Honeybees use the skyline in orientation

Ants and bees, the best-known navigators among the invertebrates, rely on two primary navigational strategies, path integration and view-based navigation. In path integration, the insects use a celestial compass and some form of odometer to continually reckon their net displacement from the nest, which allows them to return directly home at any time (Wehner and Srinivasan, 2003; Srinivasan, 2015). In view-based navigation, by contrast, the insects capture panoramic views of the landscape and then compare these remembered views with current views in order to orient in certain directions or move toward certain destinations (Zeil et al., 2003, 2014; Zeil, 2012; Cheng, 2012; Collett et al., 2013; Wystrach et al., 2016). View-based navigation is robust under a variety of circumstances (Baddeley et al., 2011, 2012; Wystrach and Graham, 2012; Schwarz et al., 2014) and even allows successful homing from places that the insects themselves have never visited but from where at least some familiar landscape features can be seen, albeit from new perspectives (Dyer, 1991, 1996; Dyer et al., 1993; Durier et al., 2004; Collett et al., 2007; Graham et al., 2010; Wystrach et al., 2012; Narendra et al., 2013; Zeil et al., 2014; Cheung et al., 2014). View-based navigation does not necessarily require the animals to distinguish individual objects or landmarks as such; all objects could merely be incorporated into the broad, panoramic views that are stored and compared (Zeil et al., 2003; Graham and Cheng, 2009b; Basten and Mallot, 2010; Cheng, 2012; Philippides et al., 2011; Wystrach et al., 2011a, 2016; Wystrach and Graham, 2012). Wasps, too, use view-based navigation much like ants and bees (Stürzl et al., 2016).

Using view-based navigation over long distances works best when the captured views encompass distant landscape features, which, because they are distant, are recognizable from more places (Stürzl and Zeil, 2007; Collett et al., 2007); that is, views containing more distant landscape features have larger catchment areas (Cartwright and Collett, 1987; Zeil et al., 2003; Collett et al., 2013). The most distant landscape feature in any landscape, and probably also the easiest to detect, is the skyline, the panoramic silhouette of terrestrial objects against the sky (Möller, 2002; Differt and Möller, 2015). Further, simulations based on real or naturalistic scenes suggest that the skyline alone may be sufficient for robust view-based navigation in ants, guiding the ants continuously as they move through a landscape (Basten and Mallot, 2010; Philippides et al., 2011; Baddeley et al., 2012; Wystrach et al., 2016).

Wehner (1981) was evidently the first to suggest that insects might use the silhouette of terrestrial objects against the sky in orientation, a suggestion later supported by Southwick and Buchmann's (1995) observation that honeybees can successfully return home after displacements from greater distances – and, we now know, also faster (Pahl et al., 2011) – when prominent skyline features such as mountains are visible from the release sites. Later, Towne and colleagues (Towne and Moscrip, 2008; Dovey et al., 2013) suggested that honeybees orient by the panoramic skylines around their nests, as their bees seemed to confuse different landscapes that shared only similar skylines. Meanwhile, Möller (2002) and Stürzl and Zeil (2007) pointed out that the skyline could be a useful and easily detectable cue for visual navigation; Wehner, Fukushi and colleagues (Wehner et al., 1996; Fukushi, 2001; Fukushi and Wehner, 2004) presented suggestive evidence that ants use the skyline in homing after artificial displacements; and Hempel de Ibarra et al. (2009) found that the learning flights of bumblebees seem to be influenced by the skyline.

The evidence that insects actually use the skyline in visual navigation remained suggestive, however, until Graham and Cheng (2009a) showed definitively that Australian desert ants orient homeward using the panoramic skyline around a feeder. Graham and Cheng (2009a) surrounded a feeder with a facsimile of the natural skyline there, and when they rotated the artificial skyline, the ants' departures from the feeder toward the nest were rotated correspondingly. Graham and Cheng (2009b) also showed that the lower part of the panorama (below approximately 26 deg of elevation) is far more important for homing than the upper part, Julle-Daniere et al. (2014) showed that skyline height influences homing direction strongly, and Schultheiss et al. (2016a) showed that ants detect the skyline using the strong contrast between sky and ground in the ultraviolet, as suggested by Möller and Differt (Möller, 2002; Differt and Möller, 2015, 2016). Further, Collett and colleagues (Lent et al., 2013; Collett et al., 2014; Buehlmann et al., 2016; Woodgate et al., 2016) have exploited the convenient experimental system offered by ants navigating in small artificial arenas to analyze how black-and-white panoramas are perceived and processed by navigating ants.

