What you see is where you go: visibility influences movement decisions of a forest bird navigating a three-dimensional-structured matrix

Abstract

Animal spatial behaviour is often presumed to reflect responses to visual cues. However, inference of behaviour in relation to the environment is challenged by the lack of objective methods to identify the information that effectively is available to an animal from a given location. In general, animals are assumed to have unconstrained information on the environment within a detection circle of a certain radius (the perceptual range; PR). However, visual cues are only available up to the first physical obstruction within an animal's PR, making information availability a function of an animal's location within the physical environment (the effective visual perceptual range; EVPR). By using LiDAR data and viewshed analysis, we modelled forest birds' EVPRs at each step along a movement path. We found that the EVPR was on average 0.063% that of an unconstrained PR and, by applying a step-selection analysis, that individuals are 1.55 times more likely to move to a tree within their EVPR than to an equivalent tree outside it. This demonstrates that behavioural choices can be substantially impacted by the characteristics of an individual's EVPR and highlights that inferences made from movement data may be improved by accounting for the EVPR.

1. Introduction

Understanding the relationships between an animal's behaviour and the environment it occupies is central to a broad range of ecological and evolutionary questions of both fundamental and applied significance [1,2]. Gaining insight into these relationships is frequently achieved either by using a behavioural response to identify influential environmental cues or by using a cue that is known to influence behaviour to quantify the response. For example, spatial characteristics of movement trajectories are used to statistically identify factors that influence movement decisions [3,4], or to determine from what distance animals start moving towards a known attractor [5]. However, correct inference of behaviour in relation to the environment is challenged by the lack of objective methods to identify what information effectively is available to an animal at a given location [6].

The information available to an individual from a specific location is determined by multiple factors. First of all, the range over which information can be acquired is constrained by the physiological properties of an animal's sensory system (i.e. its perceptual range [PR]; table 1). For example, the distance over which an animal can discriminate a visual cue against background noise is determined by the photoreceptor density of its eyes [7]. These data are lacking for most species making objective quantification of a PR generally impossible. The response distance of an animal to a known attractor (i.e. the functional perceptual range, FPR, table 1) is therefore often used as a proxy in ecological research [8,9]. However, even within an animal's FPR, visual information can be inaccessible due to the physical structure of the environment (the effective visual perceptual range [EVPR]; table 1). Hence, when the structure of the environment is spatially variable, an animal's EVPR will be a dynamic property of its location. For example, an animal's EVPR will change drastically when it moves from a valley up a mountain, or from a forest patch into a meadow.

Table 1.

Definitions of key terms and their abbreviations.

	definition
perceptual range (PR)	the range over which an animal is physiologically capable of perceiving a visual cue
effective visual perceptual range (EVPR)	the visual perceptual space available to an animal given the three-dimensional structure of the environment
functional perceptual range (FPR)	the range over which an animal tends to respond behaviourally to a visual cue

Not accounting for the EVPR carries the risk of making biased inferences. For instance, an animal's behavioural decision may appear maladaptive, while in reality it acted optimally given the information that it could effectively perceive from that location. For example, a dispersing individual may have selected a habitat patch that is of modest quality compared to other similarly close patches. However, that patch was perhaps the best quality of those visible to the individual. Recent analysis is indicative of this effect influencing the patterns of patch occupancy observed in neotropical primates [10].

Given its potential importance, it is surprising that, in the context of animal movement, few attempts have been made to explicitly account for the effect of the environment on visual information availability, neither in empirical research, nor in modelling (but see [10,11]). A possible explanation may be the limited availability of appropriate data and tools [12]. Light transmission through the air is unconstrained over distances usually considered in ecological research [7] making visual cues available to an animal in every viewing direction up to the first physical obstruction. An animal's EVPR, hence, can be modelled as a function of a three-dimensional model of the physical environment (a so called ‘viewshed’) [12,13].

While the use of viewsheds is beginning to gain some traction in ecological research (e.g. [14,15]), they have not yet been used to determine the degree to which behavioural choices are constrained to those options present within an EVPR. Here we use a unique set of movement steps of the placid greenbul Phyllastrephus placidus, an Afrotropical forest bird, to test our hypothesis that individuals navigating an agricultural matrix are constrained in selecting trees that, from our viewshed analyses, we infer that they can see (i.e. are within their EVPR).

2. Methods

(a). Greenbul behaviour

Behavioural data were represented by visually confirmed discrete movement steps of 27 adult greenbuls (e.g. a greenbul flying from one tree to the next) navigating an unfamiliar matrix. Individuals were captured in forest fragments, radio-tagged and released in an agricultural matrix within homing distance to their territories. Experiments were conducted in 2009–2010 (for details see [3]). Because animals outside their home range and native habitat are likely to be deprived of social information and memory-based spatial reference, we assume behavioural decisions to primarily reflect the response to visual cues.

