This section mainly introduces the methods used in this paper. First of all, the spatial join analysis between the road intersections (i.e., locations) and the crashes are performed to visualize the crash number at each intersection. Then, the space-time cube analysis is conducted to display the distributions of traffic crashes over time and space. Afterwards, the spatial autocorrelation analysis is proposed to analyze the intrinsic characteristics (include the spatial aggregation characteristic and statistical significance) of the crashes. Finally, the emerging hot spot analysis method is generated to explore the classification of the traffic crash hotspots and to evaluate the spatiotemporal change trends (include the changes over locations and time) of these hotspots.
4.1. Spatial Join Analysis
The crash locations have been labeled in
Figure 1, from which the crash distributions were dense along both sides of S227 but they are irregular. To explore the spatial distribution characteristic of these crashes, an accurate quantity of these crashes at each road intersection need to be counted. Thereby, a “spatial join” toolbox based on ArcGIS 10.5 is used to join the road intersections and crashes. After the “spatial join”, a new list (i.e., the join count) is added to the data list. Join count represents the accurate crash number at each road intersection, and it is counted and visualized by different colors (see
Figure 2).
The natural breakpoint classification method provided by ArcGIS 10.5 is used to classify the crashes. This classification method divides the data into several categories and the sum of the variance of each classification method is calculated, then the variance and the smallest result are selected as the optimal classification result. The advantage of this classification method is that it can find the maximum difference in data distribution status and divide it into multiple categories. Finally, the crashes are classified into five categories by the “natural breaks method” (i.e., level 1 is from 1 to 21, which denotes the crash quantity of one year belongs to this range; similarly, level 2 is from 22 to 36, level 3 is from 37 to 66, level 4 is from 67 to 104, and level 5 is from 105 to 162), as shown in
Figure 2.
In
Figure 2, the blue areas are the areas with the highest number of crashes (i.e., level 5) and the red areas show the lowest number of crashes (i.e., level 1). What can be seen is that the crashes occurred near Yunlong Road and S230 are less, and the high crash occurrence areas are distributed along S227. For example, Wujiang’s crash statistics showed that 162 traffic crashes at the intersection of S227-Yunli Road occurred, 116 crashes at the intersection of S227-Lize Road, and 108 crashes at the intersection of S227-Gaoxin Road for the year of 2016. The traffic crashes along S227 were severe, and this may be attributed to several aspects. First, S227 is a main road of Wujiang directly connected to Suzhou city, and the annual average daily traffic (AADT) in 2016 was 38105veh/day, the traffic volume was rather high, and the traffic congestion was serious at morning and evening peak hours. Under this condition, the vehicles present a “stop and go” phenomenon, which may easily lead to scraping and collisions between motor vehicles. Then, a widening and renovation project was performed on S227 in 2016, which has greatly affected the road traffic capacity. Especially at the road intersections, because part of the lanes was occupied by the construction project, the vehicles have to slow down and change lanes at upstream of the intersections, this can lead to the horizontal and vertical conflicts between the motor vehicles and motor vehicles, motor vehicles and non-motor vehicles. As a result, these conflicts were prone to traffic crashes. Finally, the large truck proportion is high on S227, and these trucks can block the view of the cars. In that case, the car drivers cannot see the signal or traffic signs clearly at road intersections, resulting in crashes.
The above analyses illustrate the high occurrence of crashes at the road intersections that distributed along S227, but it does not mean that these locations are the crash hot spots.
4.2. Time-Space Cube Analysis
Spatial joint analysis shows the spatial distributions of the crashes at road intersections of Wujiang, but it is unable to show the temporal characteristics of the crashes. Thus, the time-space cube model based on ArcGIS 10.5 is employed to analyze the temporal characteristics of the crashes. Space-time cube analysis is a 3D geo-visualization technique that maps spatiotemporal data in a cube and is useful for finding spatiotemporal patterns. In the time-space cube,
x-axis and
y-axis denotes the two-dimensional spatial range, respectively, and
z-axis represents time. Through spatial positioning, a bin time series will be obtained, and the bin time sequence will be a longitudinal column consisting of a unit cube (column). The number of data points contained in each column is the number of geographic events that occur within the unit time step. Thus, the bin time series can visually show the change of geographic events over time. The time-space cube structure [
28] is shown in
Figure 3a,b.