Although ants and bees use many similar navigational mechanisms (Collett et al., 2006b, 2013; Srinivasan, 2015), they also differ in important ways. Indeed, even closely related species of ants have navigational strategies that are adaptively specialized to some degree for the animals' visual ecologies (Schwarz and Cheng, 2010; Bühlmann et al., 2011; Cheng et al., 2012, 2014; Schultheiss et al., 2016b). And of course worker ants walk, while bees fly, leading ants to measure distance mainly by counting steps (Wittlinger et al., 2006, 2007), while bees measure optical flow (reviewed in Collett et al., 2006a; Srinivasan, 2015). And unlike ants, bees can change viewpoints by simply gaining altitude. The role of altitude in bee orientation has hardly been studied, but by gaining sufficient altitude, the bees could take in the broadest possible views – especially views of the distant skyline – which would therefore have the largest possible catchment areas (Cartwright and Collett, 1987; Zeil et al., 2003; Collett et al., 2013).

However, it has yet to be shown definitively that honeybees actually use the panoramic skyline in visual orientation. Here, then, we addressed this question using the artificial-skyline technique of Graham and Cheng (2009a) combined with a modification of Najera and Jander's (2012) method for scoring the departures of bees from a feeder. Because ants clearly use ultraviolet light in distinguishing sky from ground (Schultheiss et al., 2016a), we also ensured that the sky and ground in our artificial skyline contrasted strongly in the ultraviolet. In addition, we did the critical tests under overcast skies to ensure that the bees could use only view-based navigation, as their celestial compass is unavailable under such conditions (Dyer and Gould, 1981; Towne et al., 2005; Towne and Moscrip, 2008; Dovey et al., 2013).

We worked at the Rhein Environmental Study Area of Kutztown University of Pennsylvania with a five-frame colony of mixed-race honeybees, Apis mellifera Linneaus. We set up a test arena in a clearing 45 m from the hive (Fig. 1) and created a 360 deg panoramic photograph from the center of the arena (Fig. 2). The panoramic photograph was assembled from 20 overlapping photographs taken sequentially in portrait orientation by rotating a digital 35 mm single-lens-reflex camera (Nikon D5100, focal length 26 mm) around the vertical post of a leveled tripod, with the camera 40 cm above the ground, the height of the feeder during the experiments. The photos were assembled automatically into the 360 deg panorama in Fig. 2 using Hugin, an open-source photo stitching program (version 2014.0.0; http://hugin.sourceforge.net).

Using Adobe Photoshop, the highest terrestrial object on the panoramic photograph was marked every 2 deg, and these points were connected with a black line to mark the top of the skyline. For the silhouette in Fig. 2B, the portion of the image below the skyline was filled with black using Photoshop, but for the artificial skyline used in the experiments (see Fig. 3A), the line alone was printed on a large piece of heavy white paper (Hewlett-Packard C6569C, Palo Alto, CA, USA) 1.07 m high and 9.42 m long on a large-format printer (Hewlett-Packard Designjet Z6200 42 in Photo Printer). The portion of the panorama above the line, the artificial sky, was then painted with two coats of an ultraviolet-reflecting white paint (Flock Off! UV Reflective White Paint, Make Em Move Manufacturing, Lakewood, NJ, USA), and the portion corresponding to terrestrial objects was painted flat black (Ultra Cover Latex Paint, Flat Black, Rust-oleum Corporation, Vernon Hills, IL, USA). The black areas were then treated with two coats of an ultraviolet-absorbing finish (Aquathane Clear U.V. Absorber, Flat, T.J. Ronan Paint Corporation, Bronx, NY, USA). The sky and ground in the finished panorama contrasted strongly in the ultraviolet: using an ultraviolet light meter (Omega HHUVA1) sensitive to wavelengths from 320 to 400 nm, approximating the spectral sensitivity of the bees' ultraviolet receptors (Peitsch et al., 1992), the white and black areas of the panorama were compared. With the meter held 10 cm from the vertical panorama in diffuse daylight, the white artificial sky reflected, on average, 8.6-fold more ultraviolet light than the black silhouette (N=6 measurements; s.d. 0.68-fold), which is within the range of differences between sky and non-sky objects observed in natural scenes by Möller (2002).

