Abstract
The execution of public roadway maintenance, rehabilitation, and restoration activities disturb normal traffic flows, resulting in roadway capacity reduction, inducing travel time delays, and promoting traffic safety concerns. While they improve public roadway performance once complete, the impacts endured in executing these actions is significant. This work seeks a deeper understanding of the effects of improvement actions on traffic by juxtaposing their effects against those arising from traffic incidents that cause similar capacity reductions and related negative externalities. This is accomplished through direct and reverse comparisons with traffic incident impacts. A measure of unit delay that uses observations to determine event location extent, duration, and propagation direction was computed at both facility and corridor-wide levels to establish the degree to which improvement actions and traffic incidents are similar or dissimilar. Alternative hybrid machine-learning methods are proposed to identify and contrast those traffic characteristics that contribute greatest to correct detection of each type of downtime event. These techniques can detect traffic events and accurately distinguish between event types (whether a collision or improvement activity). The techniques were applied on seven months of data obtained from 2019 along three corridors from northern, southern, and western regions of the Commonwealth of Virginia. Those traffic characteristics that contribute greatest to correct event detection of each event type were identified and their similarities and differences were studied. General linear, multivariate regression equations were also developed for more general application.
1. Introduction
Public roadway maintenance, rehabilitation, and restoration activities (together improvement actions), like traffic accidents, disturb normal traffic flows, resulting in roadway capacity reduction, inducing travel time delays, promoting traffic safety concerns, and increasing net public cost. This capacity reduction may be due to blocked traffic lanes or even a roadway component (e.g. a bridge) that is temporarily taken out of commission during execution of improvement activities. This period of reduced capacity from such non-recurring events is considered, herein, as roadway downtime, and the inducing events as downtime events.
Reduction in travel time reliability, degradation in serviceability, increase in primary or secondary traffic incidents [1], and increase in fuel consumption are other consequences of capacity reduction due to roadway downtime. In the U.S., more than 550 million gallons of fuel and 480 million hours are lost every year due to traffic congestion brought about specifically by work zones [2]. Also, an estimated 10% of congestion and 24% of unexpected freeway delays are caused by work-zones [3]. Traffic incident-related delays form another 13 to 30% of the total congestion delay over peak periods [4]. Even after roadway capacity is reinstated, the activity is complete or the event is cleared, degradation in roadway system performance can be expected for a subsequent period of time.
Improvement actions preserve or increase serviceability and ultimately improve safety, and travel time reliability for future operations. As they are undertaken, however, they negatively impact travel times and travel-time reliability for both the facility, and, due to shifts in traffic to alternative facilities, adjacent roadways. These temporary impacts are often not explicitly accounted for in the activity planning process. Yet, in total, these actions are not infrequent and their effects not negligible; their impacts are realized during large swaths of time. Moreover, their effects are not only local but facility-level and network-wide.
This work seeks a deeper understanding of the effects of improvement actions on traffic by juxtaposing their effects against those arising from traffic incidents that cause similar capacity reductions and related negative externalities. This is accomplished through direct and reverse approaches. In the direct approach, computed vehicle unit delays are compared by event type. Computations explicitly account for location extent, duration, and negative impact propagation direction. To support the reverse engineering investigation, machine learning methods, based on concepts of Support Vector Machines (SVM) or Random Forests (RFs) with K-Nearest Neighbor (KNN), are proposed and applied to identify those traffic characteristics that contribute greatest to correct detection of both types of downtime events, i.e. roadway improvement actions and traffic incidents. Similarities and differences in these contributing factors aided in understanding their unique qualities. Additionally, findings from this investigation aided in efficiently identifying key variables for inclusion in unit delay estimation models.
Application of the developed methods and metrics on case studies allowed investigation into the: (1) most affected traffic features for event type; (2) accuracy of hybrid event detection methods and the transferability of chosen parameters; (3) influence of downtime from improvement actions on corridor performance; (4) unit delay estimates and their generalizability across corridors; and (5) relationship between event features and unit delay values. Generally, the outcomes of this work can be used to produce estimates of the public cost, whether direct through investments or indirect in the form of user costs, of highway improvement projects. These costs can then be weighed against their benefits.
A key finding of the results is that travel time reliability-related characteristics contribute more to detection of traffic incidents than to detection of improvement actions even when executed during peak hours. This indicates that improvement actions are less likely to affect traffic performance than are traffic incidents. In general, the results show that traffic incidents have more than five times the impact on traffic compared with improvement actions. This five-fold impact was found both through investigation of the contribution of traffic characteristics to correct event detection and estimated unit delays from both event types, and persists whether the analysis is conditioned on peak hours.
Also key, 60% of traffic improvement events in the case studies had nearly zero traffic impact. While not as monumental as traffic incidents, improvement actions blocking one lane and requiring approximately one hour create 9 minutes delay per vehicle on average (ranging from 0 to 16 minutes with a 90% confidence interval) when executed in the peak period. Total delays over all vehicles from improvement activities, thus, can be very substantial. However, traffic incidents blocking one lane for one hour in the same peak hours create 16 minutes delay per vehicle on average. Moreover, each additional lane blocked in a collision increases the average unit delay by 67%, while in an improvement action, the average unit delay increases 17%.
The next section reviews relevant literature and further establishes the paper’s contributions. This is followed by details of the methodological approaches used in this study (Section 2) and results from their application on three corridors within the Commonwealth of Virginia (Section 3). Additional findings and their implications follow in Section 4.
