We collected a sample of georeferenced roadkill data from two main sources: (1) monitoring studies across the country; and (2) the “Banco de Dados Brasileiro de Atropelamento de Fauna Selvagem” (BAFS). The first one consists of previously published systematic studies in roads of different regions of Brazil; such data was provided by collaborators (see Acknowledgements – Coelho et al. 2008; Freitas 2009; Caceres 2011; Teixeira et al. 2013; Dornelles 2015; Freitas et al. 2015; Ascensão et al. 2017). The second is a dataset obtained at the Brazilian Center of Road Ecology of the University of Lavras, which gathers geo-referenced and validated citizen science roadkill data from a mobile phone app and from other studies across the country. The app works as follows: the user takes a photo of the roadkilled animal, the app then records the location of the photo that is sent for identification down to species level by an expert in the taxonomic group (Castro and Bager 2019). The records used in our analysis were those with adequate species identification, which depended on the quality of the photo and the degradation degree of the carcass. We selected the roadkill data of the four species focused in this study – Cerdocyon thous, Euphractus sexcinctus, Tamandua tetradactyla, and Myrmecophaga tridactyla – ranging from 2002 to 2015. We chose these species because they present high rates of mortality in Brazilian roads. For each roadkill occurrence, we created a random point in the same road segment, thus representing the pseudo-absence of roadkill. Thereafter, we used these data to build a roadkill presence (ones) and absences (zeros) matrix for each species, with the same number of absences and presences.

For land cover and land use, we utilized the serial time data from MapBiomas (Projeto MapBiomas 2021), with a pixel size of 30 m, between 2002 and 2015. To access the exact landscape composition at the time of each roadkill occurrence, we used the land cover map of the correspondent year of the roadkill register. For example, if a C. thous was registered as roadkilled in 2006, we would collect the data from MapBiomas of the corresponding year and for the correspondent region. This process was performed to all evaluated species registers throughout the 14 years analyzed.

For each species, we considered a different influence buffer radius starting from the place of the roadkill, since each one has different home ranges, body sizes, and habits requirements. We estimated the mean home range for C. thous as 4.9 km² (Beisiegel et al. 2013), for E. sexcinctus as 0.7 km² (Silva et al. 2015), for M. tridactyla as 3.6 km² (Ohana et al. 2015), and for T. tetradactyla as 2.7 km² (Miranda et al. 2015). To assess the potential landscape influencing each individual we used a buffer with twice the radius of the home range approximated to a circle shape (A), resulting in the radius of roadkill influence (φ):

If the area of a circle is given by:

$A=\partial r^2$

where r is the radius of the circle, so the double of a radius of a circle of a given area is:

$\phi =2\left(\sqrt{\frac{A}{\partial }}\right)$

We used this radius size because the roadkill point may have occurred on the center of the home range, or on its border (Fig. 1); this way, considering a bigger radius of influence would prevent us from connecting the wrong area, or a smaller area, with the roadkilled individual. The radius for C. thous was 2.50 km, for E. sexcinctus 0.95 km, for M. tridactyla 2.14 km and for T. tetradactyla 1.85 km.

Figure 1.

Scheme exemplifying the radius of roadkill influence chosen. For a given roadkill point using the simple radius of home range (r), we might exclude some of the landscape characteristics if the roadkill occurred in the border of the home range. Including the possible home ranges (approximated to a circular shape), and doubling the radius (φ), we ensure that all landscape composition associated with the roadkill occurrence is incorporated within the analysis.

For each presence or absence of roadkills we calculated the proportion of land use and land cover inside the buffer based on MapBiomas land cover map for the corresponding year of the roadkill. The classes of land use and land cover considered in the analysis were: (1) forest; (2) savanna; (3) natural open areas; (4) forestry; (5) agriculture; (6) pasture; (7) farming; and (8) water. Farming represents the sum of agriculture and pasture in addition to mosaics or rotation of both classes in the same area. We conduct all landscape analysis and data extraction on ArcGIS v10.3 environment.

To estimate the relative chance of roadkill of each species we constructed binomial generalized linear models (GLM), considering the matrix of presences and absences as our response variables, and the proportion of the eight landscape variables inside the radius of roadkill influence as our predictive variables. We built four groups of models, one for each species, with one or two predictive variables by model, combining variables in pairs, and considering the interaction between them. We discarded models with some degree of correlation (> 0.6 or < -0.6) between predictive variables (see Suppl. material 1: Fig. S1). Models with some level of collinearity between the two predictive variables (Variance Inflation Factor > 4.0) were excluded from the analysis (Tay 2017). Overall, we had 45 models for each species (see Suppl. material 1).

