Detecting and Ranking Recurrent Bottlenecks on Urban Expressways: A Speed-Only Severity Index from Floating Car Data

Abstract

Accurate identification of traffic bottlenecks is a fundamental task in the operational analysis of urban expressways. Traditional bottleneck detection methods rely heavily on fixed-location sensors, which suffer from limited spatial coverage and high infrastructure costs. This study proposes a comprehensive Floating Car Data (FCD)-based framework for detecting, characterizing, and ranking recurrent bottlenecks on urban expressways using segment-level speed data alone. The methodology integrates normalized speed fields, a spatially aware moving window algorithm with adaptive downstream verification, and temporal persistence criteria to distinguish active bottlenecks from transient congestion events. To prioritize detected bottlenecks by their operational significance, a Speed-only Bottleneck Severity Index (S-BSI) is introduced. The S-BSI is a 0–100, five-component severity index that ranks active bottlenecks using segment-level speed data only, without requiring flow or density measurements; its components capture upstream spatial extent, temporal persistence, speed deficit magnitude, queue length, and queue growth rate. This research presents a structured methodology for identifying, characterizing, and ranking bottlenecks on urban expressways using Floating Car Data. Detection parameters are selected through a systematic grid search over 128 combinations. The framework is validated using one year of weekday commercial FCD (248 days) obtained from an urban expressway corridor. Empirical results demonstrate that the proposed algorithm reliably identifies spatially concentrated and temporally recurrent bottlenecks; in total, 32 distinct critical bottleneck locations (S-BSI ≥ 35) are detected, of which 31 (96.9%) coincide with an identifiable physical or operational congestion-generating feature of the corridor. By providing a scalable, cost-effective, and transferable detection-and-severity methodology, this research contributes to the growing body of literature on probe vehicle data applications for corridors where fixed-sensor infrastructure is limited or absent.

1. Introduction

1.1. Background and Motivation

Rapid urbanization and motorization have made traffic congestion a persistent problem worldwide [1]. Congestion increases travel times, fuel consumption, and emissions and lowers road safety and quality of life, making it a key barrier to sustainable urban development [1,2]. Consistent measurement is the first step toward management, and recent studies have moved from simple averages toward metrics that jointly describe congestion intensity, spatial extent, and duration [3]; long-term analyses further show that urban congestion follows regular, classifiable temporal patterns [2].

On urban expressways, which carry a large share of metropolitan traffic, congestion is mostly generated by bottlenecks—localized points where capacity falls below arriving demand [4]. Because queues propagate upstream, a few locations can account for a large share of total delay [4], and recent work argues that congestion management should be organized around bottlenecks rather than network-wide averages [5]. This study focuses on recurrent bottlenecks, which activate at the same locations during predictable peaks and are therefore reliable targets for planned interventions, unlike non-recurrent bottlenecks caused by incidents, weather, or events [6]. Detection alone is insufficient: with limited budgets, agencies also need to know which bottlenecks matter most, i.e., to rank them by operational severity [7]. Reliable ranking supports low-cost measures such as ramp metering, which is most effective when the active bottleneck and its queue dynamics are clearly identified [8], and targeted relief of stop-and-go operation directly reduces fuel use and emissions [7].

This paper presents such a framework which constructs a space–time speed matrix from historical FCD and applies a multi-criteria detection algorithm that simultaneously evaluates upstream congestion depth, downstream flow conditions, temporal persistence, queue propagation characteristics, and an integrated severity index. The approach is designed for urban expressways and comparable partial-access-control corridors where FCD is the sole data source and no detector infrastructure exists. It targets uninterrupted or partially interrupted flow facilities; it is not intended for signal-dominated urban arterials, where recurrent stops are governed primarily by signal timing rather than by capacity bottlenecks.

Recurrent bottlenecks are not only a mobility problem but also a major source of unsustainable transport externalities. Stop-and-go conditions at bottlenecks substantially increase fuel consumption, greenhouse gas emissions, and local air pollutants compared with free-flow operation. By enabling the systematic detection and prioritization of recurrent bottlenecks using only low-cost probe-vehicle speed data, the proposed framework supports targeted interventions that reduce congestion-related emissions and energy use. It therefore contributes to more sustainable and resource-efficient urban mobility, particularly in cities that lack dense fixed-sensor infrastructure.

Beyond raising average travel times, recurrent bottlenecks concentrate a disproportionate share of total delay at a few locations, sharply reduce travel-time reliability, and create stop-and-go conditions that raise the risk of secondary rear-end collisions at the back of queues. The same stop-and-go operation increases fuel consumption and greenhouse-gas and local-pollutant emissions relative to free-flow travel. These safety, reliability, and environmental hazards make the accurate identification of recurrent bottlenecks a prerequisite for effective and sustainable congestion management.

1.2. Limitations of Traditional Detection Methods

Traditional detection relies on fixed sensors—loops, radar, and cameras—that provide high-frequency data only at discrete points, so bottlenecks forming between stations may be missed and queue extent cannot be fully described [6]. Dense sensor networks are costly to install and maintain, limiting coverage especially in developing regions [9], and roadside speed distributions can differ systematically from probe observations, complicating method transfer [10]. Sensor-based detection is thus hard to apply on the many urban expressways lacking fixed instrumentation [9].

While these sensor technologies provide high temporal resolution data at instrumented locations, they suffer from several fundamental limitations that restrict their applicability for comprehensive bottleneck analysis such as limited spatial coverage, high infrastructure costs and limited transferability.

1.3. Floating Car Data as an Alternative

Floating Car Data (FCD), generated by GPS-equipped vehicles and navigation apps reporting position and speed at regular intervals, offers a low-cost alternative with wide spatial coverage [11]. Because probes travel with the stream, FCD gives continuous speed information along entire corridors rather than isolated points [11]. Validation studies confirm that FCD speed distributions represent prevailing conditions well [10], and FCD has been combined with fixed sensors and machine learning for network-scale flow forecasting [9], congestion prediction [12], and segment-level estimation from GPS trajectories [13]. FCD has two structural limits: commercial products typically provide only segment speeds, so flow and density must be inferred indirectly [11], and data quality depends on probe penetration, with reported optimal proportions of 3–8% [11]. Consequently, Floating Car Data (FCD) methods generally aim to extract valuable operational information primarily from speed data. While numerous studies have successfully utilized FCD for general congestion measurement or prediction [13], its application for dynamic, bottleneck-specific analysis remains underexplored. Specifically, the heterogeneous impact of dynamic critical bottlenecks on congestion diffusion has not yet been thoroughly integrated into a segment-based traffic congestion estimation framework, which strongly motivates this research.

1.4. Research Gap and Objectives

Bottleneck-oriented FCD methods remain scarce. An early bottleneck-detection study on a Beijing expressway used a speed-difference criterion between neighboring segments to identify bottlenecks; such fixed thresholds can fail during severe congestion, when both upstream and downstream speeds are depressed [14].

Ranking methods based on travel-time reliability require rich travel-time data and treat ranking as a step separate from detection [15]. Severity measures combining detection, spatial extent, temporal persistence, and queue dynamics within one speed-only framework are still missing [7], and threshold-based detection parameters are seldom calibrated systematically and reproducibly [1].

The main objective is to develop and validate an FCD-based framework that detects, characterizes, and ranks recurrent bottlenecks on urban expressways using segment-level speed data only. The specific objectives are: (1) to define active bottlenecks from spatiotemporal speed patterns using normalized speeds; (2) to design a moving-window detection algorithm with adaptive downstream verification and temporal persistence criteria; (3) to introduce a Speed-only Bottleneck Severity Index (S-BSI) that ranks detected bottlenecks on a 0–100 scale using five components derived from the same speed data; and (4) to validate the framework using one year of weekday commercial FCD (248 days) from the D-200 urban expressway corridor in Bursa, Türkiye.

1.5. Paper Organization

The remainder of this paper is organized as follows. Section 2 reviews the related literature. Section 3 presents the detection algorithm and the severity index. Section 4 describes the study area and data. Section 5 reports and discusses the results. Section 6 concludes and outlines future research directions.

2. Literature Review

This section reviews recent work related to the proposed framework in four areas: (1) traffic congestion measurement and patterns, (2) Floating Car Data for traffic monitoring, (3) bottleneck identification methods, and (4) severity quantification and ranking. The section closes by summarizing the research gap that this study addresses.

2.1. Traffic Congestion Measurement and Pattern Recognition

Early area-wide indicators such as the Roadway Congestion Index and the Congestion Severity Index (CSI) related demand to supply and remain useful precedents for composite measures, but they describe whole areas and cannot show where or when congestion is worst [16]. Recent studies develop richer metrics from large mobility datasets. An analysis of six U.S. metros using distance–time measures from the space–time cube showed that city rankings change with the congestion level considered, so single-number comparisons can mislead [3]. Pattern-recognition approaches add the temporal dimension: 699 daily Traffic Performance Index profiles from Beijing have been clustered into eight characteristic patterns using an improved self-organizing map [2], and spatiotemporal patterns have been encoded as causal relation graphs for interpretable retrieval from large highway datasets [17].

At the city level, TomTom FCD benchmarks have been combined with network and contextual indicators into a multidimensional diagnostic framework, on the argument that single rankings hide the causes of poor mobility [7].

These studies show a clear move from simple averages toward spatial–temporal pattern recognition, but they work mainly at the network or city scale. A separate and equally important group of studies has developed indicators that measure congestion severity at the level of individual road segments, which are more directly related to bottleneck analysis.

Following a common classification, segment-level severity indicators can be grouped by the type of data they use [3]. The first group is speed-based. The Speed Performance Index (SPI), defined as the ratio of the mean segment speed to the maximum permitted speed, together with related speed-reduction ratios, measures segment congestion directly from speed [13]. The second group is travel-time-based. The Travel Time Index (TTI), the ratio of the actual travel time to the free-flow travel time, is one of the most widely used indicators [13], while travel-time reliability measures based on the statistical distance between observed and free-flow travel-time distributions have been used to rank freeway segments by recurrent bottleneck severity [15]. The third group is delay-based. The Congestion Severity Index (CSI) expresses total delay per unit of travel and was one of the first measures used to rank freeway systems by severity [16].

