Predicting cassowary–vehicle collision in the Wet Tropics of Australia

Cassowary–vehicle collision data

We assembled a databased of cassowary–vehicle collisions (both fatal and non-fatal) from across the Wet Tropics bioregion in Queensland, Australia (Fig. 1). The data assessed for this work covered the years from 1991 through to 2018. The data came from three primary sources: (1) the Queensland Department of Environment and Science (QDES) cassowary strike database, (2) data collected by C4 (Community for Coastal and Cassowary Conservation) based on observations from members of the public, and (3) unpublished data provided by Hamish Campbell (Charles Darwin University; last update Aug 2018). Collated data underwent quality control for inclusion in subsequent analyses. A minimum threshold for record inclusion required a date of observation and a reliable location that could be localised to within 50 m by a georeference (or occasionally a detailed locality description). Duplicate records of the same event were checked for and removed. Most records contained additional information including animal age, sex, cause of injury and if the animal was deceased. Time of collision was only available for a minority of records (n = 32), so was not pursued further in the analyses. Data quality control identified 215 records (QDES: 143, C4: 46, Campbell: 26) from between 12 March 1992 and 19 December 2018 that were considered suitable for inclusion. Further record exclusions were required depending on the specific analysis (see below).

Fig. 1.

The Wet Tropics bioregion of Queensland, Australia, showing remnant habitat of the southern cassowary (Queensland Department of Environment and Science 2018) and major roads (Roadtypes 2 and 3; Queensland Department of Transport and Main Roads).

Landscape-scale correlates of vehicle collision

We compiled a list of open-source data on ecological, physical and anthropogenic predictor variables that have the potential to either positively or negatively influence cassowary–vehicle collision at the landscape scale (Table 1). To inform this selection we considered previously published work on animal–vehicle collisions within the Wet Tropics bioregion (e.g. Dwyer et al. 2016;Goosem et al. 2011) and the authors’ expert knowledge. We conducted quality control on the Queensland Department of Environment and Science cassowary sightings database (Table 1), improving the accuracy of georeference data for 43 of the 545 records based on associated written information and local knowledge.

To prepare the data for analysis, we created a 5 km × 5 km fishnet (i.e. square) grid in ArcGIS (ver. 9.3) overlaying the region of interest. For the 215 cassowary–vehicle collision records and the set of predictor variables (Table 1), we calculated a value for each grid cell. The result was a dataset of values for 1210 grid squares across the Wet Tropics bioregion. We attempted to minimise ‘false zero’ collision values by excluding 5 km × 5 km grid cells that (1) had <20% land area (along coastal zones), (2) had no recorded roads in any of the road categories 2–5, or (3) had 100% cassowary-absent habitat. These exclusions reduced the number of grid cells in the dataset from 1210 to 569, and the total number of collisions from 215 to 207.

Location-specific correlates of vehicle collision

From the assembled cassowary–vehicle collision database, we used 174 records that had high reported spatial accuracy (i.e. ~10 m). Using results from previously published works on animal–vehicle collisions (e.g. Dwyer et al. 2016;Goosem et al. 2011) and the authors’ expert knowledge, we compiled a list of ecological, physical and anthropogenic variables that potentially influence cassowary–vehicle collision at a particular road location (Table 2). We parameterised each of the variables for the 174 locations in three ways. First, we used Google Earth imagery, matching the dates of vehicle collision to Google Earth historical imagery as best as we could. Second, we made site visits to inspect and record locations where Google Earth imagery was not able to generate robust insight. Third, we used our own local knowledge of these locations, particularly for interpreting Google Earth imagery.

