporting 100 birds for 1 d; see refs. 55 and 56). Future studies could investigate species’ weekly abundance estimates using eBird data to assess stopover duration (57). Finally, a basic limitation of this or any radar-based study is that it cannot provide information on the particular species that comprise the images of migrating birds. This also underscores the value of using eBird data, which are species specific, in tandem with radar to develop more targeted conservation strategies.

Conclusion

We mapped the distribution of stopover densities of migratory landbirds across the eastern United States using radar data. We found that eastern landbirds migrate across a broad front far different than the pattern exhibited by many shorebirds that congregate massively in a few key sites. At finer scales, we identified stopover hotspot pixels that harbor unusually large numbers of migrants and therefore deserve particular attention from conservationists. However, our general finding of a broad-front migration means that conservation efforts, too, must proceed on a broad front, targeting key habitats throughout the eastern half of the continent. We found that deciduous forest patches and riparian forests, especially those in largely unforested regions, are of particular importance to migrating landbirds. We also flag the agricultural Midwest as likely being a human-created barrier for autumn migrants. Forest remnants within this altered landscape may be of particular importance to migrants. We conclude that the challenge of conserving migratory landbirds must be understood and undertaken at multiple spatial scales if ongoing declines of these birds are to be halted or even reversed.

Methods

Data Sources.

We downloaded level II radar data from the NEXRAD archive on Amazon Web Services Cloud (58) for 60 radar stations along the Eastern and Central migration flyways (59) covering 5 y (2015 to 2019) of autumn migration in the eastern United States. Because it is already well established that certain areas along the Gulf of Mexico can support large numbers of migrants (e.g., refs. 25, 30, and 34), we excluded radar sites near the Gulf of Mexico and focused instead on inland areas. To capture the period of peak landbird migration across the continent, we sampled data from August 1 to November 15 each year (60).

Each radar station completes a set of horizontal 360° sweeps of the atmosphere at different tilt angles every 4 to 10 min. We selected the lowest angle (0.5°) to capture the on-ground takeoff of birds shortly after sunset. This would correspond to birds departing from the places they had used during stopover. Within each sweep, radar measures aggregate target density and size within sampling volumes through the strength of returned radiation (reflectivity). It also measures target velocity relative to the radar station through Doppler radial velocity (18). We vertically integrated the resulting volumetric measurements into a two-dimensional surface of stopover density (details in the following sections) and thus converted the 3-D radar data unit of sample volumes into a 2-D map of pixels. The level II reflectivity and radial velocity data used for stopover mapping have a range resolution of 250 m and an azimuth resolution of 0.5° (area coverage: 0.4 to 21.8 ha), with data measured along each of the 720 radials, thereby covering a disc area of up to 100 km from the radar center (SI Appendix, Fig. S91; note that the effective coverage area for bird migration after data processing is smaller, with an average radius of around 80 km per radar).

We obtained land cover data for the eastern United States from the 2016 National Land Cover Database (NLCD) at 30-m resolution consisting of 16 land cover types (61). To investigate how land cover measured at different spatial scales is related to stopover density (20), we calculated the proportions (0 to 1) of different land cover types at three spatial scales for each pixel: i) local scale: within each pixel (0.4 to 21.8 ha), ii) landscape scale: within a 5 km radius of each pixel (30, 62), and iii) regional scale: within a 100 km radius of the radar station.

Radar Data Screening.

Since most passerines are nocturnal migrants (63), we screened radar data within 1 h before and after sunset on a nightly basis for each radar station during the autumn migration season following established protocols by Buler and Dawson (18) and McLaren et al. (20). In total, we screened 31,547 radar nights and 0.5 million radar scans for all 60 sites across five autumn seasons (2015 to 2019).