A more recent line of research argues that a single indicator is not enough, because congestion severity has three separate dimensions: intensity (the depth of the speed drop or delay), spatial extent (how far the queue reaches), and duration (how long it lasts) [18]. Vaziri proposed a comprehensive congestion index that combines these three dimensions for a roadway section into a single value [18], and Jun and Lim developed a composite congestion severity measure that multiplies the percent of congested time by the duration of continuous congestion; they showed that a single measure can rank the same segment very differently, so a combined measure gives a more reliable severity order [19]. At the bottleneck level, the FHWA bottleneck-intensity measure and the widely used RITIS impact factor rank locations by combining the duration and the spatial length of congestion over a long observation period [20]. However, most of these segment-level indicators need data that speed-only floating car data cannot provide, such as full travel-time distributions [15], traffic volume or occupancy to compute delay [16,20], or the fusion of speed, flow, and density [13]. Few indicators measure bottleneck severity from segment speeds alone, and even fewer add queue dynamics, such as queue length and queue growth rate, to the classical intensity–extent–duration description. This gap motivates the speed-only severity index proposed in this study, which is discussed further in Section 2.4.

Speed-based indicators have also been used to characterize road performance and level of service; operating-speed models, for instance, relate speed to geometric and operational conditions and thereby to service quality [21]. In parallel, speed-oriented traffic-management measures carry measurable sustainability benefits, and methodologies have been proposed to quantify the environmental gains of advanced traffic-management systems [22]. These strands support the use of speed as the primary operational indicator adopted in this study.

2.2. Floating Car Data in Traffic Monitoring and Prediction

FCD research has matured rapidly. A bibliometric review of 2000–2023 publications identified traffic flow state, safety, and route planning as the main clusters, reported an optimal probe investment of 3–8% of the fleet, and showed that sampling frequency mainly affects speed and heading accuracy [11]. A comparison of complete FCD speed distributions with radar using the Wasserstein distance found that probe data captured location-specific variability well, strengthening the case for speed-based analysis [10].

A second stream uses FCD and sensor data for prediction: statistical, machine-learning, deep-learning, and ensemble methods have been reviewed [1]; ensemble tree models with weather data have been shown to perform well for 24 h prediction [12]; and segments have been clustered by pattern variability, with specialized models assigned to forecast flows from FCD at scale [9]. These works confirm that FCD supports accurate network-level monitoring and prediction, but prediction answers when and how much congestion occurs, not where it is generated.

2.3. Bottleneck Identification Methods

Sensor-based methods established the basic detection logic. A well-established systematic algorithm identifies bottlenecks from loop-detector data via an upstream speed drop together with downstream speed recovery, and remains a key precedent [23]. Loop-detector data of different cycles have been fused—a large cycle to detect occurrence and a small cycle to fix location and duration—matching the classical cumulative-curves method on a Californian expressway with less processing [6].

Such methods work on instrumented freeways but cannot transfer to corridors without dense coverage [6]. FCD- and trajectory-based methods address this. A speed-difference criterion verified by a speed-at-capacity condition, applied in Beijing, showed that a large speed difference alone does not guarantee an active bottleneck, so secondary verification is needed [14]. Congestion diffusion has been modeled with percolation theory to identify dynamic critical bottlenecks, leading to a Systemic Congestion Index that captures spatial–temporal variation better than speed or the Travel Time Index [13]. Tracking of jam life cycles has shown that dissipation durations follow power laws, that jams dissolve nearly twice as slowly as they grow, and that early growth speed correlates strongly with maximal jam size [4]. A bottleneck-based management approach combining real-time detection, cost estimation, and decentralized control has been reviewed [5], and a demand-estimation method based on the linear speed–density relationship has been proposed, reducing travel-time errors at congested bottlenecks in Beijing and Los Angeles [24]. Overall, identification has progressed from point-sensor algorithms to network-level trajectory analytics, but most studies stop at detection or prediction; operational ranking of detected bottlenecks—especially from speed alone—has received little attention.

2.4. Severity Quantification and Ranking

Detection alone does not tell agencies which bottlenecks to treat first. Simple measures such as total delay capture only part of the problem [16]; a comprehensive measure should reflect how long congestion lasts, how far it spreads upstream, how deep the speed drop is, and how fast the queue grows [7]. The statistical distance between observed and free-flow travel-time distributions has been used to rank freeway bottlenecks, with three distance measures giving consistent rankings for recurrent bottlenecks [15]. Street-view imagery has been combined with a YOLOv5 model to identify slow-moving bottlenecks in Wuhan at 98.9% mean average precision [25]. Industry practice relies on simple speed rules, flagging bottlenecks when speeds fall below about 65% of a reference speed [26]. However, most severity measures require inputs unavailable in speed-only FCD—full travel-time distributions [17], street-view imagery [25], or detector flow and occupancy [6]—and few integrate the severity score with detection and with queue dynamics from the same data [7]. The proposed S-BSI responds to this need: it uses only segment-level FCD speeds, is computed inside the detection framework rather than after it, and adds two dynamic components—queue length and queue growth rate—that static composite measures miss.

2.5. Traffic Flow Theory Foundations for Speed-Only Analysis

Speed-only analysis needs a theoretical bridge from observed speeds to unobserved traffic variables. In the classical linear speed–density model, speed decreases linearly with density between free-flow speed and jam density [27].

Within this first-order framework, the upstream propagation, length, and growth rate of a queue can be estimated directly from a spatiotemporal speed matrix without direct flow or density measurements. Finally, normalizing observed speeds by the free-flow speed, following established speed-based detection practice [23,26], makes congestion thresholds transferable across segments with different speed limits. Classical transport-system modeling links macroscopic traffic variables through the fundamental diagram, and recent work shows that fusing heterogeneous data sources—for example loop detectors and floating car data—can improve estimation of the fundamental diagram and, hence, of prevailing traffic states [28]. The present study operates at the speed-only end of this spectrum, inferring queue dynamics from segment speeds under first-order theory; richer data fusion is a natural extension, which we pursue in future work.

2.6. Research Gaps and Contribution of This Study

The literature leaves four gaps. First, most FCD studies target general congestion measurement or prediction rather than bottleneck-specific detection, so the spatial sources of congestion remain hidden [1,13]. Second, speed-based methods use absolute or fixed relative thresholds and do not adapt downstream verification to the observed congestion level, so they can fail when severe congestion depresses both upstream and downstream speeds [14,23]. Third, severity quantification rarely integrates queue length and growth rate estimated from the same speed data; detection and ranking are usually separated with different data requirements [7,15]. Fourth, detection parameters are seldom selected through a systematic, reproducible search [1]. These gaps are most restrictive in the cities that most need bottleneck management, because dense detector infrastructure is rare outside a few well-instrumented corridors [9].

This study addresses these gaps with a unified, speed-only framework. It builds a normalized spatiotemporal speed matrix from one year of commercial FCD, applies a moving-window detection algorithm with adaptive downstream verification and temporal persistence filtering, and ranks detected bottlenecks with a five-component S-BSI on a 0–100 scale, where queue variables are estimated through a speed-based, first-order-consistent approximation of shockwave propagation—rather than a full LWR solution—from the same speed data. All detection parameters are calibrated through an exhaustive grid search over 128 candidate combinations, making parameter selection transparent and reproducible. Because every component derives from the same speed matrix, the method applies to any corridor covered by commercial FCD without additional instrumentation, and its outputs can directly feed ramp-metering feasibility assessments and corridor investment decisions [8]. The framework is validated on the D-200 urban expressway corridor in Bursa, Türkiye, using 248 weekdays of segment-level FCD. To our knowledge, this is the first framework combining adaptive bottleneck detection, queue-dynamics-aware severity scoring, and systematic calibration using segment-level speed data alone.

To synthesize the reviewed methods and clarify the position of the proposed framework,

Table 1. Comparison of representative bottleneck identification and severity-quantification methods in terms of data requirements, accuracy, and applicable scenarios.

Study/Method (Ref.)	Data Requirements	Reported Accuracy/Performance	Applicable Scenarios
[23]—loop-detector speed differential	Fixed loop detectors; 5-min speed and flow.	Benchmark method; 160 bottlenecks caused 64% of freeway delay.	Densely instrumented freeways.
[6]—dual-cycle loop-detector fusion	Loop detectors (30-s + 5-min); speed, flow, occupancy.	Matches the TCC method at about half the computation time.	Instrumented urban expressways.
[14]—FCD speed difference + speed-at-capacity	Link-level FCD 5-min speed; speed-at-capacity.	Found 10 recurrent bottlenecks (23.4 km, 48% of corridor). Fixed thresholds fail under severe congestion.	FCD-covered expressways; low-to-moderate congestion.
[13]—GPS-trajectory percolation (SCI)	Network-scale GPS trajectories; 5-min speed.	SCI captures congestion better than speed, RV, or TTI.	Road networks with congestion diffusion; segment-level estimation.
[15]—travel-time-reliability statistical distance	Full travel-time distributions (probe); incident records.	Consistent rankings; detects recurrent and non-recurrent bottlenecks.	Freeways with rich travel-time data.
[25]—street-view imagery + YOLOv5	Street-view images; trained YOLOv5 model.	98.9% mAP@0.5 in detecting bottleneck features.	Pedestrian/cycling bottlenecks; areas with street-view coverage.

Note: Accuracy metrics are not directly comparable, as the methods target different tasks and study sites.

compares representative bottleneck identification and severity-ranking approaches in terms of their data requirements, reported accuracy or performance, and applicable scenarios.

