1. Introduction
One of the main causes of fatalities among people and property damage worldwide is traffic crashes [
1,
2]. Global Status Report on Road Safety 2015 gathered data from 180 nations, which claims that, every year, 1.25 million people are killed in traffic crashes [
3]. Every day, 3000 people die due to traffic crashes around the world [
4]. In another study related to Iran, the number of people who are injured in traffic crashes is reported to be about ten times more than the number of deceased people, at around 240,000 cases per year. These high rates, combined with the significant share of passenger traffic, necessitate a thorough investigation into passenger safety [
5]. For an intelligent transportation system to be deployed successfully and provide appropriate levels of medical care and transportation in a timely way, in the case of a traffic incident, an accurate and quick severity prediction algorithm is crucial. Different agencies may profit from the ability to estimate the severity of a reported incident with unknown severity or the severity of future crashes. Crash severity prediction is one of the most significant aspects of crash management; it helps rescuers assess the severity of traffic crashes and their potential impact and perform effective crash management methods. Injury severity is frequently regarded as a dependent (class/response) variable when studying traffic crash injury severity using historical crash records, whereas other crash components are referred to as independent (predictor) variables. Researchers use crash data to identify risk variables and execute effective interventions to improve road safety. Driver characteristics, roadway characteristics, vehicle characteristics, crash characteristics, and atmospheric elements are all factors that influence the severity and frequency of crashes [
6,
7]. Crash data are compiled into a massive database that includes various crash-related characteristics. In order to examine the crash database, many analytical methods have been applied in the literature.
Several studies have attempted to forecast the severity of crashes in recent years [
8]. Many studies have attempted to analyze and identify the primary factors that influence the severity of road crashes [
9]. The strategies most typically employed to undertake such analyses have been linear models, nonparametric models, and data mining approaches. These methods aid in determining the cause of a traffic crash. Statistical methods have been frequently employed to forecast the severity of traffic crashes. The logistic regression model, the ordered probit model, and the mixed logit model have all been proposed by different researchers and have been used to evaluate relevant data from traffic crashes and to examine the impact of various variables on the severity of traffic crashes [
2,
10,
11,
12,
13,
14,
15,
16]. These prediction models aim to predict the severity of traffic crashes. However, according to some researchers [
7,
17], the majority of regression models have established underlying correlations between dependent and independent variables and their own set of underlying assumptions (i.e., linear relations between the variables). If these assumptions are not met, the model may produce inaccurate estimates of the chance of serious injury. According to the reported literature, machine learning outperforms traditional statistical approaches in prediction, aided by the accessibility to a vast quantity of datasets [
18]. Moreover, although crashes are random events that happen in space and time [
19], they also exhibit spatial dependency and spatial autocorrelation, which should be considered during their evaluation [
20]. Hotspot techniques should be used to manage road safety initiatives that envision minimizing traffic crashes with limited cost, as proven by [
21]. Similarly, Australia’s national hotspot program helped reduce fatal crashes by 30% [
22]. In Belgium and Denmark, hotspot applications showed similar performance [
23].
The present study is focused on three objectives. Firstly, developing an accurate model for predicting crash severity for the study area. Such a model will be helpful for medical and emergency services to provide rescue services in a timely manner and analyze crashes with unknown severity. Secondly, identifying parameters that affect and increase the severity of crashes, and thirdly, an approach that seeks to identify crash hotspots by emphasizing places where crashes are most likely to occur under specific conditions has also been determined in this study. For the prediction of road crash severity, tree-based ensemble models (random forest and gradient boosting) and a logistic regression model are examined in this work. In addition to taking into account machine learning approaches, this study also makes an effort to analyze the spatial consequences of crashes in an effort to point out and identify potential hotspots utilizing various variables. As a result, machine learning models can identify hotspots rather than just depending on the number and location of crashes by learning data patterns in variables. The current study is important because it shows the value of using statistical methods and spatial autocorrelation to identify crash-intensive-prone zones and successfully applies these methods using data from crashes on rural and urban roads in the Saudi Arabian province of Al-Ahsa, over a period from 2016 to 2018. This effort will be beneficial for traffic management authorities to devise effective policies for reducing traffic crashes and mitigating their impacts.
