Learning shapes the development of migratory behavior

Significance Migratory decision-making is often conceptualized as reducing either time or energy expenditure; however, the importance of these currencies may shift throughout life. The exploration–refinement hypothesis suggests that information is an additional currency shaping the lifetime development of migration, predicting that exploration and information gain should be favored early in life. Using a unique early-life tracking dataset, we show that white storks incrementally refine migration timing and routes by innovating novel shortcuts during migration. Storks switch from energy-efficient exploration to rapid and directed movement as they age. Together, these results suggest that learning and early-life exploration play an important role in the ontogeny of migration in a long-lived migratory bird and that information is a critical currency shaping migration.


Figures S1 to S11
Tables S1 to S15 Fig S2 .To examine the possibility that age-related differences in migration timing, flight energetics and route fidelity could be due to an increased ability of older and more experienced birds to compensate for adverse environmental conditions (i.e., strong crosswinds), we investigated how the inclusion of an interaction between crosswinds and age effected parameter estimates of models predicting migration duration (A), cumulative ODBA during migration (B) and migration route fidelity (as measured by Dynamic Time Warping; C).In all cases, except for route fidelity, including an interaction between age and crosswinds resulted in a large amount of overlap between the 95% CIs from the original models (black) and the models including the interaction (grey), suggesting that an increased ability of older birds to compensate for adverse conditions was not impacting age-related changes in migration duration or cumulative ODBA.and events from the previous year (i.e., sequential comparisons; black), between current migration events and all other migration events from the same individual (i.e., "to self"; blue), and between current migration events and all migration events from all other individuals (i.e., "to all others"; orange).The average values with 95% CI for each age class and DTW comparison are shown.DTW is a trajectory similarity metric, with smaller values representing more similar trajectories (i.e., higher fidelity) and larger values representing more dissimilar trajectories (i.e., lower fidelity).We interpret the lack of overlap between the sequential comparisons and the "all others" comparison as evidence that animals are unlikely to be following others or converging on an optimal route used by the majority.Because there was overlap between sequential comparisons and all others comparisons at age one, this indicates that some social learning may occur in early life.However, the decline in DTW values for sequential comparisons suggests that routes become more individualized as the animal ages.Sequential comparisons and comparison to self overlap during ages 1-2, but diverge after age 2. This pattern indicates that exploration likely occurs before age 3, but after age 3 individuals tend to adhere to the migration route used in the previous year.(black) and the dataset with any mortalities that occurred after age two removed (dark red), to examine the potential impact of selective mortality on the estimated effect size.Because our fidelity analysis required at least two consecutive fall migrations, the original dataset already excluded any mortalities that occurred before that period.This comparison was made for A) the model estimating the impact of age on migration duration, B) the model estimating the effect of age on cumulative Overall Dynamic Body Acceleration (ODBA) and C) the model estimating the effect of age on route fidelity.In all cases, including or excluding individuals that died after age two had no impact on the estimated coefficients in any of the models (as indicated by the overlap between the 95% CIs).This indicates that selective mortality had no measurable effect on the patterns we observed and that learning is the most likely driver of changes in migration timing, energetics and route characteristics.

Package name Version
Packages used for statistical analysis and figure creation:

Fig
Fig S3.A comparison of environmental conditions experienced by white storks during flight and non-flight periods of Fall (A, C, E) and Spring (B, D, F) migration.We compared wind support (large values represent favorable flight conditions; A-B), crosswinds (small values represent favorable flight conditions; C-D) and vertical velocity (negative values represent upwards air movement and favorable flight conditions; E-F).We used two-sided paired t-tests to examine differences across flight and non-flight periods, where a positive difference in means represents larger values experienced during the flight period in comparison to the non-flight period for the environmental variable being considered (e.g., wind support, crosswinds, or vertical velocity).

Fig S4 .
Fig S4.An examination of ontogenetic changes in decisions about when to fly or remain stationary during migration.We examined how differences in wind support (A-B), crosswinds (C-D) and vertical velocity (a proxy for uplift strength; E-F) experienced during flight and non-flight periods changed across seasons as animals aged.Larger, positive differences in wind support represent more favorable conditions during flight in comparison to non-flight periods, whereas smaller, negative differences in crosswinds and vertical velocity represent more favorable conditions during flight in comparison to non-flight periods.The numbers in parentheses above the x-axes indicate the sample size for each age class.The solid black line shows the fitted relationship and the grey polygon indicates the 95% CI estimated by semiparametric bootstrapping (n=1,000 simulations).n= 262 migration events (spring and fall migrations) from 40 individuals.

Fig S5 .
Fig S5.Changes in the Overall Dynamic Body Acceleration (ODBA) per distance traveled as individuals aged.ODBA was calculated only during migratory flight and increased with age during both Fall (A) and Spring (B) migration.Violin plots display the distribution for each age class and season with empirical data overlayed in each season-specific color (fall = purple, green = spring).Black dots show the mean ODBA per distance traveled for each age class.Numbers in parentheses show the sample size for each age class.

