Western Kentucky University

TopSCHOLAR
Geography/Geology Faculty Publications

Geography & Geology

9-2008

Kernel Density Estimation of Traffic Accidents in a

Network Space
Zhixiao Xie
Florida Atlantic University, [email protected]

Jun Yan
Western Kentucky University, [email protected]

Follow this and additional works at: http://digitalcommons.wku.edu/geog_fac_pub

Part of the Geography Commons, and the Geology Commons
Recommended Repository Citation
Xie, Zhixiao and Yan, Jun. (2008). Kernel Density Estimation of Traffic Accidents in a Network Space. Computers, Environment, and
Urban Systems, 35 (5), 396-406.
Available at: http://digitalcommons.wku.edu/geog_fac_pub/3

This Article is brought to you for free and open access by TopSCHOLAR. It has been accepted for inclusion in Geography/Geology Faculty
Publications by an authorized administrator of TopSCHOLAR. For more information, please contact [email protected].

Kernel Density Estimation of Traffic Accidents in a Network Space

Zhixiao Xie a, Jun Yan b
a Department of Geosciences, Florida Atlantic University, 777 Glades Road, Boca Raton, FL 33431, USA
b Department of Geography and Geology, Western Kentucky University, Bowling Green, KY 42101, USA

Abstract: A standard planar Kernel Density Estimation (KDE) aims to produce a smooth
density surface of spatial point events over a 2-D geographic space. However the planar
KDE may not be suited for characterizing certain point events, such as traffic accidents,
which usually occur inside a 1-D linear space, the roadway network. This paper presents
a novel network KDE approach to estimating the density of such spatial point events.
One key feature of the new approach is that the network space is represented with basic
linear units of equal network length, termed lixel (linear pixel), and related network
topology. The use of lixel not only facilitates the systematic selection of a set of regularly
spaced locations along a network for density estimation, but also makes the practical
application of the network KDE feasible by significantly improving the computation
efficiency. The approach is implemented in the ESRI ArcGIS environment and tested
with the year 2005 traffic accident data and a road network in the Bowling Green,
Kentucky area. The test results indicate that the new network KDE is more appropriate
than standard planar KDE for density estimation of traffic accidents, since the latter
covers space beyond the event context (network space) and is likely to overestimate the
density values. The study also investigates the impacts on density calculation from two
kernel functions, lixel lengths, and search bandwidths. It is found that the kernel function
is least important in structuring the density pattern over network space, whereas the lixel
length critically impacts the local variation details of the spatial density pattern. The
search bandwidth imposes the highest influence by controlling the smoothness of the
spatial pattern, showing local effects at a narrow bandwidth and revealing hot spots at
larger or global scales with a wider bandwidth. More significantly, the idea of
representing a linear network by a network system of equal-length lixels may potentially

lead the way to developing a suite of other network related spatial analysis and modeling
methods.
Keywords: Network Space, Lixel, Kernel Density Estimation, Traffic Accidents, Hot
Spots

1. Introduction
To reduce traffic accidents and improve road safety, it is crucial to understand
how, where and when traffic accidents occurred. An improved understanding of spatial
patterns of traffic accidents can make accident reduction efforts more effective. For
instance, by knowing where and when traffic accidents usually occur, law enforcement
can conduct more efficient patrols and highway departments can disseminate more
effectively to drivers the critical information about roadway conditions. In reality, the
occurrences of traffic accidents are seldom random in space and time. In most cases,
traffic accidents form clusters (known as hot spots) in geographic space. This is
because the occurrence of traffic accidents along a certain roadway segment is largely
determined by its traffic volume, which is well-known to exhibit distinct spatial and
temporal patterns (Black, 1991). There are some other important factors that may impact
the distribution of traffic accidents, including natural and environmental characteristics
such as physical environment (steep slope, sharp turn), weather (rain, snow, wind, and
fog), configuration of highway networks such as locations of access and egress points,
deficient design and maintenance of highways, etc. All of these factors more or less are
associated with distinct spatial patterns as well.
Spatial analysis of point events, known as point pattern analysis (PPA), has been
widely examined by spatial scientists and a variety of methods have been developed for
detecting hot spots of point events. The PPA methods can be classified into two broad
categories (Bailey & Gatrell, 1995; OSullivan & Unwin, 2002): (1) Methods examining
the first-order effects of a spatial process, (2) Methods examining the second-order
effects of a spatial process. The first group focuses on the underlying properties of point