To determine whether cassowary collisions were non-random in relation to the distribution of local road attributes, we created 1000 non-collision ‘control’ locations within the same study region. These locations were selected through a stratified random approach using the ‘Create Random Points’ tool in ArcGIS. We constrained selection of random locations to Roadtypes 2–5 of the Queensland Department of Transport and Main Roads network. We ensured that non-collision locations per 5 km × 5 km grid cell were selected (1) in proportion to the number of cassowary sightings in each grid cell (Queensland Department of Environment and Science unpublished), and (2) in proportion to the distribution of observed collisions within habitat types (core habitat 66%, non-core habitat 9% and absent habitat 25%). We used the first 174 of these non-collision locations and recorded data for the same variables as in Table 2, using the same three approaches to parameterisation as for collision locations. Where a random location fell on a road that had no probability of vehicle collision, for example, disused roads or agricultural access tracks (n = 46), we discarded that location and selected the next random location.

Statistical analysis

Landscape-scale hotspots and correlates of vehicle collision

The reduced dataset comprised 207 cassowary–vehicle collisions recorded between 1991 and 2018 across 569 grid cells, with 61% of these collisions occurred in just five grid cells with a collision frequency as high as 46 per cell (Fig. 2a). In the GAM model, there was strong spatial covariance structure evident in the cassowary–vehicle collision data (GAM: s(long,lat, 19.49), Ref.d.f. = 99, χ² = 204.21, P < 0.0001), highlighting a major vehicle collision hotspot in the coastal lowland from Innisfail in the north (−17.5°S) to South Mission Beach in the south (−18.0°S) and extending west up the Tully River Valley (Fig. 2b). A less prominent hotspot occurs in the north of the Wet Tropics bioregion from the Daintree River (−16.3°S) to Cape Tribulation (−16.1°S).

Fig. 2.

Landscape-scale analysis of cassowary–vehicle collisions in the Wet Tropics bioregion in Queensland, Australia. (a) the spatial distribution of data for cassowary–vehicle collisions between 1991 and 2018, (b) spatial covariance structure of collision data across the 5 km × 5 km grid cells. Darker red areas indicate areas of higher collision frequency.

Over and above the effects of spatial covariance in the collision data, the final most-parsimonious GAM model identified significant increases in vehicle collisions with increasing proportion of cassowary core habitat in the landscape (s(cass.core.hab, 3.42), Ref.d.f. = 9, χ² = 65.23, P < 0.0001; Fig. 3a), increasing highway density (s(road23.dens, 0.98), Ref.d.f. = 9, χ² = 63.74, P < 0.0001; Fig. 3b), and increasing minor road density (s(road45.dens, 1.78), Ref.d.f. = 9, χ² = 11.29, P < 0.0003; Fig. 3c), together explaining 92.8% of the variance in the data (R² adjusted = 0.928).

Fig. 3.

Landscape-scale correlates of cassowary–vehicle collision in the Wet Tropics bioregion in Queensland, Australia, (a) the proportion of core cassowary habitat, (b) density of major roads (highways and major connector roads) in m/km², and (c) density of minor and local roads in m/km². Fitted lines (±95% CI) are the model-predicted partial regression smooths from the best-fit generalised additive model. Note the log scaling on the x-axis. A value of zero on the y-axis indicates model-predicted average number of vehicle collisions across all grid squares. Where 95% CI overlaps zero, there is no significant additive effect of the predictor over that range of treatment levels.

Specifically, a significantly lower number of collisions than expected (on average, across all grid squares) occurred in landscapes with a low proportion of core cassowary habitat (<25%) whereas a significantly greater number of collisions occurred in landscapes with 50–90% core cassowary habitat (Fig. 3a).

In landscapes with more than 95% core cassowary habitat, there was an indication of a slight decrease in the frequency of collisions but at very high collision levels relative to other grid squares on average. In addition, the frequency of cassowary collisions increased linearly with increasing density of major roads (Fig. 3b) where collisions in grid squares close to 10–20 m of major roads per km² were higher than average. Finally, the frequency of collisions was higher in landscapes with greater than ~400 m/km² of minor roads (local connector and local roads). However, it should be noted that variance around predicted estimates is high at both extremely low and extremely high densities of minor roads (Fig. 3c).

No significant relationship was detected between vehicle collision and cassowary population density, proportion of non-core habitat, number of corridors or road curvature. These factors were therefore removed from the analysis during model simplification.