Most existing methods need data that speed-only FCD lacks—detector flow and occupancy [6,23], travel-time distributions [15], or street-view imagery [25]—or stop at detection without severity ranking [13,14]. The proposed S-BSI uses segment speed only, yet combines detection with a 0–100 severity ranking and queue dynamics, so it suits corridors with little or no fixed-sensor infrastructure.

3. Methodology

This section presents the complete analytical framework for identifying and assessing traffic bottlenecks from Floating Car Data (FCD). The methodology consists of six stages: data preparation and matrix construction, speed normalization, bottleneck trigger detection with adaptive downstream verification, severity scoring, queue propagation tracking with shockwave estimation, and systematic parameter calibration via grid search. Each stage is described below with its mathematical formulation.

3.1. Framework Overview

The proposed bottleneck detection framework consists of six sequential processing stages, as illustrated in Figure 1.

The framework is designed to process multi-day FCD archives, enabling identification of both individual bottleneck events and recurrent bottleneck patterns across extended observation periods.

3.2. Data Preparation and Space–Time Matrix Construction

3.2.1. Raw Data Structure

The input data consists of FCD speed records with the following attributes for each observation:

• Segment Identifier (j): Unique identifier for each road segment;
• Timestamp (t): Observation time at minute-level resolution;
• Speed (v): Measured vehicle speed in km/h;
• Geographic Coordinates: Segment centroid coordinates (latitude, longitude) or linear reference position.

3.2.2. Temporal Aggregation

Raw FCD observations are aggregated into discrete time intervals to reduce noise and ensure consistent temporal resolution. For each segment j and time interval t, the mean speed is calculated as:

$$\overline{v}\left(t,j\right)={\displaystyle \frac{1}{n_{t,j}}}{\sum }_{k=1}^{n_{t,j}}{v}_k(t,j)$$

(1)

where $n_{t,j}$ is the number of probe vehicle observations for segment j during time interval t, and $v_k(t,j)$ represents individual speed observations.

The analysis period is defined from 06:00 to 22:00 h to capture the full extent of daytime traffic operations including morning peak, midday, and evening peak periods. This 16-h window is discretized into one-minute intervals, producing 960 time steps per day.

In this study, one-minute aggregation is adopted because it provides consistent temporal resolution and sensitivity; the resulting 960-step daily profile is the direct input to the temporal-persistence test of Section 3.4.4.

3.2.3. Spatial Ordering

Accurate bottleneck detection requires proper spatial ordering of segments along the corridor to establish upstream–downstream relationships. Segments are sorted according to their geographic position along the direction of travel:

$$\mathrm{Segment} \mathrm{Order}: j_1,j_2,\dots ,j_N \mathrm{where} x_{j_1}<x_{j_2}<\cdots <x_{j_N}$$

(2)

where $x_j$ represents the linear position of segment j along the corridor and N is the total number of segments.

For bidirectional facilities, separate analyses are conducted for each travel direction (e.g., eastbound and westbound), with segment ordering reversed as appropriate to maintain consistent upstream–downstream orientation.

Here, segments are ordered by longitude because the study corridor runs east–west; this ordering defines the upstream–downstream relationships on which the moving-window detector (Section 3.4) and the upstream queue tracer (Section 3.6) directly depend, so that a consistent geographic order is a prerequisite for correct detection. Figure 2 shows the temporal and dimensional classification of the FCD data used in the study.

3.2.4. Spatiotemporal Speed Matrix

For each direction, the raw records are aggregated into a speed matrix H with dimensions $\times \mathit{N}$, where T is the number of time intervals and N is the number of road segments ordered sequentially along the corridor. Each element $H_{t,n}$ represents the average observed speed (in km/h) at segment n during time interval t.

When multiple days of data are available, the matrix is computed as the element-wise arithmetic mean across all observation days:

$$H_{t,n}={\displaystyle \frac{1}{D}}{\displaystyle {\sum }_{d=1}^D}{H}_{t,n}^{\left(d\right)}$$

(3)

where D is the number of observation days and $H_{t,n}^{\left(d\right)}$ is the speed recorded on day d. This averaging filters out non-recurring events (incidents, weather disruptions) and isolates the recurring congestion patterns that reflect systematic demand–capacity imbalances.

Missing values in the matrix are filled using linear interpolation along the temporal axis. Linear interpolation preserves the temporal continuity of speed transitions without introducing artificial discontinuities. This matrix representation enables efficient implementation of the moving window detection algorithm and facilitates visualization through spatiotemporal contour maps.

$$\mathbf{H}=\left[\begin{array}{cccc}v(t1,j1)& v(t1,j2)& \dots & v(t1,jN)\\ v(t2,j1)& v(t2,j2)& \dots & v(t2,jN)\\ \dots & \dots & \dots & \dots \\ \dots & \dots & \dots & \dots \\ v(t_T,j_1)& v(t_T,j_2)& \dots & v(t_T,j_N)\end{array}\right]$$

(4)

In the implementation, the average is computed by pooling all raw observations that fall into cell (t, n) across the D days; this is identical to the day-wise mean of Equation (3) whenever each day contributes one aggregated record per cell, and otherwise corresponds to an observation-weighted mean.

3.3. Speed Normalization

3.3.1. Rationale for Normalization

Raw speed values are influenced by segment-specific characteristics including speed limits, geometric design, and ambient traffic conditions, making direct comparison across segments problematic. Speed normalization addresses this heterogeneity by expressing observed speeds as a proportion of the Free-Flow Speed (FFS), yielding a dimensionless congestion indicator that enables consistent threshold application across the corridor.

3.3.2. Free-Flow Speed Estimation

Free-flow speed (FFS) is the average speed of vehicles under low-volume, unconstrained conditions. Under these conditions, drivers can travel at their desired speed without being affected by other vehicles [29]. FFS is a key parameter in traffic analysis. It defines the upper boundary of the speed-flow curve and is a key input to capacity analysis of a road segment [29].

In this study, FFS is set manually at 82 km/h based on field observations and the posted speed limit of the corridor. This value is applied uniformly across all segments (corridor-wide FFS mode). The decision to use a fixed FFS rather than a data-derived value was made for two reasons. First, the D-200 corridor has a relatively uniform geometric design along its mainline, which justifies a single FFS value. Second, the commercial FCD used in this study caps vehicle speeds at the posted speed limit of 82 km/h. As a result, free-flow conditions above this threshold cannot be directly observed from the data. Therefore, the posted speed limit of 82 km/h is adopted as the FFS. This decision reflects two constraints: the uniform geometric design of the corridor and the truncation of commercial FCD speeds at the posted speed limit. This truncation is a known limitation of commercial FCD products that prevents observation of actual free-flow speeds above this threshold [30].

Using a single corridor-wide FFS is appropriate when geometric conditions and speed limits are relatively uniform across segments. Standard practice permits a constant FFS based on agency policy for a given facility type or speed limit [31], which supports this approach for corridors with homogeneous design characteristics.

3.3.3. Normalized Speed Calculation

The normalized speed $S_{(t,j)}$ is calculated as the ratio of observed speed to Free-Flow Speed:

$$S_{t,j}={\displaystyle \frac{v_{t,j}}{FFS}}$$

(5)

This is with FFS = 82 km/h:

$$S_{t,j}={\displaystyle \frac{v_{t,j}}{82}}$$

(6)

A normalized value of 1.0 indicates free-flow conditions; values below the threshold y indicate congestion relative to the chosen reference. Values near 0 indicate near-standstill congestion. For example, a value of $S_{t,j}=0.60$ means that the observed speed is 60% of the free-flow speed (i.e., 49.2 km/h for FFS = 82 km/h).

3.4. Moving Window Detection Algorithm

3.4.1. Detection Parameters

For example, with $x=30$, the threshold is y = 0.70, meaning that any segment with $S_{t,n}\le 0.70$ (i.e., speed ≤ 57.4 km/h) is considered congested. With δ = 5% and y = 0.70, the persistence threshold becomes $y_{tol}=0.735$.

3.4.2. Congestion Threshold Definition

The bottleneck detection algorithm requires a threshold value y that distinguishes congested conditions from free-flow operations:

$$y=1-{\displaystyle \frac{x}{100}}$$

(7)

where x represents the percentage speed reduction from free-flow conditions that indicates congestion onset. Typical values of x range from 20% to 50%, corresponding to threshold values y between 0.50 and 0.80. This range is broadly consistent with the speed reductions associated with the transition from LOS D through LOS F in the HCM framework [31], although the HCM defines freeway LOS boundaries primarily in terms of density rather than speed ratios. A normalized speed below y indicates that the segment is operating in congested conditions.

In this study, four different x values (20, 30, 40, and 50) are included in the grid search calibration.

3.4.3. Bottleneck Definition (Condition 1 & 2)

A location is classified as an active bottleneck when it exhibits the characteristic spatial signature of upstream congestion combined with downstream recovery. This definition distinguishes true bottlenecks from spatially uniform congestion.

A bottleneck trigger is registered at space–time location $(t,j)$ if the following conditions are satisfied simultaneously:

Condition 1—Upstream Congestion: All L upstream segments ending at position j must have normalized speeds at or below the congestion threshold:

$$S_{t,i}\le y \forall \hspace{0.33em}i\in \{j-L+1,\dots ,j\}$$

(8)

This condition ensures that the detected location represents the head of a congestion zone with sufficient spatial extent, rather than an isolated low-speed reading at a single segment.

Condition 2—Downstream Recovery: The downstream segments must exhibit higher speeds than the congested upstream segments. The study employs an adaptive downstream verification mode that evaluates the speed differential between upstream and downstream regions:

$${\overline{S}}_{dn,k}\ge {\overline{S}}_{tj}+{\displaystyle \frac{p}{100}}$$

(9)

where ${\overline{S}}_{tj}$ is the normalized speed at the bottleneck boundary segment (the most downstream of the L congested upstream segments), ${\overline{S}}_{dn,k}$ is the normalized speed of the k-th downstream segment, and p is an additive offset parameter expressed in percentage points of the free-flow speed, converted to normalized-speed units through the term p/100 (e.g., p = 10 corresponds to a margin of 0.10). The condition requires every one of the K downstream segments to be faster than the boundary segment by at least p. Referencing the boundary segment rather than the average of all upstream segments keeps the criterion independent of the chosen upstream extent L.

