The effect of ecological factors on eye morphology in the western rainbowfish, Melanotaenia australis

The size, external morphology and position of the eyes can reveal a great deal about an animal's behavioural ecology (Walls, 1942; Lythgoe, 1979). Many animals rely on vision for fundamental behaviours such as foraging, mate attraction and predator avoidance, and their visual capabilities are strongly linked with eye morphology. For example, eye size has a major influence on visual resolution, as a larger eye can house more photoreceptors and will have a longer focal length compared with a smaller eye (Hughes, 1977; Land and Nilsson, 2012). Larger eyes can also enhance visual sensitivity compared to smaller eyes. The size of the aperture of an eye (the pupil) dictates how much light can enter, and animals that live in dim light conditions (e.g. nocturnal species and deep-sea animals), tend to have large eyes with large pupils (Warrant, 2004; Land and Nilsson, 2012). However, there are also costs associated with having large eyes, such as the energetic cost of manufacturing and maintaining the many millions of nerve cells within the retina, the hydrodynamic or aerodynamic costs incurred by increased weight or drag associated with large eyes, and the cost of repairing damage to the eye (Hiller-Adams and Case, 1988; Laughlin, 2001; Niven and Laughlin, 2008). Investment in the visual system relative to other body parts should therefore be specific to the visual requirements of a species (Laughlin, 2001; Niven and Laughlin, 2008), and there is an evolutionary trend towards large eyes in species that rely on vision (Walls, 1942; Howland et al., 2004; Lisney and Collin, 2007; Land and Nilsson, 2012).

Other aspects of eye morphology are also linked with an animal's behaviour. For example, the diameter of the pupil or the cornea (which sets the upper limit on the size of the pupil) relative to eye size is a consistent and useful predictor of activity patterns in a broad range of animals. Increasing the size of the pupil is an adaptation for improving visual sensitivity; therefore, if all factors other than aperture size (e.g. eye size, corneal transmission) are held constant, an eye with a relatively larger aperture will have a greater light gathering capability than an eye with a proportionally smaller aperture (Kirk, 2004). In fishes, lizards, birds and mammals, species that are crepuscular or nocturnal have relatively larger corneas/pupils than diurnal species (Kirk, 2004; Hall and Ross, 2007; Hall, 2008; Schmitz and Wainwright, 2011; Veilleux and Lewis, 2011; Lisney et al., 2012). Eye size, and the size of the aperture, can therefore be used to indicate a species' reliance on vision for behavioural tasks, and its diurnal/nocturnal activity patterns.

The visual field is a key determinant of animal vision, because it determines the volume of space that can be imaged upon the two retinas, and hence the amount of information that can be extracted at any one time (Martin, 2007, 2014). In vertebrates, the size and shape of the visual field is predominantly determined by the position of the eyes in the head, along with other factors, such as the shape of the skull, depth of the eye socket and eye mobility (Collin and Shand, 2003). Animals with frontally positioned eyes, such as primates, cats and owls, have extensive binocular overlap in their frontal visual field, so the region of highest visual acuity projects forwards (Hughes, 1977; Martin, 2007, 2014). In contrast, animals with laterally positioned eyes, such as rabbits and most birds, have extensive monocular visual fields, and a narrow region of binocular overlap, thus the regions of highest acuity project laterally (Hughes, 1977; Martin, 2007, 2014). In fishes, the position of the eyes is related to habitat, with laterally positioned eyes associated with pelagic species, and dorsally positioned eyes associated with a sedentary, benthic habit (Aleev, 1969; Hynes, 1970; Gatz, 1979; Watson and Balon, 1984; Motta et al., 1995; Frédérich et al., 2016). Fishes that locate and attack prey from below often have dorsally oriented, upward-facing eyes (Pankhurst and Montgomery, 1989; Warrant and Locket, 2004; Lisney and Collin, 2008), while benthic-feeding fishes often also have dorsally positioned eyes, which may facilitate predator detection when feeding in a head-down position (Bellwood et al., 2014). Thus, the relationship between the position of the eyes and an animal's functional requirements requires a detailed understanding of a species' behaviour and ecology.