2. Literature Review
Several studies have investigated the impact of maintenance activities on the operation of real roadway facilities and the economic costs and benefits of investments in public roadway infrastructure. These works consider the effects of maintenance actions in inducing: capacity reduction (e.g., [5, 6]), secondary incidents (e.g., [7]), and traffic delays (e.g., [2, 8]). Comprehensive reviews of capacity estimation methods and work-zone impacts can also be found in ([9, 10], respectively). Findings from these studies indicate that a large number of factors contribute to the impact of maintenance activities on roadway performance. These factors when gathered across these studies include as follows: percentage of truck traffic; pavement grade; number of lanes and lane closures; lane width, work zone layout (lanes merging, lane shifting, and crossover), construction type, duration, and time (on-/off-peak, day/night); pavement/weather condition (dry, wet, icy, sunny, rainy, snowy); and traveler familiarity (commuter/noncommuter). Some of these studies provide an equation. For example, Nassiri and Aghamohammadi [5] provide a model to predict remaining roadway capacity for a work zone. These equations require input data that may be difficult to obtain, and applying their methodologies in alternative locations requires traffic volumes and/or density, in addition to speed, data as input, which may be difficult to obtain.
These studies make important, limiting assumptions on the direction of speed reduction, and the propagation of other traffic impacts, activity durations, and extent. Specifically, they include changes in traffic measures in both traffic directions equally, use preset activity durations and presume that the location of the traffic impact is limited to the work zone and, in some cases, a preset distance upstream of the work zone. Impacts of maintenance activities were investigated either within a fixed extent of time and location (e.g. [7, 8]) or within a dynamic extent defined over a fixed discharge rate. For example, Du et al. [2] defined the delay as more than 25% drop in the normal speed and then, measured the delay cost based on such criteria. The normal speed must be fed into their model as an input. The methods proposed herein use observations from the data to determine event location extent, duration, and propagation direction. They do not require prior information on normal speeds.
While methods that have been proposed for traffic incident impact analysis have similarly accounted for incident impact propagation direction and may not require preset values of incident impact duration and extent, these other methods cannot be directly applied to work zone impact analyses. They are designed for and trained on traffic events with short durations consistent with vehicular accidents.
Numerous studies provide methods for prioritizing roadway construction projects with a goal of minimizing total costs, including monetary costs of traffic delay, and safety from inaction (e.g., [11–17]). A few prior studies include the negative effects of improvement actions during their execution (i.e. downtime) in prioritization and scheduling processes [13, 18–22]. These works replicate traffic either as a user equilibrium (UE) traffic assignment (using a cell transmission model (CTM) approach or Bureau of Public Roads (BPR) function), day-to-day route choice problem, maximum flow problem, or through microsimulation of traffic.
Using an actor-critic, deep reinforcement learning solution methodology, DCMAC [23], to solve a bilevel model in which activities are scheduled in the upper level and a UE traffic assignment captures the response of traffic in the lower level, Zhou et al. [22] found that a savings in traffic delays of between 16 and 28% with up to 15% cost savings can be achieved by accounting for the construction activity downtime effects during improvement action prioritization and scheduling. In a day-to-day setting, Yang et al. [20] found that work durations of less than 60 days significantly impact the optimal schedule; while, activities with longer duration do not.
While accounting for the impacts of work zones in prioritization and scheduling, these prioritization and scheduling works used rough estimates of improvement action impacts during their execution, e.g. a percent reduction in capacity given a duration and number of lanes blocked, as inputs to their models. Event detection methods and unit delay estimates developed herein can support these approaches, providing specific impact estimates for the considered activities.
The detection of disruptions in traffic was also considered in the literature and dates back to the 1970s [24–26]. These works developed methods to detect a change in lane occupancy values beyond a threshold. They proposed methods based on standard normal deviate, decision trees, and time-series analysis, respectively. More recent detection algorithms make use of newer technologies. Data is collected from loop detectors [27], video cameras [28], probe vehicles [29], and social media [30]. Other recent works also propose the application of more involved artificial intelligence methods (e.g., [30–33]). These machine learning techniques detect anomalies or outliers within the traffic data by considering not only lane occupancy but other factors, such as speed changes. Several works extended these concepts to detect such anomalies in real-time (e.g., [34–36]).
While methods for detection of traffic incidents are plentiful, it appears that no prior study has sought to detect maintenance, rehabilitation or other construction or improvement related activities. Despite that improvement actions are planned, and often recorded, it can be useful to have the capability to detect these events as their exact time of implementation may not be known, they may not be implemented exactly as planned due to uncertainties such as changing weather, and their occurrence may not be known to all parties. This is especially true for actions with limited duration. Moreover, it may be useful to be able to distinguish a detected potential traffic incident from an ongoing improvement activity in real-time applications. This paper fills this gap. It further proposes a reverse engineering approach for understanding differences between downtime from improvement actions and the effects of traffic incidents and introduces a concept of unit delays, computed values of which may have broad utility.
3. Methods
3.1. Event Detection and Critical Factor Identification through Reverse Engineering
The proposed reverse engineering technique uses traffic characteristics to detect the occurrence of an event. It estimates the contribution of each characteristic to detection success (correct event detection) and failure (misclassification), and with this information identifies the most critical characteristics for success under both improvement actions and traffic incidents. Investigating differences in those characteristics that most contribute to the identification of these two classes of events aids in providing a deeper understanding of how improvement actions are both similar to and different from traffic incidents in terms of their impacts on traffic performance and public welfare.
