Predictive modeling of ungulate–vehicle collision in the Republic of Korea

doi:10.21203/rs.3.rs-2754004/v1

Download PDF

Research Article

Predictive modeling of ungulate–vehicle collision in the Republic of Korea

https://doi.org/10.21203/rs.3.rs-2754004/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Context

Studies of ungulate–vehicle collision (UVC) may suffer from inadequate or scattered datasets, due to difficulties in acquiring data over vast temporal and spatial scales. Predictive modeling on UVC using a reliable data is useful to reduce the collisions.

Objectives

This study aims to understand spatial and temporal UVC characteristics by examining various parameters related to habitat, traffic, and seasonality using a UVC dataset that may be regarded as near-complete UVC data covering all road types across the Republic of Korea.

Methods

A total of 25,755 UVC points were collected between 2019 and 2021 using a standardized method by over 5,000 road menders in the Republic of Korea. Seasonal UVC predictive models of three ungulate species, Capreolus pygargus, Hydropotes inermis, and Sus scrofa, were generated using a machine-learning algorithm software, MaxEnt.

Results

The results showed that the peak UVC seasons coincided with the most active seasonal behaviors of the studied ungulates. In C. pygargus, habitat variables are most important for models across seasons, and UVC events are most likely to occur in high mountain chains. In H. inermis, habitat and traffic variables are most important for models across seasons. Although the important habitat for the models were different across seasons for S. scrofa, the maximum speed was consistently critical for models across all seasons.

Conclusions

Factors critical for UVC in the Republic of Korea were different for all three ungulate species and across seasons, indicating that seasonal behavior should be considered along with landscape and traffic features to mitigate UVC.

deer–vehicle collision

roadkill

fragmentation

spatial modeling

seasonal behavior

The expansion of linear infrastructure causes habitat fragmentation and reduces habitat connectivity. The consequent increase in collisions between vehicles and wildlife, known as animal–vehicle collision (AVC), negatively impacts human and wildlife safety. The risk to human safety increases with the body size of the wildlife involved in AVC (Neumann et al. 2012; Gren et al. 2018). The body size of ungulates is typically larger than that of other species involved in AVC (Zuberogoitia et al. 2014). Thus, ungulate–vehicle collision (UVC) is a globally important issue (Conover 2002; Seiler 2003; Malo et al. 2004). The dangers of UVC have been demonstrated in previous studies. Among the reported traffic accidents involving UVC in the United States and Canada, 2.8–9.7% resulted in human injury (Huijser et al. 2009). In Brazil, AVC with large animals, such as deer and horses, accounts for 3.3% of total vehicle accidents, and 18.5% of these accidents lead to human injury or fatality (Abra et al. 2019). In addition, 300 human fatalities and 30,000 injuries involving UVC occur annually in Europe (Bruinderink and Hazebroek 1996). Thus, understanding the patterns of UVC is important for reducing the UVC frequency for wildlife conservation as well as for human safety.

There are three categories of variables that critically impact AVC: landscape, traffic, and seasonal behavior (Kim et al. 2019a). The distribution of an animal population is often clumped, depending on resource availability, which is dictated by landscape features. Animals often require different types of resources for reproduction and survival. For example, animals often rest in a patch that may be hidden away from threats and feed on a different patch for food (Bond et al. 2002; Thomas and Taylor 2006). Disconnectedness among patches of resources forces animals to move regularly. Accordingly, landscape features are often the most important factors predicting UVC (Forman and Alexander 1998; Clevenger et al. 2003). The best model predicting the UVC of European roe deer (Capreolus capreolus) in France included landscape covariates and habitat connectivity (Girardet et al. 2015). In addition, the UVC of Odocoileus spp. is more likely to occur close to water and non-forested vegetation in the United States of America (USA), possibly because of habitat preference (Ng et al. 2008).

Road characteristics also play an important role in predicting the occurrence of AVC. Depending on the nature of the use, each road has distinctive characteristics. In general, major roads connecting large cities have many lanes with high traffic volumes, whereas roads located in the countryside around villages have fewer lanes with low traffic volumes and low maximum speeds. When an animal attempts to cross a road, its likelihood of passing the road successfully depends on various conditions, such as traffic volume, road structure, and the number of lanes on the road, and their impact varies by species. For instance, during the lockdown due to the COVID-19 pandemic, significantly fewer UVC of roe deer and wild boar were observed due to lower traffic volume (Pokorny et al. 2022). In contrast, in the USA, traffic had no impact on the UVC of the mule deer Odocoileus hemionus (Kreling et al. 2019). In addition, the number of lanes or the road width can also influence the likelihood of UVC (Girardet et al. 2015). In the Republic of Korea, most cases of UVC of the water deer Hydropotes inermis were located close to road connections, such as ramps and interchanges (Kim et al. 2021), indicating a high risk of UVC around areas where the species enters the road.

Animal movements are often seasonal, which is an important factor in AVC (Kim et al. 2021; Mayer et al. 2021; Steiner et al. 2021). Animals are often most active during their breeding season. In general, mammalian males expand their home range and/or make long-distance movements to find mates at the beginning of the breeding season (Cooper and Randall 2007; He et al. 2016; Johansson et al. 2018), making males vulnerable to AVC (Mayer et al. 2021; Raymond et al. 2021). Mammalian juveniles often disperse from their natal home range. Inexperienced juveniles often lack knowledge of high-traffic roads, which results in AVC (Kim et al. 2019b). Therefore, animal mortality due to AVC fluctuates seasonally. For instance, the number of H. inermis UVC is significantly different across months and seasons according to their seasonal behaviors (Kim et al. 2021; Raymond et al. 2021). In addition, Odocoileus hemionus UVC varied significantly by season and month because of its high activity during the mating season and reduced movement during spring owing to high food availability (Kreling et al. 2019).