Location-specific correlates of vehicle collision

For the 174 recorded collisions and 172 randomly-selected non-collision ‘control’ locations, the final most-parsimonious GAM model showed no significant spatial correlation structure in the data. There was also no significant effects of five of the six continuous predictors (verge width, road direction, road width, canopy cover over road, and land use within 250 m) or five of the nine categorical predictors (3D road visibility, physical barriers, verge vegetation type, forest condition within 250 m, and road surface) on relative collision frequency. There was a significantly higher relative frequency of collisions on straight roads rather than curves, compared to what would be expected from the non-collision control locations (Fig. 4a). There was also a significantly higher relative frequency of collisions than expected on major roads (Road types 2 and 3) and a lower frequency of collisions than expected on minor roads (Road types 4 and 5; Fig. 4b). Land use adjacent to roads had a significant effect on collision frequency at both the 50 m scale (Fig. 4c) and the 250 m scale (Fig. 4d). In both cases, relative collision frequencies were significantly lower when there was woodland or grazing land adjacent to the road on either side, compared with other types of land cover classes such as rainforest (Fig. 4c, d). Finally, there was a significant effect of the proportion canopy cover of adjacent vegetation on cassowary collision frequency, with lower-than-average collisions at low canopy cover (<10%) and higher than average collisions at higher adjacent canopy cover (>65%; Fig. 5). Together the five predictors in the final most-parsimonious GAM model explained 26.8% of the variance in the data (R²_adjusted = 0.268).

Fig. 4.

Categorical location-specific correlates of cassowary–vehicle collision at known collision locations (right panels) vs randomly located non-collision ‘control’ locations (left panels) in the Wet Tropics bioregion in Queensland, Australia. (a) road alignment, (b) road types as per the QLD Department of Transport and Main Roads categories, (c, d) land use categories across both sides of the road, specifically banana (ban), easement (eas), grazing (gra), peri-urban (per), rainforest (rai), rural-residential (rur), sugar (sug), and woodland (woo). Other land use types with low frequency were pooled as ‘other’ (horticulture, mangrove, melaleuca wetland and urban).

Fig. 5.

The proportion of canopy cover adjacent to a road as a location-specific correlate of cassowary–vehicle collision in the Wet Tropics bioregion in Queensland, Australia. Fitted lines (±95% CI) are the model-predicted partial regression smooths from the best-fit generalised additive model. Note the log scaling on the x-axis. A value of zero on the y-axis indicates model-predicted average number of vehicle collisions across all grid squares. Where 95% CI overlaps zero, there is no significant additive effect of the predictor over that range of treatment levels.

Landscape-scale correlates of vehicle collision

We showed that any length of major road passing through a high proportion of core cassowary habitat will result in an increased likelihood of cassowary–vehicle collision. The severity of this effect increases in direct proportion to the density of roads. It is therefore not surprising that there is a concentration of cassowary–vehicle collisions over a large area between Innisfail and Mission Beach, where several major transport corridors sit within cassowary habitat. Similar associations have been found for other animals vulnerable to vehicle collision elsewhere in Australia (Lunney et al. 2022a). We also observed smaller concentrations of collisions on the Kuranda Range, north-west of Cairns, where a single major road passes through intact cassowary habitat, and north of the Daintree River where a number of minor roads pass through a large area of intact cassowary habitat. These three areas undoubtedly experience the majority of cassowary–vehicle collisions in the Wet Tropics. However, high reporting rates in these areas may introduce a local reporting bias due to active community environmental groups and a strong National Parks staff presence. There are several additional areas in the Wet Tropics that have a high density of major and minor roads passing through cassowary habitat, but a low number of reported collisions in the current available data. These are the Rex Range, the upper Palmerston Range, and the Bruce Highway north of Cardwell. Our analysis suggests that these areas have a high potential for cassowary collision and an absence of reported records may reflect under-reporting rather than true absences. As such, we recommend these areas as a focus for improving the effectiveness of reporting, and that they are also included in vehicle collision mitigation strategies.