Although much is known about the relationships between eye morphology and the behaviour and ecology of a species, far less is known about how eye morphology varies within species. This is surprising because a single species can occupy a wide variety of habitats, hence requiring different types of visual specialisation. Intraspecific differences in the size and morphology of compound eyes have been reported in insects, and these differences correlate with variation in activity patterns among different strains or castes, and varies between the sexes (e.g. Roonwal and Bhanotar, 1977; Posnien et al., 2012; Streinzer et al., 2013). With regard to vertebrates, freshwater fish provide some interesting examples of intraspecific variation in eye morphology. For example, in the Atlantic molly (Poecilia mexicana), a species that has colonized caves in southern Mexico, there is a gradual reduction in eye size from the chambers closest to the surface that receive some dim light, to the deepest chambers that are in constant darkness (Fontanier and Tobler, 2009). In other freshwater fishes, such as red shiners (Cyprinella lutrensis) and the galaxid Aplochiton zebra, individuals that live in turbid habitats have relatively larger eyes than their counterparts living in clearer waters, presumably as an adaptation to improve visual sensitivity (Lattuca et al., 2007; Dugas and Franssen, 2012). However, such studies that have examined within-species variation in eye morphology have typically considered the effect of a single environmental variable (usually light availability), rather than the multiple and interacting ecological factors that will determine visual performance. Additionally, previous work has largely focused on eye diameter as a measure of eye morphology and has not considered the size of the aperture (pupil) or the position of the eyes on the head, variables that can reveal important information about the ecology, behaviour and activity patterns of a species within a given habitat.

In this study, we investigated three aspects of eye morphology – (1) eye size relative to body size, (2) pupil size relative to eye size and (3) eye position in the head – in a species of Australian freshwater fish, the western rainbowfish [Melanotaenia australis (Castelnau 1875)]. We chose this species because it occupies a large variety of freshwater habitats across northwest Western Australia, including isolated springs and billabongs (pools), ephemeral creeks, gorges and lakes (Allen et al., 2002). Previous studies have shown that this species shows extensive variation in morphological characteristics, such as body shape and lateral line morphology, in relation to environmental parameters (Young et al., 2011a,b; Kelley et al., 2017; Spiller et al., 2017). However, no previous study has specifically investigated variation in eye morphology, even though these fish are considered to rely heavily on vision for feeding, communication, sexual selection and predator detection (Brown and Warburton, 1997; Arnold, 2000; Brown, 2002, 2003; Hancox et al., 2010; Kelley et al., 2012). We collected western rainbowfish from 14 sites in a major river catchment and surveyed each site for a number of environmental and ecological characteristics that may influence eye size and morphology, including turbidity, invertebrate abundance, habitat complexity, predation risk and whether the site was partly shaded by surrounding rock formations. Geometric morphometric analyses were used to evaluate eye position and head shape, while subsequent model selection allowed us to determine which of the environmental variables best explained variation in eye size, pupil size and eye position in the head.

Adult western rainbowfish were sampled from 14 locations across the Fortescue River catchment in the Pilbara region of northwest Western Australia in May–November 2013. The upper and lower sub-catchments within our sampling region are considered to be hydrologically isolated because they are separated by a geographic barrier, the Goodiadarrie Hills (Skrzypek et al., 2013). We sampled a total of 312 individuals originating from seven sites in the upper Fortescue catchment (Coondiner Creek, Kalgan Creek and Weeli Wolli Creek), two in the mid catchment (in Karijini National Park) and five in the lower catchment (Millstream National Park) (Table S1; mean number of fish sampled per site=22.3, range=4–31). Our sampling design, and the number of fish captured, was restricted due to the availability of freshwater habitats within this semi-arid region. The Pilbara is characterised by hot summers (24–40°C) and highly unpredictable rainfall, which mainly occurs in the summer as a result of cyclonic activity (www.bom.gov.au). For most of the year, creeks within this region therefore comprise a series of discrete pools that are maintained by groundwater and become reconnected only during major flooding events (Beesley and Prince, 2010).