The test arena (Fig. 3A,B) was based on the method Najera and colleagues (Najera and Jander, 2012; Najera et al., 2012, 2014) used to record the departure directions of bees from a feeder at the center of a 1 m disk. Our technique differed from that of Najera and colleagues in that departures were recorded from above with a video camera for later analysis instead of being scored directly in real time (although Najera and Jander, 2012, made video recordings during one of their test sessions). The arena was made of 12 vertical poles spaced 30 deg apart around a circle 3 m in diameter. The paper panorama was attached to these poles by 24 straps fixed to the back of the panorama, 12 at the top and 12 at the bottom, using removable clips. In the center of the arena was a wooden frame holding a 1 m-diameter wooden disk 40 cm above the ground. The disk was painted white and marked with black radial lines dividing it into 12 sectors (Fig. 3A,C), and it was covered with a 1 m Plexiglas disk to protect it. A feeder offering unscented sugar water at the center of the disk (Fig. 3C) was made of a plastic dish 59 mm in diameter and 10 mm deep surrounded by a 7 mm bead of beeswax for the bees to stand on. Stretched across the top of the feeder was a fine nylon mesh with 1 mm openings through which the bees could drink but that otherwise kept the bees out of the sugar water. Most bees stood at the edge of the dish while feeding, but some stood on the mesh (Fig. 3C, inset).

Most bees exited the disk quickly after leaving the feeder: 94.8% exited within 2 s of departing the feeder when there was no barrier (N=87), and 72.1% did so when the artificial skyline was in place (N=710; see Fig. S1). We were not able to measure the height of the bees, but most seem to have stayed below the cone of space from which they could see the highest parts of the natural skyline over the top of the barriers (see the angled, dashed lines in Fig. 3B) before they exited the disk and were scored. Even if some bees did see the highest parts of the natural skyline (31 deg above the horizon) over the barriers, it may not have affected their orientation, as ants, at least, orient mainly using the parts of the natural panorama below approximately 26 deg (Graham and Cheng, 2009b). In any case, the bees' departure directions in the critical tests with the artificial skyline (see Results) show that they oriented by the artificial skyline alone in those tests, not the actual skyline or any other natural cues.

Eleven days before the first experiment, a group of bees was trained to visit the feeder offering unscented sugar water at the center of the test arena, 45 m from the hive. Bees were not individually marked, as the marks would not have been distinguishable on the video recordings. The bees were thereafter fed at the feeder daily for approximately 2–4 h each day, usually divided into two sessions. The concentration of the sugar solution was adjusted daily to maintain a population of about 15–20 bees standing on the feeder at any given time, which means that there were approximately four times as many bees – that is, 60–80 bees altogether – making regular round trips to the feeder. (The bees spent about 1 min at the feeder of the approximately 4 min complete circuit time – feeder to hive and back – so roughly a quarter of the bees were standing at the feeder at any given time.) Because the concentration of the food was adjusted daily to maintain a more-or-less constant number of bees at the feeder, there was little recruitment of new bees after the initial population of foragers was established, so most foragers eventually had considerable experience at the feeder.

When the bees were first trained, there was no barrier around the feeder, and for part of each feeding session thereafter, the bees foraged with a full view of the natural panorama. For the balance of the daily feedings, a barrier made of a thin, transparent vinyl sheet (110 cm high and 0.28 mm thick, Double Polished Clear Vinyl no. 1299, Nordic Shield Plastics, Oxford, MA, USA) was attached to the poles around the feeder so that the bees would learn to fly over such a barrier to get to the feeder, as they would need to do during experiments. The transparent barrier was intended to allow the bees to see the natural panorama from the feeder. However, the vinyl sheet transmitted only 68% of the ultraviolet light (based on measurements using the Omega HHUVA1 meter described above), and when the bees' departures from the feeder were later recorded with the barrier in place under overcast skies, the bees were very poorly oriented compared with when the barrier was absent under the same conditions (see Results). Thus, the vinyl barrier evidently obscured the natural panorama such that the bees could not use it for orientation under overcast skies, a difficulty probably exacerbated by the attenuation of the available ultraviolet light by the clouds. Nonetheless, the bees learned the skyline around the feeder during training, as shown below, probably when the screen was absent or as they approached or left the arena, and possibly when the screen was in place on sunny days, when the contrast in the ultraviolet would have been stronger.

The bees' departures from the feeder were recorded from above with a small video camera (GoPro Hero4 Silver, San Mateo, CA, USA) hanging 78 cm above the disk from a string suspended across the top of the arena. At that height, the field of view of the camera, which was set in wide-angle mode, encompassed the entire white disk (video still in Fig. 3C). The camera was operated remotely via Bluetooth wireless technology using a smart phone app made for the camera, and each recording session lasted 7 min, which is approximately how long the food lasted without replenishment. Two 30 s clips from the video recordings are provided in Movies 1 and 2. For two brief intervals during the 1 August trials (75 and 105 s), we detected the sun's disk through the clouds, and we excluded departures occurring during these intervals. Otherwise, no sun or blue sky was visible during the recordings.