The advantage of this method over conventional absolute-threshold approaches is its adaptivity to congestion severity. During severe congestion, when the boundary speed is very low (e.g., ${\overline{S}}_{t,j}=0.30$), the downstream requirement adjusts downward accordingly (${\overline{S}}_{dn}\ge 0.40$ for $p=10$), allowing detection of bottlenecks even when downstream conditions are also somewhat degraded. Conversely, during mild congestion (${\overline{S}}_{t.j}=0.65$), the downstream requirement rises (${\overline{S}}_{dn}\ge 0.75$), maintaining detection specificity.

This behavior is consistent with classical traffic flow theory: when a bottleneck is active, the upstream and downstream sections occupy distinct branches of the flow–density relationship—queued and unrestricted, respectively—thereby producing a measurable speed differential across the bottleneck location [32].

3.4.4. Temporal Persistence Filtering

To eliminate transient speed fluctuations, a temporal persistence criterion is applied. A candidate bottleneck at segment j must remain congested for the trigger interval plus the following m intervals (m + 1 intervals in total).

$$S_{t+\tau ,j}\le y_{tol} \forall \hspace{0.33em}\tau \in \left\{\mathrm{0,1},\dots ,m\right\}$$

(10)

where $y_{tol}=y\times \left(1+{\displaystyle \frac{\delta }{100}}\right)$ is the relaxed threshold that accounts for minor speed fluctuations. This tolerance follows established practice: a small relaxation of the persistence criterion reduces the false negatives caused by measurement noise [33]. By construction, triggers cannot be registered at the first L−1 segments, the last K segments, or the last m time intervals of the analysis window. Given the corridor length (747 segments) and the 960-step window, this boundary effect is negligible.

3.5. Speed-Only Bottleneck Severity Index (S-BSI)

Each detected bottleneck trigger is assigned a severity score on a 0–100 scale through a weighted multi-criteria index. The S-BSI integrates five normalized components that capture different dimensions of bottleneck impact.

Corridor-adaptive normalization: Rather than normalizing the spatial and temporal components by the search parameters (which would make every detection score near 1) or by the full corridor length (which would make every score near 0), the S-BSI normalizes them by data-driven references derived from the corridor’s own congestion distribution (the 95th percentile of consecutive below-threshold runs). This places the components in a meaningful [0, 1] range and yields severity scores that are comparable both across parameter settings and, in principle, across corridors with similar reference statistics.

3.5.1. Component Definitions

All five components are mapped to the [0, 1] range before aggregation.

Table 2. The parameters used in the algorithm.

Parameter	Symbol	Description
Congestion threshold	y	Defined as $y=1-x/100$, where x is the percentage speed reduction below FFS
Upstream spatial extent	L	Minimum number of consecutive upstream segments that must be congested
Downstream verification window	K	Number of consecutive downstream segments examined for flow recovery
Temporal persistence	m	Minimum number of consecutive time intervals during which the bottleneck must remain congested
Persistence tolerance	δ	Relaxation of the congestion threshold during the persistence check, defined as $y_{tol}=y\times (1+\delta /100)$; held constant at δ = 5% throughout this study.
Downstream recovery offset	p	Minimum normalized-speed margin by which each of the K downstream segments must exceed the boundary segment, expressed as a percentage and applied as S-dn,k ≥ S-tj + p/100.

gives the complete specification; the definitions and rationale follow.

The complete, self-contained specification of the proposed index is given in

Table 3. The S-BSI specification (self-contained definition).

Purpose	Rank Active Bottlenecks by Operational Severity on a 0–100 Scale.
Input	A space–time matrix of segment-level mean speeds (km/h). No flow or density data are required. Reference speed. A single corridor-wide free-flow speed (FFS); here FFS = 82 km/h. Congestion threshold. $y=1-{\displaystyle \frac{x}{100}},$ where x is the percentage speed reduction below FFS
Five components (each normalized to [0, 1])	1. Spatial extent: $c_L=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,{\displaystyle \frac{l_{count}}{L_{ref}}}\right)$ 2. Temporal persistence: $c_m=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,{\displaystyle \frac{m_{count}}{m_{ref}}}\right)$ 3. Congestion intensity: $c_g=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,\hspace{0.33em}{\displaystyle \frac{(y-S_{t,j})}{\mathrm{y}}}\right)$ 4. Queue length: $c_{ql}=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,\hspace{0.33em}{\displaystyle \frac{Q_L}{L_{ref}}}\right)$ 5. Queue growth rate: $c_{qg}=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,\hspace{0.33em}{\displaystyle \frac{(Q_L/m_{\mathrm{c}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{t}})}{R_{max}}}\right)$$R_{max}=1.0$ segment per time step
Corridor-adaptive references	$L_{ref}$ and $m_{ref}$ are the 95th percentiles of the lengths of consecutive below-threshold spatial and temporal runs in the speed matrix, using a congestion criterion of 0.5 × FFS (= 41 km/h). They are computed once per travel direction, so all parameter combinations share the same reference and the scores remain comparable.
Aggregation (weighted sum, weights w):	S-BSI $= {\displaystyle \frac{w_L\cdot c_L+w_m\cdot c_m+w_g\cdot c_g+w_{ql}\cdot c_{ql}+w_{qg}\cdot c_{qg}}{\sum w_i}}\times 100$ with $w_L=2.0,$$w_m=3.0,$$w_g=3.0,$$w_{ql}=1.5,$$w_{qg}=1.5$ (Σw = 11.0).
Critical classification	A bottleneck is labeled critical when S-BSI ≥ 35.

.

Component 1—Spatial Extent ($c_L$): measures how far upstream the congestion extends at the trigger location.

$$c_L=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,{\displaystyle \frac{l_{count}}{L_{ref}}}\right)$$

(11)

Here l_count is the number of consecutive congested segments immediately upstream of the trigger (counted under the operational threshold y), and L_ref is an empirical reference computed once per direction as the 95th percentile of the lengths of consecutive below-threshold spatial runs in the speed matrix, where a cell is treated as congested if its speed is at or below 50% of the free-flow speed (0.5 × 82 = 41 km/h). This 0.5 ratio is a pragmatic benchmark for severe congestion (travel time ≈ twice free-flow) and is not drawn from the HCM level-of-service scheme, which defines freeway LOS by density rather than speed ratio [31]. Because L_ref uses this fixed, severity-anchored criterion while y varies across the grid, the severity scale stays identical across all parameter combinations: y decides whether a segment is congested, and 0.5 × FFS is the common yardstick against which spatial extent is scored. A value approaching 1.0 indicates a queue as wide as the most severe extents observed on the corridor.

Component 2—Temporal Persistence ($c_m$): Measures how long the bottleneck remains active:

$$c_m=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,{\displaystyle \frac{m_{count}}{m_{ref}}}\right)$$

(12)

where $m_{count}$ is the number of consecutive intervals during which the trigger segment stays congested—measured with the relaxed threshold $y_{tol}$ of Equation (12), so brief fluctuations slightly above y do not truncate the duration—and $m_{ref}$ is the temporal analogue of $L_{ref}$ (95th percentile of consecutive below-threshold temporal runs under the same 41 km/h criterion). A value near 1.0 indicates a duration comparable to the longest recurring episodes on the corridor.

Component 3—Congestion Intensity ($c_g$): Measures the magnitude of the speed drop below the congestion threshold:

$$g=\max \left(0,\hspace{0.33em}y-S_{t,j}\right)$$

(13)

$$c_g=\mathrm{m}\mathrm{i}\mathrm{n}\left(1,\hspace{0.33em}{\displaystyle \frac{g}{\mathrm{y}}}\right)$$

(14)

Dividing the gap g by the threshold y spreads the score smoothly from 0 (speed exactly at the threshold) to 1 (complete standstill). This threshold-relative normalization keeps the intensity component sensitive across the full congested range and avoids the early saturation that a small fixed constant would cause.

Component 4—Queue Length ($c_{ql}$): Measures the spatial extent of the queue that forms upstream of the bottleneck:

$$c_{ql}=min\left(1,\hspace{0.33em}{\displaystyle \frac{Q_L}{L_{ref}}}\right)$$

(15)

where ${ Q}_L$ is the number of segments in the queue—the bottleneck segment plus the congested segments traced upstream of it by the queue tracker (Section 3.6)—normalized by the same $L_{ref}$ as Component 1, keeping the two on a common scale.

Component 5—Queue Growth Rate ($c_{qg}$): Measures how rapidly the queue extends upstream:

$$c_{qg}=min\left(1,\hspace{0.33em}{\displaystyle \frac{R_g}{R_{max}}}\right)$$

(16)

where $R_g={\displaystyle \frac{Q_L}{m_{count}}}$ is the queue growth rate (segments per time step) and $R_{max}$ = 1.0 segment per time step is a normalization constant. $R_{max}$ is defined directly in segment-per-time-step units rather than being converted to a fixed kinematic-wave speed; a queue that extends at or above this rate receives the maximum growth score.

3.5.2. Weighted Aggregation

The five components are combined through a weighted sum:

$$S-BS{I}_{raw}={\displaystyle \frac{w_L\cdot c_L+w_m\cdot c_m+w_g\cdot c_g+w_{ql}\cdot c_{ql}+w_{qg}\cdot c_{qg}}{\sum w_i}}\times 100$$

(17)

where the weights are: $w_L=2.0$ (spatial extent), $w_m=3.0$ (persistence), $w_g=3.0$ (congestion intensity), $w_{ql}=1.5$ (queue length), and $w_{qg}=1.5$ (queue growth), giving a total weight sum of 11.0.

