EXPLORING PATTERNS AND FACTORS RELATED TO DEER-VEHICLE COLLISIONS IN CENTRAL NORTH CAROLINA
Timothy Mulrooney, Ph.D.
Assistant Professor, Department of Environmental, Earth and Geospatial Sciences
North Carolina Central University
Erik Green
Graduate Student, Department of Environmental, Earth and Geospatial Sciences
North Carolina Central University
The purpose of this poster is to analyze geographic patterns of deer-vehicle collisions in the state of North Carolina and explore both human and natural factors that lead to these spatial patterns. This analysis is important so traffic engineers and analysts can better understand a phenomenon that costs North Carolinians millions of dollars in damage as well as lives and injuries as a result of these collisions. Using GIS techniques and data provided by the North Carolina Department of Transportation (Traffic Analysis Unit), the locations of all 723 deer vehicle collisions that occurred in the North Carolina counties of Alamance, Orange and Durham are mapped. Grouped using a rectangular enumeration unit based on quadrat analysis, other human and natural factors were also collected at this quadrat level, mapped and analyzed to determine which of these explanatory factors correlated most with deer vehicle collisions. While
many qualitative and immeasurable factors also contribute to deer-vehicle collisions, these explanatory factors include population density, proximity to highways, proximity to bridges, various permutations of land cover (as provided by NLCD data), land cover variety and average speed limit.
Introduction
Deer-vehicle collisions have many costs in the state of
North Carolina. Based on data provided by the North Carolina Department of Transportation, there were more than
21,000 deer-vehicle collisions in 2010. Of those, 251 resulted
in evident injuries at the scene, 22 were classified as
‘disabling’ and 6 of these collisions were fatal (NCDOT
2011). Using GIS, we can get a better idea as to where these
collisions occur and more importantly the natural and human
factors that contribute to the spatial distribution of these phenomena.
In the ESRI Spatial Analyst toolbar, the Point Density
feature was used to calculate “a magnitude per unit area that
fall within a neighborhood around each cell.” In this case, the
point feature was the deer-vehicle collision feature class and
the neighborhood was represented by a circle around the pixel
in question. This resulted in a surface highlighting pixels
within proximity of high-density deer-vehicle collisions (Map
2).
Factors Affecting Deer-Vehicle Collisions
We began to look at explanatory factors at the quadrat
level such as population density, proximity to highways,
proximity to bridges, various permutations of land cover (as
provided by NLCD data), land cover variety and average
speed limit. 21 different attributes were derived to determine
factors that contribute to deer-vehicle collisions. Some of
these factors were derived from the NLCD (National Land
Cover Dataset) through the USGS (Map 5).
Study Area
While information about deer data was collected for the
entire state, this study focused on the Central North Carolina
Counties of Alamance, Orange and Durham Counties. We
felt this was important for a number of reasons. This area has
a high concentration of east-west flowing traffic from the Research Triangle through Greensboro and points south in addition to north-south running state roads with less traffic. In addition, there are a variety of land cover types that are incorporated within various levels of urbanization and suburbanization. This occurs in larger cities such as Durham, Chapel Hill
and Burlington, as well as smaller communities such as
Mebane, Hillsborough, Graham and Burlington and outlying
smaller communities off of the interstate highways such as
Alamance, Elon (formerly Elon College) and Carrboro.
There were more than 700 deer-vehicle collisions in this 3
county area in 2010. The location of these collisions are
shown in Map 1.
Methodology
While point pattern analysis could done, deer-vehicle
collisions were grouped into enumeration units. However,
many of these metrics are fairly general an merely offer a description of the entire data set.
Another type of analysis that can be performed on point
data is quadrat analysis. Quadrat analysis divides a study area
into equal sized areas called quadrats and uses probability
analysis to determine the actual frequency of points within
each quadrat versus an expected value for each frequency.
This study will take this one step further and measure the deer
density/proximity for each quadrat versus a number of variables to explain these patterns.
The use of the quadrat in this study is unique because sub
-county enumeration units such as census tracts and block
groups already exist. However, the irregular shapes of these
human-defined units imply a sense of spatial bias that may
skew polygonal analysis when grouped at this level. The
size of the quadrat is important as to be large enough to capture a distinct spatial pattern, but not small enough to have
many quadrats with zero frequency (Mitchell 2005).