Artificial neural network (ANN), support vector machine (SVM), and random forest (RF) techniques are classical machine learning methods that have been widely applied in traffic incident detection studies (e.g. [36–38]). Xiao [39] noted that well-performing models, such as ANNs, did not perform as well when applied to a second data set. Motivated by this, Xiao proposed an ensemble learning method that integrates SVM and K-Nearest Neighbor (KNN) methods for incident detection. This work employs similar concepts to Xiao’s ensemble methodology by creating a hybrid of SVM (or RF) and KNN procedures, but with a different structure and application. Here, SVM (or RF) is used first for detection and KNN second for refinement; whereas, Xiao’s method seeks confirmation of correct detection by considering whether the events are selected by both methods (SVM and KNN) giving more weight to those events identified in SVM. Moreover, Xiao’s method does not correct for discontinuities in space and time arising in the detection of a single event. Thus, it may identify one event as a set of small events. The proposed hybrid methodology employs four detection classes to incorporate and distinguish between traffic incidents and improvement actions. Initial runs were made with a method based on the ensemble design of SVM and KNN in Xiao, but the sequential version proposed herein outperformed the ensemble design.
While methods for detection of traffic incidents are plentiful, it appears that no prior study has sought to detect maintenance, rehabilitation or other construction or improvement related activities. Despite that improvement actions are planned, and often recorded, it can be useful to have the capability to detect these events as their exact time of implementation may not be known, they may not be implemented exactly as planned due to uncertainties such as changing weather, and their occurrence may not be known to all parties. This is especially true for actions with limited duration. Moreover, it may be useful to be able to distinguish a detected potential traffic incident from an ongoing improvement activity in real-time applications.
As Xiao [39] found in developing an ensemble learning method for traffic incidents, combining methods was found here to enhance detection accuracy. In application to real-world data (Section 5), the RFF-KNN method outperforms the SVM-KNN method in terms of detection accuracy. Both were shown to reduce misclassification errors as compared with the proposed single-phase SVM and RF methods. The following subsections present these hybrid learning methods and their components.
3.2. Hybrid Learning Methods
Traffic event detection can be regarded as a classification problem as suggested in [29]. The proposed hybrid learning methods are aimed at classifying each data instance (a row or record from a dataset—here, the traffic characteristics associated with one roadway segment and one point in time) as falling into a nonevent or event class (improvement action, traffic incident, both). Each hybrid method is composed of two learning phases. Phase I starts by training a classification model on a portion, typically half, of the data. The trained SVM/RF classifier detects the type of downtime event (class) (at every location (roadway segment) and time increment (1-minute interval)). Phase II employs the KNN algorithm to further refine the results, reducing the number of misspecified instances. Gaps across time or location for a single event are identified and their classifications modified for increased consistency by maximizing homogeneity within and heterogeneity between downtime event classes.
Let represent an element of a matrix of traffic characteristics (e.g., speed and volume) at time interval and location , and be corresponding elements of matrices of predicted downtime event types resulting from the phase I and II classifiers for the specified and . The hybrid learning method, thus, applies a SVM or RF classifier to in phase I and refines the results using the KNN algorithm in Phase II. The final (refined) prediction, , uses values of , …, , …, , for a parameter of the KNN algorithm, from Phase I.
An overview of the proposed hybrid method is depicted in Figure 1. Figure 2 illustrates the steps of this method for a generic location. Applying a classifier in Phase I of this location resulted in detection of an ongoing incident for time periods 5 to 13, 16 to 21, and 28 to 29. In Phase II, KNN refines these time intervals to a contiguous interval by adding time instances 14 and 15 to the detection results and removing time instances 28 and 29. Descriptions of the SVM, RF and KNN subprocedures follow.
3.3. Support Vector Machine (SVM)
SVM is a supervised machine learning method that was developed for binary [40] and extended for multi-class [41] classification applications. The multiclass SVM approach breaks the classes into multiple binary (two-class) classification problems that each included the dominant class versus all alternatives. For each, SVM generates an optimal separating hyperplane that maximizes the separating margins of the two classes (e.g. traffic incidents as the dominant class against a class that includes all other event types, here, improvement action, nonevent or both) over a linearly separable sample dataset (here, speed-related changes) [42].
Given a dataset of samples, , …, , …, where , is a training vector containing all traffic characteristics for each data sample, and is the class type (dominant or alternative) associated with SVM generates separating hyperplanes as formulated in [40]:
is the normal vector to the hyperplane (hyperplane function: ), the intercept of the hyperplane, is a penalty term, and a manually chosen kernel function.
As traffic data is not generally linearly separable, nonlinear mapping functions, i.e. kernel functions, transform the traffic data into a higher dimensional space through which data points can be linearly separated. A radial basis kernel is used here and is formulated as , where is a parameter that defines the amount of influence given to a single training data sample. The performance of SVM relies heavily on the choice of the kernel function and its parameters ( and ). As in [43], to improve the classifier when applied to other datasets and avoid overfitting the model to the test data, incremental grid-searches on (ranging from 1 to 10 with step size of 1) and (ranging from 0.1 to 1 with step size of 0.1) were completed. Values of and were chosen to provide the highest average -fold cross, with , validation accuracy. See [44] for background on cross validation method. Once optimal hyperplanes are generated for each two-class problem, the predicted class of a data sample is determined according to the two-class problem with the highest prediction score.
3.4. Random Forests (RFs)
RFs, originally proposed by Breiman [45], are constructed on a collection of tree-structured classifiers similar to decision trees. Each tree casts a single vote (i.e. identifies one of the possible classes (event types)) toward the most popular class via mode or mean for each data element of a given data set. A RF is formed from an ensemble of randomly generated tree classifiers. Inserting randomness into constructing the tree classifiers makes for a more robust method that is less influenced by outliers and unbalanced datasets and is protected some from overfitting. RFs have been used in traffic incident management and safety studies in recent years (e.g., [46]). Through variable elimination, it has a secondary benefit of providing a measure of variable importance [47, 48].