The video recordings were analyzed according to a set of rules developed in pilot studies to maximize the number of measurements while excluding bees that were evidently not departing (for example, bees engaged in prolonged searching or circling) and bees that might have been influenced by other bees leaving immediately before them. The eight scoring rules were as follows. (1) An exit from the disk was scored when a bee that had been feeding at the feeder flew from the feeder and subsequently crossed the outer edge of the white disk. The sector from which the bee exited the disk was scored as the bee's exit direction. (2) The great majority of bees took flight directly from the feeder, but if a bee walked (usually a short distance) from the feeder, she was considered not to have departed until she took flight. (3) Any bee that departed from the feeder but returned to make contact with the feeder again without leaving the area was considered to have not yet departed the feeder. (4) If a bee crossed the outer edge of the disk but immediately returned, the first crossing was ignored, and only the bee's final exit was scored. (5) If a bee exited the disk exactly on one of the radial lines, the sector toward which the bee was moving was scored. (6) When two bees departed the feeder simultaneously, the bee that crossed the outer edge of the white disk first was scored; the other was ignored. (7) As soon as a scored bee completely exited the disk, the next bee to depart the feeder was scored, ignoring any bees that departed while the first bee was still in flight over the disk. However, if the first bee took more than 3 s to clear the disk, the next bee to depart 3 s after the first had departed was scored. In the latter case, any bees departing in the intervening 3 s were ignored. (8) Bees that took more than 7 s to completely clear the disk after departing from the feeder were ignored (these were rare; see Fig. S1).

We tested the inter-observer reliability of this scoring method by having two different observers (W.F.T. and A.E.R.) independently analyze five different recordings totaling 35 min of recordings and 360 departures. The results are shown in Fig. S2. With an average sample size of 72 departures, the absolute difference between the two observers in the measured mean direction was 3.3 deg (range 0.7–6.1 deg in the five recordings), the average difference in the circular concentration (r, a measure of dispersion) was 3.1% (range 0–6.8%) and the average difference in sample size was 3.1% (range 1.4–5.5%). All of the mean directions fell well within the relatively narrow 95% confidence intervals (range 17.1–28.1 deg wide) of the corresponding distributions from the other observer. Overall, then, the differences between observers were very small, especially when compared with the differences between predicted directions in the critical tests (90 deg or more; see Results). Therefore, our two observers were essentially equivalent for the purposes of this analysis, and only one observer analyzed the remaining recordings. For further analysis involving the five recordings analyzed by both observers, the distributions scored by A.E.R. (who also scored most of the others) were used.

Circular statistics were calculated using Oriana (Kovach Computing, Pentraeth, Isle of Anglesey, UK; www.kovcomp.co.uk), using mean vectors and ±95% confidence intervals to characterize distributions. The Rayleigh test was used to determine whether distributions were randomly or significantly oriented, and for significantly oriented distributions, the 95% confidence intervals were used to determine whether mean vectors coincided with predictions. Finally, the chi-squared test was used to determine whether two distributions differed from each other.

The experiments occurred over two overcast days, 1 and 6 August 2016, and included three control tests as well as tests with the artificial skyline.

In one control test, the bees had an unobstructed view of the natural panorama. The mean direction of the bees' departures (Fig. 4A, ‘no screen’) was 14.1 deg clockwise of the hive direction. The orientation was significant (P<<0.001, Rayleigh test, N=87), but the 95% confidence interval (3.0–25.3 deg) did not include the hive direction, indicating that the bees' rightward bias was significant. Indeed, as we show below, a similar rightward bias occurred in all experiments in which the bees were orienting by the full natural panorama (Fig. 4A, no-screen condition) or by the artificial skyline (Fig. 4B). The nature of this bias is considered further in the Discussion. Meanwhile, in presenting and analyzing the results of the critical tests with the artificial skyline (Fig. 4B), we incorporated the bias by expecting the bees to depart not directly toward the fictive hive but 14.1 deg to its right, based on the control test.