The component weights were not assigned arbitrarily but were derived through the Analytic Hierarchy Process (AHP) [34], based on the judgments of five transportation experts. Each expert carried out pairwise comparisons of the five severity components using Saaty’s fundamental 1–9 scale, and the aggregated comparison matrix was solved for its principal eigenvector to obtain the relative priority of each component. The reliability of the expert judgments was verified through a consistency check of the comparison matrix. The principal eigenvalue was λ_max = 5.190, yielding a Consistency Index of CI = 0.048 for n = 5 criteria. With the corresponding Random Index RI = 1.120, the Consistency Ratio was CR= 0.042. Because this value is well below the commonly accepted threshold of 0.10 [34], the expert judgments are regarded as highly consistent and the resulting weights as reliable. For operational convenience, the priority vector obtained from the principal eigenvector was scaled and rounded to the working weights w = (2.0, 3.0, 3.0, 1.5, 1.5), preserving the rank order and the approximate ratios of the expert-derived priorities (Σw = 11.0).

3.5.3. Critical Bottleneck Classification

Critical bottlenecks represent priority candidates for traffic management intervention. Critical bottleneck classification threshold separates high-impact bottlenecks that warrant priority attention from moderate or minor ones. This is not a universal constant but an adjustable filter value that reflects the sensitivity preferences of the analyst or the transport agency. Raising the threshold produces a smaller, more selective set of high-severity bottlenecks, whereas lowering it admits a larger set of moderately severe locations. Because this filter is applied to the fully specified 0–100 S-BSI scale, the selection remains transparent and fully reproducible, and the threshold can be readily adapted to different corridors, operational priorities, or budget constraints without altering the underlying severity computation.

3.5.4. Spatial Aggregation of Critical Segments

Because the analysis is performed on a segment-by-segment basis, a single physical bottleneck can cause several consecutive segments to exceed the critical threshold. When two or more adjacent segments are simultaneously classified as critical, they are treated as belonging to the same bottleneck rather than as separate events, and the bottleneck is attributed to the most downstream segment in the direction of travel—that is, to the front of the congested group. This rule follows directly from active-bottleneck theory: a bottleneck is a localized point of capacity reduction at which congestion originates, while the resulting queue forms and propagates upstream [32]. The downstream-most critical segment therefore represents the actual location of the capacity constraint, whereas the upstream critical segments represent the spatial extent of the queue generated by it.

This merging is applied only to uninterrupted sequences of critical segments. If a non-critical segment lies between two critical segments, the two groups are kept separate and reported as distinct bottlenecks. The reason is that the input data are segment-level average speeds: a segment that does not meet the critical threshold indicates that, on average, traffic recovers along that segment. Such a spatial gap implies that the congestion is not continuous and that the two groups are driven by separate capacity constraints rather than by a single propagating queue. Merging them would incorrectly combine two independent events and overstate the spatial extent of the queue.

3.6. Queue Propagation Tracking

For each detected bottleneck trigger, the framework traces the upstream propagation of the resulting queue directly on the space–time speed matrix.

Upstream Queue Tracing

Starting from the bottleneck location (t,j), the queue is traced upstream by iteratively checking whether each preceding segment is also congested. The tracing proceeds as follows: beginning at segment $j-1$ and moving upstream, each segment n is checked at a progressively later time step $t+\mathsf{\Delta }$ (where Δ increments by 1 for each segment traversed) to account for the finite propagation speed of the backward-traveling wave. A segment is included in the queue if its normalized speed satisfies $S_{t+\mathsf{\Delta },n}<y$.

This diagonal tracing through the space–time matrix follows a backward-and-downward direction that is qualitatively consistent with the upstream propagation of congestion described by first-order traffic flow theory. Because the trace advances exactly one segment per time step, it constitutes a conservative approximation of the instantaneous queue extent rather than an exact reconstruction of the kinematic wave trajectory. The procedure continues until a segment with $S_{t+\mathsf{\Delta },n}\ge y$ is encountered or the upstream boundary of the corridor is reached.

The output is a propagation path—an ordered list of (t,n) pairs representing the space–time footprint of the queue—and the total queue length, measured both in number of segments and in kilometers using actual segment lengths.

Queue propagation constraints: the queue-length $c_{ql}$ and queue-growth $\left(c_{qg}\right)$ components of the S-BSI are computed for every detected trigger; consequently, the persistence parameter m in Equation (12) is the only duration control anywhere in the detection and scoring pipeline. For the descriptive queue-event report, adjacent events on the same segment separated by no more than one time step are merged into a single event so that one physical queue is not counted several times; this merging affects only the report and not the severity scores. When events are merged, the persistence component ($C_m$) of each constituent trigger retains its own individually measured duration in the severity computation.

Because the floating car data provide segment-level speeds without direct flow or density measurements, the queue indicators used in this study are derived from speed information under simplifying assumptions. The upstream queue extent is approximated from the spatial and temporal pattern of below-threshold segment speeds rather than from a measured flow–density relationship, consistent with classical first-order traffic-flow theory. Accordingly, these indicators should be interpreted as speed-based proxies for queue dynamics rather than as exact kinematic-wave measurements.

3.7. Grid Search Parameter Calibration

The detection framework involves several interacting parameters whose optimal values depend on corridor-specific characteristics. Rather than fixing these by expert judgment alone, an exhaustive grid search evaluates all combinations of candidate values and reports their trade-offs, so that the final configuration is calibrated to the corridor rather than chosen a priori.

3.7.1. Search Space

The total number of parameter combinations is 4 × 4 × 2 × 2 × 2 = 128. Each combination is evaluated independently for the selected eastbound direction. The combination features included in the grid search in this study are given in

Table 4. Parameter dimensions examined in grid search.

Parameter	Criterion	Values Tested	Levels
Speed deficit threshold	x (%)	20, 30, 40, 50	4
Upstream spatial extent	L (segments)	1, 2, 3, 4	4
Downstream verification window	K (segments)	1, 2	2
Temporal persistence	m (time steps)	5, 10	2
Downstream additive offset	p (%)	5, 10	2

Note that the persistence tolerance δ is not treated as a search dimension; it is held constant at δ = 5% across all 128 combinations. The five searched parameters are therefore x, L, K, m, and p.

.

Congestion threshold y ∈ {0.80, 0.70, 0.60, 0.50}: This range spans threshold values commonly used in the literature and in operational practice. Commercial bottleneck detection uses approximately 65% of free-flow speed as the activation threshold [26], and one commercial speed-bucket scheme classifies speeds between 32% and 62% of free-flow as heavy congestion and 63–92% as moderate [35]. Academic work has likewise emphasized that the threshold is facility-specific and should be calibrated rather than fixed a priori [33,36].

Upstream extent L ∈ {1,2,3,4}: The minimum number of consecutive congested segments. Higher L values produce fewer but more spatially significant detections.

Downstream window K ∈ {1,2}: The number of downstream segments checked for speed recovery.

Temporal persistence m ∈ {5,10}: The minimum duration in time steps (minutes).

Downstream recovery offset p ∈ {5,10}: The additive downstream margin, expressed as a percentage; for example, p = 10 corresponds to a 0.10 margin in normalized-speed units.

3.7.2. Evaluation Procedure

For each combination, the detection algorithm is executed on the normalized speed matrix, producing a set of triggers with associated severity scores and queue-propagation metrics. Summary statistics per trial include the number of triggers, the mean and maximum severity, the mean congestion gap, contiguity (the degree to which triggers cluster in space–time rather than appearing in isolation), and queue-propagation statistics.

The grid search does not optimize a single scalar objective. Instead it produces a comprehensive results table that lets the analyst weigh detection sensitivity (number of triggers) against specificity (severity distribution) and select the combination most appropriate for the corridor. This multi-criterion philosophy reflects the established view that traffic-model calibration should be judged against several goodness-of-fit measures rather than a single metric [37].

4. Case Study: Data and Study Area

This section describes the empirical application of the proposed bottleneck detection framework to an urban expressway corridor in Bursa, Türkiye. The study area characteristics, FCD source, preprocessing procedures, and data quality considerations are presented to establish the context for the results discussed in Section 5.

4.1. Study Area Description

4.1.1. Geographic Context

Bursa is the fourth-largest city in Türkiye, with a metropolitan population of over 3.1 million. It is located in the Marmara Region, roughly 100 km south of Istanbul. The city is a major industrial center, particularly in automotive and textile manufacturing, and sits on the İstanbul and Ankara transportation corridor. This geographic position generates heavy inter-city traffic alongside daily urban commuting.

4.1.2. D-200 Urban Expressway Corridor

The study corridor is a section of the D-200 state highway that crosses the Bursa metropolitan area in an east–west orientation. The D-200 serves as the primary urban expressway, linking eastern industrial zones to western residential and commercial districts.

Table 5. Key characteristics of the D-200 study corridor.

Characteristic	Value
Total Length	31.254 km
Orientation	East–West
Traffic Flow Direction	Eastbound
Lane Configuration	2–3 lanes per direction
Facility Type	Urban expressway with uninterrupted mainline flow and partial access control (grade-separated interchanges and ramps). A single signalized intersection at the eastern terminus is included only as a validation reference point.
Speed Limit	82 km/h
Access Control	Partial access control and expressway-style ramps
Number of Segments	747 Segments

summarizes the key characteristics of the corridor. Figure 3 shows the general view of the study corridor.

The study corridor is classified as an urban expressway for several reasons. Along the analyzed section there are no signalized intersections, so the traffic stream is not interrupted by signal control. The corridor has one of the highest posted speed limits within the Bursa metropolitan area and functions as the main east–west backbone connecting the two sides of the city. According to 2025 data, the average annual daily traffic on the western section of the route is 77,128 [38]. Continuous flow is maintained mainly through grade-separation structures such as grade-separated interchanges, vehicular overpasses, and pedestrian overpasses. Nevertheless, the facility does not provide full access control along its entire length: some sections contain parking bays and bus stops, others have a high density of side-street connections, and one portion is bordered by commercial establishments with direct accesses to the mainline. These characteristics place the corridor between a fully access-controlled freeway and a conventional urban arterial, which is consistent with the partial access control reported in Table 5.