It is reasonable to conclude that the density of roads within a habitat will exert a strong influence on animal–vehicle interactions (Goosem 2007; Laurance et al. 2008, 2009). This has been demonstrated repeatedly in a wide range of global studies (e.g. Philcox et al. 1999; Fabrizio et al. 2019). Nor is it unexpected that a higher density of major roads carrying high numbers of vehicles at high speed through an animal’s habitat will result in significantly increased frequency of collisions if no mitigation measures are in place (Shima et al. 2018; Collinson et al. 2019; Find’o et al. 2019; Filius et al. 2020). Therefore, if any major roads, or moderate densities of minor roads (i.e. >400 m/km²) are constructed through cassowary habitat without effective mitigation strategies, then cassowary mortality should be expected.

We show that vehicle collision frequency increases significantly when core cassowary habitat increases above 50% of the landscape. This reflects the habitat specificity of cassowaries and suggests that while vehicle collision can occur when animals are moving between core habitat and agricultural, urban and non-core habitat areas, vehicle collision is more likely to occur while animals are moving within continuous rainforest or between large fragments of rainforest. We also show that we cannot simply assume that higher cassowary densities at a landscape scale will result in more cassowary–vehicle collisions. There are likely to be numerous other variables that influence the likelihood of a cassowaries coming into proximity with and crossing a road, including home range size (Campbell et al. 2012), number of dispersing animals, topography (Mack 1995), localised cyclonic disturbance and resource availability.

Location-specific correlates of vehicle collision

Location-specific correlates of animal–vehicle collision are to some extent species specific but are also related to characteristics of the roads themselves (Gonçalves et al. 2018; Pagany 2020). Our location-specific results show that a combination of road features, high traffic volume and vehicle speed, and adjacent vegetation leads to a higher likelihood of cassowary–vehicle collisions. That is, cassowaries are more likely to be struck by vehicles while crossing a straight section of major road where forest canopy cover is high on both sides of the road. This is likely a function of two separate processes. First, on straight roads drivers feel more confident (Hansen and Chang 2019) and tend to drive faster than on curved roads (Kang et al. 2019). Accidents are more common as speed increases (Richter et al. 2006) and for vehicle–pedestrian collisions, fatalities increase dramatically as speed increases (Hussain et al. 2019). Second, the visual fixation point of a driver is highly correlated with vehicle speed and road curvature (Wilkie and Wann 2003). On curved roads a driver tends to fix their vision on the inner edge of the curve at a closer distance, becoming closer with greater curvature (Land and Lee 1994; Hu et al. 2020). As such, on straight roads a driver’s ability to perceive hazards close to the vehicle is impaired at the same time that a vehicle’s speed reduces the driver’s ability to avoid a collision.

Cassowaries show a tendency to use regular road crossings within rainforest (Goosem et al. 2011) and cassowary sightings on roads have been shown to be spatially clumped (Dwyer et al. 2016). Consequently, there is likely to be a strong relationship between cassowary sightings and vehicle collision (Dwyer et al. 2016). We show that greater canopy cover adjacent to the road is a positive location-specific predictor of collision. However, there is no evidence that any particular value of high canopy cover land use within 50 m or 250 m of the road, such as rainforest, increases the likelihood of collision. There was, however, evidence that low canopy cover land uses adjacent to the road, such as open woodland and grazing land, result in lower-than-average collisions. This likely indicates that cassowaries are avoiding low cover land use types, and drivers can detect and avoid animals more easily when there is minimal vegetation beyond the road verge, regardless of verge width. While this result could be interpreted as support for clearing or thinning vegetation beyond the road verge, such a response would amplify fragmentation and reduce connectivity with negative consequences for biodiversity conservation (Laurance et al. 2009).