In a second control test, the feeder was surrounded by an opaque, unpainted all-white paper screen the same size as the artificial skyline (Fig. 4A, ‘white screen’). The orientation was very scattered but not random and was directed, on average, 2.3 deg clockwise of the hive (P=0.004, Rayleigh test, N=92), and the 95% confidence interval (−31.2 to +35.7 deg) included both the hive direction and the rightward-biased mean direction from the no-screen trial (14.1 deg). The distribution was significantly different from that obtained in the no-screen trial (chi-squared test, P<<0.001), so obscuring the natural panorama degraded the bees' orientation. We cannot rule out that some bees detected celestial cues through the cloud cover, although no sun or blue sky was visible to us during the trial. Nor can we rule out that some bees flew high enough above the feeder to see the highest parts of the natural panorama over the top of the screen before they exited the disk (see Materials and methods and Fig. 3B). However, the results of tests with the artificial skyline, discussed below (Fig. 4B), seem to rule out that the bees used any cues other than the artificial skyline during those tests.

In a third control test, a clear vinyl sheet the same size as the other barriers surrounded the feeder (Fig. 4A, ‘vinyl screen’). Here, the orientation was not statistically distinguishable from random (P=0.17, Rayleigh test, N=178; mean vector: −34.5 deg). The vinyl barrier was transparent to us, but it blocked 32% of the ultraviolet light in the range to which bees are sensitive (see Materials and methods), which therefore considerably reduced the contrast of the skyline in the ultraviolet as viewed by the bees through the screen, a contrast that was probably already reduced by the cloud cover. If bees detect the skyline using its strong contrast in the ultraviolet, as do ants (Schultheiss et al., 2016a), then this could explain why the bees' orientation was poor. However, it is puzzling that the orientation here was worse than it was with the opaque, all-white screen. Regardless, the results of these two tests show that the all-white and vinyl screens both degraded the bees' orientation compared with when these barriers were absent.

To critically test the hypothesis that bees orient by the panoramic skyline, we measured the bees' exits when the artificial skyline was set up in various orientations around the feeder. These tests were intermingled with the control tests under the same overcast skies on 1 and 6 August 2017.

The results of three trials in which the artificial skyline was rotated 5 deg clockwise relative to the natural skyline are shown in Fig. 4B (top), where the histogram is rotated so that the predicted departure direction, accounting for the bees' 14.1 deg rightward bias (from Fig. 4A, no-screen condition), is directly upwards (0 deg, red dashed line). This makes the direction toward the actual hive (black dashed line) 19.1 deg left of vertical, because of the bees' 14.1 deg clockwise bias plus the 5 deg clockwise rotation of the artificial skyline. The orientation was highly significant (mean vector, 0.2 deg; P<<0.001, Rayleigh test, N=196), and the 95% confidence interval (−5.4 to +5.8 deg) did not include the hive's actual direction (−19.1 deg), confirming the bees' rightward bias. The 95% confidence interval included both the predicted direction based on the bees' use of the artificial skyline (0 deg) and the predicted direction based on the bees' use of celestial or natural terrestrial cues (−5 deg), as the two predictions were only 5 deg apart. Overall, then, this trial confirmed the bees' rightward bias but did not effectively test the hypothesis that bees orient using the skyline.

In three further tests, therefore, the artificial skyline was rotated an additional 90, 180 or 270 deg. The results (Fig. 4B, right, bottom and left) show that the bees' departure directions were rotated almost exactly as predicted by the orientation of the artificial skyline in each case. The direction toward the hive (black dashed lines) remains at −19.1 deg throughout (as explained above), but the predicted directions (red dashed lines) rotate 90, 180 and 270 deg clockwise moving clockwise through the panels. In all three cases, the orientation was strong and significant, and the 95% confidence intervals were narrow (≤17.8 deg) and encompassed the directions predicted by the artificial skyline. Table 1 shows the relevant statistics for all three experiments, which together consisted of 7 trials and 514 individual measurements.

The experiments with the artificial skyline rotated 90 deg or more (Fig. 4B, right, bottom and left) confirmed the bees' approximately 14 deg rightward bias and showed that the bees oriented using the artificial skyline. Further, very few bees in these tests could have used celestial cues or views of the natural skyline seen over the barrier, as bees using such cues would have exited 14.1 deg to the right of the hive (5 deg left of directly upwards in the histograms), which very few bees did except when the prediction based on the artificial skyline was near that direction (Fig. 4B, left). Therefore, celestial cues or views over the barrier were either unavailable or unimportant to the bees in these tests.

We observed bees as they departed from a feeder under overcast skies to test the hypothesis that they can use the panoramic skyline in view-based navigation. In control tests, we found that bees with a full view of the natural panorama oriented homeward, although with a small (14.1 deg), repeatable rightward bias (Fig. 4A, no-screen condition). When terrestrial cues were obscured by an opaque, all-white screen or a semi-transparent vinyl sheet (Fig. 4A), the orientation was degraded. And when an artificial, black-and-white replica of the natural skyline that contrasted strongly in the ultraviolet was rotated into various orientations around the feeder, the bees' departures were rotated correspondingly, with their usual rightward bias (Fig. 4B). This shows that bees use the panoramic skyline in orientation.

