Flying insects face numerous challenges in natural environments, including physical clutter and variable wind, and most insects rely heavily on visual feedback to stabilize themselves and navigate through complex landscapes (Taylor and Krapp, 2007). Our understanding of how insects accomplish these tasks is based primarily on laboratory studies in which insects are compelled to fly in a challenging scenario imposed by the researcher, such as maneuvering through obstacles (Crall et al., 2015; Lecoeur et al., 2019; Ravi et al., 2020), flying upwind through unsteady air flow (Crall et al., 2017; Ortega-Jiménez and Combes, 2018; Ortega-Jiménez et al., 2013; Ravi et al., 2013), or contending with clutter and wind simultaneously (Burnett et al., 2020). However, in outdoor settings, insects typically have some freedom to choose among alternative flight conditions; for example, by flying higher or lower to the ground, flying through or above obstacles, or altering their flight path to spend more time flying upwind (i.e. into a headwind), downwind (with a tailwind), or in crosswinds (along a path perpendicular to wind flow).

In addition to navigating these physical challenges, central-place foragers that fly over long distances in search of food require some mechanism of regulating their flight speed regardless of external wind and gauging the distance they have traveled, in order to return to their nest. Antennal sensing of air speed contributes to the regulation of flight speed in insects, particularly in the absence of strong visual cues (Khurana and Sane, 2016). But antennal sensing alone can only provide a measure of air speed (flight speed with respect to the surrounding air), and so provides inaccurate distance information if wind is present. Thus, many flying insects, including central-place foragers, rely strongly on visual mechanisms to control their ground speed (flight speed with respect to the ground) and measure the distance they have traveled.

Translational optic flow, or the angular velocity at which surrounding objects or surfaces move past an animal's eyes as it moves through the environment, can be used by flying insects in a variety of ways. When flying through corridors or obstacles, bees balance the translational optic flow on their left and right eyes to maintain position in the center of the corridor or gap (Kirchner and Srinivasan, 1989), and they use optic flow to estimate their distance from lateral walls or obstacles (Srinivasan et al., 1991). A variety of insects use optic flow to regulate air speed (reviewed in Baird et al., 2021), and fruit flies and bees also use optic flow to maintain constant ground speed when flying in the presence of wind (Baird et al., 2021; Barron and Srinivasan, 2006; David, 1982). Laboratory experiments have shown that honeybees (Apis mellifera) can maintain fixed ground speeds and optic flow in a variety of external flow conditions, including when flying upwind with headwinds greater than 3.5 m s−1 (Barron and Srinivasan, 2006) and when flying downwind with tailwinds up to 2 m s−1 (Baird et al., 2021). When flying upwind, bees increase their air speed beyond the velocity of the oncoming flow to maintain a preferred ground speed. Monitoring and controlling their ground speed allows bees to estimate the total distance they have flown, based on optic flow cues (Esch and Burns, 1995; Riley et al., 2003; Srinivasan et al., 1996).

Although bees are equally likely to encounter headwinds, tailwinds or crosswinds in natural environments, most laboratory-based flight studies (whether focused on sensory cues or flight kinematics) have focused on performance in still air or headwinds, as these conditions can most easily be simulated in the lab (e.g. by motivating insects to fly upwind in a wind tunnel). A few recent studies have explored honeybee flight in tailwinds (downwind) as well as headwinds (upwind), but the primary focus of these experiments was the role of visual cues (Baird et al., 2021) or the combined challenge of wind and physical obstacles (Burnett et al., 2020, 2022), rather than the effects of wind direction on the flight performance of bees. In addition, because insects are typically compelled to fly in a single environmental condition prescribed by the researcher, we do not know whether flying with wind coming from a particular direction is preferable to bees, whereas wind from other directions makes flight more challenging.

