Luminance-dependent visual processing enables moth flight in low light

See allHide authors and affiliations

Science  12 Jun 2015:
Vol. 348, Issue 6240, pp. 1245-1248
DOI: 10.1126/science.aaa3042

Not too fast and not too slow

Moths are typically active during dawn and dusk when light levels are low and vision is challenging. Slower visual response times can allow for greater light sensitivity, but flying insects are both moving and tracking moving targets, making such tradeoffs potentially problematic. Using a combination of modeling and experiments, Sponberg et al. show that moths are able to avoid this potential decrease in visual acuity (see the Perspective by Warrant). This is because the point at which their perception of movement would be compromised is just above the natural frequency at which flowers sway. Thus, insect vision is precisely adapted to the light and movement conditions of their environment.

Science, this issue p. 1245; see also p. 1212


Animals must operate under an enormous range of light intensities. Nocturnal and twilight flying insects are hypothesized to compensate for dim conditions by integrating light over longer times. This slowing of visual processing would increase light sensitivity but should also reduce movement response times. Using freely hovering moths tracking robotic moving flowers, we showed that the moth’s visual processing does slow in dim light. These longer response times are consistent with models of how visual neurons enhance sensitivity at low light intensities, but they could pose a challenge for moths feeding from swaying flowers. Dusk-foraging moths avoid this sensorimotor tradeoff; their nervous systems slow down but not so much as to interfere with their ability to track the movements of real wind-blown flowers.

Most animals rely on vision to maneuver through complex environments. Similarly, many engineered systems use machine vision to sense their surroundings. Yet all biological or synthetic visual systems operating in natural environments must deal with high variability in ambient light intensity. Animals can encounter light intensities that vary over 10 billion–fold within a single day (13). Low-light specializations can extend an animal’s sensory capabilities but frequently produce tradeoffs, (such as increased sensitivity but reduced resolution) (46). Revealing the mechanisms that enable nervous systems to adapt to this vast range of light, as well as their functional tradeoffs, are fundamental to understanding the versatility of vision.

We investigated visual processing in dim light—and its functional consequences—in the hawkmoth Manduca sexta, an agile flyer that extracts nectar from moving flowers while hovering even in very dim light (Fig. 1A) (79). The moth’s compound eye has an adjustable pseudo-pupil that allows a large number of the highly refractive eye facets (ommatidia) to deliver light to a single photoreceptor in low light, thereby increasing sensitivity (1, 2). Early in visual processing, neurons may also pool inputs from multiple photoreceptors providing spatial summation of light (1, 10). These spatial adjustments improve sensitivity, but they are insufficient to account for the range of low-light sensitivity exhibited by neurons in the moth’s visual system (10). A hypothesized complementary mechanism is that moths may progressively slow their visual processing as light intensity (luminance) decreases (1, 10, 11). In effect, as light diminishes, moths might temporally integrate light for a longer period of time, in addition to summing over larger spatial regions (Fig. 1B) (10, 12). This luminance-dependent neural processing predicts a tradeoff: The increased sensitivity that comes with increased processing time should also slow motion detection. This predicts that moths should lag behind the motion of objects that they are visually tracking. Because moths track and feed from moving flowers (79), we can test not only whether the moth’s motion response is slowed, but also what potential tradeoffs this poses for tracking natural flower movements.

Fig. 1 The effect of light intensity on moths tracking robotic flowers.

(A) Human photoreceptors, like those of all animals, are capable of detecting even single photons (28). However, human color vision (colored arc), as well as our ability to resolve motion and spatial detail, deteriorates below the photopic visual threshold (1 to 10 cd m−2) corresponding to light levels at dusk (11, 29). This is also true for diurnal insects such as the blowfly Calliphora (30). Human scotopic vision (gray arc) is strictly monochromatic (29). The hawkmoth Deilephila is truly nocturnal, with color vision throughout much of the scotopic range (2). [In (A), “*” indicates experimental light levels. “?” indicates that data are not available.] Manduca is crepuscular and is hypothesized to adjust its visual processing (B) in order to visually track flowers over its large range of light intensity (>106 cd m−2) (10, 27). (CNS, central nervous system.) Robotic, three-dimentional printed flowers generated repeatable moth flight maneuvers (C) (movies S1 and S2). We moved the flowers with a trajectory of many superimposed sinewaves to sample many of the frequencies of movement simultaneously (D). Fourier transformations (E and F) of the flower’s (green) and moth’s (blue) movements show high coherence [(E), gray line], which is the normalized cross-power spectral density (18, 24). Flower movements were prescribed to have equal peak velocities at each frequency (F), which helps avoid saturation in the moth’s ability to track.

We used robotic artificial flowers to explore the moth’s behavior under different light levels (Fig. 1, A and C, and movies S1 and S2) (13). Moths tracked flowers moving side-to-side with a linear combination of sinusoidal trajectories with frequencies spanning two orders of magnitude (0.2 to 20 Hz, Fig. 1, D to F) (14, 15). Hovering and maneuvering in midair is demanding, inherently unstable, and energetically costly (9, 16, 17). Nonetheless, moths are able to track and feed from moving flowers even at 14 Hz (Figs. 1E and 2A). In contrast to tethered moths (open-loop) tracking visual images (18), flower tracking by freely flying (closed-loop) moths is highly linear (coherence in Fig. 1E; explicit test in fig. S2).