What might account for the bees' consistent rightward bias as they departed the arena for the hive? Both bees and ants sometimes deviate from the shortest possible beeline routes between their nests and food sources because they are attracted toward conspicuous vertical landmarks, or beacons, along the routes (bees: von Frisch, 1967, p. 331–333; Chittka et al., 1995a; ants: Graham et al., 2003), a behavior therefore called beacon-aiming (Collett, 1996; Graham et al., 2003) or beaconing (Cheng, 2006, 2012). The largest object near our test arena was a vine-covered, chain-link fence to the bees' right as they departed for the hive (see Fig. 1). Vines and plants growing inside the fence made it look like a solid wall. In the panoramic photograph in Fig. 2A, the hive is indicated with a white arrow, the part of the fence nearest the bees' flight route is a 2.9 m-high corner post marked by a gray line (Fig. 1 also shows the fence corner near the flight route), and the bees' mean departure direction with a view of the natural panorama (Fig. 4A, no-screen condition) is shown with a gray circle. The bees' mean departure direction was just 2 deg to the left of the corner post, the part of the fence nearest their flight route. Thus, the departure bias probably arose from beacon-aiming. However, the rightward bias here occurred even when the bees oriented by the artificial skyline alone (Fig. 4B), where the beacon was not evident because the fence was too low to contribute to the skyline. How could the bees orient toward a beacon without seeing it?

Beacon-aiming can divide a route into a series of segments punctuated by the beacons, which probably makes the route easier to follow and to recover if lost (Chittka et al., 1995b; Collett, 1996; Collett et al., 2003; Graham et al., 2003). Once a route is established, bees learn the sequence, lengths and compass directions of the individual segments (Chittka et al., 1995a,b; Collett et al., 1996, 2002; Srinivasan et al., 1997; Vladusich et al., 2005), called local vectors to distinguish them from the homeward global vectors continually tallied by path integration (Collett et al., 1998). Bees and some ants show very similar behavior in this respect (see reviews by Collett and Collett, 2002; Collett et al., 2003, 2006b; Cheng, 2012). Ants learn the distances and directions of the local vectors in relation to the surrounding visual panorama, so that they can later follow the vectors even in the absence of the landmarks (Graham et al., 2003; Legge et al., 2010; Wystrach et al., 2011b). Bees, too, have been shown to associate panoramic visual cues with the lengths (Collett et al., 2002) and directions (Collett et al., 1996, 1997) of route segments in tunnels and mazes (reviewed in Collett and Collett, 2002), but the steering of our bees toward the corner of the fence using only the artificial skyline (Fig. 4B) seems to represent the first clear case of bees orienting a local vector by natural panoramic cues.

The current experiments show that bees use the skyline in relatively short-range orientation, but as we discussed in the Introduction, there is strong, suggestive evidence that they use it in medium- (Towne and Moscrip, 2008; Dovey et al., 2013) and long-range (Southwick and Buchmann, 1995; Pahl et al., 2011) orientation as well. Also, bees' eyes, like those of many other insects with similar visual needs, seem well suited for detecting small skyline features, as they have a horizontal band of increased resolution looking out at the horizon, with closer angular spacing of ommatidia in the vertical than in the horizontal plane (Land, 1989, 1997a; Horridge, 2005; see especially fig. 6 in Land, 1989, and figs 1 and 4 in Horridge, 2005, for the case of honeybee workers). Indeed, as Zollikofer et al. (1995, p. 1645) have suggested for Cataglyphis ants, these horizontal foveal bands may be useful for ‘tracking distant landmarks on the horizon’.

To use the skyline in long-distance, view-based homing from unfamiliar sites, bees must recognize landscape features first learned closer to the nest (Dyer, 1991, 1996; Dyer et al., 1993; Southwick and Buchmann, 1995; Pahl et al., 2011; Cheung et al., 2014). The skyline would be especially useful in this context because it is the most distant landscape feature and therefore has the greatest possible catchment area (Cartwright and Collett, 1987; Zeil et al., 2003; Stürzl and Zeil, 2007; Wystrach et al., 2016). Further, viewing the skyline from high above the ground – above trees, small hills and other obstructions – effectively filters out all but the most distant features, maximizing the usefulness of the view for long-range homing (Cartwright and Collett, 1987; Collett et al., 2013) without increasing the amount of visual information that needs to be stored. Bees do learn the local terrain rapidly on their first orientation flights (von Frisch, 1967; Capaldi and Dyer, 1999; Capaldi et al., 2000; Degen et al., 2015, 2016), but we know little about how high they fly or whether this height is affected by trees, hills or other objects, aside from Dovey et al.’s (2013) observation that bees from a hive newly transplanted to a forested site spiraled 25–30 m directly upward toward the canopy on their first flights there. Bees do travel relatively short horizontal distances from the hive during their earliest orientation flights (Capaldi et al., 2000; Degen et al., 2015), but we do not know if they travel high above the nest during these flights.