In this study, the space-time cube model is created to display the spatiotemporal distributions of traffic crashes of Wujiang’s road intersections. The time step is one month, so z axis bin includes twelve time steps, then the bin time series of crashes from bottom to top for the year 2016 is “January 1 to 30”, “January 31 to February 29”, “March 1 to 30”, “March 31 to April 29”, “April 30 to May 29”, “May 30 to June 28”, “June 29 to July 28”, “July 29 to August 27”, “August 28 to September 26”, “September 27 to October 26”, “October 27 to November 25”, and “November 26 to December 25”, respectively. These time steps can denote the 12 months of 2016. Then, the natural breakpoint classifies the crash frequency as four categories: From 0 to 3 times/month, from 3 to 9 times/month, from 9 to 20 times/month, and from 20 to 39 times/month. These four types of crash frequency are labeled by different colors (i.e., red, yellow, cyan and blue, see
Figure 4), and blue and red represents the high and low crash locations respectively. Finally, the time-space cube visualization result is shown in
Figure 4.
In addition, based on the “identify” toolbox of time-space cube, the geographic feature can be identified by clicking the bins of the cube. As a result, the high crash occurrence locations from spatiotemporal aspect can be identified by this toolbox. For example, we click on a bin at random, and then the crash location and time series can be displayed.
Figure 5 shows spatiotemporal distribution of the road intersection “Liuhong Road-Jinye Road”, and it indicates that there occurred 22 traffic crashes at location ID 35 (i.e., the road intersection of Liuhong Road-Jinye Road) from 1 to 30 January 2016, and this road intersection corresponds to the high crash rate area (i.e., the fourth crash category, see
Figure 4). The crash spatiotemporal situation of other locations can also be identified by this way.
Notably, the high crash rate not necessarily the crash hot spot. The following sections will excavate the crash hot spot by analyzing the spatiotemporal features and trends of the high crash rate locations.
4.3. Spatial Autocorrelation Analysis
The time-space cube analysis qualitatively characterizes the crash situation of each location (i.e., each intersection) from spatiotemporal aspect, but failed to characterize the intrinsic spatial pattern of the crash. In other words, the spatial change features of these crashes (i.e., the spatial aggregation characteristics and the statistical significance) are not explored. Actually, the locations with high crash rate may not necessarily be the hot spot. As the spatiotemporal crash hot spot requests not only have the high crash rate of the current location, but also the high crash rate of its proximity locations. That is to say, the crash point has a significant agglomeration trend in the spatial range. Based on this analysis, the quantitative method (i.e., the spatial autocorrelation analysis) to analyze the spatial features of the crashes is proposed.
Spatial autocorrelation analysis is the action of determining consistent groups of objects according to the amounts of their attributes [
29]. Two types of spatial autocorrelation analysis methods are generally known: Global spatial autocorrelation analysis and local spatial autocorrelation analysis [
30]. ArcGIS 10.5 can provide the two kinds of spatial autocorrelation analysis toolbox, and with these toolboxes, the spatial autocorrelation analysis of Wujiang’s traffic crashes can be realized.
4.3.1. Global Spatial Autocorrelation Analysis
The global autocorrelation analysis method is used to describe the overall distribution of a certain phenomenon and determine whether this phenomenon has spatial agglomeration characteristics as a whole, and this can be conducted using global Moran’s I statistical analysis. The global Moran’s I method is based on the covariance relationship of the statistical correlation coefficient. Moran’s I index can be described as following.
where,
,
denote the
ith and
jth spatial observed value, respectively,
,
wij represents the elements of a spatial binary contiguity matrix and weights are the neighborhood relationships exist between location
i and its adjacent location
j,
S0 refers to the summation of all elements of
wij.
The Moran’s I value must be between -1 and 1, and Moran’s I value greater than 0 denotes a positive correlation at spatial distribution. The value approaches 0, which represents the random distribution at spatial aspect. Then, the value less than 0 indicates that the same area has a large difference in attributes, and the data space distribution shows a state of high and low interval distribution. The Moran’s I index is usually transformed to the z score with a standard normal distribution, and a positive z score means that the adjacent features have similar values, while a negative z score shows that the adjacent feature have dissimilar values [
31]. Correlation between the significant level, confidence, and z score is shown in
Table 1.
Therefore, this study performs the global spatial autocorrelation analysis of traffic crashes in Wujiang’s road intersections, with an expectation to determine the spatial crash characteristics (i.e., to judge if it is the agglomeration distribution, discrete distribution, or random distribution, at spatial aspect) the whole study area. The spatial autocorrelation toolbox provided by ArcGIS 10.5 is performed to guide this analysis. The results are shown as
Figure 6.