Here, RF is proposed for detecting event types (traffic incident occurrence, improvement action, both, neither) given a data set containing traffic characteristics, such as speed changes. Tree classifiers are grown to produce a forest. Each tree classifier is grown by applying conditional statements to a randomly chosen, fixed number (, a parameter) of the event’s data items (e.g. traffic characteristics from a row (a sample) in the dataset) creating a decision tree. RF constructs such decision trees, each predicting a sample’s event type. It votes accordingly. Here, the result of each decision tree is one of the four event types (traffic incident, improvement action, both, and neither). The event type with the largest mode over decision tree predictions is taken as the final outcome.
The final chosen type for a given sample is obtained from the total collection of votes, one from each tree in the RF for the test dataset. The accuracy of classification is affected by the chosen (tuned) parameters, and .
3.5. K-Nearest Neighbor (KNN)
KNN is a classification method that labels a data instance according to the classes of the nearest instances. The data instance is labeled based on a majority vote from these neighbors when weighted by their respective distances. That is, those neighbors closest to the data instance will have greater contribution.
Here, KNN is used to refine classification determinations made from either SVM or RF classifier creating two hybrid techniques: SVM-KNN and RF-KNN. Consider a dataset of event classifications organized by segments over time created from running SVM or RF. For the subset of data records associated with roadway segment , classification is obtained from its nearest (in time) records: . In the voting process, each neighbor is given a weight as a function of its time difference from time : for and the time of data instances and , respectively. This is illustrated on an example in Figure 3 for =10.
4. Case Study
Three case studies were conducted involving corridors in Virginia, each with a freeway segment and parallel, alternative arterial. The corridors are between 12 and 30 miles in length and each varies in terms of number of lanes (capacity), speed limit, average daily traffic (ADT), percentage of truck traffic, and presence of horizontal and vertical curves. This section describes the case study locations with relevant input data, analysis outcomes, and ultimate findings.
4.1. Locations and Data Sources
Case studies were selected from northern, southern, and western regions of the Commonwealth of Virginia. Their locations are depicted in Figure 4 and are specified as follows:
Case study I: A 12-mile stretch of the I-66 Westbound corridor between Leesburg Pike (Exit 66) and US-29 (Exit 52) and adjacent roadway sections of US-50 Westbound (between Graham Rd and US-29/VA-237/Old Lee Hwy) and US-29 (between US-50/VA-237/Old Lee Hwy and I-66 (Centerville)) in Fairfax County.
Case study II: A 30-mile stretch of the I-81 Southbound corridor between I-64/Exit 221 and I-64/Exit 191 and adjacent roadway sections of US-11 Southbound between VA-262 and I-64 in Augusta and Rockbridge Counties.
Case study III: A 25-mile stretch of the I-64 Westbound corridor between Croaker Rd/Exit 231B and US-60/Exit 200 and adjacent roadway sections of US-60 Westbound between Croaker Rd and I-295 in James City and Henrico Counties.
4.1.1. Traffic Data
Traffic data, including information related to speeds and travel times, were obtained from the I-95 Vehicle Probe Project (VPP) II contract under the INRIX suite. The data, which provide space-mean speeds, were collected through the INRIX traffic message channel (TMC) monitoring platform. INRIX reports traffic data by road segment, each referred to as a TMC segment. Selected stretches of I-66, I-81, and I-64 include a total of 16, 16, and 14 TMC segments, respectively. Lengths of these TMC segments range from less than 1 mile to 6 miles. INRIX data are widely utilized in transportation studies (e.g. [51, 52]).
The case studies relied on seven-month historical TMC segment data from April 1 to October 31 in 2019. Traffic data were retrieved and aggregated at 1-minute increments (as in [53] for each TMC segment. This choice of a 1-minute resolution for event detection aligns with the reporting frequency of data in the event dataset, and is of the highest level of granularity achievable with the datasets that are used. Traffic data of each TMC segment includes a reference speed (speed limit), current speed, historical average speed, and time required to traverse the segment. For the studied TMC segments, the number of 1-minute traffic data records collected in this 214-day period is 4,930,560, 4,930,560, and 4,314,240 in total for I-66, I-81, and I-64, respectively. Additional data associated with TMC segments 3 miles upstream of the study location start points and 1 TMC segment downstream of the end points were also obtained. These additional data were included to catch event impacts that may persist beyond the study locations. The format of the collected traffic data by TMC segment is shown in Table 1. Table 2 provides a snapshot of additional information from the INRIX dataset that is used in identifying locations associated with events and their impacts.
To minimize random variations and impacts from random errors in the traffic data, the traffic data given in 1-minute increments were smoothed using a weighted and centered moving-average that weights the data according to INRIX-supplied observation confidence scores. Less than 0.5% of the data were missing. Since the missing data were limited to a small number of observations and rarely occurred over consecutive periods of any extended length, the missing data were replaced by the mean values for prior and subsequent time increments. Such instances of missing data were assumed to result from temporary failures arising within the TMC system. An approach that imputes missing values from their nearest neighbor values while considering time, location, hour of the day, weekend/weekday, and annual average daily traffic was tested, but resulted in replacement values that were inconsistent with prior and subsequent data values. To obtain speed profiles under recurrent conditions and further calculate unit delays, Recurrent Speed Profiles (RSPs) were computed through methods presented in [54].
4.1.2. Event Data
An event dataset was created that includes traffic incidents of vehicular collisions only, with subcategories of single-vehicle accidents, multiple-vehicle accidents, and tractor trailer accidents, and improvement actions with work zone details. Improvement actions include as follows: bridge inspection work, guardrail repairs, ITS equipment repairs, paving operations, pothole patching operations, rehabilitation project activities, resurfacing operations, survey work, median/jersey wall repair/installation, new roadway construction, and road widening project activities. The events’ data used in the case studies were retrieved from a nation-wide traffic event data archiving system Regional Integrated Transportation Information System (RITIS). The Virginia Department of Transportation (VDOT) manages information about work zones, incidents and other types of traffic disruptions arising in Virginia with the aim of providing travelers with traffic information and facilitating roadway services. Table 3 synthesizes this database of studied collisions and improvement actions for the study locations during the study period. The events’ data include basic information related to timing to the minute and location in terms of geographic coordinates, duration, maximum number of lanes closed, and brief text with description.