events and measures the variation in the mean value of the process. It includes methods
such as quadrat count analysis, kernel density estimation and etc. The second group
mainly examines the spatial interaction (dependency) structure of point events for spatial
patterns, and includes methods such as nearest neighbor statistics, G function, F function,
K function and etc. Out of these two categories of methods, Kernel Density Estimation
(KDE) is one of the most popular methods for analyzing the first order properties of a
point event distribution (Silverman, 1986; Bailey & Gatrell, 1995) partially because it is
easy to understand and implement. Some KDE tools are already made available in some
leading commercial GIS software, e.g. the Spatial Analyst Extension of ESRIs ArcGIS,
as well as some popular spatial statistical analysis software, such as CrimeStat (Levine,
2004). The planar KDE has been used widely for traffic accidents hot spots analysis
and detection. The recent examples include study of urban cyclists traffic hazard intensity
(Delmelle & Thill, in press), pedestrian crash zones detection (Pulugurtha, Krishnakumar,
& Nambisan, 2007), wildlife---vehicle accident analysis (Krisp & Durot, 2007), highway
accident hot spot analysis (Erdogan et al., 2008) and etc. The purpose of KDE is to
produce a smooth density surface of point events over space by computing event intensity
as density estimation. In planar KDE, the space is characterized as a 2-D homogeneous
Euclidian space and density is usually estimated at a large number of locations that are
regularly spaced (a grid). However, in analyzing the spatial pattern of traffic accidents,
which usually occur on roadways and inside a network, the assumption of homogeneity
of 2-D space does not hold and the relevant KDE methods are not readily applicable.
Many other types of human-induced point events also exhibit similar property in that
their distributions are constrained to only the network portion (network space) of the 2-D

Euclidean space, such as residential houses, commercial sites, street lights, moving
vehicles, and etc. In short, the uniformity of 2-D space is basically too strong an
assumption for the analysis of point events occurring in 1-D infinite space (Miller, 1999a).
Special considerations are thus needed for measuring such point events occurring in
network spaces1.
In the early 1990s, spatial scientists started to realize the limitations of spatial
methods originated in the 2-D Euclidean space when applying them directly to networkconstrained phenomena. A large number of studies have been conducted since then in a
variety of application domains, attempting to extend the conventional 2-D spatial
methods to network spaces, including network autocorrelation (Black, 1992; Black &
Thomas, 1998), network Huff model of market area analysis (Miller, 1994; Okabe &
Kitamura, 1996; Okabe & Okunuki, 2001), network distance-decay (Kent et al., 2005),
space-time accessibility measures (Kwan, 1998; Miller, 1999b), space-time clustering
(Black, 1991) and etc. In particular, increased attentions are recently paid to the
applications of standard spatial statistical methods in analyzing spatial point events in a
network space (Okabe et al., 1995; Okunuki & Okabe, 1998; Okabe & Yamada, 2001;
Yamada & Thill, 2004; Lu & Chen, 2007; Yamada & Thill, 2007). For instance, the Kfunction (another popular PPA method), has been extended to network spaces. A network
version of K-function and its computational implementation are described in Okabe and
Yamada (2001). To examine the advantages of network K-function, Yamada & Thill
(2004) compared three versions of K-functions in their ability of analyzing traffic

In this paper, a network space is defined as a simplified and abstracted representation of the real world
road network, which occupies a portion of the actual 2-D geographic space. It is represented as 1-D lines
and line-intersections. The roadway width, traffic direction, and multi-lane properties are not considered for
simplicity of concept demonstration.

accident patterns, namely planar K-function, network-constrained K-function and

network K-function. Their findings indicate that standard planar K-function tends to
over-detect clusters as it searches for the clustered patterns by comparing with the
random patterns over the entire 2-D space instead of the network space, which itself often
exhibits clustering tendency (e.g. the streets in the central city are often denser than those
at the outskirts). In another study, Lu & Chen (2007) reach a similar conclusion that
planar K-function is likely to produce false alarms of clusters when being applied to
detect hot-spots of vehicle thefts in the San Antonio, Texas area. Due to the increasing
popularity of the network K-function, an ArcGIS-based software tool, known as Spatial
Analysis on a NETwork (SANET)2, was recently developed by a group of researchers at
the University of Tokyo, Japan (Okabe et al., 2006). SANET offers network version of
both global and local K-functions as well as some additional utility tools for data
processing.
In contrast to the new developments of the network K-function, few studies have
attempted to extend the KDE methods to a network space. Recently, Borruso (2005)
analyzed patterns of point events distributed on a network with a modified KDE, termed
as Network Density Estimation (NDE) in his paper, which considers the kernel as a
density function based on network distances. Borruso (2005) pointed out the possibility
of extending the standard 2-D KDE to network spaces for identifying potential linear
clusters along roadways, however in his study, the kernel is still area based (using
network service area) and the outcome (point density) is still mapped onto a 2-D
Euclidian space. In essence, it is still a KDE in a planar pace instead of a network space.
2

SANET software tools can be obtained free of charge upon written request.

See Yamada & Thill (2007) for a typology of situations involving planar and network
spaces in spatial analysis. In addition, the spatial pattern of such kind of point events is
better measured with density values per linear unit over a network instead of per area unit
over a 2-D space. Indeed, in real-world applications, the density of traffic accidents is
often reported as the number of accidents over a defined linear unit (e.g. per mile) rather
than per area unit (e.g. per square mile).
This paper presents a novel network KDE approach to estimating the density of
traffic accidents strictly over a network space. As secondary objectives, the study also
investigates the impacts on density calculation from two different kernel functions, lixel
lengths, and search bandwidths. Although developed initially for traffic accidents, this
new approach could be used to examine the spatial pattern of any point events, as far as
their distribution is limited within network spaces. The remainder of the paper is
organized as follows. The basic concepts of network KDE are discussed in Section 2. The
computational algorithm is detailed in Section 3 along with discussions of some
implementation issues. In Section 4, we present a case study with a real road network and
traffic accident data. Discussions are made and some conclusions are drawn in Section 5.