All models were ranked by Akaike Information Criteria (AIC) and selected by their corrected AIC value (AICc), with lower values of AICc representing the best models (Burnham and Anderson 2002). Models with AICc distance equal or lower than two (ΔAICc ≤ 2), and evidence ratio lower than 2 are considered equally plausible. The statistical analysis was performed on the software R 4.0.2 (R Core Team 2020) using the packages ‘bbmle’, ‘numDeriv’, ‘effects’, ‘ggplot2’ and ‘corrplot’ (Fox and Hong 2009; Wickham 2016; Wei and Simko 2021).

We collected a total of 2698 georeferenced roadkill records across the country (Cerdocyon thous (N = 1282); Euphractus sexcinctus (N = 589); Myrmecophaga tridactyla (N = 422) and Tamandua tetradactyla (N = 405)) (Fig. 2). Overall, the distribution of the observations reflects a potential occurrence and abundance of the four species. Some biases can occur, since the data came from previous studies and citizen science, but to assess the landscape composition in the radius of roadkill influence, we had enough data in terms of quantity and spatial distribution, given the potential distribution of each species.

Figure 2.

Roadkill samples distribution for the species studied. Cerdocyon thous represents the majority of roadkill samples, covering the entire territory. The other three species have samples aggregated in central Brazil, mostly in the Cerrado ecosystem.

For Cerdocyon thous and Tamandua tetradactyla only one model was selected as the best model by AIC criteria (ΔAICc ≤ 2 and evidence ≤ 2), while the remaining studied species had two equally plausible models. For Cerdocyon thous the best model shows a positive effect of agriculture and pasture proportion inside the buffer on the chance of roadkill: for each 10% of pasture cover in landscape the roadkill risk increases by 2.7%, while for agriculture it increases by 4.6%. (Table 1, Fig. 3). For Euphractus sexcinctus one selected model shows a positive relationship of roadkill risk with farming and forestry. On the other hand, the other models show a positive relationship with pasture and agriculture, these variables made up the farming class, so we only considered the first model, in which the relative risk of roadkill of Euphractus sexcinctus increases 6,5% and 14,7% for each 10% of land cover increment of farming and forestry, respectively (Table 1, Fig. 5). Both anteaters’ best models include habitat land cover, matrix land cover and the interaction between them, which will be discussed in the next section (Table 1). For Myrmecophaga tridactyla, the best model includes an interaction between pasture and forest areas (Fig. 6), while for Tamandua tetradactyla the selected model includes the interaction between savanna and agriculture areas (Fig. 7). The other selected model for Myrmecophaga tridactyla includes an interaction between forest and savanna areas, but the effect of both variables on roadkill risk was not significant (p > 0.05).

Figure 3.

Best model selected by AIC for C. thous and its estimated coefficients. A model with two variables responded better to C. thous roadkill risk, being pasture and agriculture positively related to roadkill risk.

Cerdocyon thous – The roadkill of this species responds positively to two matrix land-uses, pasture, and agriculture in the landscape (Fig. 3). Since it is a generalist species, it can occur in habitat borders, explore vast areas of human-modified landscapes and even use those regions as home ranges (Ferraz et al. 2010; Bueno et al. 2015). The use of those human-modified habitats by C. thous serves to show that habitat quality can also be an important factor when predicting roadkill risk. In those modified regions, the resources availability is scarcer and diffuser in the space than in preserved landscapes, making individuals move more in search of them (Regolin et al. 2021), thus increasing the chance of road encounter and, consequently, the roadkill risk.

Besides giving information on the studied animal mortality, roadkill records are also useful for assessing a species occurrence. We found registers of C. thous roadkill occurrences out of its original geographical distribution (Lucherini 2015) (Fig. 4). Originally, the crab-eating-fox had its habitat range limited by the dense Amazon forest, but the high-intensity of land use and land cover changes in the region led to the conversion of this forest into open areas, such as pasture and agriculture. That way, Cerdocyon thous may have had its geographical distribution expanded by the land use changes and deforestation in the Amazon forest. The growing agribusiness in the region has synergies with road expansion, which leads to several impacts on local ecosystems (Laurance et al. 2002), leading to fish-bone deforestation patterns (Laurance et al. 2002; Pfaff et al. 2007). These might be related to the savannization phenomena of the Amazon Rainforest (Sales et al. 2020). The fish-bones in the landscape resulted from roads being an arrow of habitat loss, and it was exactly in such regions that our data collection found roadkill registers of Cerdocyon thous (Fig. 4), as did other works on this species’ roadkills (Gumier-Costa and Sperber 2009; Turci and Bernarde 2009).