Data from studies on long-range migration or dispersal of insects provides some indirect information about insects' preferences for flight direction relative to wind. Radar studies reveal that many migrating insects rise far above the ‘flight boundary layer’ (FBL, i.e. the height at which wind speeds are approximately equal to the insect's own powered flight speed; Taylor, 1974), sometimes flying as high as 2–3 km above the surface. This presumably allows the insects to take advantage of strong winds that push them at speeds well beyond their maximum powered flight limits (reviewed in Chapman et al., 2011). Some of these migrating insects also display sophisticated height-selection strategies that allow them to adjust their altitude to fly with maximum tailwinds oriented in their intended direction of travel (Chapman et al., 2011). These studies on long-range windborne insect migrations show that migrating insects nearly always choose to fly downwind (i.e. with a tailwind).

However, a recent study on dispersal in Drosophila melanogaster suggests that flies do not simply fly downwind when released in a natural environment (Leitch et al., 2021). Instead, they choose a random direction of travel, then maintain a fixed heading (i.e. body orientation relative to celestial cues) while regulating their ground speed along their body axis, allowing them to be pushed sideways when external winds are not aligned with their flight heading. In this way, flies can disperse over large distances while maintaining the possibility of intercepting an odor plume that would lead them to an upwind food source (Leitch et al., 2021).

In a recent lab-based study on honeybee flight in headwinds and tailwinds, the authors reported that the wind speeds used in the study were limited to 2 m s−1 because this was the maximum speed at which bees would fly in a tailwind; in faster tailwinds, they would either land on the floor or exit the flight tunnel (Baird et al., 2021). This finding, along with the study on dispersal in fruit flies, suggests that insects' preference for flight direction relative to wind when they are flying within the FBL (i.e. within ∼0.5–15 m above the ground, where wind speed does not surpass powered flight capability) – a zone in which most insects spend the majority of their lives foraging and interacting with conspecifics – may differ from the preferences displayed by insects that engage in long-distance windborne migration above the FBL.

Here, we employed recent advances in automating video collection and analysis to examine thousands of foraging flights performed by hundreds of bumblebees flying in laboratory enclosures with both headwinds and tailwinds. We developed two novel experimental approaches to examine bumblebee flight in headwinds (upwind) versus tailwinds (downwind), in an effort to answer three questions about these commonly experienced flight conditions: (1) do bumblebees display a preference for flying upwind or downwind?; (2) do bumblebees maintain constant ground speed when flying downwind, as they do when flying upwind?; and (3) do bees display similar flight kinematics when flying upwind and downwind, or do these conditions impose different aerodynamic challenges?

In the first part of our study (experiment 1), we constructed a two-choice flight arena, in which a hive of yellow-faced bumblebees (Bombus vosnesenskii Radoszkowski 1862) could fly from their hive at one end to a feeder at the opposite end, which they could access via two different flight channels (Fig. 1A). The feeder contained the colony's only source of nectar (which was unscented, 50% sugar water, ad libitum); pollen was provided within the hive. Each flight channel was approximately 20×20 cm in cross-section and 1 m long, and the walls were covered in a speckled pattern to provide visual cues. Bees were allowed to acclimate to foraging in the arena for 1 week before experiments began, so that they would be familiar with the location of the feeder, the hive and the two channels.

Fig. 1.

Bees choose to fly upwind more often than downwind. (A) The two-choice flight arena used in experiment 1, in which bees could choose to fly from their hive to a feeder (and back to their hive) via one of two channels, with wind flowing in opposite directions. Flights were analyzed over 1.2 s video clips (filmed at 50 frames s−1) captured every minute over a 2 h period each day. (B) Image from one camera view of the flight arena, with several 1.2 s flight paths highlighted that were retained for analysis after removing walking bees. (C) Proportion of flights that occurred in the upwind (as opposed to downwind) direction. Over 12 days of testing, 2929 flights were recorded. The mean (±s.d.) proportion of bees flying upwind was 0.644±0.046, which was significantly greater than 0.5 (Wilcoxon test, P=0.00024). The number of flights recorded during each 2 h trial (n) is shown below the x-axis, and the different wind conditions are shown by different symbols (moderate, 1.25 m s−1; slow, 1.07 m s−1; minimal, 0.25 m s−1). (D) Proportion of flights that occurred in the right channel (as opposed to the left channel). The mean proportion of bees flying in the right channel was 0.525±0.060, which was not significantly greater than 0.5 (Wilcoxon test, P=0.076). In C and D, asterisks show results of binomial tests to determine whether each day's proportion of flights was significantly different from 0.5 (*P<0.05, **P<0.01, ***P<0.001; n.s., not significant). The solid, horizontal line shows the mean proportion over 12 days of testing, and shading shows ±1 s.d.