(b). The matrix

Greenbul movement behaviour was documented in an agricultural matrix in the Taita Hills (SE Kenya). This is a small-holder agricultural landscape characterized by plots of crops and scattered trees [16] and a pronounced topography (figure 1).

Figure 1.

The agricultural matrix greenbuls had to navigate while returning to their territories in forest fragments (photo: J. Aben, 2009).

The three-dimensional structure of the matrix was modelled at 1 m resolution (digital surface model [DSM] and canopy height model [CHM]) using airborne LiDAR obtained in 2013 (for details see [17] and electronic supplementary material, appendix A1). Trees in the matrix were uniquely identified and crowns delineated based on the CHM using the itcLiDAR function in the itcSegment package [18] (for details see electronic supplementary material, appendix A2).

(c). Selection of steps

The package amt [19] was used to transform the movement data into a track object, where consecutive locations are represented in consecutive steps. Only steps that were of length greater than 10 m and ended in a tree were retained and paired with 300 random steps to form a stratum. Random steps were generated using the function random_steps, with turning angles and step lengths sampled from a von Mises distribution (centred on the direction of the preceding step) and gamma distribution, respectively. Both distributions were parameterized by fitting the corresponding distribution to the observed data (electronic supplementary material, appendix A3). Finally, only those strata with at least 15 random steps that ended in a unique tree were retained, yielding 369 steps for 27 individuals (range 4–38 steps). If more than 15 random steps were available, 15 steps were selected randomly.

(d). Viewsheds

For each location, viewsheds were modelled as a function of the DSM using the r.viewshed command in GRASS-GIS [20] considering all possible viewing directions. A z-coordinate (observer_height parameter) was derived from the CHM (represented either by the mean of the CHM within a 2 m buffer or calculated relative to the height of a tree). Search radius (i.e. viewshed radius) was set to 200 m just over the maximum step length in the data. To prevent a bird's visibility being blocked by the tree it is in, unconstrained visibility was assumed within a radius of 5 m. This was modelled by detracting values from the CHM from those in the DSM raster within that 5 m radius prior to modelling the viewshed.

(e). Step-selection model

To infer the effect of visibility on tree selection by moving greenbuls, we modelled selection—the response being whether a step was observed or simulated—while controlling for observed movement patterns (i.e. directional autocorrelation and step lengths) and the potentially confounding effects of the percentage of field (defined as CHM less than 1 m) between the bird and a tree and tree height, using integrated step-selection functions (iSSF [21]) (figure 2, electronic supplementary material, appendix A4).

Figure 2.

Graphical overview of the analytical approach. A greenbul selects one tree to move to next from its current position (red dot; all panels). The decision on which tree to move to may be influenced by the percentage of field to cross (a), on tree height (b), and if a tree is visible (c; viewshed was truncated at 65 m for illustration). To test our hypothesis, a step-selection function is fitted to compare the covariates between the observed step and the alternative steps (d).

Tree visibility was determined by intersecting the tree crowns with the corresponding viewshed raster resulting in the binary indicator variable in_viewshed. We included a log of the step length and the cosine of turning angles, as these were sampled from a statistical distribution [21,22]. The variables percentage of field and tree height were standardized to have a mean of zero and an s.d. of 1 at the stratum level (i.e. an observed step paired with 15 random steps). We used a Poisson regression with random effects on individual strata with a fixed variance of 1e4 to be able to account for variation in selection among individuals [23]. A Poisson regression specified in such a way is the likelihood equivalent to a conditional logistic regression used in standard SSF [23]. We included a random coefficient for the percentage of field and tree height, but not for in_viewshed because individuals belonging to the same species are likely to have similar visual acuity. We fitted the model using the package glmmTMB [24].

All analyses were carried out using the R program [25].

3. Results

The EVPR (area of viewshed) of greenbuls tracked in the matrix was on average 0.063% that of an unconstrained PR (range 7.166 × 10⁻⁰⁵–0.297%).

While controlling for the effects of tree height, percentage of field and movement, we find that visible trees are substantially more likely to be selected by a moving greenbul (relative selection strength = 1.55) (table 2). In addition, greenbuls prefer tall trees over lower ones to move to.

Table 2.