2. Kernel Density Estimation in a Network Space

2.1. Planar Kernel Density Estimation

Network KDE is an extension of the standard 2-D KDE and a brief summary of
key aspects of the standard 2-D KDE is necessary. The general form of a kernel density
estimator in a 2-D space, termed as planar KDE in the rest of this paper, is given by:

1 d is
k( )
2

r
r
i =1

( s) =

(1)

where (s ) is the density at location s, r is the search radius (bandwidth) of the

KDE (only points within r are used to estimate (s ) ), k is the weight of a point i at
distance dis to location s. k is usually modeled as a function (called kernel function) of the
ratio between dis and r. As a result, rather than choosing a uniform function that gives
equal weight to all points within the bandwidth r, the KDE uses a model function through
which distance decay effect can be taken into account basically the longer the
distance between a point and location s, the less that point is weighted for calculating the
overall density. In the end, all the points within the bandwidth r of location s, weighted
more or less depending on its distance to s, are summed for calculating the density at s.
A number of forms of model functions, known as kernel functions, can be used to
measure the distance decay effect in the spatial weights k, such as Gaussian, Quartic,
Conic, negative exponential, and epanichnekov (Levine, 2004; Gibin et al., 2007). Three
forms of kernel functions are most commonly used (Schabenberger & Gotway, 2005, p.
111) and are discussed below, including:
(1) Gaussian function:

d
k ( is ) =
r

k(

1
d
exp( is 2 ) ,
2r
2

when 0 < dis <= r

d is
) = 0 , when dis > r
r

(2) Quartic function (which approximates Gaussian function):

d
d
k ( is ) = K (1 is2 ) ,
r
r

k(

when 0 < dis <= r

d is
) = 0 , when dis > r
r

Where K is a scaling factor and its purpose is to ensure the total volume under
Quartic curve is 1. The common values used for K include

and

3
, i.e.
4

d
3
d
d is
d is 2
3
k ( ) = (1 2 ) or k ( is ) = (1 is2 )
r

r
r
4
r
(3) Minimum variance function
2

k(

d is
3
d
) = ( 3 5 is2 ) ,
r
8
r

k(

d is
) = 0 , when dis > r
r

when 0 < dis <= r

A wealth of literature has examined the effects of the two key parameters of
planar KDE, i.e. kernel function k and search bandwidth r, on the resultant density pattern.
There exists a consensus that the choice of the kernel function k is less important than the
choice of search bandwidth r (Silverman, 1986; Bailey & Gatrell, 1995; OSullivan &
Unwin, 2002; Schabenberger & Gotway, 2005; O' Sullivan & Wang, 2007). It is also
agreed that the value of search bandwidth r usually determines the smoothness of the
estimated density -- the larger the r the smoother is the estimation.

2.2. Network Kernel Density Estimation

10

With the linear nature of network spaces in mind, this paper proposes to use the
following form of kernel density estimator for the density estimation of networkconstrained point events, such as traffic accidents, in a network space:
n

1 d is
)
r
r
i =1

(s) = k (

Instead of calculating the density over an area unit, the equation estimates the
density over a linear unit. Any of the three forms of kernel functions discussed previously
may be used. As shown in previous sub-sections, many earlier studies in planar KDE
have concluded that the choice of kernel function is not as important as the choice of
bandwidth r. We conjecture that the relative significance of the two parameters on
density estimation might not change much in the new network KDE, since there is no
particular reason for them to perform differently. To confirm our conjecture, we do
implement two kernel functions, Gaussian and Quartic functions, and compare their
impacts on the resultant density pattern in our case study. To verify the role of search
bandwidth in network KDE, we also examine how the density pattern will be impacted
when different search bandwidth are chosen.
It is necessary to emphasize that the network KDE differs from the planar KDE in
several aspects: (1) the network space is used as the point event context, (2) both search
bandwidth and kernel function are based on network distance (calculated as the shortestpath distance in a network) instead of straight-line Euclidean distance, and (3) density is
measured per linear unit. These differences are illustrated in a graphic form in Figure 1. It
can be noticed that network KDE is a 1-D measurement while planar KDE is a 2-D one.
As a result, the actual density values estimated by them would be very different for the

11

same point events dataset. This illustrative example also suggests that a Planar KDE
could possibly over-detect clustered pattern , with four traffic accidents falling within the
search bandwidth and hence included in density estimation for the focal point (x) in the
case of Planar KDE, while only two in the case of Network KDE.

3. Algorithm and Its Implementation

3.1. Computational Algorithm
The basic algorithm for network KDE is presented as follows. The basic terms are
defined and highlighted when they appear the first time, and some are illustrated in
Figure 2. The key implementation issues are described further in the next sub-section.
(1)

Create a segment-based linear reference system out of the original road network,
with each segment being a line segment between two neighboring road
intersections. A dangling line segment between a road intersection and a
neighboring road end point is also a segment. If there are multiple links between
two intersections, each link is treated as a separate segment.