Whether and how bees use the distant skyline in long-range homing also bears on the ongoing debate about whether bees have cognitive maps, in which landmarks are hypothesized to be organized into map-like spatial representations in the bee's brain. In the latest round of this debate, proponents of the different views disagree as to whether the orientation of bees in a study by Cheeseman et al. (2014a) can be accounted for by view-based navigation using the skyline (Cheung et al., 2014) or, instead, required bees to use cognitive maps (Cheeseman et al., 2014b). The original study assumed that the bees could not have used the skyline because the skyline features available were less than 2 deg high, too low for the bees to resolve according to some estimates of the bees' visual acuity (Cheeseman et al., 2014b; Menzel and Greggers, 2015).

However, bees might well be able to detect skyline features that protrude above the horizon by less than 2 deg. The effective resolution of an insect's visual system cannot be predicted based on the physical properties of the eyes alone (Land, 1997a; Horridge, 2003, 2005, 2009), so we rely instead on behavioral tests, such as measuring the finest black-and-white striped gratings that can be resolved. The finest such gratings resolved by bees in a detailed Y-maze study by Srinivasan and Lehrer (1988), summarized by Srinivasan (2010), had a frequency of approximately 0.35 cycles deg⁻¹ or a grating period (comprising one black and one white stripe) of 2.86 deg and thus a stripe width of 1.43 deg. Srinivasan (2011, p. 416) therefore says that ‘bees cannot resolve black-and-white stripes that are thinner than 1.4 deg’. The finest gratings resolved in similar Y-maze studies by Horridge (2003) and Warrant et al. (1996) were slightly thinner (1.25 deg) and slightly thicker (2 deg), respectively – differences that could be accounted for by differences in the methods used. But the thinnest stripes ever resolved by bees in grating studies were approximately 1 deg wide in a study by Hecht and Wolf (1929), and were presented to the bees from underneath, forcing them to use the ventral parts of their eyes, not their horizontal foveal bands.

Further, the skyline consists of large bright areas next to large dark areas, not striped gratings, and Land (1997b, p. 90), quoted in Horridge (2005, p. 262), has pointed out ‘the need for caution in extrapolating [visual resolution] from one task to another’ because, as Horridge (2005, p. 262) himself wrote, ‘the resolution depends on the task’. Regarding skylines in particular, Cheung et al. (2014) have shown that receptors with angular separations and receptive fields of 4 deg can obtain useful navigational information from skyline features lower than 2 deg, if the receptors give graded responses when their receptive fields span sky/non-sky borders. Perhaps this explains how Cataglyphis fortis ants seem able to orient by skyline features that protrude as little as 2 deg (Wehner et al., 1996) or less (Huber and Knaden, 2015) above the horizon even though their eyes have interommatidial separations in the foveal band of 3 deg – or even larger, as this 3 deg measurement is from C. bicolor, a slightly larger congener of C. fortis, and smaller ants tend to have larger interommatidial angles (Zollikofer et al., 1995). Similarly, Macquart et al. (2006) describe Gigantiops ants orienting toward small holes subtending only 1.5 deg, and although Gigantiops have the largest eyes among ants (4137±252 ommatidia; Gronenberg and Hölldobler, 1999), they are smaller than those of worker honeybees (A. mellifera, 5375±143 ommatidia; Streinzer et al., 2013). Horridge (2005) cites several additional cases of insects detecting small objects or distinguishing fine gratings using eyes with interommatidial angles much greater than the objects or grating periods detected, and he suggests (Horridge, 2009) at least one neural mechanism by which feature detectors could have smaller receptive fields than the original receptors whose inputs they use.

Further still, the resolution measured in grating experiments can depend strongly on the brightness and spectral composition of the light used (Hecht and Wolf, 1929; Warrant et al., 1996; Land, 1997a; Horridge, 2003; Hempel de Ibarra et al., 2014), and unlike natural skylines, the ambient light in Srinivasan and Lehrer's (1988) and Horridge's (2003) Y-maze studies contained almost no ultraviolet light. This is not a problem with these studies, of course, but it suggests a further need for caution in applying their findings quantitatively to predict the bees' resolution of natural skyline features. Finally, a recent physiological study of the bees' light-adapted eyes shows that individual photoreceptors in the highest-acuity part of the eyes can reliably detect objects subtending only 0.6 deg (Rigosi et al., 2017), implying considerably higher acuity than has previously been suspected. For now, the limit of the bees' ability to detect small, natural skyline features is unknown.

