Identification of local factors causing clustering of animal-vehicle collisions
ABSTRACT
Effective measures reducing risk of animal-vehicle collisions (AVC) require defining high-risk locations on roads where AVCs occur. Previous studies examined factors explaining locations of individual AVCs; however, some AVCs can form hotspots (i.e., clusters of AVCs) that can be explained by local factors. We therefore applied a novel kernel density estimation (KDE) method to AVCs for the Czech Republic from October 2006 to December 2011 to identify AVCs hotspots along roads. Our main goal was to identify local factors and their effect on the non-random (clustered) occurrence of AVCs. The remaining solitary AVCs occurred randomly and are likely induced by other human factors on the global scale. The hotspot identification method followed by the selected data mining methods (KDE+ methods) identified factors causing local clustering of AVCs. Distance from forest (<350 m) or linear vegetation were important factors for estimating presence of clusters of AVCs; in open areas, AVC clusters were absent. Further research on effectiveness of measures reducing risk of AVC should focus on clusters of AVCs, not on the individual AVC. We recommend that state transportation agencies focus mitigation actions in forested areas. © 2018 The Authors. Journal of Wildlife Management Published by Wiley Periodicals, Inc.
In recent years the proportion of animal-vehicle collisions (AVCs) to all registered accidents is rising in the Czech Republic (e.g., 6% in 2013, to >10% in 2015; Ritschelová et al. 2015). Road networks and the vehicles that use them are increasing; both limit safe movement of animals through the landscape (Adams and Geis 1983, Haddad et al. 2015). Animal-vehicle collisions are primarily recorded in cases when consequences are fatal for humans or lead to significant damage to vehicles. The number of AVCs is increasing in Europe, which has a major effect on road safety. Large mammals, mainly ungulates, are the main group involved in the conflict and most AVCs are caused by red deer (Cervus elaphus), roe deer (Capreolus capreolus), fallow deer (Dama dama), and moose (Alces alces; Waechter 1979, Groot-Bruinderink and Hazebroek 1996, Haikonen and Summala 2001). Wild boar (Sus scrofa) are also involved in AVCs in Mediterranean areas and central European countries (Saénz-de-Santa-María and Tellería 2015, Colino-Rabanal and Peris 2016). The sharp increase in AVCs observed during the last decade has been attributed to 3 main causes: demographic expansion of ungulates, increase in velocity of cars, and length of the road network (Rosell et al. 2013). During the last decade a series of environmental measures increased road safety and reduced the conflict with game, which is important to hunting associations (Malo et al. 2004).
Roads and adjacent landscaping as strict barriers to animal movement negatively affect important wildlife habitats (Meffe and Carroll, 1997) through forest fragmentation and alteration of bio-corridors between habitats (Bennett 1999). When animals are moving on a long section of a road, AVC risk is less than when they move in a narrow corridor, connecting a road body. These risky sections receive greater attention and most studies reported increased mortality (Bennett 1991, Georgii et al. 2011); several measures were proposed to reduce risk of AVCs (Harris and Scheck 1991).
Many projects involve implementing structures to keep wildlife away from roads (i.e., culverts, bridges; Rytwinski et al. 2016). Roads are also fenced at places of frequent AVCs (Forman and Alexander 1998), audio and optical signal systems are installed to discourage animals from crossing roads (Rowden et al. 2008), and speed limits, warning signals (Al-Ghamdi and AlGadhi 2004), and thermal sensors draw the attention of drivers to the presence of large animals (Hirota et al. 2004). Despite many of these measures and constructions, traffic accidents caused by animals are increasing (Iuell et al. 2003) perhaps because systems to identify AVC clusters are often determined separately for different taxonomic groups (Iuell et al. 2003, Glista et al. 2009), by different numbers of observers, or with differing frequency and duration of surveys (Teixeira et al. 2013). With differing methodologies, some road kills may be missed and mortality may be underestimated when there is a long time interval between searches for AVCs or species with soft or small bodies, which rapidly decay (Gerow et al. 2010, Guinard et al. 2012).
Animal-vehicle collisions are clustered in time and space (Beaudry et al. 2010). An understanding of where and when AVCs occur is important to avoid high-risk areas and design effective mitigation measures. Many AVCs are limited by spatial error, which may vary from several hundred meters to a few kilometers (Gunson et al. 2009) when locations are referenced to the closest landmark. When some techniques that measure clustering of AVCs are applied, locations differ spatially from randomly distributed AVCs (Coelho et al. 2014, Gunson and Teixeira 2015, Bíl et al. 2016).