Five species of ungulates are native to the Republic of Korea: the Siberian roe deer Capreolus pygargus, water deer H. inermis, wild boar Sus scrofa, musk deer Moschus moschiferus, and long-tailed goral Naemorhedus caudatus (Jo et al. 2018). Among them, M. moschiferus and N. caudatus are very rare and only regionally distributed; therefore, they are designated as endangered species. The other three species, C. pygargus, H. inermis, and S. scrofa, are comparably abundant throughout the country. Accordingly, these three species are the focus of this study.

Capreolus pygargus is distributed from the Republic of Korea mainland to southern Jeju Island. However, the mainland and island populations have different morphological characteristics and population densities. The mainland population weighs between 28 and 35 kg and has a population density of 1.9 individuals per km². Meanwhile, the population in Jeju is smaller than that in the mainland population, weighs an average of 17 kg, and has a population density of 5.3 individuals per km² (Park et al. 2011; Jo et al. 2018). Because of these differences, we focused on the mainland populations of C. pygargus for UVC. On the mainland, C. pygargus inhabits forest interiors with elevations between 400 and 600 m (Park et al. 2014). Fawns are born between May and July, and the rutting season starts in August. During the breeding season, males compete with other males to defend their territories. In doing so, they often approach the edges of their home ranges, which are typically bordered by roads and streams (Kim and Hong 2006; Jo et al. 2018).

Hydropotes inermis, weighing 16–21 kg, is the smallest ungulate species in the country (Jo et al. 2018). It is an edge species inhabiting various types of land cover, including forests, rice fields, and wetlands (Jo et al. 2018). Although this species is listed as vulnerable by the IUCN Red List, it is the most abundant wildlife and its population is rapidly increasing in the Republic of Korea (Kim et al. 2011; Seo et al. 2021). The population density of H. inermis is 7.7 individuals per km² on average, with the highest densities observed in the lowlands. Thus, it is no accident that H. inermis is the most common UVC victim in the country (Kim et al. 2019b). Regarding seasonal behavior, two to six H. inermis fawns are born in May and June in the temperate region. Yearlings spend one year with their mother and then disperse to their own territory between May and June (Cooke and Farrell 1998; Jo et al. 2018).

Sus scrofa is one of the largest terrestrial mammals in the Republic of Korea, with an average weight of 300 kg. Sus scrofa mainly inhabits the forest interiors. According to a long-term national mammal monitoring program run by the Korean government, this species has recently become numerous, from 1.3 individuals per km² in 1978 to 3.7 individuals per km² in 2021 (Seo et al. 2021). Accordingly, they are often found in urban environments (Lee and Lee 2019). Therefore, this species has been designated as a culling species because of constant conflicts with humans. The mating season for S. scrofa peaks between November and December. The species is omnivorous, mainly consuming vegetable matter, some invertebrates, snakes, and occasionally large vertebrates. The average home range size of this species is 5.13 km².

Landscape features that may be important for species occurrence are heterogeneous. Thus, it is difficult to generalize the findings from the AVC datasets that are collected in regional scales. Thus, AVC datasets that are broad enough, exceeding the ranges of species distribution, are ideal to provide information about likelihood of AVC at a large spatial scale. However, such nationwide datasets are difficult to obtain, because of high cost and difficulty of organizing collection efforts. Many AVC studies rely on the police reports due to availability, but such police data is produced only when an accident is reported to the police. Accordingly, an unknown proportion of AVC incidents may go unreported. Citizen science is another venue to collect AVC data at a larger temporal and spatial scale. However, data from a citizen science program may be opportunistic and may be biased, depending on levels of participant training. To overcome the difficulties of collecting AVC data, a standardized protocol with a great deal of surveyors at a large scale may be required.

In the Republic of Korea, the Korea Roadkill Observation System (KROS; nie-ecobank.kr/) was launched by the government to integrate and manage scattered AVC data (Kim et al. 2019b). KROS covers all road types across the country with a standardized data collection method. In addition, it is mandatory for road workers of all road authorities in the country to use KROS whenever they find animal carcasses on roads. Thus, the KROS dataset is considered as one of the largest with high data quality that can be useful to AVC mitigation policy in the country.

Recently, habitat modeling techniques have been incorporated to predict the occurrence of AVC. MaxEnt (maximum entropy; Phillips et al. 2004) is a machine-learning algorithm that predicts habitat suitability based on the response of presence points to environmental conditions relative to the response of background points. Thus, the MaxEnt modeling is used to predict the expected presence as an index of suitability. Because of its predictive power, MaxEnt can also be used to predict the probability of any type of occurrence using spatial data and appropriate predictor variables. For this reason, some recent studies have utilized MaxEnt to predict AVC risk and have concluded that it is useful in identifying important variables influencing AVC (Chyn et al. 2021; Garrote et al. 2018; Ha and Shilling 2018; Kantola et al. 2019; Mayadunnage et al. 2022; Sillero et al. 2019; Yue et al. 2019).

The aim of this study is to understand UVC characteristics by evaluating variables related to the surrounding landscape, traffic, and animal seasonality, using one of the largest and most complete country-wide AVC datasets in the world. To this end, we developed predictive seasonal models of UVC for three species across a road network in the Republic of Korea using MaxEnt. Based on our results, we identified the critical factors influencing UVC, which will be useful in reducing the threat to human safety.

Study area

In the Republic of Korea, there are 105,083 km of roads as of 2021 (Korean Statistical Information Service, 2021) (Fig. 1). There are seven types of roads in Korea: expressways, national highways, metropolitan roads, provincial roads, municipal roads, county roads, and district roads. The expressway and national highway are inter-provincial roads, whereas the rest extend only within their respective administrative levels. An expressway is a toll road with a speed limit above 100 km/h and is managed by the Korea Expressway Corporation (ex.co.kr). A national highway connects major cities with a speed limit of up to 80 km/h and is maintained by the MOLIT, Korea. A metropolitan road is located in a metropolitan city, and a provincial road is an arterial road in a province. Municipal, county, and district roads are typically small-scale roads with low-speed limits and are located inside the municipality, county, and district, respectively.