Fig. 1.

Bees choose to fly upwind more often than downwind. (A) The two-choice flight arena used in experiment 1, in which bees could choose to fly from their hive to a feeder (and back to their hive) via one of two channels, with wind flowing in opposite directions. Flights were analyzed over 1.2 s video clips (filmed at 50 frames s−1) captured every minute over a 2 h period each day. (B) Image from one camera view of the flight arena, with several 1.2 s flight paths highlighted that were retained for analysis after removing walking bees. (C) Proportion of flights that occurred in the upwind (as opposed to downwind) direction. Over 12 days of testing, 2929 flights were recorded. The mean (±s.d.) proportion of bees flying upwind was 0.644±0.046, which was significantly greater than 0.5 (Wilcoxon test, P=0.00024). The number of flights recorded during each 2 h trial (n) is shown below the x-axis, and the different wind conditions are shown by different symbols (moderate, 1.25 m s−1; slow, 1.07 m s−1; minimal, 0.25 m s−1). (D) Proportion of flights that occurred in the right channel (as opposed to the left channel). The mean proportion of bees flying in the right channel was 0.525±0.060, which was not significantly greater than 0.5 (Wilcoxon test, P=0.076). In C and D, asterisks show results of binomial tests to determine whether each day's proportion of flights was significantly different from 0.5 (*P<0.05, **P<0.01, ***P<0.001; n.s., not significant). The solid, horizontal line shows the mean proportion over 12 days of testing, and shading shows ±1 s.d.

We created air flow along each channel by embedding computer fans at both ends, with both fans blowing in the same direction (i.e. with one fan pushing air in from one end while the other fan simultaneously pulled air out from the other end). Within each channel, we could reverse the direction of flow by physically removing and re-installing the fans on each end so that they moved air in the opposite direction. In all trials, air flowed in opposite directions in the two channels (i.e. one channel had air flowing from hive to feeder and the other had air flowing from feeder to hive, with the direction in each channel varied on different days). In some trials, we turned on the fans in both channels, to create flows of moderate velocity (1.25 m s−1) in opposite directions. In other trials, we only turned on the fans in one channel, which led to slightly slower flow (1.07 m s−1) in that channel, along with minimal flow (0.25 m s−1) in the opposite direction in the other channel (due to some air circulation between channels through the open, end sections where both channels ended; Fig. 1A).

We systematically varied the direction of flow in the two channels to determine whether bumblebees display a consistent preference for flying upwind or downwind, while controlling for any preference the bees may have for flying in one channel versus the other (designated the ‘left’ and ‘right’ channels), or for any potential differences in flow characteristics or turbulence level between the channels (which we believe were minimal, because of the lack of obstructions within channels and the low flow velocity).

For each foraging trip an individual made, they were presented with two separate choices, deciding which tunnel to fly in for the trip from the hive to the feeder, and then deciding which tunnel to fly in for the return trip from the feeder to the hive. Experiments were performed over 12 days, and a single flow condition was tested on each day. Bees were allowed to acclimate to the new flow condition for 1 h before data collection began. We tested 6 different experimental conditions in randomized order, with 2 days/recording sessions per condition: (1) moderate flow (1.25 m s−1 in both channels), with flow in the left channel towards the feeder (and flow in the right channel towards the hive), (2) moderate flow, with flow in the left channel towards the hive, (3) slow/minimal flow (1.07 m s−1 and 0.25 m s−1) with slow flow in the left channel towards the feeder (and minimal flow in the right channel towards the hive), (4) slow/minimal flow with slow flow in the left channel towards the hive, (5) slow/minimal flow with slow flow in the right channel towards the feeder, and (6) slow/minimal flow with slow flow in the right channel towards the hive.