The objective of our study was to present an approach based on an innovative kernel density estimation method (KDE+ method; Bíl et al. 2016) to precisely identify places where AVCs clusters occur. We tested when the largest proportion of AVCs occurred daily and annually. We compared AVCs within clusters with those outside clusters to determine if environmental characteristics differed and if some factor may predict the presence of hotspots.
STUDY AREA
The Czech Republic, approximately 79,000 km2, is a hilly plateau surrounded by relatively low mountains. Bohemia, at the west part of Czech Republic, consists of a Bohemian Massif, which rises to heights of 1,000 m above sea level. This ring of mountains encircles a large elevated basin, drained by the Labe and Vltava rivers. Moravia, the eastern part, is also hilly and is drained by the Morava River. Climate of the Czech Republic is mild but variable locally and throughout the year depending on elevation. In the Czech Republic diverse geological composition and geomorphology, together with diversity of geographical and climatic conditions resulted in a relatively high biological and landscape diversity. There are 369 vertebrate species, over 29,800–48,000 invertebrate species, and 2,520 vascular plant species (Williams et al. 1998).
The state-owned Road and Motorway Directorate of the Czech Republic manages and maintains 1,250 km of motorways. The greatest traffic volume on the Czech motorways appear on highway D1 near Prague, where volume amounts to 98,000 vehicles/day. Average traffic volume on the Czech motorways (speed limit = 130 km/hour) is approximately 25,000 vehicles/day, whereas on national roads (speed limit = 90 km/hour) average traffic volume is 9,000 vehicles/day (Ritschelová et al. 2015).
METHODS
Measured Variables
We obtained data on AVCs for the Czech Republic from the Police of the Czech Republic from October 2006–December 2011. Police of the Czech Republic use Garmin Geko 201 (Garmin, Olathe, KS, USA) with a maximum error of 25 m and record the time and date of collisions. Data on AVCs often contained information on species of large mammal, sex, and age of individual involved in an accident.
We determined for each AVC, presence or absence in clusters according to the KDE+ method (Bíl et al. 2013). The KDE+ approach builds on the KDE method (Chung et al. 2011), which estimates the probability density function of the underlying data by using a kernel function. However, KDE produces a range of local maxima (clusters) that are not differentiated from one other because there is no objectively determined threshold. We therefore extended the framework of the standard KDE method by introducing repeated random simulations (Monte Carlo method), selecting only significant clusters and ranking them. We also gathered information on the animal (wild boar; roe, red, and fallow deer; red fox [Vulpes vulpes]; and European hare [Lepus europaeus]), age, sex (if available), traffic factors (i.e., road width, road category, presence or absence of a middle belt, guardrails), and environmental factors (Table 1).
| Variable | Metric or categories | References |
|---|---|---|
| Presence in cluster | Out of clusters, in a cluster | |
| Date of collision | dd.mm.yyyy | |
| Time of collision | hr and month | |
| Variables connected to animals | ||
| Mammal species | Roe, red, fallow deer, wild boar, other | |
| Sex | Male or female | |
| Age | Juvenile, yr-old, adult | |
| Traffic variables | ||
| Road width | m | Forman and Alexander (2008) |
| Road category | First, second, third class | Donaldson and Bennett (2004), Forman and Alexander (2008) |
| Middle belt | Yes, no | Adams and Geis (1983) |
| Guardrails | Yes, no | Clevenger and Kociolek (2006) |
| Environmental variables | ||
| Distance to other barrier | m | Beier and Noss (1998) |
| Distance to the forest | m | Trombulak and Frissell (2000) |
| Distance to the linear vegetation | m | Hubbs and Boonstra (1995), Bruinderink and Hazebroek (1996) |
| Distance to the stream | m | Hubbs and Boonstra (1995), Beier and Noss (1998), Christie and Nason (2003) |
| Distance to the built-up area | m | Trombulak and Frissell (2000) |
| Forest area | % | Ashley and Robinsons, (1996), Pocock and Lawrence (2005) |
| Industrial zone | % | Dyer et al. (2001) |
| Habitat | Closed, semi-closed, open | Murcia (1995), Eigenbrod et al. (2008) |
| Embankment | Yes, no | Barnum (2003), Cain et al. (2003), Malo et al. (2004) |
| Depression | Yes, no | Gunson et al. (2011) |
| Shrubs | Yes, no | Seiler (2001) |
| Grass belt | Yes, no | Adams and Geis (1983), Huijser and Clevenger (2006) |
| Bank vegetation | Yes, no | Clevenger et al. (2003) |
We obtained variables that described road segments from geographic information system (GIS) layers provided by The Road and Motorway Directorate. We obtained environmental variables from digitized GIS layers (e.g., boundary of forest and green areas, rivers, land cover [i.e., industrial areas, building areas]), and an orthophoto map provided by the State Administration of Land Surveying and Cadastre (SASD). The extent of the GIS layers was equal to the buffer of 300 m around each AVC location within a scale of 1:1,000 from the Web Map Service (WMS) orthophoto map (provided by SASD).
