Globally, there are dozens of web-based systems for reporting WVC. For example, the Swedish National Wildlife Accident Council maintains a website for official reporting of accidents involving animals (http://www.viltolycka.se/hem/). The system is operated by the Swedish National Police, and it is the largest agency-owned WVC-reporting system in the world with over 200,000 records of WVC in the last five years. Online reporting and data display has been in place since 2010, but data from police records of accidents are available back to 1985. The largest, longest-running system that relies on volunteer-observers reporting any vertebrate species is the California Roadkill Observation System (CROS), maintained by the Road Ecology Center at the University of California-Davis (http://www.wildlifecrossing.net/california). In the US, the Idaho Department of Fish and Game operates the Idaho Fish and Wildlife Information System – IFWIS (https://fishandgame.idaho.gov/species/roadkill). The system allows entry of observation of any carcass resulting from WVC and as of 12/2014 had >22,000 records. Many observation systems have appeared over the last five years and they vary in their specific purpose, taxonomic breadth, and use of social networks for collecting data and outreach. A few use smartphone-based applications to facilitate data entry from the field (Olson et al. 2014) and some use social media and communication tools to receive observations (e.g., Project Splatter in the UK, http://projectsplatter.co.uk/). One purpose of this study was to find out whether it is possible to use the data from web-based informatics systems containing volunteer wildlife observations, to plan for WVC mitigation at the scale of US states.

Existing WVC reporting systems can consist of tens of thousands of data points and represent a potential source of “big data” for road ecology, community ecology, transportation mitigation, biodiversity mapping, and other scientific/engineering disciplines. Big data refers to datasets that are large and usually geographically extensive, and so require novel solutions for storage, analysis, processing and visualization (Hampton et al. 2013). At the US state scale and possibly at a global level, WVC reporting systems provide the largest known, continuous source of data on the occurrence and distribution of a wide taxonomic range of wildlife whilst also providing opportunities for tissue sampling of genetics, disease, and other testing. Carefully structured informatics (i.e. collection, storage, management, and sharing) systems for these observations facilitate ecological analyses and other biological uses of the data.

One common finding with spatial analysis of WVC is that collisions are clustered, which often leads to analysis of proximate causes of clustering for individual species (e.g., road or landscape features; Gunson et al. 2011). One approach is to use previous collisions to develop predictive landscape models to find “hotspots” (Nielsen et al. 2003; Langen et al. 2009; Gunson et al. 2011; Bil et al. 2013), or seasonality models to find “hot moments” (Beaudry et al. 2010). This is often done for ungulates because collisions with ungulates are both a conservation and safety concern (e.g., Danks and Porter 2010). There are various costs associated with a collision between a deer and a vehicle; on average, a collision with a deer costs $6,671 to society (Hujser et al. 2009). This approach means that WVC can be measured in terms of their cost to society, which can matter regardless of clustering of WVC. Less well-studied than WVC clustering is the idea that for broad taxonomic groups, “sheet flow” of animals may result in WVC everywhere and statistically-significant clustering may only be found because of limitations in the study area, or data collection. For highway planning, it is important to understand the clustering for individual species in each of their habitats and landscapes, and determine the reasons why higher WVC occurs on that stretch of road.

There are many tools to measure impacts to species from WVC, to determine causes and correlations with WVC, and for finding places where transportation agencies can focus remedial action to reduce impacts to wildlife and improve driver safety. Analysis to identify non-random clusters of single or multiple species WVC’s (hotspots) has utilized GIS (Geographic Information Systems); a promising tool where statistics have been used to identify spatial clusters. Examples of analytical approaches and methods include: Nearest Neighbor Index (e.g. Matos et al. 2012); ‘Satscan’, borrowed from epidemiological studies, which looks for non-random clusters of events (i.e. disease outbreaks, Ball et al. 2008); the Getis-Ord- Gi statistic for spatial autocorrelation (Getis and Ord 1992); and the Kernel Density Estimator Plus method for estimating locations of high densities of events (Bil, personal communication).

