Artificial light at night is a top predictor of bird migration stopover density

Avian migration depends on terrestrial and aquatic stopover sites where birds rest and refuel between nocturnal flights, yet comprehensive, large-scale data on where and why migrants stop are lacking. Nearly 500 migratory species, many of them nocturnally migrating songbirds with diverse behaviors and broad-fronted strategies, use a distributed patchwork of habitats that are difficult to sample with traditional methods, and community science datasets can be biased and do not directly quantify active migration. Weather surveillance radar provides an opportunity to quantify active migration and infer stopover distributions by capturing migrants as they ascend from stopover habitats at dusk. This study aims to produce the first contiguous U.S. maps of seasonal migrant stopover density and to identify macroscale drivers of stopover use, with specific hypotheses that greater forest cover, higher artificial light at night (skyglow), and higher vegetative productivity are associated with higher stopover density. Understanding these drivers is essential for conservation in the face of rapid anthropogenic change, including the expansion of artificial light at night.

Radar has detected birds for over 75 years, and low-elevation scans have been used to reveal spatial distributions of migrants during takeoff, linking aerial signatures to underlying stopover habitats. Prior studies show associations between artificial light at night and migrant distributions in parts of the U.S., Mexico, and Israel, with potential attraction to illuminated areas that may lead to suboptimal stopover and increased mortality. Most past analyses have focused on local to regional scales, potentially missing macroscale patterns critical for conservation; recent broader studies have begun to identify large anthropogenic barriers (e.g., extensive Midwestern commodity crop regions) to stopover. The literature also highlights the potential for ecological traps driven by human-induced rapid environmental change and the need to address biases and gaps inherent in community science and visual observations for nocturnal migrants.

Data sources and study period: The U.S. NEXRAD network (159 radars; 143 in the contiguous U.S.) provided weather surveillance radar reflectivity (η) data for 2016–2020 across spring (Mar 15–Jun 15; 93 nights) and fall (Aug 15–Nov 15; 93 nights). From 3,066,623 scans, approximately 133,000 scans aligned with optimal exodus sampling were assembled. Scans from sunset to 2.5 h after local sunset were downloaded (AWS NEXRAD repository). Response variable was mean seasonal stopover density mosaicked across radars and resampled to 1 km. Clutter mitigation: Precipitation was removed using the MistNet CNN classifier; voxels identified as precipitation and voxels within 5 km of precipitation were set to NA; scans with >30% precipitation contamination (7.5–80 km range) were removed. Static clutter masks removed topographic clutter (based on beam geometry at 0.5° elevation) and persistent ground clutter using probability of detection (POD) at 10 dBZ from January clear scans (1995–2020); voxels with POD ≥20% in ≥2 years were masked. Ring-shaped bat roost departures were identified visually and buffered; 12 sites were screened and masked; one radar (KEWX, San Antonio) was excluded due to severe bat contamination (final n=142 radars). Insect contamination at the voxel scale remains challenging; however, reflectivity biases toward larger scatterers (birds), and prior work found no significant differences when filtering insect-dominated scans. Range correction: To reduce range-related sampling bias, vertical profiles of reflectivity (VPR; 0–2000 m AGL, 100 m bins) were used via bioRad integrate_to_ppi to vertically integrate biological scatterers and correct for partial beam overlap. Corrections were applied where adjustment factors <10 and ranges ≤80 km. Exodus selection time: For each radar, median reflectivity of the lowest elevation scan (−0.5°) out to 100 km across scans from sunset to 2.5 h post-sunset was modeled with a GAM (hour after sunset as smooth, k=10; date as random effect). The time of maximum rate of change in reflectivity was selected as the site-specific exodus time (mean 49 min after sunset; range 27–77 min). Linear models showed radar elevation predicted sampling time (p<0.001), not skyglow. For each season-year, range-corrected scans nearest the exodus time were averaged across nights, mosaicked across radars, and resampled to 1 km. Predictor variables (n=49): Enhanced vegetation index (EVI; 4 monthly predictors per season from MOD13A3 V061); MODIS nighttime land surface temperature (MOD11A2/MYD11A2 V061) transformed to Accumulated Nocturnal Degree Days (ANDD) with seasonal decomposition imputation; National Land Cover Database (2016, 2019): percent canopy cover, percent impervious surface, landcover classes at 1 km and within 5 km buffers (32 variables total); Daymet v4 R1 mean daily precipitation per season; skyglow from VIIRS DNB monthly cloud-free composites using the simplified all-sky light pollution model (sALR) to derive the all-sky light pollution ratio (ALR); elevation from NASADEM resampled to 1 km; distance to radar; and year. Predictors were harmonized to Albers Equal Area, 1 km resolution. Correlation assessment: Pairwise correlations (2016 predictors) between skyglow and canopy/forest/water showed weak correlations (median r between −0.1 and 0.04) with low percentages significant; canopy cover and forest class proportions were strongly positively correlated. Modeling framework: Gradient-boosted trees (XGBoost in R) related predictors to stopover density. Dataset split: 75% training, 15% validation, 10% test; hyperparameters tuned with early stopping (learning rate 0.1; best: max_depth 16, min_child_weight 1, gamma 0, colsample_bytree 1, subsample 1). To limit long-distance learning, 2,500 random centers were used: 2,000 models fit within 400 km bounding boxes and 500 within 800 km bounding boxes. Predictions at 1 km resolution fixed distance-to-radar at 35 km and used 2020 predictors where possible. Predictions were mosaicked by averaging overlaps. Relative hotspots: A circular focal window radius of 265 km (approximate nightly flight distance of Swainson’s and Hermit Thrushes) categorized each pixel’s predicted stopover density into high (≥90th percentile), medium (≥50th and <90th), and low (<50th) relative intensities; analogous state-level hotspot maps were produced. Flyways were defined by longitude bands: western (>103°W), central (103–90°W), eastern (<90°W). Variable importance and directionality: For each model, variable importance was quantified by gain. Partial dependence plots were generated; linear fits with p<0.05 were used to determine directionality, classifying variables as globally positive if >55% of significant slopes were positive, negative if <45% were positive, neutral otherwise. Holdout experiment: For each of the 2,000 400-km models, a 10×10 km box centered on the median training coordinates was held out (only areas with ≥25 sampling points). Models for this experiment used a higher learning rate (0.25), colsample_bytree 0.75, subsample 0.5. One prediction per location-year was averaged across overlaps; one year was randomly selected to avoid pseudoreplication. Predicted vs. observed stopover density within holdouts was compared to assess performance.