Statistical Methods
We first studied the relationship between AVCs and date and time of the collision. We considered 24 time intervals for each month representing time of day. We therefore had 12 × 24 cells. We consequently counted number of records belonging to a particular cell. Assuming that an AVC could occur equally likely in any cell, we would expect N/(12 × 24) AVCs in a particular cell, where N is number of AVCs in question. For each cell, we applied the exact binomial test (Hollander et al. 2014) to examine if the observed number of AVCs was significantly greater than the expected number of AVCs.
Second, we analyzed influence of explanatory variables on clustering separately by using the odds ratio (Simon 2001). We calculated the 95% confidence interval of the odds ratio to examine significance of the dependency between an explanatory variable and clustering. We accounted for all possible partitions into 2 groups for regarding continuous variables or categorical variables with >2 categories. Because odds ratios can be calculated only for 2 × 2 contingency tables, we could test directly only binary variables. Concerning continuous variables or categorical variables with >2 categories, we examined all possible partitions of the AVCs and studied only the partition with the strongest influence.
The third step in our analysis was to work with all explanatory variables simultaneously. We categorized each continuous variable and then applied a multiple correspondence analysis (Abdi and Valentin, 2007) to reduce dimensionality in the data. As a result, we obtained several new variables (i.e., dimensions), which are linear combinations of original explanatory variables. We then constructed a logistic regression model with dimensions as explanatory variables and presence in a cluster as the dependent variable. We then reconstructed the meaning of the original variables through coefficients of the logistic regression model and relations known from multiple correspondence analysis.
Consequently, we were able to express how many times clustering was more likely in the case that an original explanatory variable had a particular value. If the lower limit of the 95% confidence interval was >1.00, clustering of AVCs was affected positively by the value of an explanatory variable. In contrast, when the upper limit of the 95% confidence interval was <1.00, clustering of AVCs was affected negatively and collisions occurred more likely at random. We performed all computations using software R (R version 3.2.5, www.r-project.org, accessed 6 Apr 2016).
RESULTS
We obtained 599 records containing 164 (27.4%) AVCs that occurred in clusters (Fig. 1). Most of the AVCs were with roe deer (n = 382 [63.8%]), followed by wild boar (n = 180 [30.1%]), red deer (n = 18 [3.1%]), and other mammals such as fallow deer, foxes, and hares (n = 19 [3.2%]).
Animal-vehicle collisions are strongly affected by the relative sun height above or below the horizon (Haikonen and Summala 2001). It is therefore natural that time of day when AVCs occur changes during the year. Roe deer AVCs were the most frequent in May around twilight and dusk. In contrast, AVCs involving wild boar occurred mostly in November and October between dusk and midnight (Fig. 2). Most of all recorded AVCs occurred at night (79%).
We applied the exact binomial test to each cell (Fig. 2). Regarding roe deer, we determined that cells with ≥4 AVCs had more AVCs than expected (Fig. 2). For wild boar, the time periods with ≥3AVCs indicated more AVCs than expected (Fig. 2). Although collisions with roe deer occurred at any time of the day, most AVCs with roe deer and wild boar occurred at night.
We used the odds ratio method to separately examine influences of variables on clustering of AVCs. We determined that when distance from forest was <350 m, odds of clustering was >6 times greater than when the distance from forest was >350 m (Table 2). Any forest in the buffered area, roads ≥7 m wide, streams closer than 120 m, first-class roads, barriers farther than 120 m, shrubs, and depressions also increased the odds of AVC clustering (Table 2).