Sampling was conducted between 10:00 h and 15:00 h, the time of day at which western rainbowfish tend to be most active, and to account for any difference in diurnal activity patterns between sample sites. Fish were captured using seine nets (5 or 10 m length, 6 mm mesh size, depending on site) using multiple hauls (∼3–10 net captures) to remove all fish from the immediate sampling area (representing a surface area of approximately 150 m²). We also ensured that the region that was sampled was representative (e.g. percentage cover of vegetation, etc.) of that particular habitat, and we used a similar level of fishing effort at each sample site. After each seine haul, all fish were placed in lidded buckets (containing creek water and vegetation for cover) until the area had been thoroughly sampled. After this time, we randomly selected adults from the buckets, and photographed them out of the water on their right side using a digital camera (Olympus E-PL3, Olympus Corporation, Tokyo, Japan) mounted on a tripod. Following the photography, fish were placed in a separate holding bucket, to avoid re-photographing the same individuals. We did not use anaesthesia during the photography because the procedure took only a few seconds. Each image included a scale bar for subsequent image scaling. We determined the sex of each fish by placing them individually in a transparent container to examine the morphology of the dorsal and ventral fins. Males have pointed dorsal and ventral fins, while the fins of females are more rounded in shape (Allen et al., 2002). Once the adults in the habitat had been photographed, all fish were returned to their natural habitat (including those that were not photographed) at their location of capture. This study was approved by the University of Western Australia Animal Ethics Committee (protocol: RA/3/100/1176).

Each sample site was characterised according to a number of environmental and ecological characteristics (Table S1). This was aided by taking photographs of each habitat, and drawing detailed maps to record the substrate, details of submerged vegetation and rocks, surrounding riparian vegetation and overhanging vegetation or rock structures. We determined the amount of direct daylight each habitat was exposed to by examining the surrounding rock formations, such as whether sites were located in gorges or creek lines, and allocating each site a ‘gorge factor’ score between 0 and 2. Sites scoring 2 had high (>10 m) rock face on either side (i.e. sites in Karijini National Park) and were therefore only exposed to direct sunlight during midday hours. Sites scoring 1 had rock face on one side (<10 m high; sites in Coondiner Creek), and therefore were partially shaded for one half of the day, and sites scoring 0 were not located in gorges or creek lines and received direct sunlight throughout the day (sites in Millstream National Park). A habitat complexity score was used to represent the level of visual habitat structure present (Lostrom et al., 2015). Habitat structure, such as substrate and submerged vegetation, as well as debris such as snags, provide navigational obstacles and shelter, but also influences the light environment by creating shade. A habitat complexity score ranging from 1 to 5 was given for each habitat, where 1 was allocated to open water habitats (e.g. large pools >1 km long) with gravel substrates, while a score of 5 represented dense aquatic vegetation and submerged debris. Habitat complexity scores were allocated by two independent observers, by referring to both the habitat maps and the photographs of each site. The scores were then collated and adjusted (for one site) to reach a consensus. This procedure was performed before the eye morphology data were collected.