(2)

Divide each segment into basic linear units of a defined network length l. A basic
linear unit is equivalent to a cell in a 2-D raster grid. For description simplicity,
we call it lixel (a contraction of the words linear and pixel) in the paper. The use
of lixel not only facilitates the systematic selection of a set of regularly spaced
locations along a network for density estimation, but also makes the practical
application of the network KDE feasible by significantly improving the
computation efficiency. The residual of the division (if there is one), i.e. the last

12

lixel with length shorter than l, is a partial lixel, the processing of which will be
detailed later. The intersection point of two lixels is called lxnode.
(3)

Create a network of lixels by establishing the network topology between lixels, as

well as between lixels and lxnodes.

(4)

Create the center points of all the lixels. They are termed lxcenters.

(5)

Select a point process (traffic accidents in this study), which has to be a type of
point events occurring within the network space.

(6)

For each point event, find its nearest lixel. The total number of events nearest to a
lixel is counted and assigned to the lixel as a property. Those lixels with one or
more accidents assigned to them are used as source lixels. Aggregating events to
lixels is one important step for improving the computational efficiency of the
algorithm. The possible impacts and future research plans are further described in
the conclusions and discussions section.

(7)

Define a search bandwidth r, measured with the shortest-path network distance.

(8)

Calculate the shortest-path network distance from the lxcenter of each source lixel
to lxcenters of all its neighboring lixels within the search bandwidth r. It should
be noted not only the first order nearest neighbor lixels, but all the neighbor lixels
with network distance not farther than r are taken into consideration.

(9)

At the lxcenter of each source lixel and all its neighboring lixels, calculate a
density value based on a selected kernel function, the network distance, and the
number of events on the source lixel.

13

(10)

At the lxcenter of each lixel within the search bandwidth of any source lixels, sum
the density values from different source lixels and assign the total density to the
lixel; for all other lixels, the density value is zero by default.

3.2. Key Implementation Issues

As shown in the basic algorithm, there are some key implementation issues to be
carefully considered in density estimation in a network space, including dividing
segments into lixels, selecting kernel functions, defining search bandwidth, and
computing lixel-based density. In a 2-D space, a lattice of grid cells can be placed to
exhaustively cover the space for a systematic selection of a set of locations for density
estimation. But in a network space, this may not be suitable and special treatments are
needed since the real-world networks are usually represented as linear features and are
often irregularly configured. Therefore, we have proposed a simple but effective
segmentation solution in this study: (1) first break the network into a series of
independent segments with each segment being a line segment between two neighboring
road intersections; (2) then divide each segment into equal-length lixels. Both steps are
done in a linear reference system (LRS). The center of a lixel is used as the density
estimation location and the estimated density at the center is used to represent the entire
lixel. In essence, like a cell in a 2-D raster representation, a lixel is treated as a
homogeneous unit and the internal variation is not considered. After segmentation, the
topological relationship is established between lixels, and between lixels and lxnodes. The
resultant network system of lixels is the basis for assigning accident points, measuring
network distance, and ultimately calculating the accident densities for lixels.

14

One issue to notice in the segmentation process is that the length of a segment (e.g.
51 meters) may not be exactly an integer number times of a defined lixel length (e.g. 10
meters). Hence, a residual lixel with length, here 1 m, shorter than the defined lixel length
often results for this kind of segment. These residual lixels do not actually present any
issues in implementation. The density values at the lxcenters of these residual lixels is
calculated in the same way as for regular lixels, based on the same kernel functions and
the actual network distance from their lxcenters to the centers of the source lixels,
although the network distance will not be integer number times of the defined lixel
length. A minor caveat is that the resolutions (length) are different between the regular
and residual lixels, however, the overall density pattern should not be affected because
residual lixels generally only amounts to a very small proportion of the entire inventory
of lixels. Further, as far as the lixel length is sufficiently short (e.g. 10 m), including or
excluding a single residual lixel in a hot spot may have trivial effects on the hot spots
detection even at local scales.
Another related issue is how to determine whether a lixel lies within the search
bandwidth from a source lixel. There are at least two different options when a lixel could
be labeled within a search bandwidth: (1) if the network distance from the lxcenter of a
source lixel to the farthest end point of the lixel is shorter than or equal to the search
bandwidth; or (2) if the network distance from the lxcenter of a source lixel to the
lxcenter of the lixel is shorter than or equal to the search bandwidth. The two options may
perform a little differently for the marginal neighbor lixels, but the impacts should be
trivial. In this paper, we adopted the second option.

15

As described in Steps (8)-(10) in the algorithm, the density computation is an

iterative process starting from a source lixel to all its neighbor lixels within the search
bandwidth. The number of accidents associated with a source lixel is used as a multiplier
of the density values for all the lixels within the search bandwidth. After computing the
density for one source lixel and all its neighbors, the algorithm repeats the same process
for another source lixel till all the source lixels are processed. For a lixel within the search
bandwidth of multiple source lixels, its density is the cumulative value of the density
derived from all the sources. Since the number of source lixels is generally much smaller
than the total number of lixels in a network space, this density computation process is
obviously much more efficient than computing density for all lixels, most of which may
never fall within the search bandwidth of a source lixel.