When the study corridor was defined in the eastbound direction, the signalized intersection located at the eastern end of the urban expressway—where the express section terminates—was deliberately included in the analysis route. This intersection was retained as a reference point in order to facilitate the comparison of the outputs of the proposed method and methodology, since an isolated signalized intersection is expected to act as a strong and recurrent capacity constraint. As the results confirm, this signalized intersection was indeed detected as a critical bottleneck, which provides additional qualitative support for the validity of the detection framework.

The corridor carries both through traffic and local commuting demand. On-ramp and off-ramp locations along the corridor create merge and diverge points that are potential bottleneck locations. The mix of at-grade intersections and grade-separated interchanges produces varying capacity conditions along the route.

4.2. FCD Source and Characteristics

4.2.1. Data Provider

The FCD used in this study was obtained from a commercial traffic data provider operating in Türkiye. Raw data includes segment-based speed and time information at one-minute intervals. The data was processed using Python (3.12.3) programming language with operations such as matching and geolocation.

The commercial FCD used in this study was provided as an aggregated segment-level traffic product. For each road segment and timestamp, the dataset includes speed and travel time information, but it does not include the number of contributing probe vehicles, fleet composition, or the penetration rate of the provider in the traffic stream. Therefore, the dataset supports segment-level traffic performance analysis, but not a direct estimation of probe sample size or market penetration.

4.2.2. Spatial Resolution

The road network is divided into segments according to the provider’s proprietary segmentation scheme. Each segment has a unique identifier (Segment ID). Segment lengths vary depending on road geometry and intersection density. On the study corridor, typical segment lengths range from approximately 2.5 to 60 m. Segment coordinates (start point, end point, and midpoint) are provided in WGS-84 geographic coordinates, allowing spatial ordering and length computation via the Haversine formula. Segments are sorted by their geographic position (longitude) to establish a consistent upstream–downstream ordering. Eastbound segments are sorted in ascending longitude order; westbound segments in descending order.

4.2.3. Temporal Resolution

FCD observations are provided at one-minute temporal resolution, enabling detailed analysis of traffic dynamics during bottleneck formation and dissipation. The temporal parameters of the FCD dataset are summarized in

Table 6. Temporal parameters of the FCD dataset.

Parameter	Value
Daily Analysis Window	06:00–22:00 (16 h)
Temporal Resolution	1 min
Time Intervals per Day	960 intervals
Research Date Range	2 January 2025–30 December 2025
Number of Working Days:	248
Number of Working Data Rows:	178,031,016 (178 million)

.

Only weekday data between 06:00 and 22:00 was used. This 16-h window captures morning peak, midday, and evening peak periods while excluding overnight hours with minimal traffic. Weekend data was excluded to focus on recurring commuter congestion patterns. Data for weekends and public holidays in 2025 were not included in the study. The study processed over 178 million rows of data for weekdays over a one-year period.

4.2.4. Data Attributes

Eastbound direction is analyzed in this study. Since coordinate data was not found in the raw data, this data was added later for use in the study. A complete spatiotemporal speed matrix is constructed as described in Section 3. The raw speed feed did not contain segment coordinates; WGS-84 coordinates for each segment (start, end, and midpoint) were obtained from the provider’s segment-geometry reference and matched to the speed records. These coordinates were then used for spatial ordering and for length computation via the Haversine formula. The FCD data attributes used in this study are listed in

Table 7. Lists of the used attributes in the FCD dataset.

Attribute	Description	Unit
Segment ID	Unique segment identifier	-
Time Stamp	Observation date and time	YYYY-MM-DD HH:MM
Speed	Estimated segment speed	km/h
Coordinate X	Segment centroid easting	meters
Coordinate Y	Segment centroid northing	meters

.

4.3. Data Preprocessing

4.3.1. Data Cleaning

Raw CSV files are loaded using a chunk-based reading strategy to handle large file sizes efficiently. During loading, the following filters are applied:

Records with missing values in Segment ID, Speed, Local Time Stamp, or start_x are dropped. Timestamps are parsed into a standard datetime format for consistent processing.

4.3.2. Spatial Filtering

The analysis is restricted to segments within the defined study corridor. Segments on connecting local streets, service roads, and parallel facilities are excluded to focus on the mainline expressway. The geographic coordinates embedded in the data allow automatic spatial ordering without manual intervention.

4.3.3. Temporal Aggregation and Matrix Construction

Speed observations are aggregated into 1-min intervals. For each segment and time interval, the arithmetic mean of all available speed observations is computed. The aggregated data is organized into the spatiotemporal speed matrix H (dimensions T × N, where T = 960 time intervals and N is the number of segments). This matrix serves as the primary input for all subsequent analyses. An average of 248 days of data is calculated, and then a matrix with dimensions of 960 × 747 is created.

4.3.4. Multi-Day Averaging

The framework operates in “average” processing mode: speed matrices from all 248 weekdays are averaged into a single representative matrix for each direction. For each cell (t, j), the mean speed across all available days is computed:

$${\overline{v}}_{t,j}={\displaystyle \frac{1}{D}}{\displaystyle {\sum }_{d=1}^D}{v}_{t,j}^{\left(d\right)}$$

(18)

where D is the number of days with valid data for that cell. This averaging smooths day-to-day variability and highlights recurring congestion patterns rather than isolated incidents.

4.3.5. Missing Data Treatment

After matrix construction, remaining missing values (cells where no probe vehicle data was available) are filled using linear interpolation along the temporal axis. Interpolation is applied only within the interior of available data (i.e., leading and trailing gaps at the edges of the time series are not extrapolated). This approach preserves the temporal continuity of speed transitions, which is important for accurate bottleneck onset and offset detection.

Linear interpolation was chosen over alternatives (forward fill, mean imputation, or zero fill) because it maintains smooth speed transitions and does not introduce artificial discontinuities. Since the temporal resolution is 1 min and FCD coverage is generally dense during peak hours, there are no cells requiring interpolation. Because FCD coverage is dense during the 06:00–22:00 analysis window, no cells required interpolation in the constructed matrix; linear interpolation along the temporal axis is retained only as a fallback for interior gaps and was not triggered for this dataset.

4.4. Free-Flow Speed Determination

The free-flow speed (FFS) is a critical parameter, as it defines the reference condition for speed normalization. In this study, FFS is set manually at 82 km/h based on the posted speed limit of the corridor. This value is applied uniformly across all segments (corridor-wide FFS mode).

The decision to use a fixed FFS rather than a data-derived value was made for two reasons. The D-200 corridor has a relatively uniform geometric design along its mainline. Therefore, the posted speed limit of 82 km/h is adopted as the FFS, consistent with the conservative estimation approach suggested in the literature [29,30].

A sensitivity analysis performed using the configuration selected in the study (x30_L3_K2_m10_dn5). The number of triggered bottleneck points was computed separately for free-flow speeds (FFS) of 60, 70, 80 and 90 km/h, and compared with the baseline value of 82 km/h. The results are summarized in

Table 8. The number of bottleneck points triggered based on FFS change.

No:	FFS (km/h)	Trigger Count
1	60	5752
2	70	8626
3	80	10,662
4	82	11,123
5	90	11,832

.

The analysis shows that FFS has a strong and monotonic effect on the number of triggered bottleneck points: as FFS increases, the count consistently rises. However, the relationship is not strictly linear. The marginal increase declines at higher speeds (+2874 points from 60 to 70 km/h, but only +1170 from 80 to 90 km/h), indicating a concave, saturating trend and a diminishing sensitivity of the thresholds as FFS grows. Because all normalized thresholds depend directly on this reference speed, this behaviour confirms that the outputs are sensitive to the assumed congestion speed, but that the model responds in a smooth and predictable manner rather than abruptly.

4.5. Segment Length Computation

Segment lengths are computed from the start and end coordinates using the Haversine formula:

$$d=2R\mathrm{a}\mathrm{r}\mathrm{c}\mathrm{s}\mathrm{i}\mathrm{n}\left(\sqrt{{\mathrm{s}\mathrm{i}\mathrm{n}}^2\left({\displaystyle \frac{\mathsf{\Delta }\varphi }{2}}\right)+\mathrm{c}\mathrm{o}\mathrm{s}{\varphi }_1\mathrm{c}\mathrm{o}\mathrm{s}{\varphi }_2{\mathrm{s}\mathrm{i}\mathrm{n}}^2\left({\displaystyle \frac{\mathsf{\Delta }\lambda }{2}}\right)}\right)$$

(19)

where R = 6371 km is the Earth’s mean radius, $\mathit{\varphi }$ denotes latitude, and $\mathit{\lambda }$ denotes longitude. Segment lengths are used to express queue extents in kilometers and in the spatial calculations of the queue propagation analysis.

4.6. Data Quality Considerations

4.6.1. Penetration Rate

FCD accuracy depends on the probe vehicle penetration rate. The commercial provider does not disclose specific penetration rates. However, the density of observations during peak periods on the D-200 corridor suggests adequate coverage for reliable speed estimation. During off-peak hours and on minor connecting roads, observation density is lower, which may introduce greater uncertainty. Since the analysis focuses on peak-period bottleneck detection, this limitation is mitigated: penetration rates tend to be highest during congested conditions when more vehicles (particularly commercial fleet vehicles) are on the road. A limitation of the present dataset is that the commercial provider does not disclose probe vehicle counts, fleet composition, or penetration rate. Therefore, the representativeness of the FCD cannot be evaluated directly through vehicle sample size statistics. The available data consist of aggregated segment-level speed observations rather than raw probe trajectories. For this reason, the reliability assessment in this study is based on the temporal and spatial continuity of the reported speed records, especially during the peak periods used for bottleneck detection. This limitation should be considered when interpreting the results, and future studies should validate the FCD product against independent fixed-sensor data where such information is available.