After each day's hour-long acclimation period, we collected video data over a period of 2 h (from noon to 14:00 h), recording a subsample of 1.2 s of video per minute (resulting in 120 flight clips per recording session). The entire length of both channels was filmed using two synchronized video cameras (Photonfocus MV1-D1312-160-CL) along the length of the arena, recording at 50 frames s−1. Cameras were calibrated each day using a checkerboard calibration routine in Matlab, and were automated to start, stop and save 1.2 s video clips every minute throughout the recording session.

Video data were analyzed in Matlab using motion-based multiple object tracking. This involved background subtraction to detect moving bees and a Kalman filter to assign moving points (bees) to tracks. Note that individual bees could not be uniquely identified because of the wide view of the filming area and subsequent low resolution of each individual. Given the large number of flights analyzed (which was substantially higher than the number of workers normally present in a hive) and the fact that some individuals within bumblebee hives are known to perform more foraging flights than others (Crall et al., 2018), our dataset is assumed to contain repeated measures of multiple flights by individual bees, which increases the chance of Type 1 statistical errors (see Discussion). Short tracks (less than 6 frames long) and erroneous points (points that became stationary) were removed, and we created 3D flight paths by matching tracks from different cameras and minimizing residual error (Fig. 1B). The 3D flight paths allowed us to exclude bees whose entire track was less than 1.5 cm above the floor of a channel (and thus were assumed to be walking) from further analysis.

We pooled all flights within each 2 h filming session, and classified each flight as upwind or downwind, and as left channel or right channel, depending on the location of the bee, the direction of its motion, and the direction of air flow during that trial. We then summed the total number of flights that were upwind and divided by the total number of flights to calculate the proportion of upwind flights (note that this total includes flights in both the left and right channels, as flow was upwind in each channel for one of the directions of travel, from hive to feeder or feeder to hive). We separately summed the total number of flights in the right channel (regardless of flow direction) and divided by the total number of flights to find the proportion of flights in the right channel.

Using the proportions calculated for each of the 12 days of data collection, we tested whether the proportion of upwind flights (and separately whether the proportion of flights in the right channel) was significantly greater than 0.5, using a one-sample Wilcoxon test in R (one-sided test to determine whether the proportion is greater than 0.5, n=12 days/proportions). Finally, because the total number of bees foraging each day can vary substantially (this is typical, and is seen even in the absence of experimental treatments), we tested each day's proportion of upwind (and right channel) flights to determine whether it was significantly different from 0.5 using a two-sided binomial test in R.

In the second part of our study (experiments 2 and 3), we allowed a hive of common eastern bumblebees (Bombus impatiens; Cresson 1863) to forage freely over a period of several weeks at a nectar feeder placed in the working section of a wind tunnel, traveling round-trip to the feeder from the exit/entry of their hive at the other end of the working section. As in experiment 1, individual bees could not be uniquely identified, and our dataset is assumed to contain repeated measures of multiple flights by individual bees, which increases the chance of Type 1 statistical errors (see Discussion). Bees encountered tailwinds when flying from the hive to the feeder, and headwinds when returning from the feeder to the hive (Fig. 2A). The working section of the wind tunnel was 45×45 cm in cross-section and 1.4 m long. Flow within the tunnel was unimpeded by the feeder (as this was at the downstream end of the working section), and turbulence intensity was low (<1.2%; Ravi et al., 2013). Black vertical bars 1 cm in width and spaced 2 cm apart were printed on clear film and attached to the side walls of the working section to provide visual cues. Bees were allowed to freely enter and exit the working section via a tube connecting the wind tunnel to their hive. The feeder on the downwind side of the working section provided ad libitum artificial nectar (50% sugar water) and was the only source of nectar for the hive; pollen was provided within the hive.

Fig. 2.