With 728 recorded deer-vehicle collisions (n) in an area
of approximately 1180 mi2 (A), the quadrat size was computed using the following formula:
l
The research team felt this generalized density metric
was better than deer-vehicle collisions or density for 2 reasons. Primarily, the independent variable now has granularity
for better use in regression analysis. If deer-vehicle collisions
at the quadrat level only have 13 (0 – 12) different values
(Map 3).
If deer-vehicle density (deer-vehicle collisions per square
mile) were used, there would be slightly more dependent values. Secondly, the quadrat boundary is not the ultimate delimiter when computing deer density. Given the circular search
radius, points that occur outside of the quadrat will affect a
pixel’s density. While these densities are grouped at the
quadrat level, points that are close to each other, but happen
to lie within different quadrats can be reflected in this more
comprehensive metric and give the dependent variable more
granularity for future regression analysis. The resulting map
of this generalized collision density map from Map 2 is
shown in Map 4.
Another function that we explored was the raster-based
Variety parameter using the Focal Statistics function and then
averaged at the quadrat level using the Zonal Statistics function. Within each zone (quadrat) or neighborhood, we can
count perform a number of different functions on the raster
data (NLCD) within it. This includes statistical functions
such as mean, median and mode, but these are useless because NLCD data is categorical. The variety function finds
the number of unique values in each zone is assigned to all
cells in that zone. We can use different search functions. We
looked at pixel variety for 1 mile squares or 5 by 5 pixel
squares (Map 8). In Map 8, some quadrats have an average
variety of 3.93. That means for the 25 pixels that surround
the pixel in that quadrat, there are almost 4 different NLCD
cover types in that area. Other more rural areas have a much
lower pixel variety, meaning that this is primarily only one
cover type (probably forest). This rationale is that more variety will mean that deer are more likely to encounter a road as
they move from one cover type (forest, for example) to another cover type (wetland). The pixel variety for a 1 by 1
mile square ranged from 7 all of the way up to 14.2.
The following factors were derived to help determine independent variables that explained our deer density metric (Map
4). All of these factors, derived at the quadrat level, were:
However, the Var_5PIX (pixel variety based on 5 x 5 pixel
square) also was significant at the ρ = .001 level for all models. Another variable that that was frequently significant in
the positive direction was percent forest. This was later confirmed with when a simple linear model was run without the
road characteristics metrics (road mile, highway miles, road
distance, highway distance). While the r2 value was much
lower (.2841), the percent forest and Var_5PIX metrics were
most significant (ρ = .001 and ρ = .01 respectively) in this
model.
Another interesting variable that seemed to come out was the
DIST_LOW_IMP, which represents the distance to low impact urban areas. There are categories 21 (developed, open
space) and 22 (developed, low intensity) using the Revised
Anderson Classification scheme. It was statistically significant in some of the models, but what most telling was when a
single regression model was run between DEER_DENSE and
DIST_LOW_IMP. These single regression models were also
run in case any of the factor cancelled or counteracted each
other. It was a positive correlation where ρ = 0 and r2
= .1855. Of the 17 factors that were not related to road distance and proximity, this was the strongest model.
Factors co-located to deer-vehicle collisions are most related
to highway and road mileage within each quadrat, as well as
proximity to highways and roads at the quadrat level. This
relationship is fairly self-explanatory, so other factors related
to various land covers and land cover variety were explored.
In all models, percent forest cover and pixel variety contributed significantly toward deer-vehicle collisions. Pixel variety
shows that more diversified land covers will contribute towards collisions (i.e. Habitat Fragmentation), as deer needing
to move between different covers may encounter cars. In addition, the distance to low impact urban areas were also another interesting contributory factor to deer vehicle collisions.
While very few deer-vehicle collisions occur within city centers, they occur towards the suburb where human development meets with more open areas (Map 1). In some literature, this is referred to as the wildland-urban interface in the
study of wildfires, albeit at a much smaller and less grander
scale.
Discussion
Percent Deer Habitat
Percent Water
Percent Urban
Percent Planted
Percent Wetland
Avg. Distance to
Bridge
Percent Barren Land
# of Highway Miles
Percent Forest
# of Road Miles
Percent Shrubland
Avg. Distance to
Highway
Percent Herbaceous
Avg. Distance to
Road
Distance to Low Inten- Avg. Speed Limit of
sity and Open Space
Roads
Developed Land
Pixel Variety (# of difOriginal Source
ferent pixels), 1 mile
square
Pixel Variety, 5 pixel
NLCD
square
Pixel Variety, 15 pixel
NC DOT
square
Population
US Census
Population Density

2* A
n
The l value was computed to be 9504 feet (1.8 miles),
thereby making each quadrat to be approximately a square
with 3.24 miles in area. 395 quadrats compose the study area.