Variables from our location-specific analysis only explain approximately 30% of data variance, suggesting other local or incident specific conditions may influence cassowary collisions. The most obvious of these factors are resource availability (e.g. food or water) and cassowary behaviour at that site or time. Resource availability is spatiotemporally variable and challenging to quantify, particularly when done post hoc. Moreover, a paucity of dietary resources may influence cassowary behaviour and vehicle collision as much as an abundance of resources. Cassowaries exhibit frequent movement throughout the day (Westcott et al. 2005), whether foraging or dispersing. They also have unpredictable foraging patterns, regularly returning to an area within hours when resources are scarce, but with a lower frequency when resources are abundant (Westcott et al. 2005). Movement patterns differ between sexes (Campbell et al. 2012) and between resident and dispersing animals. Encounters with vehicles may occur while foraging along a road, using the road as a movement corridor, or while crossing the road within habitat.

There are also cassowary behaviours that aid in predicting vehicle collisions. Cassowaries are overwhelmingly more active during daylight hours, with peaks of foraging activity in the early morning and late afternoon (Bentrupperbaumer 1998; Westcott et al. 2005). These times often correspond to higher traffic volumes along roads. Moreover, cassowaries are often inquisitive, easily distracted, generally slow moving, and therefore are likely to cross a road slowly (Goosem et al. 2011). Slow road crossings by cassowaries can be further aggravated by their opportunistic consumption of the remains of road-killed animals.

Management implications

Our findings that cassowary–vehicle collision is most influenced by the density of roads and the presence of rainforest is not unexpected, indicating that finding an effective solution will not be easy. First, human populations, transport corridors and agriculture are embedded within protected cassowary habitat, resulting in frequent cassowary–vehicle interactions. Second, a projected increase in the human population of this area (Cairns Regional Council 2015) will result in more roads and higher traffic volumes within cassowary habitat. Third, the increasing frequency of cassowary–vehicle collision (2016–2018 annual average of 21.0 vs 1992–2015 of 6.2) provides no evidence that cassowaries are becoming conditioned to roads and traffic. While we cannot exclude more active reporting as an explanation for increased collisions in recent years, the same data provide no evidence that humans are responding to (often emotionally driven) campaigns to change their driving behaviour. Such findings are not only specific to cassowaries, but apply widely to the problem of wildlife roadkill for conservation globally (Lunney 2013).

Given these factors, the only way of reducing cassowary–vehicle interaction is to avoid constructing any new roads within rainforest and to invest significant resources into engineering or re-engineering of existing roads to minimise the interaction between those factors that place cassowaries in danger and that increase the probability of vehicle collisions. For example, large overpasses and underpasses on new and existing roads (Goosem et al. 2011; Goosem 2012) to serve as functional road crossing points for cassowaries should be considered where roads pass through cassowary habitat, where location-specific correlates of cassowary collision exist, and where cassowary crossing locations are identified. As the density of major roads increases, so does the urgency of employing engineering measures in their construction or for the upgrading of existing roads.

The approach to corridor establishment across roads and restoration planting for cassowary habitat near roads should be re-evaluated in light of the findings of this work. Where corridor linkages cross a road, collision mitigation measures such as over- or under-passes, speed humps, guiding structures and speed limits should be incorporated (Goosem 2012; Lunney et al. 2022b). Careful consideration must be given to local context to ensure mitigation measures do not have unintended consequences (Goosem 2008) and that they are in fact effective at mitigating collision risk for cassowaries (Soanes et al. 2024).

Conclusions and future directions

Where habitat that is suitable for cassowaries intersects with roads carrying vehicles, cassowary mortality will continue to occur in the Wet Tropics region. Implementing more effective management solutions is critical to address this threat. Additionally, improved insight on how cassowaries interact with the road network will help to refine collision mitigation measures and understand their effectiveness. Cassowary sightings and vehicle collision, both fatal and non-fatal, must continue to be reported. Any opportunity to increase the likelihood of collisions and sightings being reported should be pursued, including novel ways to obtain this information. Methods to improve the quality and detail of this observation data should also be adopted. For example, greater insight on collision context and outcomes for the cassowary involved could lead to more effective solutions. More broadly, this work contributes to a growing awareness of how including local context is essential when deploying effective road risk mitigation strategies to improve conservation outcomes.