Fig
Fig S7.A comparison of Dynamic Time Warping (DTW) values between current migration events

Fig S8 .
Fig S8.The interactive effect of age and crosswinds on route fidelity (see TableS12).Route Fig S8.The interactive effect of age and crosswinds on route fidelity (see Table S12).Route fidelity was measured as the log of Dynamic Time Warping (DTW), where smaller values represent greater route fidelity.Crosswinds represent the absolute value of the wind vector perpendicular to the movement direction of storks during periods of migratory flight (where larger crosswinds represent more challenging conditions for flight).The color of each point represents the age of the individual (with the purple to yellow color gradient corresponding to ages 1 -7).

Fig
Fig S9.A visualization of the steps involved in identifying route deviations, using an example from the fall migrations of a white stork tracked in 2018 (pink) and 2019 (purple).

Fig S10 .
Fig S10.The influence of age on route straightness.(A) Standardized coefficient estimates of the fixed effects of a model examining the impact of age on straightness of the entire route, while controlling for potentially confounding effects of season and environmental conditions (i.e., wind support, crosswinds, and uplift strength estimated as vertical velocity).(B) As individuals aged, route straightness increased, suggesting that birds move more directly between migratory destinations as they age.A straightness value of one represents the most direct movement path between start and end points (i.e., the beeline).The numbers in parentheses above the x-axis in B indicate the sample size for each age class.In B, the solid black line shows the fitted relationship and the grey polygon indicates the 95% CI estimated by semiparametric bootstrapping (n=1,000 simulations).n= 262 migration events (spring and fall migrations) from 40 individuals.

Table S1 .
The locations and sample sizes of the stork tagging areas.

Table S2 .
Details on the 40 individuals used in analysis, including the month and year they were 701 tagged as fledglings, the date of death (if applicable), the status of the tag, the breeding 702 population where they were tagged and the number of migration events (including spring and fall 703 migrations) extracted from their tracking data.

Table S3 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on migration duration while accounting for potentially confounding effects of migration distance (km, natural log transformed), season (Fall is the reference category), and environmental conditions (i.e., wind support, crosswinds, uplift).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.In contrast to the standardized estimates shown in Figure1, all

Table S4 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the start of fall migration while accounting for potentially confounding effects of migration distance (calculated as the great circle distance between the start and end of migration), wind support, crosswinds, and vertical velocity (a proxy for uplift strength).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 141 fall migration events from 40 unique individuals.

Table S5 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the end of fall migration while accounting for potentially confounding effects of migration distance (calculated as the great circle distance between the start and end of migration), wind support, crosswinds, and vertical velocity (a proxy for uplift strength).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 141 fall migration events from 40 unique individuals.

Table S6 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the start of spring migration while accounting for potentially confounding effects of migration distance (calculated as the great circle distance between the start and end of migration), wind support, crosswinds, and vertical velocity (a proxy for uplift strength).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 120 spring migration events from 37 unique individuals.

Table S7 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the end of spring migration while accounting for potentially confounding effects of migration distance (calculated as the great circle distance between the start and end of migration), wind support, crosswinds, and vertical velocity (a proxy for uplift strength).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 120 spring migration events from 37 unique individuals.

Table S8 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the difference between wind support experienced during flight and non-flight periods of migration.Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 262 migration events (spring and fall migrations) for 40 individuals.

Table S9 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the difference between crosswinds experienced during flight and non-flight periods of migration.Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 262 migration events (spring and fall migrations) for 40 individuals.

Table S10 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on the difference between uplift strength (measured as vertical velocity) experienced during flight and non-flight periods of migration.Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Model estimates are based on 262 migration events (spring and fall migrations) for 40 individuals.

Table S11
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on cumulative Overall Dynamic Body Acceleration (ODBA) during migratory flight while accounting for potentially confounding effects of wind support, crosswinds, vertical velocity (a proxy for uplift strength), and season (Fall is the reference category).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Unlike the standardized estimates in Figure 2, all units are unstandardized.Model estimates are based on 244 migration events (spring and fall migrations) from 40 unique individuals.

Table S12 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on route fidelity while accounting for potentially confounding effects of wind support, crosswinds, vertical velocity (a proxy for uplift strength), and season (Fall is the reference category).Route fidelity is calculated as the natural log of Dynamic Time Warping (DTW), which is a metric of route dissimilarity where smaller values represent higher route fidelity across migration seasons.Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Unlike the standardized estimates shown in Figure3, all units are unstandardized.Model estimates are based on 174 observations from 39 unique individuals.

Table S13 .
Statistical summary of the fixed effects of a linear mixed effects model examining the effect of age on route straightness while accounting for potentially confounding effects of wind support, crosswinds, vertical velocity (a proxy for uplift strength), and season (Fall is the reference category).A straightness value of one represents the most direct movement path between start and end points (i.e., the beeline).Random effects include a random intercept for individual ID and a random slope (Age|ID) to account for repeated measures on the same individual over time.Unlike the standardized estimates shown in Figure S6, all units are unstandardized.Model estimates are based on 262 migration events (spring and fall migrations) from 40 unique individuals.

Table S15 .
The name and version of R packages used in analysis.