• Model performance: Strong correspondence between predicted and observed stopover densities in holdout tests. Spring: R²=0.85 (F1,42430=244600, p<0.0001); Fall: R²=0.87 (F1,42864=287300, p<0.0001). Flyway-specific R² ranged 0.73–0.82 (spring) and 0.78–0.84 (fall).
• Top predictors: Across seasons at 2000–400 km modeling scales, elevation, skyglow, distance to radar, precipitation, and year were most important. Distance to radar and percent cultivated crops (5 km) were negatively associated with stopover density; other top predictors (e.g., percent tree canopy cover, percent deciduous and evergreen forest at 5 km, skyglow, precipitation) showed consistent positive associations.
• Skyglow effect: In over 70% of models, skyglow was a highly influential and consistently positive predictor of stopover density; in the western flyway, skyglow was the top predictor. Variable importance summaries indicated skyglow had strong positive partial dependence slopes.
• Predictor directionality counts: Spring (of 47 predictors with sufficient data): 33 positive, 7 neutral, 5 negative. Fall: 32 positive, 7 neutral, 8 negative. Land cover proportions at 1 km generally ranked lowest by gain; perennial ice (5 km) was second least important in both seasons.
• Predictor class contributions: Broad-scale (5 km) land cover was the most important non-sampling predictor class, accounting for 33–34% of summed gain; forest-related measures were consistently important drivers.
• Macroscale spatial patterns: Spring stopover density was greatest in the central U.S.; average density was 1.5× higher than the eastern flyway and 2.9× higher than the western. States with highest mean spring densities: Arkansas, Oklahoma, Louisiana, Texas, Mississippi (descending order); highest densities occurred along the Gulf of Mexico, especially southern Texas. Fall densities were greatest in the southeastern U.S.; top states: Alabama, Tennessee, Arkansas, Mississippi, Georgia; the eastern flyway had 1.2× the central and 5.8× the western flyway.
• Seasonal differences: 70.7% of 1-km pixels had higher stopover density in fall; on average, pixel-level densities were 66% higher in fall; 32% of the contiguous U.S. showed ≥100% increase from spring to fall. Greatest fall-over-spring differences occurred in the eastern half of the U.S. and mountain west; spring>fall differences were found along the coastal western U.S., Texas/Louisiana coasts, and northern Great Plains.
• Hotspot geography: Relative hotspots (265 km radius) often occurred near coastlines, geographic barriers (e.g., mountain ranges in Colorado and California), and extensive forests. An overall eastward shift in hotspots from spring to fall aligns with looped migration trajectories of North American landbirds.
• Conservation implication: Elevated stopover densities in peri-urban illuminated areas support the hypothesis that artificial light at night may act as an ecological trap at macroscales, potentially increasing collision risk and other negative outcomes for migrants.