Water velocity was measured using a flow meter (FP111; Global Water, College Station, TX, USA) placed 10 cm below the surface and averaged from three readings. We determined the predation pressure at each site using previous records of fish predator assemblages in the sampling areas (Young et al., 2011b), augmented by our own observations at the time of sampling. Fish predators of rainbowfish have previously been categorised as ‘high risk’ (e.g. barramundi, Lates calcarifer; western sooty grunter, Hephaestus jenkinsi) or ‘low risk’ (e.g. spangled perch, Leiopotherapon unicolor; barred grunters, Amniataba percoides) (Young et al., 2011b), and we used these classifications to assign each site a value of 0 (only low-risk predatory species present) or 1 (where 1 or more high-risk predatory species have been recorded). We sampled the total number of surface invertebrates by sweeping a 250 µm mesh dip net over a 10 m length of the pool (Lostrom et al., 2015). We collected three samples for each pool and calculated the average number of surface invertebrates present. Benthic invertebrates were sampled using a 500 µm mesh D-net and trampling the sediment over a 1 m² area for 1 min. The contents of the net were then passed through 2 mm and 500 µm steel mesh sieves before counting all invertebrates present. The percentage cover of green filamentous macroalgae was also evaluated based on on-site observations and used as a measure of food availability (rainbowfish are omnivorous). A water sample (30 ml, unfiltered) was collected at each site and kept cool and in the dark until subsequently analysed for turbidity in the laboratory. A turbidity meter (Hach 2100A; Hach, Loveland, CO, USA) was used to obtain three turbidity measures for each pool; the mean value was used in subsequent analyses.

In order to measure eye size relative to body size, and pupil size relative to eye size, the horizontal diameters of the eye and the pupil, along with fork length (FL; measured from the tip of the upper jaw to the fork in the tail), were measured to the nearest 0.1 mm from the scaled photographs of each fish (Fig. 1) using ImageJ software (Schneider et al., 2012). The horizontal diameters of the eye and pupil were used to calculate a pupil/eye diameter ratio for each fish. By using pupil size as a measure of the aperture of the eye in western rainbowfish, we assumed that this species does not exhibit any significant pupil mobility, as is the case for the vast majority of bony fishes (Douglas, 2018). Indeed, throughout our own experimental work with rainbowfish we have not observed any changes in pupil size in response to factors such as changes in background illumination or handling, for example. Nevertheless, we acknowledge that a very small number of predominantly benthic bony fishes have been found to have mobile pupils (Douglas et al., 2002; Douglas, 2018), and so pupil size may not necessarily be an appropriate measure of the size of the aperture of the eye in these species.

We used geometric morphometric analyses (Zelditch et al., 2012) to quantify variation in the position of the eye and shape of the head of each fish. Images were scaled for size according to the scale bar in each image before using the software program TPSDIG (version 2.17; available at http://life.bio.sunysb.edu/morph/) to assign landmarks to each image. A total of 12 landmarks were placed on each image: two landmarks were fixed and positioned on the tip of the snout and at the top of the operculum plate, while 10 were semi-sliding landmarks and were assigned along the outline of the head and the back of the operculum (Fig. 1). The program TPSRELW was subsequently used to generate a series of relative warps (RWs) that describe overall changes in eye position and head shape, and centroids (measured as the squared distance of each landmark from the mean or central position), which represent overall head size. The data supporting this study are publicly available at the University of Western Australia's Research Repository (doi:10.26182/5e4b93d5b9a70).

Eye size scales with body size in fish (Howland et al., 2004; Lisney and Collin, 2007; Schmitz and Wainwright, 2011; Caves et al., 2017). Therefore, we initially examined the relationship between body size (FL) and both eye diameter and the pupil/eye diameter ratio. There was a significant positive correlation between body size and eye diameter (r₃₁₁=0.90, P<0.001), but no relationship between the pupil/eye diameter ratio and body size (r₃₁₁=−0.03, P=0.62). Therefore, we used the residuals of the regression relationship between body size and eye diameter in all further analyses. Use of the residuals allows us to determine whether fish have a larger (positive residuals) or smaller (negative residuals) eye relative to their body size. The residuals of eye diameter were not correlated with the pupil/eye diameter ratio (r₃₁₁=−0.05, P=0.36).