We hypothesize that volunteer-collected observations of WVC could be used to prioritize roadway sections for mitigation action. We describe the use of data from state-scale, online observational networks for roadkill/wildlife occurrences in California (CA) and Maine (ME). We found that there were sufficient data to identify statistically-significant “hotspots” for many of the states’ highways. We propose that novel online, volunteer-based systems could be used to augment the efforts of state DOTs and wildlife agencies and help inform location and type of mitigation actions.

Volunteer-collected data were downloaded for each of CA and ME from their respective online systems. Date ranges for CA August, 2009 to October, 2014 and for ME were June, 2010 to November, 2014. WVC (n = 12,064) for specific highways were selected by hand in GIS based on their proximity to the highway. Any question about which of adjacent roadways a WVC was associated with was resolved by referring to the WVC record, which includes a narrative description of the site of observation.

The California Department of Transportation (Caltrans) maintains databases for carcass retrieval by District maintenance staff and for deer-vehicle-collisions (DVC) requiring a report and attendance by the California Highway Patrol. Partially-complete data-sets were retrieved from Caltrans using a request under the California Public Records Act. Data for portions of two Districts (3 & 4), were the most complete for carcass retrieval and accident reporting. Carcass retrieval data for 1984-1997 and 2001-2009 and DVC data for 2008–2010 were obtained for District 3, I-80 and SR50, and carcass/DVC data for 2005–2012 were obtained for District 4, I-280. DVC were summarized by tenth post-mile for each highway. Data from transportation-maintenance staff in Maine were not available at the time of the study.

Two types of “hotspot” analysis were conducted: a test for spatial autocorrelation, which identifies highway segments statistically-different from their neighbors, and calculation of WVC-density (# WVC/km-year), which allows comparison of WVC against some threshold of concern (Wang et al. 2010). These approaches are complementary in that there may be interest in high-densities regardless of whether or not clustering is statistically significant; conversely there may be interest in identifying geographically-discreet areas for mitigation action.

Each highway was dissolved into one long line segment and subsequently cut into regular-length segments of 0.40 km (0.25 mi) to 1.6 km (1 mi). These lengths were chosen because of previous research indicating that these are appropriate road segment lengths for studying wildlife crossings and WVC (Malo et al. 2004; Taylor and Goldingay 2004). WVC observations were forced into co-location with their respective highways using a “snap to line” tool (https://github.com/robintw/RTWToolsForArcGIS) implemented in ArcGIS 10.1. The “spatial join” tool in ArcGIS 10.1 was used to sum the number of observations per line segment and these sums per line segment length were used as the basis for density-based analyses and for subsequent spatial autocorrelation analysis.

We used a measure of spatial autocorrelation test called the Getis-Ord Gi* z-score statistic (Getis and Ord 1992) to determine whether or not WVC observations in California and Maine were spatially clustered in “hotspots” along highways. The Getis-Ord Gi* z-score is a measure of the statistical significance of clustering for each analysis unit, in this case highway segments. The Getis-Ord Gi* z-score was calculated using the default settings in ArcGIS 10.1.

Highway-specific observations were separated by year of observation, for full years of data: 2010, 2011, 2012, and 2013. Spatial autocorrelation of observations was determined for each year of observations. Different lengths of highway segment can affect where hotspots are identified. Shorter segment lengths (e.g., 0.4 km) may result in more hotspots than longer segments (e.g., 1.6 km) because there is greater likelihood at shorter distances that there will be a difference among segments in terms of # carcasses than at greater distances. The potential effect of varying highway segment lengths on hotspot identification was analyzed by carrying out autocorrelation analysis with 3 segment lengths: 0.4, 0.8 and 1.6 km.

Caltrans WVC data were used separately from volunteer-collected data from the California Roadkill Observation System (CROS) to analyze spatial autocorrelation and carcass density. Mule deer (Odocoileus hemionus) comprised >95% of Caltrans observations for many highways and were selected from all Caltrans data (carcass retrievals and collisions) to determine density of deer-vehicle-collisions (DVC) along select highways.