The study demonstrates that macroscale patterns of migratory bird stopover density can be robustly mapped using weather radar and environmental predictors, and that artificial light at night (skyglow) is a pervasive, positive predictor of where migrants concentrate. These findings address the core question by identifying key environmental drivers—especially forest cover, precipitation, and skyglow—that explain stopover distributions across the contiguous U.S. The strong model performance in spatial holdouts indicates the approach captures transferable relationships beyond directly sampled areas. The spatial patterns—central U.S. dominance in spring, southeastern dominance in fall, and an eastward shift of hotspots—are consistent with known looped migration strategies and highlight ecological contexts such as proximity to coasts, mountain barriers, and forested regions. The positive association with skyglow raises concerns that artificial lighting may create ecological traps by attracting migrants into suboptimal or hazardous landscapes, elevating risks of collisions, predation, altered connectivity, and phenological disruption. The hotspot maps at a biologically relevant scale (nightly flight distance) provide a hierarchical framework for conservation prioritization across jurisdictions and seasons. Further integrating passage rates to compute stopover-to-passage ratios could help distinguish true ecological hotspots from light-induced aggregations and quantify the degree to which bright areas function as traps.

This work delivers the first contiguous U.S. maps of seasonal migratory bird stopover density and relative hotspots, revealing macroscale patterns and identifying key drivers, foremost among them skyglow, forest cover, and precipitation. The approach integrates multi-year NEXRAD radar with 49 geospatial predictors using spatially constrained gradient-boosted models, yielding high predictive performance and actionable hotspot delineations for conservation. The consistent importance of skyglow underscores artificial light at night as a widespread ecological stressor with potential trap dynamics for migrants. Future directions include incorporating 3D vegetation structure (e.g., GEDI-derived canopy height and complexity), leveraging the full multi-decade NEXRAD archive to assess temporal change, integrating migration passage with stopover density to quantify trap potential, and evaluating interactions among stopover density, timing, and land-use change. Given rapid increases in skyglow, coordinated policy, mitigation, and advocacy are needed to reduce light pollution while mechanistic studies elucidate why nocturnal migrants are attracted to light.

• Spatial coverage: Radar coverage directly sampled 26.2% of the contiguous U.S.; models extrapolated to the remaining 73.8%, which may introduce uncertainty where environmental contexts differ from sampled areas.
• Sensor and target ambiguities: Although reflectivity is biased toward larger scatterers (birds), species-level resolution is not possible, and insect contamination cannot be perfectly excluded at the voxel scale. One radar site (KEWX) was removed due to bat contamination; other bat roost areas were masked, potentially omitting some true bird signals near those sites.
• Range and geometry biases: Despite range correction with vertical profiles, residual biases related to beam geometry and terrain may remain, and distance-to-radar effects necessitated fixing this variable at prediction time (35 km), which may not capture local variability.
• Predictor limitations: Some predictors (e.g., fine-scale land cover at 1 km) contributed little to gain; unmodeled habitat characteristics (e.g., canopy height/complexity) and local factors (e.g., microclimate, food availability) could improve predictions. Skyglow may correlate with unmeasured urban features, complicating causal interpretation of its positive association.
• Temporal scope: Analyses span 2016–2020; interannual variability beyond this window and longer-term trends were not assessed here though feasible with historical archives.
• Causality: Observational modeling identifies associations, not causal mechanisms; the attractant effect of light and potential ecological traps require integration with passage rates and mechanistic/behavioral studies.