4.6.2. Map-Matching Uncertainty

GPS positioning errors and map-matching algorithm limitations can cause speed observations to be assigned to incorrect segments, especially near interchanges and complex geometries. Segment-level aggregation provides some robustness to isolated map-matching errors, but systematic biases could affect detection accuracy at specific locations, particularly at merge and diverge areas.

4.6.3. Speed Estimation Processing

The commercial provider’s speed estimation algorithms incorporate proprietary processing steps that are not fully disclosed. Differences between instantaneous GPS-derived speeds and the provider’s segment speed estimates may arise from temporal smoothing, outlier filtering, and traffic state inference procedures. These processing steps generally improve estimate quality but may also introduce lag in detecting rapid traffic state transitions. Despite these limitations, GPS-based probe vehicle data have demonstrated acceptable accuracy for traffic monitoring applications in validation studies—from controlled field experiments [39] to evaluations of commercial FCD products [40], which found FCD to be a useful surrogate for macro-level urban speed monitoring despite reduced accuracy during heavily congested conditions—supporting their use for bottleneck detection research.

4.6.4. Directional Separation

The bottleneck formation mechanisms differ between directions. In order to perform the calculations mentioned in the article on the matrices, the segments need to be geographically sorted. Therefore, the segments were sorted geographically before processing.

5. Results and Discussion

This section presents the empirical results obtained by applying the proposed FCD-based bottleneck detection framework to the Bursa D-200 corridor. The analysis covers spatiotemporal speed patterns, grid search calibration outcomes, detected bottleneck characteristics, severity analysis, queue propagation and shockwave behavior, parameter sensitivity, and a summary of key findings.

5.1. Spatiotemporal Speed Patterns

Figure 4 presents the spatio-temporal speed contour map for the study corridor, showing the average speed values across all segments and time intervals throughout the analysis period. The horizontal axis represents the distance from the beginning of the corridor, and the vertical axis represents the analysis time interval. The color scheme ranges from yellow (high speeds, 65–80 km/h) to green (medium speeds, 45–65 km/h) and blue to orange (low speeds, below 30 km/h), while dark red indicates lower speeds (below 10 km/h).

Figure 5 presents the corresponding normalized speed heatmap (V/FFS %) for the eastbound direction, in which each segment speed is expressed relative to the free-flow speed.

The normalized speed heatmaps provide a comprehensive view of corridor performance across the 06:00–22:00 analysis window. Several key patterns are evident:

Congestion during peak hours. The direction exhibits a significant speed reduction during peak commute times. In the eastbound direction, the majority of the speed reduction occurs during evening rush hour (16:30–19:30). This directional asymmetry is consistent with commute patterns.

Spatial heterogeneity. The speed drop is not evenly distributed along the corridor. Instead, there are distinct sections where velocities drop sharply, while adjacent sections maintain higher velocities. This spatial variation suggests the presence of local capacity constraints—bottleneck geometric, demand, or control characteristics—rather than excessive density across the corridor.

These characteristic patterns are well described by classical traffic flow theory [32]. This framework builds on earlier theoretical foundations [41,42] and has been empirically confirmed through cumulative curve analysis [43]. Together, these patterns provide visual confirmation that the corridor contains true bottlenecks, where demand exceeds local capacity, rather than simply regions of uniformly slow traffic.

Average speeds around midday. Between approximately 11:00 and 16:00, higher speeds are observed in most sections, predominantly in the eastbound direction, compared to peak hours. This observation supports the use of peak period analysis for bottleneck detection and confirms that congestion in this corridor is primarily demand-driven rather than infrastructure limitations throughout the day.

Analysis of the spatiotemporal patterns reveals distinct congestion characteristics during morning and evening peak periods:

Morning Peak (07:30–10:30): Congestion initiates at multiple locations between 07:30 and 08:30, with queue extents reaching maximum values around 08:30–09:00.

Evening Peak (16:30–19:30): Congestion patterns emerge around 16:00, with peak severity occurring between 17:00 and 19:00.

Midday Conditions (11:30–16:00): Localized congestion events of shorter duration and lesser severity compared to peak periods.

The analyses suggest that the primary capacity constraint occurs during the evening peak hours (16:30–19:30). The congestion observed during this period not only extends travel durations but also significantly increases travel time uncertainty, thereby reducing the overall reliability of the transport system.

Figure 6a shows the mean upstream queue length of the bottlenecks whose onset falls within the morning (07:30–10:30) and evening (16:30–19:30) peak windows. For each bottleneck, the queue is traced upstream from the bottleneck segment along the consecutive segments that remain below the detection threshold, and its length is obtained by summing the true lengths of these segments, which were calculated from the GPS coordinates of the segment endpoints rather than assumed to be equal. As in the previous figure, n is the number of bottleneck events over which the mean and the standard deviation were computed (n = 25 in the morning and n = 17 in the evening) and includes all detected events with an onset inside the corresponding window. The mean queue length is 2.51 km for the morning window and 4.16 km for the evening window. The error bars denote ± one standard deviation; for the morning window the lower bar extends close to zero, which suggests that the dispersion between individual queues is of a magnitude similar to the mean, and the difference between the two windows should therefore be interpreted with caution. All values are derived from the 248-day average speed profile and thus describe a representative day rather than any individual date.

Figure 6b reports the mean duration of the bottlenecks whose onset falls within the morning (07:30–10:30) and evening (16:30–19:30) peak windows. For each detected bottleneck, the duration is the length of time for which the bottleneck segment remains at or below the tolerance-adjusted detection threshold, counted in one-minute steps from the moment at which the bottleneck is first identified; the values are therefore expressed in minutes and are obtained from the 248-day average speed profile. In this figure n is the number of bottleneck events available for averaging in each window (n = 25 in the morning and n = 17 in the evening); it includes all detected events with an onset inside the window, without applying the severity-based critical filter used in other parts of the study, and it is the sample size over which the mean and the standard deviation were computed. The mean duration is higher for the morning window (564.7 min) than for the evening window (207.1 min). It should be noted that each event is assigned to a window by its onset time only, so that its duration may continue beyond the end of that window; this feature, together with the smoothing effect of averaging speeds over many days, may contribute to mean durations that exceed the three-hour length of the windows. The error bars denote ± one standard deviation and are comparatively wide, particularly for the morning window, indicating substantial variation between individual events, so the means are best read as indicative values rather than as precise estimates.

5.2. Corridor Traffic Data

This section presents an analysis of travel times for the studied corridor, focusing on daily variations, intraday temporal fluctuations, and flow reliability.

The aggregate average travel time profile derived from all observed days demonstrates a distinct bimodal distribution, a pattern typical of urban arterial roads. Figure 7 shows the daily travel-time profile of the corridor, with the thin grey lines representing the 248 individual days and the solid line representing the daily mean. The profile is clearly bimodal: a moderate morning peak reaches a mean travel time of about 49 min around 09:00, whereas the evening peak is far more severe, rising to roughly 85 min around 18:00–18:30, compared with an off-peak baseline of about 38–40 min. The much wider day-to-day spread during the evening peak (individual days exceeding 120 min) also indicates that travel-time reliability deteriorates most sharply in the evening.

Figure 8 presents the corresponding corridor-mean speed over the day, which mirrors the travel-time profile. Mean speed begins near free-flow conditions (about 67 km/h) in the early morning, falls to roughly 47 km/h during the morning dip (around 09:00), partially recovers to about 55 km/h at midday, and then declines to its daily minimum of approximately 39.5 km/h during the evening peak (around 18:30) before recovering to about 64 km/h after 21:00. The deeper and longer speed reduction in the evening confirms that the evening peak is the corridor’s critical operating period.

Figure 9 shows the spatial speed profile along the 31 km corridor at four representative times of day (08:30, 12:00, 18:30, and 21:30), together with the FFS reference of 82 km/h. The 21:30 profile remains close to free-flow speed along the entire corridor, whereas the 18:30 profile exhibits deep, localized speed troughs—falling below 10 km/h near km 4–5 and km 15–17—while adjacent sections recover to much higher speeds. This pronounced spatial heterogeneity, with speed drops concentrated at a few recurrent locations rather than spread uniformly, provides direct visual evidence of localized capacity constraints (bottlenecks) rather than corridor-wide congestion.

Figure 10 compares the spatial speed profiles averaged over the morning peak (07:30–10:30), midday (11:30–14:00), and evening peak (16:30–19:30), where the shaded area represents the speed deficit relative to FFS. The midday period shows the smallest deficit, the morning peak an intermediate one, and the evening peak by far the largest, with the deepest deficits occurring repeatedly at the same locations (around km 4–5 and km 15–17). The spatial stability of these deficits across different peak periods reinforces that the underlying bottlenecks are fixed, demand-driven features of the corridor rather than random fluctuations.

Taken together, Figure 7, Figure 8, Figure 9 and Figure 10 establish that corridor performance is dominated by a severe and highly recurrent evening peak whose speed deficits are concentrated at a small number of fixed locations. These locations are the candidate bottlenecks that the detection framework is designed to identify and rank in the following sections.

5.3. Detected Bottleneck Characteristics

Calculations were performed based on the grid combinations given in Section 3.7. Some triggered bottleneck points obtained using the defined time interval (248 days of data) are plotted on the heat maps below. A detailed list of triggered bottlenecks is given in

Table A1. The results of grid search parameter combinations 1–32.