There are multiple types of contamination affecting biological signals in radar scans, which must be reduced or eliminated in order to derive the density of migratory landbirds. Precipitation is the most common form of contamination (20), and we applied MistNet, a deep convolutional neural network (64), to automatically remove radar nights that were heavily contaminated by rain (nights with precipitation covering more than 10% of the radar domain, accounting for 28% of radar nights). We then visually screened the remaining nights of bird migration and excluded those contaminated by scattered precipitation within an 80 km radius of the radar station (accounting for an additional 12% of radar nights). We also excluded nights with anomalous propagations of radar beams (3%), nonbiological clutter such as sea breeze (7%) and biological clutter such as nocturnal feeding flights of waterfowl, local movements of resident species, and other organisms with distinctive radar signatures (12%) (18). Unlike waterfowl, however, shorebirds do not have a distinctive pattern on radar scans and thus cannot be differentiated from landbirds. Nonetheless, we believe that on the clear nights we selected, the radar was primarily capturing migratory passerines taking off from their stopover sites due to their greater abundance compared with shorebirds (4, 34).

Some contamination-free nights might be dominated by insects that typically have lower airspeeds (65). We therefore removed insect-dominated flights (8%) with mean airspeeds less than 4.5 m/s (18), which were calculated by vector subtracting wind velocities (data from the North American Regional Reanalysis, 66) from ground vectors determined from radial velocity data.

After screening, we were left with 9,022 bird-dominated radar nights (29% of total radar nights) for further analysis.

Radar Data Processing.

For each of the remaining bird-dominated radar nights screened above, we calculated the instantaneous nightly stopover density at the onset of synchronized migratory flight (exodus) near the end of evening civil twilight (67, 68). Using data from 1 h before sunset until 2 h after sunset, we fit a spline function to the change in mean reflectivity within 10 to 40 km of the radar center over time in terms of sun elevation angle (20). On a clear night of bird migration, the reflectivity is typically low prior to sunset and increases abruptly during the exodus (an example curve is provided in SI Appendix, Fig. S92). We identified the inflection point with maximum growth rate in reflectivity or 15 min after the initiation of the exodus (whichever came first) as the instantaneous sampling time (i.e., sun angle) for that night (the variation in sampling sun angles across radars is provided in SI Appendix, Fig. S93). By temporally interpolating reflectivity at peak exodus for each sampling volume, both the degree of dispersion of birds aloft from their departing place and the sun angle bias (18) are minimized, allowing us to generate geographically accurate and representative estimates of on-the-ground stopover density every night (20, 69). Note that our approach accounts for and excludes local movements of resident species, such as blackbirds heading to nocturnal roosts.

We accounted for the range bias in radar measurement (i.e., the sampling height of radar beams increases with distance away from the radar center, 69) by deriving a vertical profile of reflectivity (VPR) for the entire airspace of radar coverage (i.e., the vertical distribution of birds in the sampled airspace) at the above-identified sampling time every night. The VPR for each night was used to vertically integrate reflectivity measured at different locations (i.e., sampling heights) to produce an estimate of volumetric density across all heights above the ground and then converted into a surface density metric on the ground in units of square centimeters per hectare (18). This unit measures the total area of reflected surfaces (i.e., bird biomass) in the airspace per hectare in a given radar pixel on each sampling night. For each night, pixels that encompassed less than 10% of the VPR (usually pixels further away from the radar center) were eliminated to ensure data quality. When stacking all sampling nights for seasonal density (see sections below), the effective detection range of each radar domain was determined by including only pixels that were sampled in at least 75% of sampling nights for that radar (20).

Accounting for Temporal Sampling Bias within Seasons.

The rigorous screening protocol described above removed more than two-thirds of sampling nights, leaving on average 28 nights per season across 60 radar stations (SD = 7.7). To account for the potential temporal bias in the sampling coverage within season among radar domains, we fit in LOESS smooth curves to the radar-average nightly density (area-weighted mean of all sampled pixels within a radar domain) across Julian dates to estimate the general phenology of autumn migration intensity for each radar domain (an example curve is provided in SI Appendix, Fig. S94). When fitting the curve, we pooled data from all 5 y to ensure enough sampling points throughout the season, assuming the final curve fitted captures the overall pattern of stopover density changes with time for that specific radar domain. For each year, we calculated the proportion of seasonal migration density sampled based on the sampling nights of that year and the corresponding LOESS interpolated density relative to the total area under the curve (SI Appendix, Fig. S94). Finally, we divided the total sampled density of each year by the radar year–specific proportion metric (N = 300, 0.32 ± 0.10 SD) calculated above to generate the interpolated sum of cumulative stopover density across the entire migration season (square centimeters per hectare), representing the total biomass of migratory birds that had stopped over during autumn migration per unit area. Such calibration is a new improvement on radar data processing from previous radar stopover mapping efforts (e.g., ref. 18).