The rest of this paper Is structured as follows. A summary of the most recent research in this field is given in
Section 2. Data source and data processing are explained along with descriptive statistics in
Section 3. Machine learning models taken into account, in this paper, are described in
Section 4.
Section 5 presents exploratory data analysis, while
Section 6 discusses the study’s findings.
Section 7 concludes with a brief summary of the findings and suggestions for future research.
2. Literature Review
Traditional methods for predicting collisions and classifying their seriousness use statistical modeling [
24]. Poisson, binomial, Poisson–lognormal, negative binomial, gamma, and zero-inflated regressions, negative multinomial models, generalized estimation equations, random effects models, and random parameters models have all been applied in this situation [
25]. The Bayesian hierarchical binomial logit, Bayesian-ordered probit, log-linear model, extended-ordered logit, multinomial logit, multivariate probit, ordered probit, and ordered logit are among the models that have also been utilized [
26] for the estimation of crash severity. Despite the advances made with these techniques, statistical modeling has limitations because each model includes its predetermined correlations between the dependent and independent variables a and assumptions [
27]. Statistical modeling, according to [
28,
29], necessitates assumptions about data distribution. Such premises could be false and, as a result, violated. On the other hand, the drawbacks of this strategy have been extensively investigated, presenting an opportunity to employ fresh approaches. Machine learning methods, such as decision trees (D.T.s), nearest neighbor classification (K.N.N.), support vector machines (SVMs), evolutionary algorithms (E.A.s), deep learning models, and artificial neural networks (ANNs), have been employed to analyze data for a variety of road safety issues and devised as data analytic methods due to their capacity to handle enormous volumes of multidimensional data and overcome the shortcomings of statistical methodologies. Furthermore, the flexibility of machine learning modeling techniques, generalization and learning capabilities, and great predictive capacity led to its acceptance as an accurate, generic, and practical mathematical model in road safety.
ANN is a useful tool for solving problems in a variety of fields. It can be utilized for traffic incident detection [
30], public transportation [
27,
31], and road planning [
32]. ANN is a complex, non-linear, parallel processor with a natural tendency to store and retrieve experimental knowledge [
33]. The possible non-linear relationships between injury severity levels and crash causes were modeled by Delen et al. [
34] using a number of ANN models. Bayesian neural network (B.N.N.) and ANN models have been used to research road safety concerns for many years. Although the multilevel network architecture of ANN and B.N.N. models are similar, they differ in their capacity to predict outcome variables [
35]. Moghaddam et al. [
36] employed ANNs to predict and estimate crash severity in urban roadways, as well as to uncover major crash-related parameters. Variables, such as roadway width and head-on collisions, were found to have an impact on the outcomes.
The effectiveness of B.N.N. models for forecasting road crashes was evaluated by Xie, Lord, and Zhang [
37] using data gathered on rural areas in Texas. Ona et al. [
17] used a Bayesian network to identify the elements that influence injury severity, which was divided into two categories: slightly hurt and killed/severely injured. Simoncic [
38] created a Bayesian network for injury severity analysis. The findings revealed that Bayesian networks could be used to model road crashes. It also highlighted some benefits of using a Bayesian network over a regression model, such as the ability to include more variables and a larger data set.
The statistical learning theory underpins the support vector machine (SVM) approach [
39]. SVM models [
40,
41], a novel class of models based on structural risk reduction and statistical learning theory [
42], have lately been developed for road safety studies. Furthermore, the findings revealed that SVM models outperform B.P.N.N. models (or are at least equal) and do not overfit the data. Nearest neighbor classification (K.N.N.) is a basic yet groundbreaking machine learning algorithm. In a prediction problem, the K.N.N. classifies an observation based on the k-nearest observations [
43].