Topographically, the Republic of Korea is dominated by the Baekdu-daegan mountain chain, which has numerous peaks over 1,000 m. The mountain chain runs along the east coast and turns into the midland of the peninsula to the south. Because of this mountain chain, rivers generally flow westwards or southwards, thereby alluvial plains developed in the west and south. Large cities with dense networks of roads are well-developed in these plains. (Fig. 1).

Data collection

A smartphone application linked with KROS was developed and released for use by more than 5,000 road menders and related managers affiliated with road authorities, who were in charge of removing animal carcasses from the roads. Road menders took pictures whenever a carcass was found while patrolling the roads or when a civil complaint about AVC was received. Along with pictures, location and time information were conveyed to the KROS website in real time. The UVC data for C. pygargus, H. inermis, and S. scrofa used in this study were derived from the 2019–2021 KROS data (Fig. 1). Road data required for spatial analyses were obtained from the National Transport Information Center (www.its.go.kr/).

Habitat suitability modeling

To map the probabilities of UVC for three species of ungulates (C. pygargus, H. inermis, S. scrofa), habitat suitability models (HSMs) were created for all species. For habitat suitability modeling, we applied the MaxEnt software, which uses presence data to compute habitat suitability using machine learning. We used occurrence data obtained from the Global Biodiversity Information Facility (GBIF; GBIF.org, 2022), with a bounding area of the Republic of Korea. After removing occurrences with coordinate uncertainty greater than the resolution of our environmental layers (0.001 decimal degrees), we retained 5,178 occurrences of C. pygargus, 35,050 occurrences of H. inermis, and 13,088 occurrences of S. scrofa for modeling. The HSMs were trained using environmental layers (Table 1) with a resolution of 0.001 decimal degrees, including bioclimatic variables (Worldclim.org ver. 1.4; Hijmans et al. 2005) and habitat-related variables (land cover type (Table 2), distance to forest, distance to water, distance to urban areas, elevation, slope, ruggedness, distance to streams and coast, tasseled-cap greenness, and wetness). Duplicate presence records were removed as an option in the modeling software to reduce the spatial autocorrelation caused by sampling bias. The models were trained with five replicates using the cross-validate run type, with 80% of the points used for training and 20% used for testing. The average of the five replicates was used as the final model. Variables were reduced based on percent contribution (> 5) or permutation importance (> 5) within the initial full model. Final HSMs were validated using the area under the receiver operating curve (AUC) and true scale statistics (TSS; Allouche et al. 2006).

Table 1

Variables used for modeling habitat suitability and UVC probability of three ungulate species in the Republic of Korea.
Category	Variable	Model used	Explanation
Bioclimate	Annual mean temperature	HSM, UVC	-
Bioclimate	Mean diurnal range	HSM	Mean of monthly (max temp – min temp)
Bioclimate	Isothermality	HSM	Mean diurnal range/Temperature annual range * 100
Bioclimate	Temperature seasonality	HSM	Standard deviation * 100
Bioclimate	Maximum temperature of warmest month	HSM	-
Bioclimate	Minimum temperature of coldest month	HSM	-
Bioclimate	Temperature annual range	HSM	Maximum temperature of warmest month- Minimum temperature of coldest month
Bioclimate	Mean temperature of wettest quarter	HSM	-
Bioclimate	Mean temperature of driest quarter	HSM	-
Bioclimate	Mean temperature of warmest quarter	HSM	-
Bioclimate	Mean temperature of coldest quarter	HSM	-
Bioclimate	Annual precipitation	HSM, UVC	-
Bioclimate	Precipitation of wettest month	HSM	-
Bioclimate	Precipitation of driest month	HSM	-
Bioclimate	Precipitation seasonality	HSM	Coefficient of variation
Bioclimate	Precipitation of wettest quarter	HSM	-
Bioclimate	Precipitation of driest quarter	HSM	-
Bioclimate	Precipitation of warmest quarter	HSM	-
Bioclimate	Precipitation of coldest quarter	HSM	-
Habitat	Elevation	HSM, UVC	Digital Elevation Model (DEM; GMTED provided by USGS) smoothed using focal statistics
Habitat	Slope	HSM, UVC	Derived from DEM using Slope tool in ArcMap
Habitat	Ruggedness	HSM, UVC	Derived from DEM – standard deviation of slope using focal statistics
Habitat	Habitat Landcover	HSM, UVC	Extracted from World Land Cover 30m BaseVue 2013, Source: MDAUS
Habitat	Forest distance	HSM, UVC	Euclidean distance to forest land cover class (BaseVue 2013)
Habitat	Urban distance	HSM, UVC	Euclidean distance to urban land cover class (BaseVue 2013)
Habitat	Water distance	HSM, UVC	Euclidean distance to water cover class (BaseVue 2013)
Habitat	River distance	HSM, UVC	Euclidean distance to streams/rivers
Habitat	River-coast distance	HSM, UVC	Euclidean distance to streams/rivers or coast
Habitat	Greenness	HSM, UVC	Vegetation index, derived using Tasselled Cap transformation of corrected Landsat reflectance imagery
Habitat	Wetness	HSM, UVC	Surface and canopy moisture, derived using Tasselled Cap transformation of corrected Landsat reflectance imagery
Habitat	Corridor distance	UVC	Euclidean distance to wildlife crossings
Traffic	Lanes	UVC	Number of lanes per road (two-way)
Traffic	Max speed	UVC	Maximum road speed

Table 2

Land cover categories included in BaseVue (2013).
Category	Description
1	Deciduous forest
2	Evergreen forest
3	Shrub/Scrub
4	Grassland
5	Barren or minimal vegetation
7	Agriculture, general (cultivated crop lands)
8	Agriculture, paddy (crop lands characterized by inundation)
9	Wetland
11	Water
20	High density urban (over 70% constructed materials)
21	Medium-low density urban (30–70% constructed materials)

UVC mapping

MaxEnt was used to model the UVC probabilities of the three ungulate species in the Republic of Korea. Environmental layers included those used to build the HSMs (Table 1; excluding bioclimatic variables, except annual mean temperature and annual precipitation), final HSMs, traffic-related variables (number of lanes and maximum speed), and Euclidean distance to wildlife crossings. These layers were masked (extracted) using road layers, because UVC can only occur along roads. Therefore, background points falling elsewhere would not represent a true lack of UVC along the roads. As with the HSMs above, five replicates using the cross-validation run type were averaged for the final maps of the UVC probability. However, HSM of S. scrofa was excluded because of its high correlation with other environmental variables with higher contributions to the model. To understand the temporal effect on UVC probability, models of UVC probability for each species were created with data separated by season (spring: March–May; summer: June–August; fall: September–November; winter: December–February). Duplicate presence points were not removed because UVC occurrences could be considered independent observations (i.e., not multiple counts of the same individual). Finally, we set the default prevalence to the percentage of UVC for each group (species by season). The final models were assessed using AUC and TSS. Spatial analysis was conducted using ArcGIS Desktop 10.8.1 and ArcGIS Pro 2.8.0.