4. Case Study The Analysis of Traffic Accident Patterns in Bowling Green,

Kentucky
The proposed algorithm is implemented in the ESRI ArcGIS environment, using
Microsoft Visual C# 2005. The test dataset includes a real transportation network system
in the Bowling Green, Kentucky area, and the traffic accident data for year 2005 (Figure
3). The traffic accident data is provided by the Kentucky State Police Department. The
point location of each accident is recorded as a pair of longitude and latitude via the
carry-on GPS unit within a reporting police car. A total of 3226 traffic accidents are
analyzed in the case study. A series of tests are conducted to demonstrate the
applicability of the algorithm, to compare the difference between the new algorithm and a
standard planar KDE, and to examine the impacts of different parameters on density

16

calculation, including kernel functions, lixel length, and search bandwidth. Two kernel
functions, Gaussian and Quartic, are compared at a fixed search bandwidth of 100 m for
local details and a search bandwidth of 1000 m for the overall pattern, both with a lixel
length of 10 m. Four versions of segmentation scenarios, with the lixel length at 5 m, 10
m, 50 m, and 100 m respectively, are tested at a fixed search bandwidth of 100 m and
with a Gaussian kernel function. We also examine the impacts of search bandwidth at
local and larger spatial extents. Total six search bandwidths are used, including 20 m, 100
m, 250 m, 500 m, 1000 m, and 2000 m, all with the same lixel length (10 m) and a
Gaussian kernel function.

4.1. Comparison of Planar KDE and Network KDE

An intuitive approach to present the spatial pattern of accidents is to map the
accident locations or the simple accident count per lixel after each accident is assigned to
its closest lixel. Figures 4-a and 4-b illustrate the spatial pattern of traffic accidents for a
small part of the study area with these two intuitive approaches respectively. Figures 4-c
and 4-d show the density values calculated for the same area using a standard planar
KDE (10-m raster cell) and the new network KDE (10-m lixel) respectively, both with a
Gaussian kernel and a 100-m search bandwidth. Figures 5-a and 5-b present the overall
spatial pattern of density values computed with the network KDE (10-m lixel) and the
planar KDE (10-m raster cell) respectively, both with a Gaussian kernel but a wider
search bandwidth of 1000 m. A wider bandwidth is used to better reveal the differences
of the two KDE approaches and the role of search bandwidth will be discussed further in
section 4.4. By a simple visual comparison, both KDE based density values present a

17

much informative pattern of accident likelihood. However the density maps of the planar
KDE and network KDE are rather different, although some common hot (high value)
regions can be identified in both maps. The density surface by the planar KDE
exhaustively fills out the entire study area (although some areas with value of zero) while
that by the network KDE only covers the space occupied by the street network. In
addition, the maps indicate that the planar KDE is likely to overestimate the density
values in comparison to the network KDE. In Figures 4-c and 4-d, the density value for
the planar KDE can reach as high as over 2000 (per km2) (Figure 4-c), in comparison, the
highest value for the network KDE is only about 140 (per km) (Figure 4-d).

4.2. The Impacts of Kernel Functions

Figures 6-a and 6-b illustrate the spatial patterns of density values computed in a
small part of the study area with the proposed network KDE algorithm based on kernel
functions of Gaussian and Quartic respectively, both at a 10-m lixel length and with
a100-m search bandwidth. It is obvious that the two kernel functions result in very
similar local density variations. Figures 7-a and 7-b show the overall density pattern for
the study area using a Gaussian and a Quartic function respectively, at the same lixel
length (10 m) and with the same search bandwidth (100 m). It appears that the choice of
kernel functions also make little difference in the overall density pattern. Only that the
density values estimated with Quartic kernel are higher than those with Gaussian kernel.
These observations corroborate our conjecture that kernel functions are least important in
structuring the density pattern in network KDE, similar to the case in planar KDE.

18

4.3. The Impacts of Lixel Length

Lixel length, like raster cell size in a planar KDE, significantly affects the local
variation details of the network density value pattern. Figure 8(a)-(d) show the network
KDE estimated density values, with the lixel length at 5 m, 10 m, 50 m, and 100 m
respectively. Although the kernel function is the same (Gaussian), and with the same
search bandwidth (100 m), the density values along roads lose local variation details as
lixel length increases. The larger lixel lengths effectively hide the detailed structures
shown at finer resolutions. For example, for the same network segment from A to B in
Figures 8(a)-(d), we can notice Figure 8-a shows much more local details of variation in
the density values than Figure 8-d. As a trivial point, the density values remain in almost
the same range between 0 to about 200 per kilometer even when different lixel lengths
are used.