We also used estimates of the total cost of deer-vehicle collisions to provide estimates of the cost per mile segment per year from deer-vehicle collisions (Hujser et al. 2009). Deer-vehicle collision data were from both Caltrans and CROS databases and were summarized to the tenth post-mile. There are various costs associated with a collision between a deer and a vehicle. On average, a collision with a deer costs US$6,671 (Hujser et al. 2009). We used this estimate of the total cost of DVC and segment-specific densities of DVC to provide estimates of the cost per mile segment per year from DVC. This provides another way to prioritize areas for mitigation, including both spatial location and economic benefits from mitigation action.

The total number and length of statistically-significant clusters (p < 0.05), or “hotspots”, were determined for highways and interstates in each of California and Maine (Table 1, Figure 1). Twenty-eight percent (6,940 km) of California’s 25,041 km of state highways and interstates and 9% (947 km) of Maine’s 10,900 km of state highways and interstates were analyzed for hotspots, though not all highways had WVC observations along their entire length. The length of individual hotspots varied considerably, from 0.8 km to 17.7 km. The total length of all hotspots increased significantly with length of highway analyzed (p < 0.02) at a rate of 10%, 0.10 km/km (Figure 2). If this rate held for all highways, the total length of hotspots would be 2,504 km in California and 1,090 km in Maine.

Figure 1.

Locations of hotspots on California and Maine highways. The Gi* statistic, Z-score indicates the statistical significance of WVC clusters. A score of >1.96 indicates a statistically-significant cluster (p < 0.05); scores lower than 1.96 are not significant (p > 0.05).

Figure 2.

Relationship (CA and ME) between length of hotspots and highway length. The formulas and R² values are for the combined ME and CA data.

A few highways had sufficient data to conduct year-specific cluster analysis for 2010, 2011, 2012, and 2013. For one example highway, CA-49, certain hotspots persisted throughout the 4 years of data collection (Figure 3A) and the majority were present in one or several years (Figure 3A). For another highway, CA-13, locations of hotspots varied from year to year (Figure 3B). One example highway (CA-190) was segmented into varying-lengths for analysis, from 0.40 km to 1.6 km (Figure 3C). There was a tendency for shorter segments to result in a greater number of identified statistically-significant clusters (0.4 km: n = 20 clusters) and longer segments to result in fewer and longer clusters (1.6 km: n = 4 clusters). For many of the highways, the statistically significant hotspots often overlapped at these different scales.

Figure 3.

Geographic variation in hotspots among years and with varying highway-segment lengths. Annually-specific hotpots for A CA-49 and B CA-13 C Variation in position and number of hotspots along CA-190 with varying segment lengths: 0.4, 0.8, and 1.6 km segments.

The vast majority of Caltrans observations were of mule deer (Odocoileus hemionus). For example, during one reporting period along I-80 (1967 to 1992), there were observations of 906 mule deer, 5 black bear (Ursus americanus), 1 beaver (Castor canadensis) and 1 raccoon (Procyon lotor). This dominance of observations of deer is likely to be different for more urban areas. In comparison, observations from the CROS for I-80 (2009 to 2014) included 679 individuals from 63 species, with 69 individuals being mule deer. For the highways where state agency and volunteer-collected data were available, the carcass counts from each source for the most part did not overlap (Figure 4A, B). In other words, where carcass counts from CROS were high, carcass counts from Caltrans were often low or nonexistent, and vice-versa. Similarly, the hotspots calculated using each source of data did not overlap with each other (Figure 4C, D). State agency data were dominated by mule deer carcasses, which were primarily collected at higher elevations and away from urban areas. Although data collection by volunteers also occurred in these areas, hotspots from their data were primarily identified near developed urban and agricultural areas.

Figure 4.