4.2. Data Preparation
4.2.1. Relating Traffic Event Occurrence to Traffic Performance
The collected traffic data from INRIX was sorted to create a time-ordered and TMC-segment sequenced set of records. In an event occurrence, upstream traffic will often travel at slower speeds while downstream speeds may be higher in comparison. Other traffic characteristics may be similarly affected up and downstream of the event. To investigate such possible changes over the course of an event’s impact, 17 additional traffic-related variables were developed. 16 relate to traffic characteristics and the remaining is associated with time. These variables build on the speed profiles of TMC segments and relate either to traffic or the event impact’s temporal characteristics. A reduction in the speed ratio (SpR) can be expected at the time and location of an event. At the very moment of the event, the average speed ratio in both upstream and downstream segments is expected to be greater than that in the location of the event. At later time increments, as delays propagate, this will not be the case. Observations of the dynamics aid in event detection.
Several of the listed traffic-related variables are defined based on a metric for travel time reliability, Extra Buffer Time Index (EBTI), previously introduced by Tavassoli Hojati et al. [55]. In general, extra buffer time is defined as the extra delay caused by an event. It indicates the extra travel time needed to arrive at a destination on time with 95% certainty in a traffic event. Details of EBTI calculation can be found in [55].
Temporal characteristic, including peak hour and weekend indicators, acts as a dummy variable. Table 4 consolidates the potential explanatory variables, including spatiotemporal traffic characteristics and time-related variables, for event detection. The response variable to these explanatory variables is the event class for a given location-time pair .
4.2.2. Balancing the Data
It is reasonable to expect that for any location, there will be a disproportionate number of location-time pairs in which no event (neither traffic incident nor improvement activity) has occurred. This leads to an imbalance in data supporting to the four detection outcomes (no event, traffic incident, improvement activity, or both traffic incident and improvement activity). Consider the case study on the I-66 W corridor. With its 4,930,560 data records, only 231,711 data records, 4.7% of the data set, relate to an event. This data imbalance can lead to issues of bias, misclassification, or low accuracy/sensitivity in detection outcomes.
To avoid issues of low sensitivity or accuracy, methods for balancing the data are often applied. In works in the literature, oversampling and undersampling techniques were used to balance the data by duplicating instances of the minority class and removing random instances of the majority class, respectively. A downside to these techniques is that they are known to cause issues of over- or underfitting to the data. To address this concern, [56] introduced Synthetic Minority Oversampling Technique (SMOTE). SMOTE generates synthetic data records of the minority class for the K-Nearest Neighbors of a minority sample.
In this work, the dataset is split into training (including 2,773,440 records for 9 continuous segments of I-66 W, 2,465,280 records for 8 continuous segments of I-81 S, and 2,465,280 records for 7 continuous segments of I-64 W) and test (including 2,157,120 records for 7 continuous segments of I-66 W, 2,465,280 records for 8 continuous segments of I-81 S, and 2,465,280 records for 7 continuous segments of I-64 W) datasets. Parameters of the machine learning techniques are set on the training data and then applied within the machine learning method in the analysis completed on the test data. Table 5 provides a count of the event records and final number after applying SMOTE by event type for each of the three case study locations. Event records are between 0.01 and 7% of the records in the original training datasets. Records from the three event classes are considered as minority classes, while nonevent records form the majority class.
A mixed technique of adding synthetic minority records and removing majority records was applied to achieve the best compromise between data size and algorithmic performance. First, SMOTE with was applied to the training datasets to create synthetic minority records for inclusion. The application of SMOTE was completed separately for each minority class. Second, a random set of the majority class (nonevent) records was removed producing a roughly 1 : 4 ratio of event to nonevent records. This ratio is considered a golden standard and, thus, follows standard methods from case-control studies [57]. The parameters of the event detection models were calibrated on the balanced training dataset.
5. Results and Discussion
The importance of traffic features to the accurate detection of event occurrence by type is analyzed in this section. Event-based unit delays at facility- and corridor-levels were estimated for each of the case study locations. Commonalities and differences between improvement actions and traffic incidents can be discerned from event detection model results. Factors affecting these unit delays were identified through multivariate statistical analysis. Before proceeding to apply the detection methods, their performance on the case study locations is analyzed.
To evaluate the performance of the SVM, RF, SVM-KNN, and RF-KNN methods in classifying each , each data element for each time interval and location as being associated with one of the four event (or nonevent) types, a confusion matrix is created from the results of each method that includes classification errors by event type, and an overall prediction accuracy is computed from the matrix. The confusion matrix concept was first introduced in [58] and is illustrated for this application in Table 6. In the confusion matrix, gives the number of instances of event type that are predicted to be of event type and gives the ratio of instances of event type that were predicted to be of event type to the total number of instances of type , : =. Thus, is known as the detection rate for type . The overall predictive capability of a method for a given application then is determined by .
Either of the developed hybrid techniques can be designed around various input traffic characteristics. Those characteristics that contribute in the method’s specification with the greatest accuracy provide additional insights into which characteristics from the data contribute most to the event type. Different characteristics may contribute more to one or another event type.
In applying either proposed hybrid method, some improvement actions were not detected. The occurrence of false positive and negative findings, in part, can be explained by the differences in the effects of improvement actions on traffic performance as a function of lanes blocked, duration of implementation, and other factors. Understanding which features make an activity less detectable can provide further understanding.