Our first set of analyses used ANOVA/MANCOVA to test whether there was any variation among sample sites, or any variation attributable to sex, in our measures of eye size and morphology. We then conducted a further set of tests, using linear mixed models, to examine the role of the ecological and environmental predictors in explaining variation in eye morphology (while controlling for sampling design). For the ANOVA/MANCOVA tests, we tested the effect of sex (entered as a fixed effect) because western rainbowfish are sexually dimorphic (Allen et al., 2002; Lostrom et al., 2015) and because sexual dimorphism in eye size has been reported in a number of fish species (Echeverria, 1986; Cooper et al., 2011; Webster et al., 2011; Dugas and Franssen, 2012; Záhorská et al., 2013). We tested for an effect of collection site (entered as a fixed factor), sex and the site by sex interaction on the eye diameter residuals and the pupil/eye diameter ratio using ANOVA. We used MANCOVA, with centroid entered as the covariate to account for body size, to investigate the effect of these factors on the first five RWs (RW1–RW5) describing variation in eye position in the head. Centroid and sex were retained in subsequent models if they had a significant effect in the MANCOVA.

We determined the importance of the ecological and environmental variables on RW1 (which accounts for the most variation in eye size/shape; 34.5%) using linear mixed models. These models allowed us to evaluate the importance of the predictors while controlling for the spatial component of the sampling design. A map of the sampling region (Fig. 2) shows that the collection locations were clustered into three discrete regions (sub-catchments). We therefore accounted for this in our analyses by entering the term ‘sub-catchment’ (three levels) as a random effect in all models. We then considered a priori which of the ecological and environmental variables might influence eye morphology (Table S2); these included gorge factor (three levels), predation risk (two levels), habitat complexity (ordered factor with 5-levels), invertebrate abundance (benthic+surface invertebrates: a covariate) and turbidity (a covariate). Prior to running the linear mixed models, we performed data exploration, following the methods outlined in Zuur et al. (2010) to explore the structure of our data, to examine relationships among the variables, and to test the validity and assumptions of our models. Thus, in addition to examining the plots of the residuals versus the fitted values, we constructed conditional plots to verify homogeneity of each of the factors, and histograms (for each factor) to investigate skewness in the data. We found that some of the environmental characteristics were correlated: surface water velocity was negatively correlated with both turbidity (r₃₁₁=−0.16, P=0.005) and filamentous macroalgae cover (r₃₁₁=−0.41, P<0.001), and the level of filamentous macroalgae was positively correlated with habitat complexity (Spearman's rank test: ρ=0.60, n=312, P<0.001). To avoid potential issues associated with collinearity, we dropped the covariates surface water velocity and filamentous macroalgae as these were deemed less biologically relevant for visual tasks than turbidity and habitat complexity. One site (Kalgan) had unexpectedly higher turbidity values than the rest, and only four fish were captured (Table S1), increasing the probability that these samples are not representative of the population. We chose to exclude these samples from our turbidity analyses; we confirmed that removal of these four fish had no effect on the model outcomes for either the turbidity analyses, or on the modelling for each of our dependent variables. We also noted that treating habitat complexity as a fixed effect (low complexity=scores 1–2; high complexity=scores 3–5) had no outcome on the model rankings. Our model selection approach also included fitting a null model, which contained only the intercept, and we compared the top-fitting model with the null model using the likelihood ratio (LR) test. We used Akaike's information criterion for smaller sample sizes (AICc) to evaluate the fit of each model, where models with lower AICc values are considered more parsimonious (Symonds and Moussalli, 2011). Models with a change in AICc (ΔAICc) >10 were excluded from the model set, those with 6<ΔAICc<10 were considered unlikely and those with ΔAICc<2 were considered equal best models (Symonds and Moussalli, 2011). All statistical analyses were performed using the software program R (https://www.r-project.org/) using the software packages lme4 (Bates et al., 2015) and AICcmodavg’ (https://CRAN.R-project.org/package=AICcmodavg).

The ANOVAs revealed a significant effect of collection site on both the eye diameter residuals (F_13,284=16.8, P<0.001) and the pupil/eye diameter ratio (F_13,284=14.4, P<0.001; Fig. 3; Table S3), but no effect of sex on either of these parameters (eye diameter residuals: F_1,284=2.24, P=0.14; pupil/eye diameter: F_1,284=3.47, P=0.064) and no significant site×sex interaction (eye diameter residuals: F_13,284=0.44, P=0.95; pupil/diameter size: F_13,284=0.56, P=0.89). The sex factor was therefore not included in subsequent models.