4.4. The Impacts of Search Bandwidth

Search bandwidth plays the most significant role in structuring the network
density pattern. As shown in a local part of the study area (Figure 9(a)-(d)), the density
pattern gets smoother with increasing search bandwidth (20 m, 100 m, 250 m, and 500 m
respectively), even when the kernel function is the same (Gaussian) and at the same lixel
length (10 m),. The density values are almost invariant with a 500-m search bandwidth
(Figure 9-d), whereas the density variation pattern is quite bumpy with a 20-m bandwidth
(Figure 9-a).
The Figure 9 only shows the response of density value to the search bandwidth at
a small part of the study area with relatively narrower bandwidths. Figures 10(a)-(f)
19

present the impacts of the search bandwidth (20 m, 100 m, 250m, 500 m, 1000 m, and
2000 m respectively) on the overall density pattern for the whole study area. It appears
that the narrow bandwidths (20 m, 100 m, and 250 m) may produce patterns suitable for
presenting local effects or hot spots at smaller scales. As the search bandwidth
increases from 20 m to 2000 m, the local hot spots are gradually combined with their
neighbors, and larger clusters appear. The maps at wider search bandwidths (500 m, 1000
m, 2000 m) seemingly give better sense of locations of the hot spots at larger spatial
scales.

4.5. Density Visualization

In a 2-D space represented by regularly spaced grid cells or points, a density
surface can be easily displayed in a raster GIS. The visualization of density values across
a network needs to adopt different strategies to reflect the linear nature of the network
space. Usually three visual variables are associated with drawing line features, including
color, width, and height. The visualization can be implemented with one of the three
variables or in the combination of two and even all three of them. We already show
density patterns with the width variable in Figures 4, 6, 7, 8, 9, and 10. Figure 5-a uses
color (gray tone) to show the overall density pattern estimated by the network KDE.
Figure 11 presents the spatial pattern of density using 3-D symbols (gray tone and height)
in ESRI ArcScene. The visualization methods presented in these figures appear to be
effective in presenting the network constrained accident density pattern. As a quick and
general observation, it appears that major corridors in the study area tend to have

20

relatively higher estimated density values of traffic accidents. And several major
intersections can be visually identified as hot spots of traffic accidents.

5. Conclusions and Discussions

Recognizing the limitations of applying standard 2-D planar KDE methods in a
network space, this paper develops a network KDE approach to characterizing the spatial
patterns of traffic accidents on roadways. The basic operation unit of the new KDE
algorithm is a lixel of a defined network length along linear road segments. All the
accidents are assigned to their nearest source lixels. The density value at the center point
of a lixel is computed as the sum of kernel-function derived densities from all the source
lixels within a specified search bandwidth measured by the shortest-path network
distance. As a case study, the new KDE algorithm is successfully applied in a real world
road network system and traffic accidents dataset. The operational success of the
application demonstrates the applicability of the algorithm. Visual comparison of the
resultant density patterns of a standard planar KDE and the network KDE shows that the
planar KDE covers space beyond the network context and over-estimates the density
values. The case study also examines the impacts on network KDE density calculation
from different kernel functions, lixel lengths, and search bandwidths. Although we only
implement two kernel functions and cannot make sound conclusions, it does appear that
the kernel functions have less important role in structuring the density patterns over
network space. The lixel length, like the cell resolution in raster representation, impacts
the local variation details of the spatial pattern of density. It is also found that the search
bandwidth imposes the highest influence by controlling the smoothness of the spatial

21

pattern. A narrower search bandwidth could reveal local effects, whereas a wider
bandwidth makes hot spots much more obvious at large spatial scales.
It should be pointed out that the proposed algorithm only calculates the density
value at the center point of each lixel and uses that value to represent the whole lixel.
Although we can certainly compute density values for as many points along a linear
segment as possible, we argue that it is appropriate to use the center point density value
to approximate the density value over the lixel of a reasonable length, mainly because
there are infinite points on a lixel of any length and a certain degree of simplification and
abstraction is always needed for a practical application. Similar simplification and
abstraction are not uncommon. For example, in the 2-D space raster representation, the
center value of a grid cell has often been used to approximate the value of the whole cell.
In addition, an equal-length segmentation approach is used in the current algorithm. In
the future study, it may be interesting to examine other segmentation scenarios with a set
of varied lixel lengths, e.g. shorter length for central urban, longer for rural areas, and so
on.
It is very tempting to recommend an optimal lixel length. However, we avoid
doing so since a universally applicable lixel length may not exist, like the cell resolution
in a 2-D raster representation. In remote sensing, imagery of a particular resolution is
chosen to suit specific application context, and it should be so in the network KDE. The
length of lixels must be carefully chosen based on application context. The inherent data
quality of the road network and accident locations is also an important factor to take into
consideration. In traffic accident, a 10-m lixel length should be sufficiently short if not
optimal given the length of a vehicle, the direct impact area of an accident, and the

22

location accuracy in measurement process. This lixel length may not suit other point
events. We also refrain from recommending an optimal search bandwidth for similar
reasons and believe search bandwidth selection should also be application dependent. It
may be wise to use a set of search bandwidths to reveal hot spots from very local
effects to global scale. In essence, this would lead to a multi-scale exploration, which is
actually needed for many systems with hierarchical structure.
In the implementation, we aggregate all the accident points associated with a focal
source lixel into an accident count of the focal lixel, and their locations on the focal lixel
are not further distinguished. It is from the center point of focal source lixel that the
network distance to each neighboring lixel center is calculated. The aggregation has the
advantage of computational efficiency and more importantly it is closer to the reality in
the sense that a real accident seldom occurs precisely at a dimensionless point location
but occupying a certain length along a roadway. The location accuracy is also affected by
measurement process and contributes uncertainties to the event position. Regardless, the
impacts of aggregation on the measured distance and subsequently the density value
should be minor. For comparison purpose, we will implement the network KDE with unaggregated events in the future study.
In the current algorithm, road segments as well as lixels are also not further
distinguished and are treated equally. In reality, road segments are different in their
traffic flow capacity (e.g. the number of lanes and speed limit), directionality (one way or
two ways, left turn or right turn) and etc. The same accidents will have different
implications depending on the type of road segments they are on, and may have different
impacts on the traffics ahead of or behind the accident vehicles. Such information are not