Figure 4.

Roadkill records of C. thous in Amazon ecosystem. Red dots represent roadkill records of C. thous and the hashed area is the original species’ range on Amazon. It is possible to notice some registers out of the crab-eating-fox’s range according IUCN (Lucherini, 2015), specifically in areas where forests (dark green areas) were converted into farming areas (yellow areas).

As a generalist species, Cerdocyon thous occurs, and is roadkilled in fragmented human-modified landscapes with agricultural and pasture uses. As reported for Chrysocyon brachyurus (Maned-wolf) in the Atlantic Forest (Bereta et al. 2017) and in Amazon Rainforest (Silva-Diogo et al. 2020), the occurrence of carnivores from open environments in forest areas is a result of deforestation, that causes savanna species to expand its occurrence to previously dense forested regions. The creation of those novel ecosystems (Lindenmayer et al. 2008) is related to human-induced modification, and can lead to biotic homogenization (McKinney and Lockwood 1999), when forest dependent species can be replaced for generalist species, compromising the ecosystem functioning. C. thous is a generalist species, and its occurrence in Amazon Rainforest can compromise, by competition, populations and the conservation of other niche-equivalent carnivores, like Atelocynus microtis (short-eared-dog), a more habitat specialist and forest dependent species (Pitman and Beisiegel 2013).

Euphractus sexcinctus – Like C. thous, for this species two land use matrixes are included in the selected model, showing a positive relationship between farming and forestry with roadkill risk (Fig. 5). As all selected models have the same variables (Table 1), we considered, as the most appropriate model, using the one with less complexity to explain the roadkill risk (Table 1, Fig. 5).

Figure 5.

Best model selected by AIC for E. sexcinctus and its estimated coefficients. For the six-banded armadillo the best model represents forestry and farming positively related to roadkill risk, showing that the roadkill of this species is related to human modified landscapes.

This species inhabits a vast number of natural formations, but also human-modified landscapes, such as sugar cane plantations (Dalponte and Tavares-Filho 2004) or even pastures (Anacleto 2007). In the analysis of its threats, the expansion of road network and roadkill is cited as a risk factor (Silva et al. 2015). Since it has generalist habits, its occurrence and roadkill can be associated with commercial tree species in forestry areas, hence its roadkill elevated rates. Forestry regions can be areas of dry soil and low understory coverage, resembling the open dry forests of its original occurrence. The species shows a high density of individuals in the north of São Paulo State, where there is a high landscape coverage of sugar cane, pasture, Pinus sp. and Eucalyptus sp. cultivation (Dalponte and Tavares-Filho 2004; Silva et al. 2015), reflecting the areas where these animals are roadkilled. Furthermore, armadillos provide plenty of ecosystem services (Rodrigues et al. 2020), such as soil bioturbation, seed dispersal, and the construction of borrows, which show an important role in sheltering other species for heat control, nesting, and movement (Rodrigues et al. 2020). That indicates the importance of understanding the occurrence and threats to E. sexcinctus which can affect other species.

Myrmecophaga tridactyla – the best model to predict the relative risk of roadkill for this species matches with its behavior, including its relationship with pasture, forests and the interaction between these variables (Fig. 6). Among the studied species, this is the only one that is threatened according to national and international sources, considered Vulnerable (VU) in the IUCN red list (Miranda et al. 2014a) and in the national red list of Brazilian threatened fauna (ICMBio 2018). The giant anteater is locally extinct in some Brazilian states, such as Espírito Santo, Santa Catarina and Rio de Janeiro (Bergallo et al. 2000; Passamani and Mendes 2007; FATMA 2011) and it is Critically Endangered (CR) in Rio Grande do Sul (Marques et al. 2002). The main threats to this species are habitat loss and land-conversion to agriculture and pasture (Miranda et al. 2015), as well as the expansion of road network, which causes habitat fragmentation and road mortality. Despite the giant-anteater being frequently associated with savannas, like the Cerrado biome, this animal inhabits a wide range of formation types, such as forests, grass-fields and even pasture and agricultural fields, mainly because its main diet consists of ants and termites, which are very abundant in open areas, but it also has a dependency on shaded areas, like forests and understories (Miranda et al. 2015). Camilo-Alves and Mourão (2006) related this activity to a thermoregulatory behavior (Rodrigues et al. 2008), using open areas in milder temperature during the day and sheltering from the sun in warmer weather (Miranda et al. 2015).