The model selection analyses (Table 1) revealed that habitat complexity was the top model (ΔAICc<2) for the residuals of eye diameter and was significantly better than the null model that contained only the random intercept (LR test: χ²=47.1, d.f.=4, P<0.001; Table 1). The residuals of eye diameter differed according to habitat structure such that fish occurring in habitats with the lowest complexity score (i.e. 1=open water) tended to have smaller eyes than those occurring in habitats with higher complexity scores (i.e. dense vegetation; Fig. 4A). Variation in the pupil/eye diameter ratio was best explained by the gorge factor (comparison to null model: LR test: χ²=27.21, d.f.=2, P<0.001; Table 1). Fish occupying gorge habitats had significantly smaller pupil/eye diameter ratios than those occupying half gorge sites or open habitats (Table S3; Fig. 5).

The first five RW scores (RW1–RW5) explained 85.8% of the total variation in shape (RW1=34.5%, RW2=20.9%, RW3=15.1%, RW4=8.5%, RW5=6.7%). Visualization of the relative warps using ordination plots (Fig. 6) revealed that negative RW1 scores were associated with an elevated eye position, reduced slope of the head and broadening of the operculum plate relative to positive scores. RW2 described the overall bend of the head (upwards tilt for negative values), while RW3 accounted for the overall widening of the posterior region of the head (broader for negative scores). Note that RW1–RW3 were correlated with the pupil/eye diameter ratio (RW1: t₃₁₀=2.50, P=0.013; RW2: t₃₁₀=2.57; RW3: t₃₁₀=−2.37, P=0.018). Centroid size was significantly negatively correlated with RW1 (r₃₁₁=−0.31, P<0.001) and RW3 (r₃₁₁=−0.15, P<0.010), and there was a negative trend for an association between RW2 and centroid size (r₃₁₁=−0.10, P=0.067). These findings suggest that negative allometry is associated with eye and head morphology and all subsequent models thus included centroid as a covariate.

The MANOVAs revealed an overall effect of site (F_13,296=8.02, P<0.001) and centroid (F_1,196=23.6, P<0.001) on the combined RWs (RW1–RW5), but no effect of sex (F_1,296=0.78, P=0.57). The linear mixed models and subsequent model selection procedure were performed only for RW1, which explained the greatest proportion of variation in shape (34.5%), and excluded the factor sex. The linear mixed models revealed that habitat complexity was the top-fitting model (ΔAICc<2) in explaining variation in RW1 (Table 1). The model containing the effect of habitat complexity was significantly better than the null model that contained only the random intercept and the centroid (LR test: χ²=59.3, d.f.=5, P<0.001; Table 1). Fish from sites with a habitat complexity score of 2 (low complexity) had significantly lower RW1 scores (post hoc Tukey tests; P<0.01; Fig. 4B), meaning that they had more dorsally located eyes, less sloped heads and broader operculum plates than those from sites with higher levels of habitat complexity.

Investigations of intraspecific variation in eye morphology are relatively rare, but have the potential to reveal how environmental and ecological variables contribute to visual specialisations. In this study, we investigated eye morphology in a single species of Australian freshwater fish, the western rainbowfish, which is found in a range of habitats with diverse ecological and environmental characteristics (Allen et al., 2002; Lostrom et al., 2015). We found significant variation in all three aspects of eye morphology that we investigated – eye size relative to body size, pupil size relative to eye size, and the position of the eye in the head – among individuals collected from different habitats in a single river catchment. Our modelling approach, which allowed us to evaluate the relative role of the environmental and ecological variables, revealed that the size of the eye and pupil, and the position of the eye in the head, are related to two habitat components: the structural complexity of the habitat and the presence of surrounding rock formations (i.e. gorges). Our findings reveal that multiple, interacting factors influence eye morphology in western rainbowfish, suggesting that visual systems are not only species-specific, but dependent on an individual's habitat and behavioural tasks.