Given the shape of the study area, some quadrats had to be
Clipped to extent of the study area. 307 of the 395 quadrats
(77.7%) were entirely contained within the study area, with
the remaining 88 quadrants (or parts thereof) having areas
less than 3.24 mi2. With respect to deer-vehicle collision frequency, 63% of quadrats had at least 1 deer-vehicle collisions,
with a high quadrat of 12 collisions. While the number of
deer-vehicle collisions or collision density at the quadrat level
could be used as the dependent variable in this model, a more
robust metric was devised.
It uses the Anderson Classification scheme to represent
what occurs on the surface of the earth. NLCD data were instrumental in helping to develop attributes at the quadrat level
about urban areas potential deer habitat (non-urban), water,
urban areas (Map 6), barren land, forest, shrub land, planted
agriculture, wetland and low intensity urban areas. The spatial resolution for these data is 30 meters and a variety of different techniques are used to assign each pixel a distinct category. There are about 8 general categories and 21 distinct
NLCD classes among them (http://www.mrlc.gov/
nlcd06_leg.php). NLCD data for the entire state of North
Carolina can be extracted from LANDSAT data and is provided by the USGS at http://www.mrlc.gov/nlcd06_data.php.
GIS Data provided by the NC DOT (North Carolina Department of Transportation) can allow us to view different
permutations of transportation issues relation to deer-vehicle
collisions, which lie at the very heart of this problem. At the
quadrat level, we can use GIS functions such as Clip, Summarize and Join to compute the number of total highway
(Map 7) and road miles on a quadrat by quadrat basis. In addition, we can use the raster-based Euclidean Distance function to see how close deer collisions are to both highways and
road. It is generally assumed that these functions will correlate highly with deer-vehicle collisions. The distance to
bridges attribute was also computed. The rationale was that
bridges may serve as a safe haven to cross a road or may traverse a river or stream, which may increase the likelihood of a
collision.
Patterns of Deer-Vehicle Collisions
The patterns of deer-vehicle collisions are highlighted in Map
4. They generally mimic the patterns seen in Map 2 and to a
lesser extent Map 3. Both in terms of the raw number of deer
-vehicle collisions and the generalized density, the area
around Hillsborough experiences the most deer-vehicle collisions. While other factors can be explored, this is a high traffic areas at the confluence of Interstate routes 40 coming from
Raleigh and 85 coming from Durham and places Northeast.
While deer-vehicle collisions generally avoid the center of
larger cities such as Durham, Chapel Hill and Burlington,
they occur around the periphery of these cities. In rural areas,
deer-vehicle collisions are most prominent in areas of higher
traffic along state routes 86, 49, 87 and 54.

It impossible to predict the exact patterns (both temporal
and spatial) of deervehicle collisions
with any certainty or
accuracy.
Results
Using the statistical programming application R, a few
models were used to find quantifiable explanatory factors
to help explain the deer density metric (Map 4) derived
from all 700+ deer collisions in the study area. A linear and
exponential multiple regression models were used with all
21 independent variable to predict the deer density metric.
All models resulted in an r2 value between .421 and .5491.
In all of the models, variables that contributed most significantly (ρ = .001) to these models were road miles (# of road
miles within quadrat) and highway distance (average distance to highway). These are fairly intuitive given that
most if not all deer-vehicle collisions occur on roadways.
However, we can use GIS to get a better idea to areas that
may be predisposed to having these occurrences based on
traffic, land cover, population and various permutations of
each.
This study looked to find quantifiable factors to help explain deer-vehicle collisions. Using the models in this
study, factors related to roads, land cover, ancillary transportation factors and population distribution were used to
explain their relationship to deer-vehicle collisions using a
generalized metric at the quadrat level. Given time constraints, there are many other factors that could be used in
these models. This includes traffic patterns, soil types, the
location of farms, edge density and migration patterns.
Credits: Many special thanks go to the many peoples who have contributed data for this study. Data were provided by the
University of North Carolina Highway Safety Research Center (http://www.hsrc.unc.edu/index.cfm) who provided data in a
GIS-compatible format. GIS data representing roads were provided by the North Carolina Department of Transportation
(http://www.ncdot.gov/it/gis/). Census data were provided by Environmental Research System Institute (ESRI)