Bees fly less frequently and along more sinuous flight paths in higher flow velocities. (A) In wind tunnel experiments, bees were allowed to fly freely from a hive entrance at the upstream end of a wind tunnel working section to a feeder at the downstream end, flying downwind from the hive to the feeder and upwind from the feeder to the hive. Flow velocities were alternated for hour-long periods between 0, 0.75 and 2.0 m s−1, and bees were filmed with either four 100 Hz cameras over the working section (experiment 2) or one 5000 Hz camera capturing a lateral view (experiment 3). (B) The proportion of total flights recorded in experiment 2 was highest during periods with no flow (0 m s−1) and lowest during periods with 2.0 m s−1 flow. Proportions were calculated separately for downwind and upwind flights. A total of 1662 flights were captured over 6 days, with three hour-long periods of filming each day. (C) Flight path sinuosity (total distance traveled divided by linear distance from the start to the end point) in experiment 2 increased with flow speed, for bees traveling in both directions. Notched box plots show the median, 25th and 75th percentiles, and circles show individual data points. Upwind and downwind flights were analyzed separately (see Materials and Methods); asterisks indicate significant differences (one-way ANOVA with Tukey's HSD, **P<0.01, ***P<0.0001; n.s., not significant).

Fig. 2.

Bees fly less frequently and along more sinuous flight paths in higher flow velocities. (A) In wind tunnel experiments, bees were allowed to fly freely from a hive entrance at the upstream end of a wind tunnel working section to a feeder at the downstream end, flying downwind from the hive to the feeder and upwind from the feeder to the hive. Flow velocities were alternated for hour-long periods between 0, 0.75 and 2.0 m s−1, and bees were filmed with either four 100 Hz cameras over the working section (experiment 2) or one 5000 Hz camera capturing a lateral view (experiment 3). (B) The proportion of total flights recorded in experiment 2 was highest during periods with no flow (0 m s−1) and lowest during periods with 2.0 m s−1 flow. Proportions were calculated separately for downwind and upwind flights. A total of 1662 flights were captured over 6 days, with three hour-long periods of filming each day. (C) Flight path sinuosity (total distance traveled divided by linear distance from the start to the end point) in experiment 2 increased with flow speed, for bees traveling in both directions. Notched box plots show the median, 25th and 75th percentiles, and circles show individual data points. Upwind and downwind flights were analyzed separately (see Materials and Methods); asterisks indicate significant differences (one-way ANOVA with Tukey's HSD, **P<0.01, ***P<0.0001; n.s., not significant).

We performed experiments 2 and 3 in the wind tunnel on two separative hives of bumblebees. In experiment 2, we filmed bees with four overhead video cameras (Photonfocus MV1-D1312-160-CL), which imaged overlapping regions covering the full length of the working section, to obtain recordings of bees' overall flight velocities and trajectories while traveling upwind or downwind. Videos were motion triggered throughout the filming period and recorded at 100 Hz. Flow velocity was varied over three levels: 0, 0.75 and 2 m s−1. We allowed bees to acclimate to the wind tunnel for 3 days prior to performing wind experiments. The three flow velocity treatments were presented each day between the hours of 13:00 h and 16:00 h, and each treatment lasted for 1 h. We performed flight trajectory experiments over 6 days and modified the order of treatments to account for all possible combinations.

In experiment 3, we used a high-speed video camera (Phantom v410, Vision Research) to capture high-resolution videos at 5000 Hz, to analyze details of bees' body and wing kinematics during upwind and downwind flights. The high-speed camera was placed on the side of the wind tunnel to capture a lateral view of bees flying upwind or downwind, and a calibration object was used to convert video data from pixels to centimeters. The camera filmed an area of 10×10 cm, and was automatically triggered by bees flying through a laser aimed at a photoresistor. In this experiment, we varied flow velocity over the same three levels (0, 0.75 and 2.0 m s−1) throughout the day over the course of 2 weeks, performing additional trials at some velocities until enough video clips in each condition were captured.

Video data from both wind tunnel experiments were tracked using custom code in Python that incorporated the OpenCV package (https://github.com/nickgravish/Tracker). The image processing procedure consisted of: (1) computing the background from the median pixel values over time, (2) background removal and thresholding to isolate foreground objects (i.e. bees), (3) contour identification and ellipse fitting of foreground objects. After these processing steps, we had a set of bee contours (ellipses) for every video frame. In the next step, we performed contour association to link bee observations across frames. This step is unnecessary when there is only one bee in the video; however, in cases where multiple bees are present (which did occur), this is a necessary step to properly link tracks across video frames. To perform data association, we used a modified Kalman filter that linked objects across frames by minimizing the positional error between frames. This association step resulted in a list of flight track information for each frame, including body position and orientation (from the fitted ellipse), body size (from the number of thresholded pixels and a pixel to centimeter calibration), and velocity (estimated for each frame as the output of the Kalman filter). The final video processing step was to refine body orientation by removing fast-moving objects (the wings) and retaining slow-moving objects (the body).