UVC data

Between 2019 and 2021, there were 970 C. pygargus (mainland only), 24,268 H. inermis, and 517 S. scrofa datapoints in the georeferenced UVC dataset collected through KROS. In general, the UVC frequencies of each species were seasonal (Fig. 2). The UVC frequency of C. pygargus was the highest in summer and was 1.7 times higher in summer than that in the lowest season—winter. Summer is the rutting season of the species, which may increase the probability of UVC of males. Unfortunately, sex information of UVC victims was not collected in the KROS. The UVC frequency of H. inermis peaked in spring and was the lowest in winter, with a difference of 2.9 times. Spring is the dispersal season of juveniles, which may lead to individuals that have not experienced the dangers of roads to the roads, resulting in a high risk of UVC. For S. scrofa, the UVC frequency was the highest in fall and the lowest in spring, with a difference of 3.4 times. The fall is the beginning of the rutting period of S. scrofa, as well as the high-activity season to prepare for overwintering. Thus, the UVC peak seasons coincided with the most active season for all three species.

Habitat suitability models

The habitat suitability models showed a good fit for C. pygargus (average AUC values of 0.802 ± 0.007) and S. scrofa (0.782 ± 0.012) but a poor fit for H. inermis (0.602 ± 0.004). However, this model was generally acceptable for H. inermis, which is a habitat generalist. For C. pygargus, elevation had the highest contribution, followed by the maximum temperature of the warmest month, and the minimum temperature of the coldest month had high permutation importance. For H. inermis, the mean temperature of the wettest quarter had the highest contribution, followed by ruggedness, and the slope had a high permutation importance. For S. scrofa, elevation had the highest contribution, followed by land cover and precipitation of the coldest quarter, and greenness had high permutation importance. In short, important variables for the habitat model of C. pygargus were mainly related to climate, those of H. inermis were related to climate and topography, and those of S. scrofa were highly variable.

UVC probabilities

The predictive powers of the models evaluated using AUC values were considered acceptable (> 0.75) for all 12 models (Table 3). In general, all UVC maps of the three species showed a common pattern of low probability in urbanized regions throughout the country. In C. pygargus, UVC probabilities were high around the Baekdu-daegan mountain chain in spring and summer, whereas they were low in fall and winter throughout the country (Fig. 3). Important variables in the UVC predictive models of C. pygargus were highly related to habitats but were not related to traffic across seasons (Fig. 4 and Table 4). In the spring and summer models, the variables with high permutation importance (> 10) were habitat suitability and river distance. As habitat suitability and distance to the river increased, the likelihood of UVC also increased monotonically. In the fall model, habitat suitability had the highest permutation importance, followed by river distance and ruggedness. The likelihood of UVC probability was bimodal, with peaks of very low and high ruggedness. For winter, only river distance showed high permutation importance. In C. pygargus, habitat-related variables were the most important for UVC models across seasons, and UVC events were most likely to occur in the high mountain chain.

Table 3

Area under the receiver operating curve (AUC) and true scale statistics (TSS) of 12 UVC probability models. Values (mean ± standard error) for spring, summer, fall, and winter UVC in each species.
	Spring	Summer	Fall	Winter
C. pygargus
AUC	0.877 ± 0.032	0.910 ± 0.026	0.923 ± 0.032	0.962 ± 0.020
TSS	0.597 ± 0.091	0.625 ± 0.038	0.804 ± 0.102	0.845 ± 0.072
H. inermis
AUC	0.750 ± 0.007	0.752 ± 0.007	0.785 ± 0.010	0.791 ± 0.010
TSS	0.518 ± 0.020	0.515 ± 0.018	0.516 ± 0.013	0.506 ± 0.012
S. scrofa
AUC	0.785 ± 0.065	0.817 ± 0.055	0.794 ± 0.035	0.754 ± 0.067
TSS	0.434 ± 0.180	0.444 ± 0.241	0.436 ± 0.114	0.320 ± 0.090

Table 4

Permutation importance of each variable included in the seasonal models of *C. pygargus*. The highest values (> 10) for each model are presented in bold.
Parameter	Spring	Summer	Fall	Winter
Annual mean temperature	3.4	4.3	0.9	1.5
Annual precipitation	1.2	0.4	0.3	1.1
Corridor distance	0.9	0.8	2.1	0.5
Elevation	2.2	1.5	0.9	0.9
Forest distance	6.3	2.5	0.4	0.4
Greenness	5.7	3.3	1.5	0.4
Landcover	2.1	2.8	1.6	0.1
River distance	18.3	17.9	29.4	82.9
River-coast distance	1.8	0.2	1.5	0.2
Habitat suitability	42.1	58.4	39.0	8.2
Ruggedness	5.2	1.0	10.5	0.8
Slope	2.1	1.2	3.2	0.3
Urban distance	1.3	1.6	1.2	0.6
Water distance	1.2	0.3	1.7	0.1
Wetness	0.8	1.1	0.9	0.4
Maximum speed	2.5	2.1	2.8	0.6
Lanes	3.0	0.6	2.1	1.2

Unlike the UVC probability of C. pygargus, a high UVC probability of H. inermis was largely distributed outside the high mountain chain throughout the country (Fig. 5). Furthermore, there were few changes in UVC probabilities throughout the seasons. Thus, both habitat- and traffic-related variables were important for the UVC models of H. inermis across the seasons (Fig. 6 and Table 5). For H. inermis UVC predictive models, the variables with the highest permutation importance were habitat suitability, maximum speed, and number of lanes for all seasons. For the traffic-related variables, UVC probabilities were high on roads with three to five lanes and with maximum speeds between 40 and 50 km/h or higher than 100 km/h.