Regarding visual resolution, Wystrach et al. (2016) have shown with simulations that, perhaps surprisingly, low visual resolution sometimes outperforms higher resolution in helping an ant to recover a route from which it has been displaced. This is because the most useful landscape features for such purposes – large, distant ones – usually have low spatial frequencies, and views containing high spatial frequencies can yield more spurious matches between current and remembered views. However, in the case of insects that sometimes need to navigate using very low skyline features, like bees (e.g. Cheung et al., 2014) and some desert ants (Wehner et al., 1996; Huber and Knaden, 2015; Cheng et al., 2014), the highest practical resolution in the vertical dimension might well be needed to maximize the detection of such features, although high resolution in the horizontal dimension is probably much less important.

The white sky and black ground of the artificial skyline used here contrasted strongly in the ultraviolet because of the ultraviolet-reflecting white paint used for the sky and the ultraviolet-absorbing top coat covering the black silhouette. These treatments more than doubled the skyline's contrast in the ultraviolet (see Materials and methods). In 2014 and 2015, we performed some preliminary tests with a paper panorama lacking the ultraviolet treatments, and while the bees showed significant orientation overall in the predicted directions (W.F.T., A.E.R., A.E. and D.F.B., unpublished observations), the orientation was very much more scattered than we report above for the treated panorama. Other methodological differences preclude any firm conclusions based on this comparison, but our observations are at least entirely consistent with an important role for ultraviolet light in the bees' separation of sky from non-sky, as in ants (Schultheiss et al., 2016a).

Bees and ants normally use path integration and view-based navigation together, making it difficult to study these systems separately. In the current experiments, we isolated view-based navigation by waiting for overcast skies to block celestial cues. An alternative way to isolate view-based navigation is to use foragers captured at the nest at the end of a foraging trip, as their homeward path-integration vector will have been run down to zero by the return trip. Such zero-vector foragers, lacking a homeward vector, rely on view-based navigation alone to orient homeward upon release from distant sites (Wehner and Srinivasan, 1981; Wehner et al., 1996). Zero-vector ants have been used very often in experiments (e.g. Fukushi and Wehner, 2004; Narendra, 2007; Graham and Cheng, 2009a,b; Collett, 2014; Legge et al., 2014; Wystrach et al., 2011b, 2012; Schwarz et al., 2014, 2017) but zero-vector bees only rarely (Wehner et al., 1990, 1996), probably because the visual panorama available to ants is more easily manipulated, and ants are more easily tracked upon release. Bees have, however, been tracked with radar (e.g. Cheeseman et al., 2014a; Degen et al., 2016) or had their vanishing bearings (Dyer, 1991; Dyer et al., 2002; Chittka et al., 1995a) or homing speeds and success rates (Capaldi and Dyer, 1999) measured, techniques that are effective but cumbersome compared with the arena technique (Najera and Jander, 2012; Najera et al., 2012, 2014; Palikij et al., 2012).

The arena technique could be used with zero-vector bees to study visual navigation in isolation, then, by capturing bees on their return to the hive, transporting them passively to the arena, and scoring their departures in the presence of any visual panorama one might wish to present there. For example, one could replicate and extend the cloudy-day, artificial-skyline tests reported here (Fig. 4B) but on sunny days using zero-vector bees. This, in fact, is how Graham and Cheng (2009a) showed that ants could orient by an artificial skyline. Or one could ask whether bees use the skyline to determine their location – not just their orientation, as we did here – as visual panoramas contain both directional and positional information (Zeil et al., 2003; Zeil, 2012). For this, one would need to present a skyline appropriate not for the arena location, as we did here, but for a nearby site. If bees use the skyline to determine location, then such bees would be expected to depart as if from the nearby site, not from the arena.

We thank Ed Vitz for helping to build the apparatus; Ed Vitz, Jordan Kemfort and Ryan Geisler for help with developing methods; Robyn Underwood and Steve Finke for supplying bees and help with beekeeping; Fred Esposto and Ezry St Iago-McRae for help with constructing one of the paper panoramas; Rich Courtney for help with printing skylines; Chris Sacchi for advice on statistics; and two anonymous reviewers for helpful suggestions on the manuscript.