Comparison of state agency and volunteer-collected data-based hotspots. Carcasses reported in the CROS system (inner blue-range segments) overlaid with carcasses reported in the Caltrans system (outer red-range segments) along CA-80 (A) and along CA-50 (B).

Figure 4.

Continued. Locations of statistically-significant clusters from CROS data (purple segments) and Caltrans data (red segments) for CA-80 (C) and for CA-50 (D).

Identifying locations of WVC clusters is one type of information useful for transportation mitigation planning. Identifying locations of high-cost from deer-vehicle collision (DVC) is another type. For one highway (CA-50), there was some overlap of hotspots identified from volunteer observations of all species of WVC and 2 locations of high estimated cost of DVC from volunteer and DOT observations (Figure 5). For CA-50, the estimated annual cost of DVC ranged from <$500 to >$10,000 per mile (Figure 5). For another highway (I-280), according to Caltrans databases, there have been 362 collisions with deer between January, 2005 and July, 2012, or roughly 48/year. For I-280, there was very little overlap between the single hotspot identified from volunteer observations and the longer stretches of high estimated annual cost from DVC (data not shown). Also, the estimated annual cost of DVC was higher than for SR 50, reflecting a higher rate of DVC, and varied from <$1,000 to >$40,000 per mile.

Figure 5.

Locations of potential, cost-effective areas for mitigation. Statistically-significant clusters using volunteer observations (outer purple segments), annual rate of DVC per post-mile (points), and associated costs ($/mile, gray segments) of DVC for CA-50.

Animals die as result of collisions with vehicles because of traffic speed, traffic volumes, seasonal changes in movement, separation of important habitat areas, occluded line-of-sight, and other factors (Barthelmess 2014; Hobday and Minstrell 2008; Litvaitis and Tash 2008). Most of the observations of dead animals made using the online, state systems described here were opportunistic and thus do not reflect actual rates of WVC on a particular roadway. For certain highways analyzed in the present study (CA-13, CA-190, CA-94, I-280) known observers have consistently and frequently made observations of WVC, thus in these cases the reported rates are a closer approximation of actual rates, especially for larger animals that are both easily observed and more difficult for scavengers to displace. WVC may occur and not be observed, be removed by highway maintenance crews, or be scavenged by other animals. Scavenging rates can be very high for roadkilled animals, affecting confidence in estimates of total impact of WVC on populations (Antworth et al. 2005; Barthelmess and Brooks 2010).

The observations in the current study do reflect the presence of particular species at particular times of year and thus are a presence-only type of record. These data are useful in understanding wildlife distribution and movement, and for roadkilled animals, proximate causes of the collision (Barthelmess 2014) or, as demonstrated here for frequently-driven roads, spatial-aggregation of collisions. Large-extent databases of WVC observations provide a tool for developing and testing predictive models for contributing factors to WVC. Because of unevenness in sampling and the unknown level of effort going into opportunistic reporting in the systems described here, we are not in a position to rank risks to wildlife among highways. However on single routes with high and/or regular rates of observation, local hotspots (blind curves, riparian crossings) may be located and calibration made of observations per unit effort, relative visibility and reporting rates found for different species, and other bias-correction rates calculated.

We demonstrate that volunteer-observations of WVC can contribute to understanding locations of WVC clusters that could be suitable for mitigation action at US state scales. We found that the length of highway segments analyzed had an effect on the position and occurrence of clusters. This is similar to the finding for bird species richness, where geographic clustering was found to depend on analytical scale (Ma et al. 2012). Because of this, the best approach for mitigation planning would be to carry out cluster analyses for multiple segment lengths, depending on the taxonomic group or process of interest. Previous research indicates that road segment lengths of 0.4 to 1.6 km area appropriate for studying wildlife crossings and WVC (Malo et al. 2004; Taylor and Goldingay 2004). Representation of cluster locations across multiple segment length classes may indicate places of particular importance from a collision point of view. For future studies, it would be worth formalizing segment-analysis lengths that reflect a combination of consideration of ecological processes (e.g., species-dependent movement distances) and transportation-planning (e.g., segment scales for planning).