Table 5

Permutation importance of each variable included in the seasonal models of *H. inermis*. The highest values (> 10) for each model are presented in bold.
Parameter	Spring	Summer	Fall	Winter
Annual mean temperature	8.8	7.1	5.5	8.2
Annual precipitation	1.2	1.3	1.1	3.7
Elevation	2.2	1.7	1.8	3.7
Forest distance	0.9	0.9	0.6	0.6
Greenness	2.8	2.3	5.7	3.1
Landcover	5.7	5.2	8.9	8.7
River distance	0.3	0.6	1.5	0.9
River-coast distance	0.7	1.4	0.5	1.1
Habitat suitability	30.5	32.9	25.0	23.2
Ruggedness	0.4	0.6	0.7	0.1
Slope	0.4	0.4	0.2	1.1
Urban distance	0.8	1.1	0.9	2.3
Water distance	0.9	0.4	0.4	0.4
Wetness	1.3	1.3	2.1	2.6
Corridor distance	0.4	0.6	0.8	1.3
Maximum speed	24.3	26.7	27.4	22.8
Lanes	18.3	15.5	16.8	16.3

The UVC probabilities of S. scrofa fluctuated between seasons (Fig. 7), and the important variables for predicting the UVC of S. scrofa varied by season (Fig. 8 and Table 6). The variables with the highest permutation importance were elevation, maximum speed, and urban distance for spring. In summer, the variables with the highest permutation importance were wetness, maximum speed, forest distance, and urban distance. The variables with the highest permutation importance were maximum speed, number of lanes, and forest distance for fall. The variables with the highest permutation importance were river distance, maximum speed, ruggedness, and land cover for winter. Notably, although habitat suitability was not included in the modeling due to high correlations with other variables, variables related to habitat were highly influential in S. scrofa UVC models across seasons. Moreover, maximum speed was critical for the UVC models across all seasons. As the maximum speed increased, the UVC probabilities generally increased with a peak of approximately 90 km/h. The UVC probabilities were highest close to the forest and the river and then tended to decrease as the distance from the forest and the river increased. UVC probabilities peaked in areas with low ruggedness and a few hundred meters from urban areas in summer and then slightly decreased as the ruggedness and distance from urban areas increased. However, UVC probabilities peaked very close to urban areas and then dropped rapidly as the distance to urban areas increased in spring. UVC probabilities were generally higher in areas with low wetness. To summarize, important variables predicting UVC of S. scrofa fluctuated greatly by season as well as the spatial distribution of the high probability of UVC in the prediction map.

Table 6

Permutation importance of each variable included in the seasonal models of *S. scrofa*. Habitat suitability is excluded due to high correlations with other variables. The highest values (> 10) for each model are presented in bold.
Parameter	Spring	Summer	Fall	Winter
Annual mean temperature	1.2	0.7	0.4	0.3
Annual precipitation	1.0	1.3	1.5	2.9
Corridor distance	0.2	0.2	3.9	1.3
Elevation	28.7	4.0	5.7	4.8
Forest distance	8.7	13.1	12.7	7.4
Greenness	0.7	0.3	4.2	0.3
Landcover	1.6	5.7	7.7	11.2
River distance	9.7	7.6	4.8	17.4
River-coast distance	3.3	0.1	4.6	2.0
Ruggedness	3.5	2.1	5.2	13.6
Slope	3.8	1.8	1.2	1.1
Urban distance	14.2	12.8	3.8	6.0
Water distance	1.1	8.9	3.1	1.5
Wetness	4.8	22.7	8.4	9.4
Maximum speed	14.8	16.9	18.1	14.8
Lanes	2.8	1.6	14.7	5.9

Our results confirmed that active-season behavior was correlated with peak UVC seasons in three ungulate species (C. pygargus, H. inermis, and S. scrofa) in the Republic of Korea. The UVC frequencies peaked in rutting season of C. pygargus and S. scrofa and peaked in dispersal season of H. inermis. This pattern is applicable to other species. For example, in the UK, overall AVC occurrences peaked in summer, when the abundance of young mammals and birds is the highest (Raymond et al. 2021). Therefore, we conclude that seasonal behaviors impact the temporal patterns of UVC and that high activity seasons lead to high UVC rates.

The variation in the predictive power of the habitat suitability model using AUC seemed to be related to the nature of habitat use in the three ungulate species. Habitat suitability models are generally a better fit for habitat specialists, which require in a narrow range of environmental conditions, and a poorer fit for habitat generalists, which thrive in a wide variety of environmental conditions (Elith et al., 2006; Lobo et al., 2008) C. pygargus and S. scrofa are known as habitat specialists. These two species are interior species inhabiting internal spaces in forests. In contrast, H. inermis is a habitat generalist that exploits wetlands and adjacent areas even in modified habitats. Generalist species utilize various types of habitats without distinctive preferences, often resulting in poor model fits in the HSMs (Evangelista et al. 2008). Thus, low values of model fits reflect the generalist habitat use of H. inermis, not negating the validity of models.

We highlight that the UVC predictive models for each species had distinctive patterns. The variables important for predicting C. pygagus UVC across seasons were related to habitat characteristics only. In contrast, the UVC models for H. inermis and S. scrofa included habitat-related and traffic-related variables. Notably, the UVC probabilities of these two species were highest for roads with a maximum speed of 100 km/h. The maximum road speeds were critical for the UVC probability models of H. inermis and S. scrofa in all seasons. For both species, UVC probabilities increased drastically at maximum speeds of 70–80 km/h or around 90–100 km/h. Vehicle speed is a critical factor that affects UVC by reducing the response time of drivers to the emergence of an animal in front of the vehicle (Seiler 2005; Danks and Porter 2010; Meisingset et al. 2014). Danks and Poter (2010) discovered that UVC probability increased drastically with an increase in speed, especially from 72 km/h onward.