Figure 6 shows that Moran’s I value is 1.261786 (it is greater than 0), which suggests that the traffic crashes present a positive correlation at a spatial distribution. P-value is 0.000 (it is lower than 0.01), which denotes the random distribution probability of the traffic crashes is less than 1%. Conversely, the spatial aggregation distribution (i.e., the clustered status in
Figure 6) probability is more than 99%. Furthermore, z score is higher than 2.58, which demonstrates that the spatial aggregation trend of the crashes at global distribution is significant.
4.3.2. Local Spatial Autocorrelation Analysis
Global autocorrelation analysis proves that the traffic crashes at road intersections of the study area are aggregated from the spatial aspect, and this aggregation phenomenon is significant. However, the global autocorrelation analysis does not indicate exactly where the crashes gathered (i.e., the specific aggregated locations). The local spatial autocorrelation analysis method can provide the specific aggregation ranges of the crashes and preliminarily identify the hotspots. Getis-Ord Gi* is a technique for local spatial autocorrelation analysis [
32], and the equation is as follows [
33].
where
is the property value of
jth element,
is the spatial weight of
ith element and
jth element, and
is the total number of element.
Getis-Ord Gi* can be regarded as the index of local spatial autocorrelation analysis and shows that the high or low values are clustered. The value is involved in spatial analysis, which indicates where hot spots (representing high values clustered in a location) or cold spots (representing low attribute values clustered in a location) appear in the area. The “cluster and outlier analysis” toolbox based on ArcGIS 10.5 is utilized to make the analysis. The results are shown as
Figure 7. Four types of points are shown in
Figure 7: HH (i.e., the aggregation between high and high values), HL (i.e., the high values are surrounded by low values), LH (i.e., the low values are surrounded by high values), and LL (i.e., the aggregation between low and low values). HH (i.e., the pale pink color) may represent the hot spot, while LL (i.e., light gray points) seems to denote the cold spots.
Notably, z score is the return value of each Gi* statistics. For positive z scores with significant statistical significance (confidence level= 99%, 95%, and 90%), the higher the z score, the closer the high value (hot spot) clustering. Likewise, for negative z scores, the lower the z score, the closer the low value (cold spot) clustering. Grading the degree of clustering of hot or cold spots, as shown in
Table 2.
The initial hot spot identification results are shown in
Figure 8. From
Figure 8, the red locations are crash significantly increased locations (i.e., the hot spot with 99% confidence level), and the blue area are crash significantly decreased locations (i.e., the cold spot with 99% confidence level). Specifically, the crash locations along S227 (i.e., the road intersections of S227-Gaoxin Road, S227-Youche Road, etc.), Yongkang Road (i.e., the road intersection of S227-Yongkang Road), Youche Road (i.e., the road intersection of Youche Road-Luxiang Road), and Gongnong Road (i.e., the road intersection of Zhongshan Road-Gongnong Road) are the crash hot spots, and the crashes in these road intersections are significantly increasing. The crash locations along Tiyu Road, Changban Road, and Lianyang Road are cold spots, and the trends are significantly decreasing.
4.4. Emerging Hot Spot Analysis
The above local spatial autocorrelation analysis has preliminarily identified the crash hot spots. While the emerging hot spot analysis can make a more detailed interpretation and classification of these accident hot spots, they also present the changes of different kinds of hotspots. The emerging spatiotemporal hotspot analysis toolbox provided by ArcGIS 10.5 can identify the trends of crash density in the time-space cube, and classified these hot and cold spots according to
Table 3. Accordingly, this study analyzes the spatiotemporal hotspot of the time-space cube (the time step is one month), and excavates the characteristics of cold hot and hot spots over time. In the process of spatiotemporal hot and cold spots analysis, ArcGIS’s default hotspot is used to analyze the neighborhood distance, and monthly bins were generated for time units. Then, the visualization results of crash hot spots over time for the year of 2016 (from January to December) is shown in
Figure 9. The warm color represents the hot spot. The hot spots illustrate that the crash frequency is high at the current locations, and also with a significant agglomeration trend from the spatial perspective. The cold spots are characterized by cold color, and the cold spots indicate the crash frequency is low at the current locations, and the low situation is also gathering on a spatial perspective.
Three types of hot spots (i.e., consecutive hot spot, intensifying hot spot, sporadic hot spot) are distributed at different locations of the study area for the year of 2016 (see
Figure 9). Compared to the persistent hot spots, sporadic hot spots, oscillating hot spots, and historical hot spots, the spatiotemporal change trends of those new hot spots, consecutive hot spots, intensifying hot spots, and diminishing hot spots may deserve more attention, especially the new hot spot and intensifying hot spot (see
Table 3).