Determination of Stopover Hotspot Pixels.

At a finer scale within radar domains, we generated yearly maps of cumulative stopover density among radar pixels (SI Appendix, Figs. S3–S7) using the same method above, except that the radar-specific LOESS proportional metric was used to calibrate the density measured at each of the 8.9 million pixels. We used a 90% quantile threshold to identify high-density pixels each year. To assess the interannual consistency of stopover use, we stacked 5 y of data and summed the number of years that a radar pixel was classified as “high density.” Hotspot pixels were identified as those where the cumulative stopover density was in the top 10% in all 5 y.

Data Analysis.

We assessed multiscale habitat associations of the above-identified hotspot pixels by constructing generalized additive models (GAMs) to capture the nonlinear responses of stopover density to land cover composition (20, 40). We incorporated 10 major land cover types, including deciduous forest, evergreen forest, mixed forest, shrub, grassland, woody wetland, pasture, cultivated crops, and urban development as habitat covariates (25). The four classes of urban development in the NLCD classification were regrouped into low- and high-intensity urban development based on percentages of impervious surface. Distance from radar was included as a corrective variable to account for any remaining range bias after VPR correction. We also included relative elevation to the radar installment, distance to the coast (including the Great Lakes, set to a maximum value of 150 km, 20), longitude, and latitude as geographical covariates in the model. Considering that stopover site selection might be affected by regional-scale habitat availability (e.g., small habitat patches in a generally inhospitable matrix, 11), we incorporated additional interaction terms between landscape- (5-km buffer) and regional-scale (within 100 km radius) land cover percentages to assess scale dependency of habitat associations. Apart from the logistic GAMs on hotspot probability that is determined by both migrant abundance and interannual consistency, we also fit GAMs with Gaussian error distribution on log-transformed seasonal cumulative stopover density (average across 2015 to 2019) to assess habitat associations of cumulative migrant abundance only.

In addition to the single GAM that captures the overall habitat associations of hotspot pixels, we divided our study area into three avifaunal biomes based on delineations in Rich et al. (70) and ran separate GAMs for each biome to capture more region-specific habitat preferences. The division of avifaunal biomes is based on shared habitats and similar bird communities (70). The three biomes that comprise our study area are eastern forests, northern forests, and, in the western part of our study area, prairie. Given the apparent concentration of migrants at the prairie–forest biome boundaries (Fig. 2), we calculated the distance to the neighboring biome boundaries (set to a maximum value of 150 km) and included them as additional covariates in the biome-specific GAMs.

We fitted the GAMs and checked the model diagnostics using the mgcv package in R (71). We examined concurvity among covariates and used a threshold of 0.95 for the “worst-case scenario” (the largest index value for any coefficient vector) in the mgcv package to drop variables (e.g., relative elevation, local shrub cover, grassland cover, and high-intensity urban development, and most landscape-scale covariates other than deciduous forest cover) whose smooth term can be approximated by a combination of the others (71). This generous threshold was chosen given the large sample size and the fine-scale covariates that vary smoothly over space (72). We applied the double-penalty approach for smoothing and variable selection via the “select = TRUE” argument in the GAMs (73). The final models were selected based on the Akaike information criterion (SI Appendix, Table S1).

Sensitivity Analyses.

Finally, we performed sensitivity analyses by varying pixel-scale quantiles (top 5% and 20%) and consistency thresholds (3 y and 4 y) for the identification of hotspot pixels. We constructed separate GAMs for each combination and compared the results (SI Appendix, Figs. S22–S29) with our main findings. We also performed spatial sensitivity analyses by repetitively running the GAMs 60 times, excluding one radar each time, and compared the results (SI Appendix, Figs. S30–S89) with the main findings.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

All radar data used in this study are publicly accessible through the “noaa-nexrad-level2” Amazon S3 bucket (58) (https://aws.amazon.com/public-datasets/nexrad/).