UVC probabilities were largely influenced by landscape variables in C. pygargus, which prefers high-elevation habitats above 200 m (Kim et al. 2017), as demonstrated in this study by habitat suitability modeling with elevation as an important parameter. A river is formed typically in the lowlands, away from the preferred habitats of C. pygargus. Thus, UVC probabilities of C. pygargus linearly increased as distance to river increased. UVC probability increased with an increase in habitat suitability for C. pygargus. High habitat suitability indicates high-quality habitat and food availability to wildlife. The number of AVC and population size were considered to be correlated (Schwartz and Shilling 2020), which could result in a high probability of UVC. From this perspective, C. pygargus is highly dependent on its habitat. The UVC probability also increased at low and high ruggedness and decreased at intermediate ruggedness in the fall. However, the effect of ruggedness on the UVC probability of C. pygargus could not be determined and further studies are necessary.

UVC predictive seasonal models of H. inermis included high permutation importance of the number of lanes, maximum speed, and habitat suitability for all seasons and showed similar patterns across seasons. The UVC probability of H. inermis was highest on four-lane roads and decreased after an increase in the number of lanes. This may be because the KROS data have been collected mainly from expressways and national highways (Song et al. 2020). In these road types, the main number of lanes is four (58.9%), followed by two (29.3%), six (7.9%), eight (3.4%), and more than ten (0.4%) (Korean Statistical Information Service, 2021). For maximum speed, the predicted probability of UVC was highest at 40–50 km/h and around 100 km/h but tended to decrease at a maximum speed of less than 40 km/h, higher than 100 km/h, or between 50 km/h and 100 km/h. Roads with maximum speed limits of less than 50 km/h are usually located in the countryside. Roads in the countryside, except expressways and highways, are level with the surrounding areas, making them easily accessible to H. inermis. As a habitat generalist, H. imermis often cross roads near or in agricultural lands. This observation does not apply to C. pyrugus and S. scrofa, because these two species are habitat specialists and are typically found in forest interiors. In addition, the UVC probability of H. inermis increased with an increase in habitat suitability until intermediate suitability and decreased toward high suitability. This pattern could be related to the general traits of the species in habitat selection, as mentioned above. As a habitat generalist, the species did not seem to rely on habitat suitability; therefore, the likelihood of UVC events is not necessarily high in high-quality habitats.

Compared with the other two species, the UVC seasonal models of S. scrofa were more complex. Except for the maximum speed, the UVC models had different variables that were important by season. Sus scrofa is highly flexible in habitat use (Boitani et al. 1994). This species is highly affected by the environment, food availability, seasonal changes, and anthropogenic disturbances (Lee 2013). These characteristics make this species susceptible to environmental changes. For instance, although S. scrofa typically inhabits the forest interior, the UVC probability peaked around grasslands and agricultural lands in winter, when food is scarce. Additionally, the predicted probability of UVC was highest at approximately 90–100 km/h. Another study also concluded that vehicle speed is an important factor in predicting the UVC probability of S. scrofa, and an increase in speed increases UVC probability (Peris 2012). Moreover, such a high-speed limit implies an expressway in the Republic of Korea, and the UVC frequency of S. scrofa on the expressway significantly increased between 2004 and 2019 (Kim et al. 2021). Therefore, we suggest using the maximum speed limit for regulating the UVC occurrences involving S. scrofa.

The three target species in this study had the largest body size among all vertebrates inhabiting Republic of Korea, which means that human safety is also critically threatened when a collision occurs. Therefore, it is necessary to understand the spatio–temporal characteristics of each species to maintain human safety on roads and for wildlife conservation. In this study, UVC was predicted using various factors related to traffic, species habitat, and spatial distribution with high probability of UVC fluctuation according to the seasonal behavior of the focal species. To reduce such threats to humans and to conserve wildlife, it is important to consider the environmental features surrounding roads and traffic factors. Utilizing and applying the distinctive variables important to predict UVC across seasons for each species will be helpful in planning road safety and conservation actions.

Acknowledgments

We are grateful to the National Institute of Ecology (https://www.nie.re.kr) for providing KROS data for analyses.

Funding

This work was supported by a grant, titled “Research on mobility of small animals,” by SL Corporation to YJ.

Competing Interests

All authors have no relevant financial or non-financial interests to disclose.

Author Contributions

Kyungmin Kim conceptualized the structure of the manuscript and wrote orginal draft. Desiree Andersen performed formal analyses and reviewed/edited the manuscript. Yikweon Jang supervised and reviewed/edited the manuscript.

Data Availability

The datasets generated during and/or analysed during the current study are available from the National Institute of Ecology on reasonable request.