Our finding that rainbowfish eye size is linked to the structural complexity of the habitat is supported by other studies showing that fish from structurally complex habitats tend to have larger eyes than those from less complex habitats (Dobberfuhl et al., 2005; Willacker et al., 2010). Larger eyes may be associated with more complex habitats because these environments favour increased visual sensitivity and/or the need for higher spatial resolution (Land and Nilsson, 2012). Complex benthic habitats may have reduced light availability owing to attenuation of light with depth (Lythgoe, 1979), and because of shading from macrophytes or other submerged structures (e.g. tree branches). Further reductions in light, along with shifts in spectral composition, can occur if complex habitats also contain high amounts of suspended sediments, phytoplankton and dissolved organic matter (Kirk, 1994). Complex habitats may therefore favour increased visual sensitivity, which can (along with other visual adaptations) be attained by increasing the diameter of the eye (Land and Nilsson, 2012).

A larger eye can also increase spatial resolution, which may be advantageous for visually mediated behaviours such as prey detection and navigation in structurally complex environments (Hughes, 1977). A recent comparative study of 159 species of ray-finned fishes found that a large eye size was associated with increased visual acuity, and that species in complex habitats tended to have higher visual acuity than expected after accounting for eye size relative to body size (Caves et al., 2017). A positive relationship between eye size and visual acuity is also supported by a study of closely related cichlids (from the same clade) that found that a species from highly structured rock habitats (Asprotilapia leptura) had higher visual acuity (measured behaviourally) than species from sandy habitats (Xenotilapia flavipinnis) and intermediate sandy/rocky habitats (X. spiloptera) (Dobberfuhl et al., 2005). Visual acuity can also be evaluated by mapping retinal topography, and in coral reef fishes, differences in the number and density of ganglion cells are linked with a species' habitat (Collin and Pettigrew, 1988a,b). Specifically, the sampling region of the visual field is specialised depending on whether the species views a broken horizon or an unobstructed horizon, resulting in species-specific variation in visual acuity (Collin and Pettigrew, 1988a,b, 1989). Collectively, these previous studies lead us to predict that rainbowfish in structured habitats have relatively large eyes in order to increase visual acuity and to enhance behavioural performance in these environments. It would therefore be interesting to assess visual acuity (either behaviourally or from retinal structure) and examine corresponding visual behaviours (e.g. foraging tasks) in western rainbowfish to determine whether there are any population differences.

In this study, the structural complexity of the habitat also influenced the position of the eye in the head, as well as the slope of the head and the width of the operculum plates. Specifically, rainbowfish from complex habitats tended to have more ventrally located eyes, along with more sloped heads and narrower operculum plates compared with those from less complex habitats. Most studies that have reported a link between habitat complexity and morphology in fishes have focused on the divergence between two classic morphotypes – benthic/littoral morphs that feed on invertebrates in sediments or on macrophytes, and limnetic forms that feed mainly on zooplankton in the open water (Schluter and McPhail, 1993). This results in a strong relationship between trophic morphology and diet (e.g. Svanbäck and Eklov, 2002, 2003), which may also determine the location of the eyes in the head. In sticklebacks (Gasterosteus aculeatus), for example, benthic forms have more dorsally located eyes than limnetic forms (Behm et al., 2010), which could contribute to their increased foraging success on susbtrates relative to limnetic forms (Bentzen and McPhail, 1984). We similarly anticipate that the differences in head shape and eye position in the present study will relate to variation in behaviour in rainbowfish; in complex habitats, more ventrally located eyes may facilitate detection of predators or prey that are located below the fish. Although habitat complexity was the most important predictor of variation in eye size and eye position in this study, it is important to acknowledge that other habitat variables (including correlates of habitat complexity, such as water flow and macroalgal cover) may also underlie our results. Controlled laboratory experiments are clearly necessary to establish cause and effect. We also acknowledge that our findings may be influenced by sampling bias, if, for example, the spatial and temporal activity patterns, habitat preferences and survival of fish are differentially affected by an individual's eye morphology.