From this flight track information, we calculated several kinematic variables. For experiment 2 (flight paths viewed from above), we restricted our analysis to trajectories within the central 30 cm of the tunnel's length, during which all bees were in motion (i.e. not taking off or landing). We calculated the sinuosity of each flight trajectory as the total distance along the 2D flight path divided by the linear distance between the start and end points of the trajectory. We noted that in a small number of flights, bees reversed direction, flew in a loop, or performed other maneuvers that interrupted their progress from one end of the tunnel to the other, resulting in high path sinuosity. Bees flying along more sinuous paths would experience varied, fluctuating optical flow, which could affect our comparison of optic flow regulation in upwind versus downwind flights; thus, we removed flights with high sinuosity (defined as sinuosity >1.1) and restricted our analyses to relatively direct flights with path sinuosity of 1.1 or less. We also excluded trajectories in which mean ground speed (see below) was less than 0.02 m s−1, as these likely represented bees walking on the bottom of the working section rather than flying (speeds along the tunnel were bimodal, with the low-speed peak occurring below 0.02 m s−1).

From the remaining trajectories, we calculated the mean and standard deviation of ground speed (the bee's speed relative to the ground, regardless of flow velocity), based on the instantaneous speed of the bee along the tunnel's long axis (i.e. speed along the x-axis, defined as the dimension aligned with the walls of the tunnel). We also calculated the mean and standard deviation of air speed (the bee's speed relative to the surrounding air), by adding the flow velocity to the bee's ground speed (when bees were flying upwind) or subtracting the flow velocity from the bee's ground speed (when bees were flying downwind).

To determine whether flights from the hive to the feeder (downwind when flow was present) and from the feeder to the hive (upwind with flow) could be analyzed together, we used a two-sample Wilcoxon test to compare bees' mean ground speed, standard deviation of ground speed, and sinuosity of flights in the two directions with 0 m s−1 air flow. Based on the outcome of these tests (see Results), we performed further analyses on flights in the two directions separately. To determine how flow velocity in the wind tunnel affected the measured kinematic variables, we performed one-way ANOVA on each variable (mean and standard deviation of bees' ground speed, mean and standard deviation of bees' air speed, and path sinuosity) with flow velocity (0, 0.75 or 2 m s−1) as a factor, analyzing flights from the hive to the feeder (the ‘downwind’ direction) and flights from the feeder to the hive (the ‘upwind’ direction) separately. Post hoc testing for significant variables was performed with Tukey's HSD test. Because some of the data did not meet assumptions of normality and homogeneity of variance, we also performed an equivalent non-parametric test on each set of data (a Kruskal–Wallis chi-squared test) to verify our results.

For experiment 3 (lateral high-speed videos), we used the orientation of ellipses fitted to the bees' bodies to calculate pitch angle, as the angle between the body axis and the horizontal. For each trajectory, we found the mean body pitch angle as well as the standard deviation of body angle. Finally, we calculated the average flapping frequency for each flight by measuring the frequency component of the instantaneous velocity along the tunnel axis (the x-axis). The velocity along this axis is calculated from the lateral bee silhouette, which has a slow component associated with center of mass movement and acceleration, and a fast component associated with the rapid forward and backward shift of the silhouette due to the wing motion. We performed a fast Fourier transform (FFT) on the x-velocity time series and determined the frequency of the maximum power signal of the FFT to estimate flapping frequency.

As in experiment 2, we tested the data to determine whether flights from the hive to the feeder and from the feeder to the hive differed, by performing a two-sample Wilcoxon test to compare mean body pitch angle, standard deviation of body angle, and mean flapping frequency in the two directions with 0 m s−1 air flow. Based on the outcome of these tests (see Results), we performed further analyses on flights in the two directions separately. To determine how flow velocity in the wind tunnel affected the measured kinematic variables, we performed one-way ANOVA on each variable (mean and standard deviation of body angle, mean flapping frequency) with flow velocity (0, 0.75 or 2 m s−1) as a factor, analyzing flights from the hive to the feeder and from the feeder to the hive separately. Because some of the data did not meet assumptions of normality and homogeneity of variance, we also performed an equivalent non-parametric test on each set of data (a Kruskal–Wallis chi-squared test) to verify our results.