Fig. 2 Visual processing slows in low light.

The relative amplitudes of moth and flower give the tracking gain (A), and their relative timing gives the phase difference (B) with means ± 95% confidence intervals (CIs). The qualitative shape of the response is consistent with the drift-compensation response in the diurnal hawkmoth Macroglossum (8). Regressing phase onto frequency [(B), dash-dotted lines] estimates the best-fit time constant across all frequencies (C) or just those <10 Hz (D). We created a prediction for the low-luminance response (light blue) from a closed-loop model of the high-luminance response (orange) with a delayed sensory gain (fig. S3). This prediction successfully captures the overshooting (E) and phase shift (F) evident in the actual low-light response (dark blue) replotted from (A) and (B) on a log axis.

Moths lag farther behind flower movement in dim moonlight [0.3 lux (lx)] than in the brighter light of early dusk (300 lx), as predicted for a slowing visual system (Fig. 2B; F = 46.3, df = 349, P < 0.0001). At both light levels, moths produce the largest motions (highest gain) to flower movements at intermediate frequencies (1- to 2-Hz bandpass, Fig. 2A and fig. S1). However, in low light they begin overshooting the flower (gain > 1, which is 0 dB) (all frequencies: F = 16.4, df = 349, P < 0.0001; at 1.7 Hz: t = 4.5, df = 20.5, P = 0.0002). Because the mechanics of flight are the same at both light levels, the differences in tracking must arise because the moth relies on a luminance-dependent neural process.

To quantify the change in processing time, we can estimate a single time constant from the phase response at each light level (Fig. 2B). Overall, the moth’s tracking response is 17% slower in dim moonlight than at early dusk: 83.3 ± 2.4 ms as compared to 71.5 ± 3.5 ms, respectively (mean ± SEM; Fig. 2, B and C; t = 2.82; df = 21; P = 0.01). We obtained similar results when we considered only frequencies below 10 Hz, which was the highest frequency that all individual moths could track (67.2 ± 3.1 and 80.6 ± 3.3 ms; Fig. 2D).

Next we tested whether the slower tracking responses were consistent with a luminance-dependent delay within the moth’s visual system. Moths track flowers in closed loop, meaning that their visual input is the difference between their own motion and the flower’s (visual “error”; fig. S3) (13, 15). We applied a luminance-dependent delay to the visual response within this closed feedback loop (fig. S3) (15). From this, we predicted the moth’s response in low light, Embedded Image, for each frequency component of the flower’s movement (defined by the Laplace variable s). This prediction (Fig. 2, light blue lines) depends only on the measured high-luminance response, Embedded Image, and a single parameter τ, representing the difference in processing time (delay) between the two light levels

Embedded Image(1)

A delay (τ) of 10.4 ms best fits the change in response from the bright to dim light levels. This single parameter captures the empirical phase differences as well as the overshooting that occurs under low light (Fig. 2, E and F). Hence, the hypothesized luminance-dependent visual processing in the moth’s nervous system can account for the differences in the closed-loop tracking response at different light intensities.

The above model treats the change in processing as a single time delay. Changes in spatial processing are likely to help shape the moth’s tracking response, especially the gain (10, 19). Temporal processing might also be distributed across multiple regions of the visual system, including in the photoreceptors and motion detection circuits (1, 10, 19). Most importantly, insect motion detection itself is thought to arise by the correlation of adjacent photoreceptors (20, 21), and the process of correlation is nonlinear. What remains unclear is how these more complicated processes would affect moth motion lag if their time constants were luminance-dependent. To be consistent with our simplified models, a more detailed model would need to produce a lag in its input-output relationship that increases linearly with the time constant inside the model. To test for this relationship, we implemented a longstanding model proposed for visual motion detection, the elaborated Hassenstein-Reichardt (HR) correlator (see the supplementary text). We simulated this model’s response to an oscillating visual input and explored how this response depends on the model’s time constants. (10, 12, 20, 21). These simulations show that the phase shifts we observed in experiments are consistent with the HR model of insect visual processing, as well as with simple closed-loop delay models (fig. S5).