Because light intensities in shaded habitats can be several orders of magnitude lower than those in open habitats (Ovington and Madgewick, 1955; Endler, 1993; Fleishman et al., 1997), we predicted that fish in the shaded gorge habitats would have larger eyes with relatively larger pupils, in order to increase the light-gathering capabilities of the eyes. For example, nocturnal reef fishes have relatively larger eyes and larger, more rounded pupils than diurnal reef species (Pankhurst, 1989; Schmitz and Wainwright, 2011). However, our results contradict this pattern as we found that fish from sites surrounded by rock formation (i.e. reduced light availability) had relatively smaller pupils for a given eye diameter than those from semi-gorge or open sites. A more detailed consideration of eye morphology, for example, incorporating the diameter of the lens and the shape of the pupil (e.g. the optical ratio), might reveal subtle morphological adaptations for enhanced visual sensitivity (Schmitz and Motani, 2010). Future studies could also determine whether individuals from the dimmer gorge habitats possess retinal adaptations for improving visual sensitivity, such as larger photoreceptors, a higher proportion of rod photoreceptors or higher photoreceptor to retinal ganglion cell summation ratios (Marshall, 1979; Wagner, 1990; Land and Nilsson, 2012), and whether these adaptations are associated with improved visual performance.

Turbidity can significantly hinder vision underwater, causing a reduction in visual resolution (Wells, 1969), a reduced visual range and decreased target contrast (Lythgoe, 1979; Utne-Palm, 2002). In contrast to other studies with fishes (Kotrschal et al., 1991; Huber and Rylander, 1992; Huber et al., 1997; Lisney and Collin, 2007; Caves et al., 2017), we found no evidence that fish in turbid habitats have smaller eyes. In addition to investing in non-visual senses, such as chemoreception (Kotrschal et al., 1998), this finding may be due to the fact that turbidity values were generally similar at all sites at the time of sampling. Although many fishes rely on vision for predator detection and evasion (Chivers et al., 2001; Fischer et al., 2017), we also found no relationship between the position of the eyes and predation risk. One explanation for this is our broad characterisation of risk (i.e. presence/absence of predator species) rather than using a more stringent measure, such as predator density or diel activity patterns. Finally, because large eyes are associated with zooplanktivory and insectivory in fishes (Motta et al., 1995; Huber et al., 1997; Behm et al., 2010), we predicted a relationship between invertebrate abundance and eye size. However, as suggested by these previous studies (Motta et al., 1995; Huber et al., 1997; Behm et al., 2010), habitat use is probably a more important contributor to morphological variation in eye size in rainbowfish than diet, even though the two are clearly linked.

A variety of environmental and ecological factors have been invoked to explain interspecific variation in eye size and shape, but few studies have examined variation in the visual system of a single species. Our findings suggest that individuals vary in the degree to which they invest in their visual system, which we predict will translate to differences in visual sensitivity and visual acuity among populations, with corresponding implications for behavioural performance. The extent to which the sensory systems of fishes can be adapted to specific ecological and environmental conditions means that they may also be highly susceptible to rapid changes caused by human activity (Collin and Hart, 2015; Kelley et al., 2018). Such impacts may be particularly significant in species that exhibit habitat-specific variation in eye morphology because visual specialisations may no longer enhance fitness and may be detrimental to survival.

We would like to thank an anonymous reviewer and Mateo Santon for comments on the manuscript that greatly improved the scope and quality of our work. We are grateful to S. Lostrom, P. Grierson, J. Evans and P. Davies for advice and support and to S. Luccitti, S. Wild, A. Storey, J. Delaney, J. Iles, A. Siebers, K. Bowler and G. Skrzypek for assistance in the field. We would like to acknowledge the rangers at Karijini and Millstream National Parks for information and access to sampling sites.