From
Figure 9a, the intensifying hot spot locations in January are location ID 26, 27, 28, 37, 38, 48, and the corresponding road intersections are Youche Road-Luxiang Road, Zhongshan Road-Gongnong Road, S227-Gongnong Road, Liuhong Road-Gongyuan Road, S227-Yongkang Road, and S227-Chuihong Road, respectively, which declares that the traffic crashes at these intersections show a significantly increasing trend. As time goes by, the intensifying hot spot locations also increased. For example, on the basis of January, location ID 36 (i.e., road intersection of Liuhong Road-Luxiang Road), location ID 58 (i.e., road intersection of Huayuan Road-Lize Road), location ID 59 (i.e., road intersection of S227-Gaoxin Road) and location ID 69 (i.e., road intersection of Changan Road-Gaoxin Road) are added as the intensifying hot spots one after another from February to March. In April, location ID 68 (i.e., intersection of Huayuan Road-Tiyu Road) also become the intensifying hot spot, and then after it the intensifying hot spot locations keep stable (i.e., intensifying hot spots contain location ID 26, 27, 28, 36, 37, 38, 48, 58, 59, 68, and 69 from April to December). Thus, all these intensifying hot spots are not only statistically significant hot spots within each time step (i.e., one month), but also have a significant increase trend in spatial agglomeration. Therefore, these road intersections such as Youche Road-Luxiang Road, Zhongshan Road-Gongnong Road, S227-Gongnong Road, Liuhong Road-Luxiang Road, Liuhong Road-Gongyuan Road, S227-Yongkang Road, S227-Chuihong Road, Huayuan Road-Lize Road, S227-Gaoxin Road, Huayuan Road-Tiyu Road, and Changan Road-Gaoxin Road should be paid more attentions when taking preventive measures against traffic crashes.
In addition, the consecutive hot spot location in January only contains location ID 17 (i.e., the road intersection of Jiangxing Road-Xitanghe Road). As time goes, the consecutive hot spot locations increase from February to May, and these added locations include location ID 16, 46, 57, 67, 76, 77, and 78, the corresponding road intersections are Jiangxing Road-Luxiang Road, Lize Road-Wumo Road, Zhongshan Road-Tonghong Road, Zhongshan Road-Tiyu Road, Kaiping Road-Luxiang Road, Kaiping Road-Zhongshan Road, and Huayuan Road-Kaiping Road. Then from June to August, location ID 79 (i.e., the intersection of Changan Road-Easttaihu Street) is added to the consecutive hot spot locations. Afterwards, location ID 24 (i.e., the road intersection of Suzhouhe Road-Youche Road) has become consecutive hot spots from September to December, and the consecutive hot spots distribution trend is stable at this period. The above analyses illustrate the traffic crash followed a seasonal change, there are weak hot spots in the coldest months of January and February, then the crash hot spots are gradually increasing from summer to autumn, which aligns with the study in Reference [
22].
With respect to traffic crash cold spot, four types of traffic crash cold spots (i.e., sporadic cold spot, persistent cold spot, consecutive cold spot and intensifying cold spot) are identified in the study area. The cold spots in the first half of 2016 are mainly the sporadic cold spots (i.e., from January to March) and persistent cold spots (i.e., from April to June). The sporadic cold spots are not the research focus, we only analyze the persistent cold spots (i.e., the locations include location ID 62, 63, 71, 72, 80, 81, 82, 91, 93, 104, 97, 98, 106, 107, 108, 116, and 117 from April to June). These persistent cold spots represent that the traffic crash rate is low at these locations, and spatial agglomeration of them present no obviously increasing or decreasing trend. As times go, most of the above persistent cold spots became the intensifying cold spots from July to December (see
Figure 9g–l), and this indicates the low crash agglomeration trend is increasing, so the traffic crashes at these locations are less and less from July to December. The above analysis demonstrates the traffic crash rate is low at location ID 62, 63, 71, 72, 80, 81, 82, 91, 93, 104, 97, 98, 106, 107, 108, 116, and 117 from April to December, especially in the second half of year 2016, the traffic safety condition at the above locations is obviously getting better.
Overall, the traffic crash hot spots are mainly distributed in the northeast of Wujiang’s major urban area, while the crash cold spots are concentrated in the southwest of Wujiang. This finding is consistent with the existing reference in [
27], which indicates that the road intersection crash in the Northeast part of Wujiang is more serious and should be prevented and administered in priority.