In all these analyses, a simple luminance delay term accounts for the closed-loop response (Fig. 2, E and F). The increased phase lag in low light supports the hypothesized luminance-dependent visual processing. However, in order to consider overall tracking performance, we cannot consider phase lags and gain separately, because they each describe only one aspect of the response. To accomplish our second aim of assessing a potential performance tradeoff, we combined gain and phase into one metric of tracking error (ε). Specifically, we used the distance in the complex plane between the moth’s responses and the ideal tracking conditions (gain = 1, phase = 0°; fig. S4) (13, 14)Embedded Image (2)where Embedded Image

Although tracking performance depends on light intensity, the effect is frequency-dependent and separates into three distinct frequency bands (Fig. 3A). Below 1.7 Hz, tracking error is relatively low (ε < 1) and does not differ between high- and low-luminance conditions (F = 0.49, df =137, P = 0.48). However, tracking error increases rapidly at higher frequencies, even when gain remains high. This is because the moth is more than 90° out of phase with the flower between 1.7 and 8 Hz. In this frequency band, moths track significantly worse in low light (F = 17.6, df =192, P < 0.0001). However, at the very highest frequencies (>8 Hz), tracking error reverts toward unity because the gain is very small.

Fig. 3 Tracking performance and real flower movements.

Tracking error [(A), mean ± 95% CI] is a function of frequency for both luminance conditions. Three distinct frequency bands result, denoted by dashed lines. Behavioral performance of moths (B) was scored as no flight (~F); flight, but not tracking (F, ~T); or flight and tracking (T). The power spectra (C) of hawkmoth-pollinated flowers blowing in breezes from 0.1 to 2.7 m/s are normalized to the total power. 94% of the cumulative power in the flower’s movement (black: mean) occurs in frequencies below 1.7 Hz, where ε ≈ 0.2 to 0.4 (D).

Counterintuitively, moths are still much more likely to forage and track moving flowers in dim light levels, despite poorer tracking performance (higher ε; Fig. 3B). This suggests that moths incur worse performance in order to obtain the increased sensitivity afforded by luminance-dependent neural processing. However, moths might avoid this tradeoff if the flowers they track only move at frequencies below ~2 Hz, below where the performance tradeoff occurs.

We filmed five species of hawkmoth-pollinated flowers in natural wind (see the supplementary text, fig. S6, and movie S3) and found that these flowers move at frequencies where the moth does not suffer a performance cost for tracking in dim moonlight conditions. The vast majority (94%) of motion in all the flowers measured was at frequencies below 1.7 Hz (Fig. 3, C and D, and fig. S6). So whereas moths benefit from the increased sensitivity of slower visual processing in low light, they also avoid negative consequences, because natural flowers usually oscillate in the wind at speeds below those that incur a performance tradeoff.

This matching of frequencies between pollinator vision and flower movement indicates that the relationship between the two organisms probably constrains the slowing of neural processing in variable light conditions. Notably, moths’ tracking error increases dramatically above the frequencies of normal flower motion, regardless of light level (Fig. 3C). When the robotic flower moved at frequencies above 2 Hz, the moth would have tracked better by remaining stationary (ε > 1; Fig. 3, A and D). If moths’ visual processing was even slower, they would begin to experience adverse tracking performance in the frequency band where natural flowers move. One reason moths might rely on both temporal and spatial adjustments to deal with low light (10) is to limit the costs to both motion-tracking performance and spatial resolution.

The frequencies with which a moth can maneuver could provide a selective pressure on the biomechanics of flowers to avoid producing floral movements faster than those that the moth can track in low light (22). The converse interaction—flower motions selecting on the moth—could also be important, suggesting a coevolutionary relationship between pollinator and plant that extends beyond color, odor, and spatial features (23) to include motion dynamics.

The emerging use of system identification to connect open- and closed-loop experiments (6, 14, 15, 18, 24, 25) provides a useful paradigm for exploring sensorimotor strategies in many systems. Robotic models enable rapid, repeatable experiments that extract critical features of the biological system (26) and extend the physical modeling toolkit that has been useful for teasing apart pollinator-plant interactions (7, 9, 23). Here the robotic flower enabled us to test predictions about closed-loop behavior from open-loop electrophysiological results and models of neural processing (6, 15).

The dual demands of acquiring reliable sensory information and maintaining motor performance are a general challenge, especially for animals such as Manduca, which operate in impoverished sensory environments (46, 27) and on the edge of flight instability (6, 25). Matching the requirements of the motor system to constraints imposed by the dynamics of the environment can provide strategies that enable more extreme sensory performance while averting tradeoffs in motor performance.

Supplementary Materials

Materials and Methods

Supplementary Text

Figs. S1 to S6

Tables S1

References (3165)

Movies S1 to S3

References and Notes

  1. See the supplementary materials and methods.
  2. Acknowledgments: J. Riffell kindly provided the artificial Datura wrightii scent. We are also grateful for valuable contributions from J. Lockey, M. Salcedo, S. Kazi, A. Molback, A. Hinterwirth, A. Mountcastle, E. Roth, and N. Cowan. Data are available on Dryad, accession no. 10.5061/dryad.jd7b9. This project was supported by the Komen Endowed Chair, an Air Force Research Laboratory grant (FA8651-13-1-0004), an Office of Naval Research Multi University Research Initiative grant (N00014-10-1-0952), an Army Research Office grant (W911NF-13-1-0435), and an Air Force Office of Scientific Research grant (FA9550-14-1-0398) to T.L.D, and by NSF fellowship DBI-1103768 to J.P.D.
View Abstract

Stay Connected to Science

Navigate This Article