For classification problems, the D.T. technique is highly beneficial. A training set of inputs and outputs (i.e., classes) is produced during the development of a tree. D.T.s are data mining techniques [
44,
45], which can be utilized to address classification difficulties. One of the key advantages of D.T.s is that their structure allows for the extraction of ‘if-then’ decision rules (D.R.s). Such DRs may reveal behaviors that are unique to a particular data collection. They are not only practical but also simple to comprehend from the standpoint of safety analysis [
46]. Evolutionary algorithms (E.A.s) are based on natural selection processes, in which only the strongest individuals survive [
47], and use stochastic search strategies. The two fundamental types of evolutionary algorithms are genetic algorithms and genetic programming, and they are particularly useful for optimizing problems typically related to other approaches.
Using vast data gathered on unsignalized crossings in Florida, multivariate adaptive regression splines (MARSs), a recently created machine learning technique, was used by Abdel-Aty and Haleem [
48] to predict vehicle angle collisions. The data showed that MARS performed better than the N.B. models. After using R.F. to screen the covariates, the suggested MARS models produced promising results. According to the data, MARS appears to be an effective method for forecasting collisions at unsignalized intersections.
Deep learning is a novel approach that emerged from the advancement of machine learning. It examines many layers of non-linear information, both supervised and unsupervised, to interpret or categorize patterns [
49]. A kind of machine learning called “deep learning” creates models that can extract traits from the most fundamental to the most apparent layers [
50]. This capability eliminates the difficulty of raw data processing, which is a typical issue with many machine learning algorithms. Deep neural networks (D.N.N.s), convolutional neural networks (CNNs), and recurrent neural networks (R.N.N.s) are the most common examples of this approach, which are used in speech recognition, visual object recognition, and object detection [
44]. In contrast to earlier machine learning techniques, which were constrained in their ability to read natural data in its raw form, deep learning develops computer models to extract inherent qualities from data at various levels. Deep learning applications in the field of transportation are limited, with only one study focusing on traffic flow prediction [
51], despite the fact that deep learning methods have proven exceptional performance in a number of applications [
52].
Moghaddam et al. [
36] employed ANNs to predict and estimate crash severity in urban roadways, as well as to uncover major crash-related parameters. The results showed that the most important factors that increase crash severity on urban highways include variables such as highway width, head-on collisions, the type of vehicle at fault, disregarding lateral clearance, keeping track of distance, inability to control the vehicle, exceeding the permitted speed, and driver deviation to the left. Sharma et al. [
51] used SVM and multi-layer perception for predicting crash severity. They only used a small number of data samples to test their hypothesis. They looked at only two variables: speed and alcohol as major factors in car crashes. The SVM outperformed with a 94% accuracy rate. They stated that high-speed driving while inebriated was the cause of the crash. Tiwari et al. [
53] used decision tree (D.T.), N.B., and SVM, as well as S.O.M. and k-modes, for clustering. They obtained superior results with the clustered dataset.
For traffic agencies, AlMamlook et al. [
54] employed N.B., AdaBoost, R.F., and logistic regression (L.R.) to locate highways with high crash risk. They used A.U.C., R.O.C., recall, precision, and F-measure to evaluate their models. The R.F. outperformed with a 75% accuracy rate. In another study, Beshah and Hill (2010) [
55] investigated crucial roadway-related characteristics that can influence the severity of road crashes. They created decision rules for road safety measures using D.T., NB, and K.N.N. They were primarily concerned with drivers and pedestrians, with no regard for other factors such as time or speed. They did not report the accuracy parameters of their machine learning models in detecting crash severity risks.
A developing nonparametric tree-based model called the boosted regression tree model can capture the non-linear effects of both discrete and continuous variables without preprocessing the data. Using 5-year crash information for provincial highways in Ontario, Canada, Lee and Li (2015) [
56] employed the boosted regression tree model to examine the severity of driver injury in addition to other nonparametric models. The boosted regression tree model’s findings revealed a substantial correlation between the severity of driver injuries and vehicle ejection and head-on crashes. They found that for both single-vehicle and two-vehicle crashes, the boosted regression tree model predicted driver injury severity more accurately than the classification and regression tree model.