23

accessible for the study area, however in areas where they are available, it may be
beneficial to incorporate them into the density estimation. The network KDE also has one
of the same fundamental drawbacks as the planar KDE for hot spots detection. Because
no statistical significance is employed in the process, it is ad hoc and there is no
indication of a density threshold above which hot spots can be confidently declared.
Experiments with different density thresholds may be needed in real applications.
Nevertheless, we believe the network KDE method presented in this paper could be
useful and are readily applicable in real world setting for traffic accident decision making
by different agencies. More significantly, although designed for KDE estimation, the idea
of segmenting a linear road network into a network system of equal-length lixels may
potentially lead the way to developing a suite of other network related spatial analysis
and modeling methods and have a larger impact than what is shown in this initial
application.

Acknowledgement
The authors greatly appreciate the helpful comments of three anonymous
reviewers and the detailed comments by the Editor, Dr. Jean-Claude Thill.

24

References
Bailey, T. C., & Gatrell, A. C. (1995). Interactive Spatial Data Analysis. Essex: Longman.
Black, W. R. (1991). Highway Accidents: A Spatial and Temporal Analysis.
Transportation Research Record, 1318, 7582.
Black, W. R. (1992). Network Autocorrelation in Transport Network and Flow Systems.
Geographical Analysis, 24, 207222.
Borruso, G. (2005). Network Density Estimation: Analysis of Point Patterns over a
Network. Lecture Notes in Computer Science: Computational Science and Its
Applications, 3482, 126132. Erdogan, S., Yilmaz, I., Baybura, T., & Gullu, M.
(2008). Geographical information systems aided traffic accident analysis system case
study: city of Afyonkarahisar. Accident Analysis and Prevention, 40(1), 174181.
Delmelle, E. C., & Thill, J.-C. (2008). Urban Bicyclists A Spatial Analysis of Adult and
Youth Traffic Hazard Intensity. Transportation Research Record, in press.
Gibin, M., Longley, P., & Atkinson, P. (2007). Kernel Density Estimation and Percent
Volume Contours in General Practice Catchment Area Analysis in Urban Areas. In
the Proceedings of the GIScience Research UK Conference (GISRUK) 2007.
Maynooth - Ireland.
Kent, J., Leitner, M., & Curtis, A. (2006). Evaluating the Usefulness of Functional
Distance Measures When Calibrating Journey-to-Crime Distance Decay Algorithms.
Computers, Environment and Urban Systems, 30(2), 181200.
Krisp, J.M., & Durot, S. (2007). Segmentation of lines based on point densities An
optimisation of wildlife warning sign placement in southern Finland. Accident
Analysis and Prevention, 39(1), 3846.

25

Kwan, M-P., (1998). SpaceTime and Integral Measures of Individual Accessibility: A

Comparative Analysis Using A Point-Based framework. Geographical Analysis,
30(3), 191216.
Levine, N. (2004). CrimeStat III: A Spatial Statistics Program for the Analysis of Crime
Incident Locations. Ned Levine & Associates, Houston, TX, and the National
Institute of Justice, Washington, DC.
Lu, Y., & Chen, X. (2007). False Alarm of Planar K-Function when Analyzing Urban
Crime Distributed along Streets. Social Science Research, 36(2), 611632.
Miller, H. J. (1994). Market Area Delineation within Networks Using Geographic
Information Systems. Geographical Systems, 1(2), 157173.
Miller, H. J. (1999a). Potential Contribution of Spatial Analysis to Geographic
Information Systems for Transportation (GIS-T). Geographical Analysis, 31(4),
373399.
Miller, H.J. (1999b). Measuring SpaceTime Accessibility Benefits within
Transportation Networks. Geographical Analysis, 31(2), 187212.
Okabe, A., Yomono, H., & Kitamura, M. (1995). Statistical Analysis of the Distribution
of Points on a Network. Geographical Analysis, 27(2), 15275.
Okabe, A., & Kitamura, M. (1996). A Computational Method for Market Area Analysis
on a Network. Geographical Analysis, 28(4), 330349.
Okunuki, K., & Okabe, A. (1998). A Computational Method for Optimizing the Location
of a Store on a Continuum of a Network When Users' Choice Behavior Follows the
Hu_ Model. Discussion Paper, Center for Spatial Information Science at the
University of Tokyo, No.19.