| Variable | Odds ratio | 95% CI | |
|---|---|---|---|
| Distance to the forest <350 m | 6.41 | 2.54 | 16.17 |
| Forest area >0% | 4.81 | 2.37 | 9.78 |
| Road width ≥7 m | 2.68 | 1.42 | 5.07 |
| Distance to the stream <120 m | 2.40 | 1.62 | 3.57 |
| First-class road | 2.05 | 1.35 | 3.11 |
| Distance to other barrier ≥120 m | 1.72 | 1.19 | 2.50 |
| Shrubs | 1.68 | 1.01 | 2.81 |
| Depression | 1.59 | 1.07 | 2.40 |
We categorized continuous explanatory variables (Table 1) and performed multiple correspondence analysis. As a result, we obtained 3 dimensions explaining 93.8% of the inner variability of the dataset. We consequently constructed a logistic regression model with 3 explanatory variables (3 dimensions) and presence in a cluster as a dependent variable. The dimension that contained all explanatory variables (dimension 2) was not statistically significant (P > 0.05). Dimension 1 (P = 0.005) contained all variables except distance to stream and bank vegetation and explained 54.26% of the variance. Dimension 3 (P < 0.001) contained all variables except distance to built-up area, middle belt, and industrial zone and explained 10.12% of the variance. Thus, we included dimensions 1 and 3 as explanatory variables in our logistic regression model. We were then able to estimate the probability of clustering at any location where explanatory variables were known. Furthermore, we reconstructed the meaning of original variables through the coefficients of the logistic regression model and relations known from multiple correspondence analysis (the dimensions are linear combinations of original variables).
The meaning of an original variable X taking a particular value x (e.g., X = road class, x = first class) can be expressed as o(X = x) = p/(1–p), where p is the probability of clustering under the assumption that X = x. Number o(X = x) quantifies how many times clustering is more likely when X = x compared to the case X ≠ x (e.g., clustering is 1.06 times more likely for first class roads than in the case of other roads). We also computed the 95% confidence intervals for each o(X = x). Hence, we divided presence of an attribute (X = x) into 3 groups (Fig. 3): positively affects clustering (CI > 1.00), no effect on clustering (CI contains 1.00), negatively affects clustering (CI < 1.00). Distance from the forest and bank vegetation were the most important factors affecting presence of clusters of AVCs. In open areas, AVC clusters were absent. Presence of semi-closed habitat, roads ≥9 m wide, streams, shrubs, embankments, and first- and third-class roads and absence of guardrails, grass, and middle belts also positively affected clustering (Fig. 3).
DISCUSSION
Temporal patterns of wildlife mortality have been summarized in several studies (Smith and Dodd 2003). We confirmed peaks of mortality related to diurnal or annual animal movements. Collisions with roe deer occurred at any time of the day, whereas AVCs with wild boar often occurred at night. The greatest number of collisions with roe deer was recorded between May and July. Stache et al. (2013) reported roe deer collisions were correlated with increasing daytime activity. However, in our results, peak of AVCs was pronounced at dusk. Longer periods of wild boar activity in autumn reported by Podgórski et al. (2013) were not reflected by our data regarding the diel AVC distribution. Most AVCs in our study occurred just after sunset during October and November. Annual AVC distribution was linked with pre-noon for roe deer or weaning in year-old females, and the wild boar main rut (Groot-Bruinderink and Hazebroek 1996).
Spatial collision patterns (AVC cluster positions along roads) may be used to monitor longer periods of time. For instance, such a significant spatial association was reported between hotspots sampled for subsequent years (Garrah et al. 2015). Santos et al. (2015) suggested that numbers of hotspots are negatively correlated with an increasing time interval between surveys, due primarily to missing hotspots. The sampling interval and carcass rate persistence (e.g., correlated with body size) affect, among other factors, accuracy of hotspot identification (Teixeira et al. 2013). Our models, however, are mainly based on large mammal-vehicle collisions whose carcasses remain on roadways longer than smaller animals.
Road mortality surveys are often evaluated separately for different taxonomic groups. Different sampling for each taxonomic group only increases variability among taxons and complicates determining hotspots using local characteristics (Langen et al. 2009). Such findings emphasize that by adding multiple taxons for multi-year studies, managers can identify locations where security features may be achieved (Garrah et al. 2015).
Our dataset did not perform a systematic road-kill survey because AVCs collected by the Police of the Czech Republic are usually limited to reported accidents. However, when we determine the factors affecting only clustered AVC spatial distribution patterns, they are well suited for the analysis we performed. In addition, to consider the risk (AVC) by an insurance company, any person involved in a collision may call the Police. Therefore, the police database includes nearly all road-kill accidents that were fatal for humans or caused damage to vehicles, and includes accidents when the driver rolled off the road to avoid a collision with the animal. The coordinated police dataset accompanied with awareness campaigns have the potential to maximize the amount of AVC data collected and integrated into a single system to mitigate planning.
Most road-kill studies have been conducted on a local scale with variable buffer sizes around the location points (Malo et al. 2004, Seiler 2005, Grilo et al. 2009). Unfortunately, this concept has caused high heterogeneity in the datasets. We therefore applied a new statistical concept and methodologies when clusters of AVCs were compared with other collisions, which were not located within a cluster and were randomly distributed (Bíl et al. 2013). In our study, we first established genuine hotspots (AVCs that were clustered in time and space), and then explained their presence by environmental characteristics.