In order to determine the variables influencing the severity of pedestrian crashes in the Melbourne metropolitan region, where mid-block collisions accounted for 46% of all pedestrian collisions and mid-block collisions were the scene of 49% of pedestrian fatalities, Toran Pour et al. (2017) [
57] developed three models employing various decision trees (D.T.s). They employed bagging and boosting approaches to enhance the D.T.s’ accuracy, stability, and resilience. According to the study’s findings, the boosting strategy increased each D.T. model’s accuracy by 46%. Additionally, the results of raising D.T.s demonstrated that neighborhood social traits were just as significant in impacting the severity of pedestrian collisions as traffic and infrastructural factors.
A gradient boosting decision tree (G.B.D.T.) model was presented by Wu et al. (2019) [
58] to examine the combined effects of crash-causing elements on four road crash indicators (i.e., injuries, deaths, number of crashes, and the financial loss). The economic, demographic, and road network conditions of Zhongshan, China, from 2000 to 2016, are studied using a total of 27 detailed influential elements. The findings demonstrate that the G.B.D.T. outperforms other conventional machine learning approaches in terms of prediction accuracy, handling multicollinearity across explanatory variables and, more crucially, ranking the influential factors on road crash prediction. The outcomes also demonstrate that there are parallels and variances among the most important determining factors for the crash indicator. Moreover, Wu et al. (2021) [
59] applied a scaled stacking gradient boosting decision tree (SS-GBDT) for predicting bus passenger flow, which can also be used for road crash prediction. To analyze non-intersection accident severity data, Pande and Abdel-Aty (2009) [
60] used association rule (or market basket analysis) learning directly for the first time. The analysis’s findings provide straightforward criteria that show which crash features are related to one another. Using non-intersection crash data from the state of Florida for 2004, the application is illustrated. Later, De Ona et al. (2013a) [
9] used association rules to draw out practical decision rules from the output of D.T. modeling. The percentage of samples that fit the rule among those that solely fit the left side, confidence, and lift was used to establish the importance of each generated rule. Support measures how often the rule occurs in the dataset (the statistical dependence of the rule). Other researchers (Montella, Aria; Montella, De O’a, Mauriello, Riccardi, and Silvestro, 2020) [
61] extracted decision rules in the form of IF–THEN rules from CART results, where “IF” refers to the statuses of various independent variables and “THEN” is the status of the corresponding crash severity class variable. They contrasted the rules produced through association rule learning with the IF–THEN rules. The outcomes demonstrate the consistency of the two sets of regulations. Discovering relationships between independent factors and crash severity outcomes can be conducted by applying association rule learning, which is a straightforward process. The influence of independent factors on the seriousness of crashes in each rule cannot be measured, and its forecast accuracy is restricted in comparison to other ML techniques.
The breadth of traffic crash severity prediction is the subject of this research. The goal of traffic crash data mining research can be separated into two categories: (1) predicting the severity of traffic crashes and (2) identifying relevant elements determining the severity of crashes. The severity of road crashes is reliably predicted using ensemble learning algorithms in this article. Gradient boosting machine and RF are two tree-based ensemble learning models employed in this investigation. Additionally, this study not only considers machine learning methods but also makes an effort to assess the spatial effects of crashes in order to highlight and pinpoint possible hotspots using various variables. Thus, the research contributes to the field of road safety by offering a model that would identify the important features for estimating, preventing, and forecasting crashes, particularly in the study region.
3. Data Collection and Data Processing
Al-Ahsa, the largest oasis in Saudi Arabia and the largest oil field in the world, is the largest governance of the eastern province of Saudi Arabia that stretches from Kuwait at 29°20′ N. to the southern tip of the Gulf of Bahrain at 25°10′ N (
Figure 1). The case study is conducted in the Al-Ahsa region. This city was chosen for this study due to the city’s high crash rate, which has been documented in recent literature [
62]. Al-Ahsa recorded 31.9 percent of crashes in the Eastern Province on average between 2009 and 2016, the highest crash rate among other cities in the eastern region (
Figure 2). Other cities reported fewer than 5% crashes. This unmistakably shows that cities in Al-Ahsa are significantly vulnerable to traffic crashes. It is noteworthy that UNESCO has recently listed the Al-Ahsa as a heritage site in Saudi Arabia. Thus, the city was recently recognized as a UNESCO-listed heritage site in Saudi Arabia, giving this city tremendous potential to become a popular international tourist destination. However, the high crash rate in this city may create adverse effects on such potentiality. Thus, it requires a proper response.