26

Okabe, A., & Okunuki, K. (2001). A Computational Method for Estimating the Demand
of Retail Stores on a Street Network Using GIS. Transactions in GIS, 5(3), 209220.
Okabe, A., & Yamada, I. (2001). The K-Function Method on a Network and its
Computational Implementation. Geographical Analysis, 33, 271290.
Okabe, A., Okunuki, K., & Shiode, S. (2006). SANET: A Toolbox for Spatial Analysis
on a Network. Geographical Analysis, 38, 5766.
OSullivan, D., & Unwin, D. J. (2002). Geographic Information Analysis. John Wiley,
Hoboken, New Jersey.
O' Sullivan, D., & Wong, D. W. S. (2007). A Surface-Based Approach to Measuring
Spatial Segregation. Geographic Analysis, 39(2), 147168.
Pulugurtha, S.S., Krishnakumar, V. K., & Nambisan, S. S. (2007). New methods to
identify and rank high pedestrian crash zones: An illustration. Accident Analysis and
Prevention, 39(4,) 800811.Schabenberger, O., & Gotway, C. A. (2005). Statistical
Methods for Spatial Data Analysis. Chapman & Hall/CRC, Boca Raton, Florida.
Silverman, B. W. (1986). Density Estimation for Statistics and Data Analysis. Chapman
Hall, London.
Yamada, I., & Thill, J.-C. (2004). Comparison of Planar and Network K-Functions in
Traffic Accident Analysis. Journal of Transport Geography, 12, 149158.
Yamada, I., & Thill, J.-C. (2007). Local Indicators of Network-Constrained Clusters in
Spatial Point Patterns. Geographical Analysis, 39(3), 268292.

27

List of Figures

Figure 1. Illustration of the basic differences between the Planar KDE and Network KDE
for the same point event dataset. To estimate the density value at a focal point x, the
planar KDE treats the whole 2-D space as the context and finds 4 accident points (solid
dots) within a search bandwidth r, whereas the Network KDE only finds 2 accidents
within the same bandwidth in the network space based on network

28

distance.

Figure 2. Illustration of the basic terms used in the proposed network KDE algorithm.
The line segment between the two road intersections A and B is called a segment, so is
the dangling line segment AC between a road intersection A and a road end point C. Each
segment is divided into lixels of a defined network length l. Here, five lixels are created
for segment AB, with lixels 1, 2, 3 and 4 being regular lixels with length l, and lixel 5
being the residual lixel with length less than l. The dark dots are the lxnodes, the
intersection points of lixels. Note the original road intersections (A, B) and end points (C)
are always lxnodes.

29

Figure 3. The study area, Bowling Green, Kentucky. Test dataset include a road
transportation network system and traffic accident data for year 2005.

30

Figure 4. Illustration of different ways of presenting accident spatial patterns in a local

part of the study area: (a) accident point locations, (b) number of accidents per 10-m lixel,
(c) a standard planar KDE (10-m raster cell), (d) the proposed network KDE (10-m lixel).
For both (c) and (d), a Gaussian kernel and a 100-m search bandwidth are used. Both
KDE are more informative in presenting the density pattern, but only the network KDE
estimates density in the event context, the network space.

31

Figure 5. Illustration of different overall patterns produced by the network KDE and
planar KDE in the study area: (a) the proposed network KDE (10-m lixel), and (b) a

32

standard planar KDE (10-m raster cell). A Gaussian kernel and a 1000-m search
bandwidth are used.

Figure 6. Illustration of the local impacts of kernel functions in the network KDE: (a) a
Gaussian kernel based network density (per km), and (b) a Quartic kernel based network
density (per km). For both (a) and (b), a 100-m search bandwidth and a 10-m lixel length
are used. Both kernel functions result in similar local details.

33

Figure 7. Illustration of the impacts of kernel functions on the overall density pattern in
the network KDE: (a) a Gaussian kernel based network density (per km), and (b) a

34

Quartic kernel based network density (per km). It appears that the kernel functions make
less difference in the overall pattern.

Figure 8. Illustration of the impacts of t lixel length on the calculated density values for
the network KDE. (a)-(d) are the results for the lixel length of 5 m,10 m, 50 m and 100 m
respectively, with a Gaussian kernel and a 100-m search bandwidth. The shorter the lixel
length, the more detailed location variation can be revealed, as shown in the segment
from A to B.

35

Figure 9. Illustration of the impacts of narrower search bandwidths on the calculated

density values (per km) at local scales. (a)-(d) are the results for the search bandwidth of
20 m, 100 m, 250 m, and 500 m respectively, with a Gaussian kernel and a 10-m lixel
length. The local variations of density gradually get lost with increased search
bandwidths.

36

Figure 10. Illustration of the impacts of different search bandwidths on the overall
density pattern. (a)-(f) are the results for the search bandwidth of 20 m, 100 m, 250 m,
500 m, 1000 m, and 2000 m respectively, with a Gaussian kernel and a 10-m lixel length.
37

As the search bandwidth increases from 20 m to 2000 m, the local hot spots are
gradually combined with their neighbors and larger clusters appear.

Figure 11. Three-Dimensional visualization of the spatial pattern of density estimated by

the proposed network KDE (10-m lixel length, Gaussian kernel, 100-m search
bandwidth), with ESRI ArcScene.

38