Figure 6.

Best model selected by AIC for M. tridactyla and its estimated coefficients. The best model shows an interaction between forest and pasture, both variables positively related to roadkill risk, but the size and effect of direction changes according to the proportion of the other variable.

Depending on the amount of matrix in the landscape, the direction of the effect of habitat on roadkill risk changes. In other words, when there are small quantities of forest, the effect of pasture is positive to predict the roadkill, while when there is an increased proportion of forest in the region the effect changes, and the roadkill risk decreases with the increase of pasture areas. This could be related to the species’ habits: when the landscape is mostly composed of pastureland, the animals need to move more in search of shaded shelter. This movement decreases in frequency when there are some forested areas in the landscape, allowing the individuals to rest, and therefore, decreasing the chance of encountering a road and consequently being roadkilled. This shows the importance of maintaining habitat patches in the landscape, such as riparian forests or even native vegetation fragments inside private rural property, as established by the Brazilian Forest Code (Metzger 2010). Thereby, we hypothesize that the giant-anteater roadkill risk is associated with larger extensions of monoculture or pasture, without natural formations or habitat patches (Noonan et al. 2021; Versiani et al. 2021).

Tamandua tetradactyla – The roadkill risk of the collared-anteater is strongly related to savanna and agriculture patches, and its interaction (Fig. 7). This species, considered as common in Brazil, is listed as Least Concerned in the IUCN (Miranda et al. 2014b), but in Rio Grande do Sul state it is considered Vulnerable (VU) (Marques et al. 2002). Despite occurring in a wide range of habitats (it can be found in all Brazilian vegetation formations), it has a dependence on areas with tree formation, since it is a semi-arboreal animal that feeds on termites and ants in tree canopies, and finds shelter in hollow trees (Ohana et al. 2015). It is considered a forest dependent species, unlike C. thous and E. sexcinctus, and is more closely related in terms of habitat needs with M. tridactyla (Desbiez and Medri 2010).

Figure 7.

Best model selected by AIC for T. tetradactyla and its estimated coefficients. The roadkill risk is negatively related to agriculture. However, the interaction between agriculture and savanna was significant, which changes the effect direction of agriculture in the presence of more savanna areas, increasing the risk of roadkill with the increment of agriculture, in other words, landscapes with savanna and agriculture mosaics are more likely to have collared-anteaters roadkills.

This habitat dependence reflects on the best model selected to predict the collared-anteater roadkill risk: the presence of savanna formations modulates the effect of agriculture. When a landscape has no natural formations cover, the effect of agriculture is negative, since the species probably do not occur in the area; and with the increment of habitat areas, the roadkill risk increases rapidly, reaching our model’s peak when we have at least 40% of savanna and 50% of agriculture.

On the other hand, the roadkill risk when the landscape is entirely comprised of savanna, without agriculture, is very low, and it increases very fast when there is an increase of agricultural coverage. As it is a forest dependent animal, it was unexpected that its roadkill response was better suited to savanna than to dense forests, but that can be explained by this animal’s movement pattern. In areas with continuous dense forests the locomotion of individuals occurs mainly through canopies, but in areas with low density of trees, as open areas, savannas and monocultures, it moves by ground. It can also move more often in search for sheltering trees, therefore increasing the chance of being roadkilled.

It is already known that many factors affect the roadkill risk of a species, such as species density and movement patterns (Rytwinski and Fahrig 2013). We also recognize that the home range for each species varies according to each study area, mainly due to differences in habitat quality (Ofstad et al. 2016; Viana et al. 2018). In this research, we aimed to relate roadkill risk with landscape composition in a national scale in a country with continental properties – Brazil, so there is a limitation on the availability of data to better control those factors. We preferred to use the most accurate data we had, which was mean home range for each species, rather than other inaccurate or non-existent data. Others aspects can also affect the roadkill, as nocturnal or diurnal habits; the flux and speed of vehicles in roads; the presence of crossing opportunities, as underpasses and overpasses; the season and weather; and as discussed above, the population density and habitat quality. All those factors can be considered in further research that can compare roads or road segments between themselves, use roadkill rates and model space and time to have more accurate responses. Here, we search for general patterns in large scale considering only space; future research in smaller scales should make an effort to include these variables in the modelling process and analyze how each of them influence the roadkill risk.