The crash data for Al-Ahsa was collected from the Traffic Control Authority and used in this study from October 2014 to May 2018. A total of 3994 crashes were reported in this city during that time. Vehicle collisions were the most frequent type of mishap, accounting for around 8 and 30%, respectively, of incidents that resulted in fatalities and injuries, according to an overall review of the data set (
Figure 3). Another common crash type that resulted in 6.5 percent of death and 12.5 percent of injury incidents was vehicle overturning. Additionally, 12.5 percent of injury crashes and 2 percent of fatal crashes reflect the vulnerability of pedestrians. The probability of injury crashes is demonstrated by a number of incident types, such as collisions with parked cars, road railings, motorcyclists, and stationary objects, where fatal crash rates were very low.
Data preprocessing and cleaning is an essential preliminary step for crash analysis using machine learning algorithms. The raw data had 8 different features with 8027 data points. The data were filtered in several steps. First, all the dates had to be converted from Hijri to Gregorian, and then three different columns were made for day, month, and year. Secondly, one of the features from the feature containing similar information was considered. One of the downsides of the data was that it had only two classes of injuries, namely, fatal and serious injuries. Because there was an imbalance in the target class, under-sampling technique was applied for the logistic regression model. The other two classification models did not require treatment for data imbalance. The feature’s number of deaths and injuries had some missing values. These data could not be imputed as they might have produced misleading inferences with the model accuracies. Missing data points were discarded from the final data set used for modeling.
Descriptive Statistics
A total of 4093 crashes from 2015 to 2018 made up the dataset used for the analysis. A total of 9031 people were hurt in the collisions. Nineteen percent were fatal crashes, and the remaining was severe injuries. The three leading causes of collisions—sudden turning, speeding, and failure to yield—accounted for roughly 47 percent, 18 percent, and 15 percent of collisions, respectively. Within the available dataset, the pattern of crashes over time demonstrates a decline in serious injury and fatal crashes, as illustrated in
Figure 4. In their study, which similarly focused on Saudi Arabia, the authors of [
64] ascribed the decline in serious crashes due to the installation of roadside cameras.
Fatality and serious injury R.T.C.s involving moving vehicles made up about 38 percent of the total number of crashes in the dataset that was made accessible. Numerous vehicle crashes (21 percent of R.T.C.s) and running over the road were two more common collisions (14 percent R.T.C.s). The authors of [
62] cited the first two crash categories as being typical of those reported in the Eastern Province. This Saudi Arabian governorate includes the study region from which the data for this study was gathered.
The available dataset’s descriptive statistics for the number of people who suffered fatalities and significant injuries in R.T.C.s are shown in
Table 1. Because practically all R.T.C.s involve, on average, at least one fatality victim, the severity index is relatively high. Approximately 1.15 million people reside in the Al-Ahsa region [
66]. Comparing
Table 1’s numbers to those from other research in the literature, they are high. According to the authors of [
67], there were just 3.3 fatal crashes per year in Iowa city of U.S.A., as opposed to 352 in Al-Ahsa, in the available data that were made accessible.
Another study [
68] found that the study area had 311 deaths per 100,000 persons, compared to 14.93 deaths per 100,000 in Columbia. It is interesting to note that the crash rates per year for the research area were not found to be more significant when compared to earlier studies. The yearly fatal crash rate in a study conducted by the authors of [
67] was roughly 350, which is close to the number for the research area. Although the authors of [
69] did not mention serious injury and fatal crash rates per year, the following studies further supported their argument. For fatal and significant injury R.T.C.s, Abdi et al. [
70] recorded an annual rate of 1463 crashes for Addis Ababa, while [
69] recorded an annual rate of 3631 for Torino (Italy). On the other side, there were 1333 fatal and serious injury R.T.C.s annually in the research area.
This illustrates how R.T.C.s are greater in severity and involve more victims in the research area, although having lower occurrence rates than in other regions of the world. Further details about the crash dataset can be found in [
63,
65].