5.1. Event Detection Method Performance
The proposed detection methods were executed on balanced training datasets. The default value of needed for the RF-based procedures suggested in the literature can be obtained from the square root of the number of explanatory variables in the training data set [59]. Larger values of , though, have been found to lead to better performance (e.g., [46, 60]), but larger values result in greater computation time. Runs with between 3 and 8 were completed as 8 variables were chosen for inclusion. It was found that an of 6 performed best. For the SVM-based methods, parameters (set to 10 for all locations) and γ (set to 0.6 for locations I and III and 1 for location II) were set to achieve the highest average 10-fold cross validation accuracy for all locations. Additional parameter settings include (for KNN) set to 10 in all runs and RF’s at 200 for location I and II and 150 for location II.
Table 7 reports detection accuracies. The accuracy of all tested methods was very high with the best performing method ranging from approximately 91 to 99%. Results of the event detection methods are reported in a confusion matrix (Table 8) giving prediction counts and detection rates for each event type and case study location with parameter settings as identified earlier. Incorporating KNN to create the hybrid approaches increased detection rates by as high as 7 percentage points (81.5 to 88.5%), and, thus, decreased misclassifications similarly. The KNN addition also improved the method’s accuracy for all case studies and nearly all event types when added to both SVM and RF methods. Generally, the RF-based approaches outperformed the SVM-based methods with and without incorporating KNN.
Note that the parameters applied in Case Studies I and III are identical and differ only minimally from those used in Case Study II. If these parameters are used on Case Study location II, the results diminish in accuracy only minimally.
5.1.1. Misclassified Improvement Event Instances
Between 26 (Case Study I) and 41% (Case Study III) of improvement actions were misclassified as nonevents. These results were further analyzed by conditioning on time of day (Figure 5), maximum number of lanes closed (Figure 6) and improvement type (Figure 6). From Figure 5, it can be noted that improvement actions were most likely to be misclassified when they were executed in off-peak hours. In Case Study I, the misclassification rate when aggregated over the day is 27%, however, when conditioned on the time of day, this value ranges between 8% and 45%, 45% occurring in off-peak hours. The variability over the day in misclassification rates is not as severe in the other two, more rural case study locations.
The procedures are generally able to distinguish between improvement actions and traffic incidents, implying that impact traffic differently. However, the ability to distinguish between this event types may be because improvement actions tend to be executed in off-peak hours and traffic incidents are more likely to occur in peak hours. Thus, the increased detection rates of improvement actions in off-peak hours may be from taking advantage of knowledge of the time of day in which the event occurs. That is, if the event arises in the off-peak hours, it is best to guess that it is an improvement event. Conditioning on time of day, density plots and means of AvgR (Figure 7) and EBTI (Figure 8) traffic performance metrics were further studied. Note that density gives the probability scaled by bin width. This is proceeded by investigation into the travel time delays from each event type. Statistics of AvgR and EBTI conditioned on type and time of occurrence are summarized in Table 9.
Sharp peak values around a ratio of 1 for AvgR and very low EBTI values with some difference in AvgR, but no significant difference in EBTI between peak and off-peak time periods, were observed. This indicates that improvement actions have little impact on travel time reliability at any time of day, but have higher impact on speeds in peak hours. Additionally, when misclassified, regardless of time of day, the improvement activity was presumed to be a nonevent.
The results in Figure 6 from studying improvement detection rates conditioned on type and number of lanes blocked shows that the greater the reduction in roadway capacity caused by the event, the more likely the event is to be correctly classified. Pothole patching, which typically has low impact on capacity, was further studied by comparing results in Figure 5 with the timing of events as illustrated in Figure 9. These events were misclassified as nonevents in between 50% and 70% of all occurrences for the case studies. The results indicate that time-of-day is not a relevant factor. This finding was useful in designing the unit delay estimation models.
5.1.2. Feature Importance in Event Detection by Event Type
Features that contributed most to correct event occurrence and type detection for events involving either traffic incidents or improvement actions were identified to understand how improvement actions differ from traffic incidents in terms of their impact on traffic. For this purpose, feature importance scores were computed through a permutation-based approach employed in training the detection portion of the RF-KNN hybrid method. In this approach, a baseline model is created on a given dataset with various traffic features and its accuracy and detection rates are recorded. Then, values from one feature are randomly shuffled, and the modified dataset is passed to the model update accuracy and detection rates. The feature importance scores are computed from the difference between detection rates of the baseline and permuted models. This difference is known as the Mean Decrease Accuracy (MDA). The higher the MDA value of a traffic feature, the more important is that feature. Table 10 reports the MDA values computed for each of the case study locations and highlights the top 7 features for each event type and case study location.
Eight features were of greatest importance over all locations for improvement and collision event types are as follows: UpSdR, U/CAvgR, D/CAvgR, AvgR, SdR, D/CEBTI, DownAvgR, and UpAvgR. These features were supplied as candidates for inclusion, and various combinations of six () of the eight were included in the developed improvement action/collision detection methods.
The results indicate that EBTI-related metrics were not among the top important features for almost any event type and location. Additionally, these metrics were among the least important features for improvement activities. This confirms the earlier claim that improvement actions have little impact on travel time reliability.
Also, of note is that the feature importance score is generally greater (almost five times on average) for all metrics associated with collisions than for those associated with improvement actions. As expected, it was observed that traffic conditions on upstream segments were more important than for downstream segments. On the other hand, the ratio of changes of traffic conditions in downstream segments to changes of traffic conditions in the segment containing the event (D/CAvgR and D/CEBTI) played a more important role than a similar ratio for upstream segments (U/CAvgR and U/CEBTI). This may indicate that the rise in speed downstream of the event location is more significant than the drop in speed upstream of the event location. This difference was greater for collisions than for improvement activities.