Estimated coefficients of the integrated step-selection model. Effects with a p-value < 0.05 were considered as significant.

variable	estimate	s.e.	p-value
tree height	0.275	0.055	<0.001
per cent field	−0.100	0.066	0.130
in_viewshed	0.440	0.178	0.013
log(step length)	−0.417	0.119	<0.001
cos(turn angle)	0.625	0.103	<0.001

4. Discussion

Greenbuls returning to their territories after translocation had to navigate a heterogeneous matrix in the topographically complex terrain. Their movement typically consisted of short flights interspersed with pauses in vegetation [3]. Visual observations indicated that flights were generally straight, suggesting that greenbuls typically decided where to go from one location and did not update their decision while in flight. By modelling viewsheds from these locations and quantifying visibility of potential target trees, we were able to demonstrate that, indeed, greenbul movement reflected behavioural responses to information within their effective visual perceptual ranges; trees that were within a greenbul's EVPR had a substantially higher probability to be chosen. Notably, while it is most often the trees within the EVPR that are selected by individuals, there are still movement steps where individuals move to a tree that is not. We suggest at least four potential reasons for this. First, because the LiDAR data we used are not fully three-dimensional, modelled visibility is likely conservative compared to what a greenbul could perceive in reality, which may result in a tree being wrongly classified as being outside an individual's EVPR. Second, greenbul movement data were collected in 2009–2010, while LiDAR data were obtained in 2013. While in this period, there were virtually no changes in the trees present in the landscape (only one tree visited by a bird in 2009–2010 was no longer present in the LiDAR data in 2013), changes in the matrix over that period of time may result in some viewsheds not perfectly reflecting visibility as experienced by tracked greenbuls. Third, in some cases, birds may update their choice on the destination while in flight and as new trees come into view. Finally, some decisions may have been driven by information acquired through other sensory modalities (e.g. a bird avoiding a road because of hearing vehicles).

Our finding that greenbuls selected relatively tall trees could reflect a positive association with protection or food availability [26]. It could, however, also represent an adaptive behavioural strategy in order to increase their EVPR. Field observations may support this hypothesis as, on multiple occasions, greenbuls were found to move to the top of a tree prior to their next move. Vertical movements influence an animal's EVPR and hence its ability to collect information [27]. By selecting high-viewing points in the matrix, greenbuls may increase navigation efficiency and reduce predation risk due to more effective vigilance. Our results, hence, suggest that spatial behaviour may not only be driven by a response to external information, but also by the need to collect it. Future studies that seek to elucidate how the environment shapes spatial behaviour are encouraged to also consider visual information acquisition as a potential driver [28].

Animal movement is often presumed to reflect the behavioural response to environmental cues, and characteristics of movement trajectories are frequently used to quantify this relationship. Movement steps considered in our study consisted of discrete flights from one location to the next, and viewsheds modelled from intermittent locations were found to correspond well to the operational scale of decision-making. Movement steps typically considered in habitat selection analysis, however, reflect the effective displacement of an individual over a given period of time [29], and locations may therefore not correspond to the operational scale of decision-making limiting opportunities for progress to be made by accounting for EVPR. Hence, the identification of the relevant scale of decision-making should be a research priority. In concordance with our results, behavioural decisions are often made by animals during pauses in the actual movement process [30]. The development and increased availability of dedicated biologgers [31] may allow these pauses in movement data to be identified statistically allowing development of more refined models of movement that account for sequentially updated viewsheds along a movement trajectory.

5. Conclusion

Our key message, that individuals are more likely to move to desirable locations within a landscape that they can see than to similarly desirable locations that they cannot see should probably not be surprising! What is more surprising is that as far as we are aware it has not previously been demonstrated. Furthermore, across the broad range of modelling approaches that are applied to understand animal movement by analysing movement trajectories, we have found none that accounts for an animal's EVPR. Not accounting for EVPR within movement modelling risks reducing our ability to correctly infer behaviours and for some species and in some landscapes, this impact may be substantial. Here we have introduced the effective visual perceptual range to movement modelling. By gaining the necessary movement and environmental data to expand its use into habitat selection analysis and predictive modelling, we can improve predictions and make better management recommendations.

Data accessibility

The data and R code used for the statistical analysis is available at [32].

Authors' contributions

J.A. conceived of the study, designed the study, coordinated the study and wrote the manuscript; J.S. carried out the statistical analyses and drafted parts of the Methods section; J.H. and P.P. produced the digital surface model and canopy height model and contributed to the Methods section; J.T. conceived of the study and contributed to the writing. All authors gave final approval for publication and agree to be held accountable for the work performed therein.

Competing interests

We declare we have no competing interests.

Funding

Marie Sklodowska-Curie grant, grant/award no.: 661211 and NERC, grant/award no. NE/J008001/1.