Abra FD, Granziera BM, Huijser MP, Ferraz KM, Haddad CM, Paolino RM (2019) Pay or prevent? Human safety, costs to society and legal perspectives on animal-vehicle collisions in São Paulo state, Brazil. Plos One 14(4):e0215152.
Adhikari P, Park SM, Kim TW, Lee JW, Kim GR, Han SH, Oh HS (2016) Seasonal and altitudinal variation in roe deer (Capreolus pygargus tianschanicus) diet on Jeju Island, South Korea. J Asia Pac Biodivers 9(4):422–428.
Allouche O, Tsoar A, Kadmon R (2006) Assessing the accuracy of species distribution models: prevalence, kappa and the true skill statistic (TSS). J Appl Ecol 43(6):1223–1232.
Baek SY, Lee SM (2021) Temporal patterns and hotspot identification of vehicle collisions with the roe deer (Capreolus Pygargus) on Jeju Island, South Korea. Pak J Zool 1–6.
Boitani L, Mattei L, Nonis D, Corsi F (1994) Spatial and activity patterns of wild boars in Tuscany, Italy. J Mammal 75(3):600–612.
Bond BT, Burger Jr LW, Leopold BD, Jones JC, Godwin KD (2002) Habitat use by cottontail rabbits across multiple spatial scales in Mississippi. J Wildl Manag 1171–1178.
Bruinderink GG, Hazebroek E (1996) Ungulate traffic collisions in Europe. Conserv Biol 10(4):1059–1067.
Danks ZD, Porter WF (2010) Temporal, spatial, and landscape habitat characteristics of moose—vehicle collisions in western Maine. J Wildl Manage 74(6):1229–1241.
Chyn K, Lin TE, Wilkinson DP, Tracy JL, Lawing AM, Fitzgerald LA (2021) Fine-scale roadkill risk models: understanding the intersection of wildlife and roads. Biodivers Conserv 30(1):139–164.
Clevenger AP, Chruszcz B, Gunson KE (2003) Spatial patterns and factors influencing small vertebrate fauna road-kill aggregations. Biol Conserv 109(1):15–26.
Conover MR (2002) Resolving wildlife conflicts: the science of wildlife damage management. CRC Press, Boca Raton, Florida, USA.
Cooke A, Farrell L (1998) Chinese Water Deer. Mammal Society, London
Cooper LD, Randall JA (2007) Seasonal changes in home ranges of the giant kangaroo rat (Dipodomys ingens): a study of flexible social structure. J Mammal 88(4):1000–1008.
Elith, J., H. Graham, C., P. Anderson, R., Dudík, M., Ferrier, S., Guisan, A., J. Hijmans, R., Huettmann, F., R. Leathwick, J., Lehmann, A., Li, J., G. Lohmann, L., A. Loiselle, B., Manion, G., Moritz, C., Nakamura, M., Nakazawa, Y., McC. M. Overton, J., Townsend Peterson, A., … E. Zimmermann, N. (2006). Novel methods improve prediction of species’ distributions from occurrence data. Ecography, 29(2), 129–151. https://doi.org/10.1111/j.2006.0906-7590.04596.x
Evangelista PH, Kumar S, Stohlgren TJ, Jarnevich CS, Crall AW, Norman III JB, Barnett DT (2008) Modelling invasion for a habitat generalist and a specialist plant species. Divers Distrib 14(5):808–817.
Forman RT, Alexander LE (1998) Roads and their major ecological effects. Annu Rev Ecol Evol Syst 1:207-C2.
Forman RT, Sperling D, Bissonette JA, Clevenger AP, Cutshall CD, Dale VH, Fahrig L, France R, Goldman CR, Heanue K, Jones JA, Swanson FJ, Turrentine T, Winter TC (2003) Road ecology: science and solutions. Island Press, Washington, DC.
Garrote G, Fernández–López J, López G, Ruiz G, Simón MA (2018) Prediction of Iberian lynx road–mortality in southern Spain: a new approach using the MaxEnt algorithm. Anim Biodivers Conserv 41(2):217–225.
GBIF.org (2022) GBIF Occurrence Download. The global biodiversity information facility. https://doi.org/10.15468/dl.dvrut.
Girardet X, Conruyt-Rogeon G, Foltête JC (2015) Does regional landscape connectivity influence the location of roe deer roadkill hotspots? Eur J Wildl Res 61(5):731–742.
Gren IM, Häggmark-Svensson T, Elofsson K, Engelmann M (2018) Economics of wildlife management—an overview. Eur J Wildl Res 64(2):1–16.
Ha H, Shilling F (2018) Modelling potential wildlife-vehicle collisions (WVC) locations using environmental factors and human population density: a case-study from 3 state highways in Central California. Ecol Inform 43:212–221.
He X, Chen M, Zhang E (2016) Home range of reintroduced Chinese water deer in Nanhui East Shoal Wildlife Sanctuary of Shanghai, China. Anim Prod Sci 56(6):988–996.
Hijmans RJ, Cameron SE, Parra JL, Jones PG, Jarvis A (2005) Very high resolution interpolated climate surfaces for global land areas. Int J Climatol 25(15):1965–1978. https://doi.org/10.1002/joc.1276
Huijser MP, Duffield JW, Clevenger AP, Ament RJ, McGowen PT (2009) Cost–benefit analyses of mitigation measures aimed at reducing collisions with large ungulates in the United States and Canada: a decision support tool. Ecol Soc 14(2).
Jo YS, Baccus JT, Koprowski JL (2018) Mammals of Korea. National institute of biological resources, Inchoen.
Johansson Ö, Koehler G, Rauset GR, Samelius G, Andrén H, Mishra C, Lhagvasuren P, McCarthy T, Low M (2018) Sex‐specific seasonal variation in puma and snow leopard home range utilization. Ecosphere 9(8):e02371.
Kantola T, Tracy JL, Baum KA, Quinn MA, Coulson RN (2019) Spatial risk assessment of eastern monarch butterfly road mortality during autumn migration within the southern corridor. Biol Conserv 231:150–160.
Kim A, Lee JM, Jang GS (2017) Modeling the spatial distribution of Roe Deer (Capreolus pygargus) in Jeju Island. Journal of the Korean Association of geographic information studies 20(4):139–151.
Kim BJ, Oh DH, Chun SH, Lee SD (2011) Distribution, density, and habitat use of the Korean water deer (Hydropotes inermis argyropus) in Korea. Landsc Ecol Eng 7(2):291–297.
Kim DH, Hong YK (2006) A systemic approach for Roe Deer in Jesudo. Korean System Dynamics Review 7(2):191–213.
Kim K, Serret H, Clauzel C, Andersen D, Jang Y (2019a) Spatio-temporal characteristics and predictions of the endangered leopard cat Prionailirus bengalensis euptilura road-kills in the Republic of Korea. Glob Ecol Conserv 19:e00673.