5.2. Unit Delay Estimates
Using traffic data collected by probe vehicles, traffic conditions are assessed to determine if an event has arisen and has impact. An approach proposed in [54] that uses -means clustering was employed to classify a location- (roadway segment) time pair as having or not having a significant change in speed as per the speed ratio. Following rules of contiguity, the spatial and temporal extent of change in speed ratio as a consequence of the event is delineated and is referred to as the event’s impact area. Vehicle unit delays, a measure of extra travel time incurred due to a reduction in speed, are computed based on speed differences over the event’s impact area. Specifically, the extra travel time incurred per vehicle along each roadway segment that falls in the delineated impact area is calculated as the difference between the inverse of both average observed and average recurrent speeds multiplied by the length of that segment. Summing the extra travel time per vehicle incurred over all segments falling within the impact area and dividing by the event’s impact duration gives the vehicle unit delay for the event.
Unit delays were computed for improvement action and collision event types at both facility and corridor levels, as well as by case study location and in the aggregate. Unit delays were also computed for the minor, adjacent facilities to the main study roadways where the events took place. To study the network-wide impacts of an event, unit delay calculations are completed over a more inclusive impact area that incorporates not only the impacted segment(s) of the facility on which the incident or activity arises but portions of a parallel facility and relevant connectors (including on- and off-ramps) to this facility to which traffic diverts. Thus, the vehicle unit delays are computed across a broader geographical area.
Similar to the speed-ratio matrix developed in [54], a second speed-ratio matrix was developed over observed traffic conditions at time intervals after the event on all potentially impacted segments of the minor facility. In the minor facility, the extent of the event’s impact over space and time was restricted to only those segments of the facility that fall within a 2- or 3-mile radius of the event’s location. The same time intervals as used in studying the event impact on the major facility were applied in studying the impact on the minor facility. Traffic conditions at the on- and off-ramps in affected segments of both facilities were further analyzed. Since the lengths of the segments that contain the ramps are unequal, unit-delays normalized to segment length are reported.
Figure 10 shows the process of identifying the potentially affected segments of US-50 W (minor facility of Case Study I). Segments 10 to 15 of US-50 W fall within the extent boundary (i.e. the 2-mile radius). Segment H of I-66 W and 13 of US-50 W contain interchanges that fall within the bounded area. It is hypothesized that traffic may detour, leaving the major facility (I-66 W) at segment H and entering the minor facility (US-50 W) at segment 13. Normalized unit-delays of each minor facility (US-50 W) segment are also plotted in this figure. Speed ratio matrices associated with a single collision event occurring on 9/4/2019 at 6 : 27 : 00 are given in Figure 11 for both major and minor facilities.
Normalized unit-delays associated with impacted segments (in both major and minor facilities) for a randomly selected set of events arising in Case Study I are reported in the plots of Figure 12. This case study was selected for this additional analysis, because the roadway segments were small enough to be able to investigate traffic changes near the intersections. From this additional analysis, it was found that higher normalized unit-delays in collision events occurred along the upstream segments of the segment containing the intersection connecting the minor facility (US-50 W) to the major facility (I-66 W). This supports the hypothesis that drivers are diverting to an alternative roadway in collision events. The same was not noted for improvement actions.
The results indicate that major and minor facility unit delays were greater for collisions than improvement actions. Table 11 and Figure 13 show the statistics and histograms of major and minor facility unit-delays, respectively. Figure 13 further indicates that approximately 60% of the improvement actions have almost no impact (near zero unit delay) at the facility level while 30% of collisions had similar limited traffic impact. By comparing aggregate mean unit delays on the facility itself against the delays in the adjacent facility, the relative difference between the roadways for collisions is less than that of improvement actions. This is reasonable as drivers may choose their routes with foreknowledge of improvement actions. This indicates the importance for studies of improvement activity impact on roadway performance to investigate the impacts beyond the facility in which the activity is executed.
Broadening the area of investigation from a radius of 2 miles to one of 3 miles produced greater values of unit-delays with an increase of 26% for collisions and 46% for improvement actions. It may be that the wider radius catches more of the delays caused by the event on the major road, but it may also catch delays from other events occurring on the minor roadway or other roadways in the larger roadway network. Additionally, speed data may not be as accurate along secondary roadways, causing potential inaccuracies in the computations. Further investigation on the appropriate radius to use might consider dependencies on roadway geometry (e.g. number of on- and off-ramps) and event characteristics (e.g. severity of a collision).
To estimate the specific impact of the event type on traffic performance, an analysis of the relationship between event features (e.g. number of blocked lanes, time of day) and resulting delays at the facility level was completed.
Censored multivariate regression models, specifically Tobit regression [61], for estimating facility level unit-delays (minutes/vehicle per 1-minute interval of event duration) for improvement actions and collisions were developed to explore the significance of chosen independent variables, such as number of lanes blocked, event duration, time-of-day, percentage of trucks, and day-of-week. These variables were chosen to use insights gained from application of the event detection methods, results of which highlighted the most important features. Independent variables of significance included roadway and traffic characteristics (Annual average daily traffic (AADT), K-factor, and number of lanes) and event features (e.g. number of blocked lanes, event type, and time-of-day). A difference in parameters of these models and which independent variables are significant for traffic incidents versus improvement actions provides important additional insight into improvement action impacts. Results of the Tobit regression are reported in Table 12.
With these equations, vehicle unit delays can be estimated for a facility with similar characteristics. These equations can also be used for future planning for the same or other roadway segments with similar traffic characteristics.