6. Results and Discussion
A total of 8027 automobile collisions occurred throughout the study period, of which 21.7% resulted in fatalities and 78.3% in serious injuries. To create models that predict the degree of injury severity in collisions, eight predictor variables (attributes) were combined with the class variable of injury severity. The work also seeks to assess the importance of qualities (relative variable importance) in causing a crash.
The confusion matrix produced by each classification approach on the test data is summarized in
Table 5. The confusion matrix for each class displays how instances belonging to that class are classified. The diagonal of the contingency table contains all of the correctly classified samples. Consequently, it is possible to visually check the matrix for errors.
The confusion matrix was used to determine the various values of the classifiers’ accuracy measure, which were then listed in
Table 4 and
Table 5 to demonstrate how well the classification algorithms performed in relation to the crash in Al-Ahsa. The prediction accuracies for each class for all three classification algorithms are listed in
Table 6. It expressly displays the sensitivity, specificity, accuracy, and recall results for all three machine learning approaches that were acquired using 10-fold cross-validation. The classifiers’ performance metrics are provided in
Table 6. The models were developed in a Python environment with Scikit-learn library. Random forest classifier achieves an accuracy of 94.00%, with a precision of 0.95 and 0.94 for fatal and serious injuries, respectively. The accuracy is 93.00%, with a precision of 0.91 and 0.93 for fatal and serious injuries when XGBoost was used, whereas using the logistic regression, the accuracy is 79.00% with a precision of 0.59 and 0.79 for fatal and serious injuries, respectively.
Comparing the accuracy parameters of these models reveals that the R.F. model had the best prediction accuracy in terms of recall, precision, and F1 score. The random forest model outperformed the other models by a small margin (F1 score—R.F. model: fatal 0.85 and serious injury 0.96; XGBoost: fatal 0.81 and serious injury 0.85). Technically, it can be inferred that the important feature of R.F. was better than the XGBoost; however, the difference in the feature list was not very significant. With slightly superior performance metrics than logistic regression, which came in last in the performance metrics ranking, XGBoost was the next-best classification algorithm. This clearly advocates the use of machine learning techniques in place of statistical techniques for predicting crash severity.
According to the gain ratio evaluator derived using random forest and XGBoost models,
Figure 6 and
Figure 7 depict the relative priority ranking of each feature. Higher feature relevance is implied by an evaluator with a high value. The R.F. model suggests that explanatory variables that heavily contribute to the crash injury severity outcome are: faulty tires, hitting a moving vehicle, not giving way, running over, and sudden turning. Whereas the XGBoost model also produced similar important features with faulty tires, not giving way, failing to stop at a stop sign, and sudden turning among the top-ranked features responsible for the severe crash injury. In case of both machine learning models, age had relatively lower impact on the prediction of crash severity. However, with slight differences in the importance of other parameters. Because R.F. model showed better performance, in terms of predicting severity, hence, it would be reasonable to assume that the feature importance shown by this model is more reliable.
Finally, the classification model results for crash injury severity in Al-Ahsa were mostly in line with other regions’ research on crash injury severity. Regarding the causes of the crash, our study from the same area indicated that the type of collision, the state of the road, the illumination, and speeding have the greatest effects [
79].
Spatial Autocorrelation and Hotspot Analysis
The spatial distribution patterns of traffic accidents in the Al-Ahsa Region were identified using spatial autocorrelation (Global Moran’s I). If the characteristics are spatially clustered, dispersed, or randomly distributed, it is shown by the values of Moran’s I [
84]. Equation (14) was used to calculate the Global Moran’s I, z-score, and
p-value using the spatial autocorrelation
where
Xi is an attribute value of the target feature at location
i,
N is the total number of features,
Wi,j is the spatial weight between features at locations
i and
j, and the neighboring feature at locations
i and
j has an attribute value of
Xj [
84]. The results of Moran’s I statistic can be verified using z scores, where a confidence level is established. The temperature of cell I is similar to the crash characteristic of the nearby cells if cell I has a significant positive value (i.e., a positive number). I denote a substantial clustering range if it is a large positive value. The cell surface temperature of cell I is significantly different from the surrounding cells if the value of I is negative and significant, on the other hand, which denotes a negative spatial correlation. We can locate hotspots based on the dispersion of these sites. It assesses if the pattern expressed is random, clustered, or scattered using an associated attribute and the offered collection of features.