Kim K, Woo DG, Seo H, Park T, Song EG, Choi TY (2019b) Korea road-kill observation system: the first case to integrate road-kill data in national scale by government. J For Environ Sci 35(4):281–284.
Kim M, Park H, Lee S (2021) Analysis of roadkill on the korean expressways from 2004 to 2019. Int J Environ Res Public Health 18(19):10252.
Kim K, Seo H, Woo D, Park T, Song E (2021) The water deer on a road: road-kill characteristics of a nationally abundant but internationally threatened species. J For Environ Sci 37(1):62–68.
Korean Statistical Information Service (2021) https://kosis.kr/. Accessed on 27^th November 2022.
Kreling SE, Gaynor KM, Coon CA (2019) Roadkill distribution at the wildland‐urban interface. J Wildl Manag 83(6):1427–1436.
Lee SM (2013) Home range, diet and crop damage of wild boar (Sus scrofa) in Geochang County, Gyeongnam, Korea. Master’s thesis, Seoul National University.
Lee SM, Lee EJ (2019) Diet of the wild boar (Sus scrofa): implications for management in forest agricultural and urban environments in South Korea. PeerJ 7:e7835.
Lobo, J. M., Jiménez‐Valverde, A., & Real, R. (2008). AUC: a misleading measure of the performance of predictive distribution models. Global Ecology and Biogeography, 17(2), 145–151. https://doi.org/10.1111/j.1466-8238.2007.00358.x
Malo JE, Suárez F, Díez A (2004) Can we mitigate animal–vehicle accidents using predictive models? J Appl Ecol 41(4):701–710.
Mauget, R (1982) Seasonality of reproduction in the wild boar. Control of pig reproduction, pp 509–526.
Mayadunnage S, Stannard HJ, West P, Old JM (2022) Identification of roadkill hotspots and the factors affecting wombat vehicle collisions using the citizen science tool, WomSAT. Aust Mammal.
Mayer M, Nielsen JC, Elmeros M, Sunde P (2021) Understanding spatio-temporal patterns of deer-vehicle collisions to improve roadkill mitigation. J. Environ. Manage 295:113148.
Meisingset EL, Loe LE, Brekkum Ø, Mysterud A (2014) Targeting mitigation efforts: The role of speed limit and road edge clearance for deer–vehicle collisions. J Wildl Manag 78(4):679–688.
National Institute of Biological Resources (2017) Analysis of culling program and a study on management monitoring methodology. National institute of biological resources. Incheon.
Neumann W, Ericsson G, Dettki H, Bunnefeld N, Keuler NS, Helmers DP, Radeloff VC (2012) Difference in spatiotemporal patterns of wildlife road-crossings and wildlife-vehicle collisions. Biol Conserv145(1):70–78.
Ng JW, Nielsen C, St. Clair CC (2008) Landscape and traffic factors influencing deer–vehicle collisions in an urban environment. Hum Wildl Confl 2(1):34–47.
Oh JG (2004) Characteristics of ecological behaviour of roe deer (Capreolus pygargus tianschanicus) in Jeju Island, Korea. Doctoral dissertation, Korea National University of Education.
Park Y, Lee W, Kim J, Oh H (2011) Morphological examination of the Siberian roe deer Capreolus pygargus in South Korea. J Anim Vet Adv 10(21):2874–2878.
Park YS, Lee WS (2014) Characteristics of habitat-using of Siberian roe deer in Seoraksan (Mt.) National Park. Journal of the Korea Society of Environmental Restoration Technology 17(1):91–109.
Pokorny B, Cerri J, Bužan E (2022) Wildlife roadkill and COVID‐19: A biologically significant, but heterogeneous, reduction. J Applied Ecol 59(5):1291–1301.
Phillips SJ, Dudík M, Schapire RE (2004) A maximum entropy approach to species distribution modeling. In Proceedings of the twenty-first international conference on Machine learning, pp 83.
Raymond S, Schwartz AL, Thomas RJ, Chadwick E, Perkins SE (2021) Temporal patterns of wildlife roadkill in the UK. PLoS One 16(10):e0258083.
Schwartz AL, Shilling FM, Perkins SE (2020) The value of monitoring wildlife roadkill. Eur J Wildl Res 66(1):1–12.
Seiler A (2003) The toll of the automobile: wildlife and roads in Sweden. Doctoral dissertation. Swedish University of Agricultural Sciences.
Seiler A (2004) Trends and spatial patterns in ungulate-vehicle collisions in Sweden. Wildlife Biol 10(1):301–313.
Seiler A (2005) Predicting locations of moose–vehicle collisions in Sweden. J Applied Ecol 42(2):371–382.
Seo M, Seo J, Jeong S, Hong Y, Choi S, Hyun J, Kim Y, Hwang I, Lee S (2021) Wildlife survey. National Institute of Biological Resources.
Song E, Woo D, Kim M, Park T (2020) Intensive survey on roadkill hotspots in South Korea. National Institute of Ecology.
Sillero N, Poboljšaj K, Lešnik A, Šalamun A (2019) Influence of landscape factors on amphibian roadkills at the national level. Diversity 11(1):13.
Steiner W, Schöll EM, Leisch F, Hackländer K (2021) Temporal patterns of roe deer traffic accidents: effects of season, daytime and lunar phase. PLoS One 16(3):e0249082.
Thomas DL, Taylor EJ (2006) Study designs and tests for comparing resource use and availability II. J Wildlife Manage 70(2):324–336.
Yue S, Bonebrake TC, Gibson L (2019) Informing snake roadkill mitigation strategies in Taiwan using citizen science. J Wildl Manag 83(1):80–88.
Zuberogoitia I, Del Real J, Torres JJ, Rodríguez L, Alonso M, Zabala J (2014) Ungulate vehicle collisions in a peri-urban environment: consequences of transportation infrastructures planned assuming the absence of ungulates. PLoS One 9(9):e107713.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Predictive modeling of ungulate–vehicle collision in the Republic of Korea

Status:

Version 1

Abstract

Context

Objectives

Methods

Results

Conclusions

Figures

Introduction

Materials And Methods

Study area

Data collection

Habitat suitability modeling

UVC mapping

Results

UVC data

Habitat suitability models

UVC probabilities

Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1