Regression model coefficients indicate that for each additional lane blocked, the average change in mean unit-delay from a collision or improvement action is 1.497 (a 67% increase) or 0.067 (about a 17% increase), respectively. That is, each additional lane blocked in a collision has four times the percentage increase of an additional lane blocked in an improvement event. An improvement action event occurring in the peak hour has twice the percentage increase in average unit delay as that of a collision. Moreover, the impacts of AADT and -factor are significant only for improvement actions. Each 10,000 vehicles per day increase in AADT (for the same number of lanes) will increase the unit delay for improvement actions by 0.55 minutes on average, while a 1% increase in the -factor decreases unit delays of improvement actions by 0.15 minutes on average. Moreover, it was found that if multiple vehicles were involved in a collision, an increase in unit-delay by 1.149 minutes on average (a nearly 55% increase) can be expected. Given a typical road as in case study 1 (, , ), unit delay for an improvement activity executed in the peak period and blocking one lane is estimated to be 0.146 minutes per vehicle per minute of impact duration.
6. Conclusions
This investigation culminated in the creation of mathematical tools, insights, and outcomes. Machine learning techniques (SVM, RF, and hybrid methods SVM-KNN and RF-KNN) were introduced to detect the occurrence of both improvement actions and traffic incidents for use both off- and on-line that rely on only widely available data. With these techniques, system operators can take action to divert traffic and/or provide appropriate services to clear an event in a timely manner when events are detected in real-time. These techniques can be extended to detect multiple, simultaneous events from one or more event type. The development of algorithms for detecting improvement activities during their execution also serves as a reverse engineering approach for better understanding the impacts of these activities on traffic. Employing these methods on the case studies revealed similarities and differences in the impacts of these activities with the impacts of traffic incidents. This deeper understanding of the effects of downtime from roadway improvement activity execution can be useful in optimal activity scheduling.
A secondary contribution to the literature on traffic event (collision and improvement action) detection is also made through the proposed hybrid learning-based methods that rely on readily available speed data. The hybrid learning methods were shown to reduce misclassification errors as compared with single-phase machine learning methods of SVM and RF. Thus, they are useful tools in detecting and distinguishing between collisions and improvement activities.
Unit-delay estimates can be used to plan and prioritize improvement actions. They can also be used in scheduling models for prioritizing improvement actions. Findings of this work can also assist with comparison of maintenance and rehabilitation options across different locations considering their characteristics (e.g. K-factor and AADT). By plugging in the K-factor, AADT, and number of lanes of a roadway, the unit delay incurred for improvement action options can be assessed and compared by also plugging in the corresponding lanes blocked by the action and time of day the action is to be executed. The total delay of an action that, for example, closes 3 lanes for 2 days can be compared to another action that closes 1 lane for 6 days. The relative impact of these actions may also differ on other roadways with different K-factors and/or AADT values.
Application of these approaches on three case study locations led to numerous insights. In terms of the solution methodologies, it was found that the addition of KNN to create the RF-KNN method improved the sensitivity by up to 7 percentage points over the single-phase RF method. Improvement actions and collisions were detected 74 and 88 percent of the time, respectively. Improvement actions did not significantly impact travel time reliability, even if they are executed in peak periods, yet collisions impact travel time reliability by up to 32%. Pothole patching activities were found to be the least detectable type of improvement action, while other pavement-related operations, along with utility works, were found to be the most detectable types of the improvement actions. This is important, because the easier to detect, the greater the impact. The impact of events on traffic performance was estimated in terms of unit delays. Unit delay estimates indicate that more than 60% of improvement actions and more than 30% of collisions have no impact on traffic. Paving operations and multivehicle accidents were found to have the highest unit delays associated with improvement actions and collisions, respectively. Future work might examine additional variables, including weather and sight distance, in unit delay computation. The proposed RF-KNN detection model might be compared with other state-of-the-art anomaly detection models in future studies.
Network-level analysis uncovered greater tendency for roadway switching in collision events than during improvement activities. Consistent with this, facility- and network-level unit delays are closer in value for improvement actions than for collisions. It is likely that this occurs because drivers will more likely have foreknowledge of improvement activities than of collisions.
Based on feature importance scores from event detection models, as well as from the calibration of the Tobit regression model for projecting vehicle unit delays, the impact of traffic incidents on roadway performance (in term of delays) was found to be five times greater than that of improvement actions.
The developed event detection tools can aid field observation-based traffic event reporting, e.g. through observations by police or road patrol. These tools may also facilitate quick, automated event detection using streamed traffic data.
Specified unit-delay estimates, and estimates from proposed equations for computing average unit delays for specified roadways, can be used in construction activity planning and prioritizing improvement actions while accounting for facility- and corridor-level impacts arising from the event downtime.
Applying findings from this study, along with implementation of the developed methods, can lead to improved construction activity planning. With event detection and insights into the event’s consequences through real-time application of the proposed tools, an agency can more readily respond to the event, providing appropriate services to clear a traffic incident and diverting traffic to alternative routes during construction activities or incidents. These actions can aid in reducing event impact and providing greater travel time reliability. They may also be used in estimating the net financial or public welfare impacts of considered transportation investments when choosing between alternatives.
In this work, the machine learning methods were trained on a portion of the data and tested on the remaining data in each of three case studies. Alternatively, the techniques could be trained on the entire dataset for one case study and then, tested on the other two case studies. Degraded performance may arise with this alternative method if the geometries and traffic characteristics of the case studies greatly differ.
Data Availability
Some or all data, models, or code used during the study were provided by a third party. Direct requests for these materials may be made to the provider.
Conflicts of Interest
The authors declare no conflicts of interest with respect to research, authorship, and/or publication of this article.
Acknowledgments
This work was supported by the Virginia Transportation Research Center (VTRC) of the Virginia Department of Transportation. This support is gratefully acknowledged, but implies no endorsement of the findings. The authors are grateful to the program officers at the VTRC for their ideas and providing access to data to support this investigation.