Our automated methods of video collection and analysis in experiment 1 allowed us to examine 2929 voluntary foraging flights (both outbound and return flights to the hive) in the two-choice flight arena over 12 days of filming, with foraging sub-sampled over a 2 h period each day. This included 804 flights with moderate flow velocity (1.25 m s−1) in both channels, and 1117 flights with low flow velocity (1.07 m s−1) in one direction and minimal flow velocity (0.25 m s−1) in the other direction. The total number of flights recorded over the testing period varied between days, from a minimum of 64 to a maximum of 512 (mean±s.d. 244±132 flights day−1; Table S1). Based on the proportions calculated for each of the 12 days of data collection, we found that the mean proportion of bees flying upwind was 0.644±0.046, and the overall proportion of bees flying upwind was significantly greater than 0.5 (one-sample Wilcoxon test, V=78, P=0.00024; Fig. 1C). In contrast, the mean proportion of bees flying in the right channel was 0.525±0.060, which was not significantly greater than 0.5 (one-sample Wilcoxon test, V=58, P=0.076; Fig. 1D). The binomial tests to determine whether each day's proportion of flights was significantly different from 0.5 showed that the proportion of bees flying upwind was significantly greater than 0.5 on 10 of the 12 days (Fig. 1C; Table S1). In contrast, the proportion of bees flying in the right channel was not significantly different from 0.5 on 8 of the 12 days, was significantly higher than 0.5 on 3 days, and was significantly lower than 0.5 on 1 day (Fig. 1D; Table S1).

Experiment 2, in which we captured overhead views of flight trajectories along the wind tunnel, resulted in 1662 digitized trajectories over 6 days (with motion-triggered videos collected over a period of 3 h per day). After excluding high-sinuosity flights and low-speed walking tracks, we had a total of 1449 flights for analysis. These included 470 flights towards the feeder with 0 m s−1 flow, 283 flights towards the feeder with a 0.75 m s−1 tailwind, and 136 flights towards the feeder with a 2 m s−1 tailwind, as well as 316 flights towards the hive with 0 m s−1 flow, 173 flights towards the hive with a 0.75 m s−1 headwind, and 71 flights towards the feeder with a 2 m s−1 headwind. Despite filming bees for the same total amount of time at each flow velocity, we found that the number of flights declined sharply as flow velocity increased; thus, more than 50% of the flights captured in each direction occurred with no flow (0 m s−1) and fewer than 20% of flights occurred in 2 m s−1 flow (Fig. 2B).

We found that bees' flight behavior differed significantly when flying down the wind tunnel towards the feeder and when flying up the tunnel to return to the hive, even in the absence of external flow. Flight trajectories with no flow (0 m s−1) differed significantly between the two directions in mean ground speed (two-sample Wilcoxon test, P=6.6×10−7) and path sinuosity (P<2.2×10−16), although the standard deviation of ground speed was not significantly different (P=0.76). We therefore analyzed flights in the two directions separately.

When flying in both the downwind and upwind directions, bees' flight path sinuosity was affected by flow velocity (Table S2), with increased sinuosity in higher flow velocities (Fig. 2C). Bees' mean air speed also varied with flow velocity, in both the downwind and upwind directions (Table S2). Air speed increased significantly with flow velocity for bees flying upwind and decreased significantly with flow velocity for bees flying downwind, with bees in 0.75 m s−1 flow displaying airspeeds averaging around 0 m s−1 and bees in 2.0 m s−1 flow displaying negative air speeds (i.e. flying backwards relative to the air; Fig. 3A). Despite these large changes in bees' air speed, their mean ground speed was unaffected by flow velocity, for flights in either the upwind or downwind directions (Table S2; Fig. 3B).

Fig. 3.