Figure 8 displays the spatial autocorrelation graph using Moran’s I statistics of the z-score and
p-value for all traffic accidents between 2016 and 2018 in the Al-Ahsa region.
Figure 8 shows that the
p-value was less than 0.001, the z-score for all crashes was 6.169, and Moran’s I was 0.067. The positive Moran’s I, high z-scores, and tiny
p-values indicate that traffic crashes were spatially clustered, and there is less than a 1% chance that this clustered pattern arose by chance. As a result, this exhibited a statistically significant status, namely, densely clustered.
The Getis-Ord function is widely used in this work to examine hotspots for automobile crashes. A statistic called Getis-Ord points out high data point density where statistically significant point clusters are located in the neighborhood of a given point.
A hotspot can be identified if other features with high values are everywhere around a feature with a high-value density [
83,
84]. It is possible that features with a high-value density are not statistically significant hotspots. Equation (15) is used to construct Getis-Ord
, and
Figure 9 displays a map of the spatial distribution of crash hotspots between 2016 and 2018. The road segments with statistically significant traffic crash hotspots are identified on this map.
where
wi,j is the spatial weight matrix between objects
i and
j,
xj is the attribute value for item
j, and
n is the number of features.
A hotspot is indicated by a positive z-score, whereas a negative z-score shows a cold area. Z-score values that are more positive indicate a more intensive high-value clustering (i.e., stronger hotspots). Conversely, a lower z-score denotes a greater concentration of low values (i.e., cold spots). Accordingly, points near the center of the city present stronger hotspots, while those away from the center and on small roads show cold spots
7. Conclusions and Future Study
This study aimed to develop and compare models for predicting the severity of road crashes in Al-Ahsa city of Saudi Arabia, based on the crash data from 2016 to 2018. Two machine learning techniques, namely, R.F. and XGBoost, and the logistic regression model, as a statistical technique, were used. Among the developed models, R.F. had the best performance in terms of accuracy, precision, recall, and F1 score. The overall model accuracy in the severity predictions of RF, XGBoost, and logistic regression had accuracies of 94%, 93%, and 79%, respectively. The feature importance score from R.F. model suggests that faulty tires, not giving way, sudden turning, and running over were among the most important causes of severe crashes.
Over the past 20 years, methodologies for identifying crash hotspots have advanced, and they now play a critical role in implementing successful traffic safety management programs. In this study, the tendency of a crash pattern to cluster in space was determined using global autocorrelation analysis with Moran’s I. The results demonstrate that high spatial dependence across the research region and the spatial patterns were clustered with a distance threshold of 500 m. The precise aggregated locations of traffic crashes across the research area were then determined using the local spatial autocorrelation Getis-Ord statistic. The Getis-Ord statistic successfully identified high-value clusters with a specified distance threshold. A region was not regarded as a hotspot if a high-value point was not surrounded by other high-value points. As a result, segments with high-value features could be predicted and given priority. The study will advise decision makers on the best places to invest or put safety measures in place.
This study had some limitations with implications for future research and can be extended in the future from various perspectives. First, for the crash severity, there were only two available classes (fatal and serious injury); P.D.O. was absent due to data scarcity, which otherwise consists of the biggest share of the crash severity class. Moreover, the analysis period used is from 2016 to 2018, which can be considered a short period for this study. This is because of the unavailability of an extended data set in the study area. Future studies can utilize detailed and comprehensive datasets containing information on other explanatory variables, such as road geometric features, weather conditions, driving behaviour, road conditions, etc. One of the ML models’ criticisms is their inability to interpret the model. Machine learning interpretation techniques such as SHapley Additive exPlanations (S.H.A.P.) may be utilized to better visualize the feature sensitivity on the model output for a better model interpretation. Moreover, it is recommended for future studies to apply a larger dataset, including different cities of Saudi Arabia, along with more severe classes. In addition, we also recommend performing further analysis to investigate the temporal trends of crashes. In this regard, the use of time series analysis is recommended. Provided more data are available, deep learning techniques are also promising in crash severity studies.