Identifying locations of clusters of WVC is a common step preceding mitigation and conservation actions to protect wildlife from vehicle-caused mortality (e.g., Hobday and Minstrell 2008). In the present study, cluster locations were found to vary in position along study highways across years of observations. For one highway (CA-13), there was virtually no overlap among WVC clusters from year to year (2010 to 2013). For another highway (CA-49), there were locations where clusters were identified every year, from 2010 to 2013. There were also locations where clusters were identified during 1, 2, or 3 of the 4 years. It is likely that factors contributing to WVC, such as traffic volume and speed, land cover, and road characteristics, did not change significantly during the study period. This suggests that temporal-dependence of cluster-locations is related to changes in the behavior of individuals and species along these highways. In addition, locations of statistically-significant clusters are not the only locations for concern about WVC. Highways with high rates of WVC across many adjacent segments may have few clusters, but many areas of concern because of impacts to drivers and animals (e.g., Figure 4, CA-50). This type of finding is very important for conservation planning, because it suggests that there may not be predictable landscape “corridors” or “linkages”, with corresponding stretches of highway suitable for mitigation action to protect wildlife movement. This finding contrasts with previous findings for certain taxonomic groups. For example, Langen et al. (2008) found that locations of clusters of herpetofauna road mortality were stable over time (i.e. comparison of 2002 and 2006/07). These clusters co-occurred with ponds and wetlands, which could explain the lack of change over time. We did not have sufficient data to divide the WVC observations into individual species and years.

The hotspots identified from volunteer-observations may not align with clusters identified using Department of Transportation (DOT)-collected WVC observations, because the latter are typically of ungulate and other large species that pose a risk to drivers. The combination of high-species-diversity observations by volunteers and DOT/wildlife agency observations of large animals could provide the ideal combination of WVC data to directly inform mitigation planning that provides both conservation and driver-safety benefits. In addition, because of the taxonomic breadth of volunteer-collected WVC observations, individual species could be considered for safety (e.g., mule deer) or conservation (e.g., meso-carnivores) reasons.

The annual cost of deer collisions, varied between the two CA state-highways analyzed and ranged from <US$ 500 to >US$30,000 per mile. To put these numbers in perspective, it can cost ~US$25,000/mile to augment a 5-6 foot chain link fence to make it into an 8-foot, deer-resistant fence (e.g., deer-fence in ID, https://fishandgame.idaho.gov/content/post/i-15-mule-deer-fence-near-pocatello-complete) and up to US$100,000/mile to construct a new 8-foot, deer-resistant fence. Fences are typically associated with purpose-built crossing, or other, structures that allow wildlife passage acros a right-of-way. There were segments of high costs from deer collisions (>US$5,000/mile-year) throughout both SR 50 and I-280. Fence/crossing mitigation of certain stretches of state highway could pay for themselves in terms of avoided costs from deer collisions in a matter of 1–20 years, depending on rate of collision and existing fence infrastructure.

Many segments of the state highways studied are likely to have collisions between vehicles and any animal, including deer. These areas may or may not be predictable, but what is certainly predictable is that providing directional fencing to encourage deer and other wildlife to usable crossing structures will reduce WVC. Directional fencing and accompanying escape structures (e.g., jump-outs to allow animal escape from the road-side of a fence) and highway under or over-crossings have proven to be effective for reducing collisions between wildlife and vehicles (Hedlund et al. 2004; Seamans and Helon 2008). This utility is predictably compromised if the structures and materials are not monitored and maintained causing more animals to enter the roadways. At the scale of whole states and state highways, these structures will seem expensive, though not compared with the costs in lives, injury and property damage from collisions, or swerving to avoid collisions, with animals. Thus, strategically placing mitigation structures and showing their potential and actual cost-effectiveness will be very important for a more wide-spread adoption.