Factors potentially responsible for occurrence of AVCs are usually divided into 3 groups: connected with species ecology and behavior (sex, age, dispersal, habitat use, migration), traffic factors (vehicle velocity, visibility, density), and environmental factors (presence of natural corridors and their fragmentation; Davenport and Davenport 2006). Habitat variables, however, have been related to road fatalities elsewhere (Clevenger et al. 2003, Malo et al. 2004, Seiler 2005) and our results are consistent with the findings of these studies. Roadside vegetation coverage (Seiler 2005) and presence of corridors (streams and forest edges) can be more important in explaining the limited AVCs distributed elsewhere and explaining the aggregation of AVCs into clusters. In road-kill models, forests are an important prerequisite for large mammal-vehicle collisions in Europe (Almkvist et al. 1980, Kofler and Schulz 1987) and in the United States (Finder et al. 1999, Hubbard et al. 2000). The short distance to a forest (<350 m) was also the most important explanatory factor in our survey, followed by length of forest edge. Generally, these variables increase the chance of presence of a cluster of AVCs. Moreover, presence of these features and their co-occurrence with other variables at same locality explained road-kill patterns.
Habitat corridors may potentially mitigate some of the worst effects of habitat fragmentation (Bennett 1999), but their importance can only come from the ecological centers that they are connecting. On the one hand, edges between habitats are fundamental structures in landscape function, and hence are of central importance in conservation biology (Lidicker 1999). The high ratio of edge to area, on the other hand, might be detrimental to species using the corridor (Weldon and Haddad 2005). Human activities in the agricultural landscape produce poorly connected windbreaks or abrupt alleys. The result of animal habitat use therefore could be a preference for falsely attractive habitats. Moreover, edges are widely used by animals to move and some corridors can act as habitat sinks or ecological traps (King et al. 2009).
The second category of factors influencing number and likelihood of AVCs comprises traffic density and vehicle velocity. However, detailed information on traffic intensities is usually unknown and traffic volumes are frequently substituted by annual average daily traffic (AADT). In our current research (Z. Andrášik, CDV Transport Research Centre, unpublished data), we report that AADT is a poor and unreliable predictor of occurrence of AVC clusters. Because the relative frequencies of AVCs/hour during a day are different when compared to the hourly traffic volumes during a day, AADT does not well represent traffic intensity. Furthermore, traffic volume within an hour is not proportional to AADT in general. If we assume, however, that traffic intensity is correlated with type of road (highway, local road), a greater density of collisions is often reported on intermediate roads than on major highways or on local access roads (e.g., amphibians, Kuhn 1987; small mammals and birds, Oxley et al. 1974; carnivores, Clarke et al. 1998; and ungulates, Skölving 1987). After the quick appearance of a vehicle, large mammals return to the point of entry on the road, increasing the likelihood of collisions. Significantly fewer collisions occurred on minor county roads with reduced speed limits and on unfenced highways with traffic denser than 8,000 vehicles/day (Seiler and Helldin 2005). Highways with traffic levels >10,000 vehicles/day are therefore considered an insurmountable barrier for most terrestrial vertebrates (Rosell and Velasco Rivas, 1999).
MANAGEMENT IMPLICATIONS
Based on the KDE+ method applied in this study, we identified factors influencing the non-random occurrence of AVCs. Clusters of AVCs are caused by local factors that must be identified. In our study, these factors included proximity to forest, stream, or other barrier; road width; presence of a national road; and shrubs near the roadside. Roadside management should therefore focus on existing clusters where the combination of the most AVC-provoking factors occur. Local mitigation measures would be most effective by preventing clustered collisions (i.e., hotspots). Therefore, it would be useful to prioritize mitigation action by state transportation agencies to focus only on AVC clusters and thus protect biodiversity and driver safety more effectively than in the past. We recommend monitoring the police database every few years using the KDE+ method to get cluster locations in a region of interest. Confirmed clusters should be recorded systematically, marked with warning signs, or apply additional mitigations. Long-term verification of the presence of AVC clusters enables managers to measure the effectiveness of mitigations and to find new AVC clusters as they develop.
ACKNOWLEDGMENTS
We thank Z. Janoška for his help with data preparation and D. Livingstone for assistance with language editing. We would like to express our gratitude to P. R. Krausman, B. D. Leopold, and reviewers for their valuable suggestions, proofreading, and production of the manuscript. This work was supported by project of Transport R&D Centre (LO1610) and Masaryk University, Brno, Czech Republic (MUNI/A/1078/2017).
LITERATURE CITED
-
Associate Editor: Bruce Leopold.