Comb. No	x_Percent (%)	L	K	m	DN Speed %	Trigger Count	Avg Severity (0–100)
1	20	1	1	5	5%	18,589	54.92
2	20	1	1	5	10%	6571	53.81
3	20	1	1	10	5%	18,515	45.02
4	20	1	1	10	10%	6535	42.99
5	20	1	2	5	5%	15,839	54.83
6	20	1	2	5	10%	4299	58.08
7	20	1	2	10	5%	15,777	45.04
8	20	1	2	10	10%	4275	46.12
9	20	2	1	5	5%	18,354	55.32
10	20	2	1	5	10%	6505	54.24
11	20	2	1	10	5%	18,285	46.38
12	20	2	1	10	10%	6473	44.87
13	20	2	2	5	5%	15,609	55.29
14	20	2	2	5	10%	4292	58.13
15	20	2	2	10	5%	15,552	45.48
16	20	2	2	10	10%	4268	46.16
17	20	3	1	5	5%	17,383	56.85
18	20	3	1	5	10%	5874	57.18
19	20	3	1	10	5%	17,319	47.81
20	20	3	1	10	10%	5846	47.53
21	20	3	2	5	5%	14,659	57.06
22	20	3	2	5	10%	4190	59.16
23	20	3	2	10	5%	14,607	46.94
24	20	3	2	10	10%	4170	47.32
25	20	4	1	5	5%	16,956	57.53
26	20	4	1	5	10%	5850	57.31
27	20	4	1	10	5%	16,892	48.22
28	20	4	1	10	10%	5822	47.66
29	20	4	2	5	5%	14,486	57.39
30	20	4	2	5	10%	4175	59.25
31	20	4	2	10	5%	14,434	47.29
32	20	4	2	10	10%	4155	47.37

,

Table A2. The results of grid search parameter combinations 33–64.

Comb. No	x_Percent (%)	L	K	m	DN Speed %	Trigger Count	Avg Severity (0–100)
33	30	1	1	5	5%	15,015	46.04
34	30	1	1	5	10%	5538	44.43
35	30	1	1	10	5%	14,885	39.3
36	30	1	1	10	10%	5500	38.33
37	30	1	2	5	5%	12,705	45.78
38	30	1	2	5	10%	4021	44.61
39	30	1	2	10	5%	12,586	39.12
40	30	1	2	10	10%	3988	38.48
41	30	2	1	5	5%	14,430	47.06
42	30	2	1	5	10%	5471	44.81
43	30	2	1	10	5%	14,327	41.55
44	30	2	1	10	10%	5442	40.22
45	30	2	2	5	5%	12,165	46.87
46	30	2	2	5	10%	3960	45.09
47	30	2	2	10	5%	12,072	40.78
48	30	2	2	10	10%	3935	38.86
49	30	3	1	5	5%	13,368	48.9
50	30	3	1	5	10%	5259	45.73
51	30	3	1	10	5%	13,283	43.42
52	30	3	1	10	10%	5236	41.7
53	30	3	2	5	5%	11,202	48.77
54	30	3	2	5	10%	3754	46.35
55	30	3	2	10	5%	11,123	42.54
56	30	3	2	10	10%	3735	40.82
57	30	4	1	5	5%	12,654	50.58
58	30	4	1	5	10%	5124	46.33
59	30	4	1	10	5%	12,616	44.98
60	30	4	1	10	10%	5109	42.46
61	30	4	2	5	5%	10,516	50.7
62	30	4	2	5	10%	3619	47.22
63	30	4	2	10	5%	10,482	44.24
64	30	4	2	10	10%	3608	41.87

,

Table A3. The results of grid search parameter combinations 65–96.

Comb. No	x_Percent (%)	L	K	m	DN Speed %	Trigger Count	Avg Severity (0–100)
65	40	1	1	5	5%	10,117	38.12
66	40	1	1	5	10%	4329	38.39
67	40	1	1	10	5%	10,008	33.62
68	40	1	1	10	10%	4287	34.32
69	40	1	2	5	5%	8101	38.28
70	40	1	2	5	10%	2883	40.6
71	40	1	2	10	5%	8005	33.7
72	40	1	2	10	10%	2850	36.17
73	40	2	1	5	5%	9503	39.49
74	40	2	1	5	10%	4116	39.42
75	40	2	1	10	5%	9438	35.3
76	40	2	1	10	10%	4085	35.51
77	40	2	2	5	5%	7534	39.89
78	40	2	2	5	10%	2689	42.25
79	40	2	2	10	5%	7479	35.43
80	40	2	2	10	10%	2666	37.47
81	40	3	1	5	5%	8852	40.49
82	40	3	1	5	10%	3824	40.16
83	40	3	1	10	5%	8802	36.29
84	40	3	1	10	10%	3801	35.93
85	40	3	2	5	5%	6953	40.92
86	40	3	2	5	10%	2399	43.77
87	40	3	2	10	5%	6913	36.45
88	40	3	2	10	10%	2384	38.38
89	40	4	1	5	5%	8066	41.63
90	40	4	1	5	10%	3586	40.71
91	40	4	1	10	5%	8026	38.24
92	40	4	1	10	10%	3568	36.81
93	40	4	2	5	5%	6191	42.37
94	40	4	2	5	10%	2161	45.07
95	40	4	2	10	5%	6160	38.74
96	40	4	2	10	10%	2151	40.09

and

Table A4. The results of grid search parameter combinations 97–128.

Comb. No	x_Percent (%)	L	K	m	DN Speed %	Trigger Count	Avg Severity (0–100)
97	50	1	1	5	5%	4897	33.69
98	50	1	1	5	10%	2156	36.54
99	50	1	1	10	5%	4769	30.3
100	50	1	1	10	10%	2101	32.07
101	50	1	2	5	5%	4244	33.91
102	50	1	2	5	10%	1873	39.73
103	50	1	2	10	5%	4167	30.44
104	50	1	2	10	10%	1858	34.62
105	50	2	1	5	5%	4613	34.67
106	50	2	1	5	10%	2010	38
107	50	2	1	10	5%	4500	31.32
108	50	2	1	10	10%	1963	33.32
109	50	2	2	5	5%	3998	34.85
110	50	2	2	5	10%	1753	41.21
111	50	2	2	10	5%	3935	31.2
112	50	2	2	10	10%	1746	35.71
113	50	3	1	5	5%	3747	35.56
114	50	3	1	5	10%	1465	37.99
115	50	3	1	10	5%	3657	31.38
116	50	3	1	10	10%	1419	31.35
117	50	3	2	5	5%	3146	35.9
118	50	3	2	5	10%	1213	42.52
119	50	3	2	10	5%	3106	31.19
120	50	3	2	10	10%	1207	34.37
121	50	4	1	5	5%	2779	37.33
122	50	4	1	5	10%	1062	38.41
123	50	4	1	10	5%	2718	32.02
124	50	4	1	10	10%	1020	29.41
125	50	4	2	5	5%	2216	37.85
126	50	4	2	5	10%	836	44.15
127	50	4	2	10	5%	2197	31.6
128	50	4	2	10	10%	832	32.6

. Depending on the combination parameters, the number of triggered bottlenecks was at least 832 in the 50_L4_K2_m10_dn10 (in the combination labels, dn denotes the downstream offset p) combination and at most 18,589 in the 20_L1_K1_m5_dn5 combination.

This trade-off between sensitivity and specificity is a fundamental feature of any threshold-based detection algorithm: relaxing the threshold increases the detection rate but also raises false positives, while tightening it reduces false alarms at the cost of missed events. This phenomenon has been documented in bottleneck detection contexts: the optimal parameter settings identified for a Portland, Oregon, corridor differed from those originally selected for a San Diego, California, corridor [23], and researchers in other cities were advised to perform a similar calibration analysis for their own networks [33]. It has also been addressed more formally within a Bayesian framework that represents the detection threshold as a probability distribution rather than a fixed value [36].

The persistence requirement of m = 5 min, combined with the 5% tolerance, ensures that detected bottlenecks represent sustained congestion events rather than transient fluctuations caused by signal cycles or measurement noise.

Figure 11, Figure 12 and Figure 13 show space–time speed heatmaps for the study corridor, with the identified bottleneck triggers for three of the 128 grid-search combinations.

The detected bottlenecks are spatially concentrated at a small number of corridor locations. This concentration pattern is consistent with bottleneck theory: capacity constraints tend to occur at locations where freeway geometry changes—such as sag vertical curves, tunnel entrances, and lane drops—rather than at arbitrary locations [32].

The fact that a small number of segments account for most detections indicates that the algorithm is identifying recurring capacity constraints rather than randomly distributed speed anomalies.

The spatial distribution of bottleneck detections reveals clustering near interchange areas and merge/diverge points. The segments with the highest detection frequencies are located near on-ramp merge areas, which is expected because merge traffic adds demand at specific points along the corridor. This observation is consistent with classical bottleneck theory, which holds that capacity constraints develop at fixed locations where freeway geometry changes, such as sag vertical curves, tunnel entrances, and lane drops [32]. Consistent with this, recurrent bottlenecks have been attributed to fixed physical geometries in FHWA work [20]. A systematic analysis of seven San Diego freeways similarly found that recurring bottlenecks were concentrated at a small number of fixed locations and accounted for the majority of corridor-wide delay, which is consistent with the spatial concentration observed in the present corridor [23].

Bottleneck activations show strong temporal regularity. This temporal consistency across multiple observation days confirms that the detected bottlenecks are demand-driven and recurring, which are the necessary conditions for bottleneck identification in traffic flow theory [44].

The temporal regularity also has practical implications: highly predictable bottleneck activation enables proactive traffic management strategies that anticipate congestion rather than merely reacting to it.

5.4. Bottleneck Severity Analysis

The five-component S-BSI enables differentiation between bottlenecks that are spatially localized and short-lived (low scores) and those that produce extensive queues persisting over long periods (high scores). This differentiation is one of the main advantages of a multi-component index over a simple threshold-based severity measure.

Bottlenecks with S-BSI ≥ 35 are classified as critical. The critical bottlenecks are concentrated at a small number of locations, with the same segments appearing repeatedly across different time intervals. This recurrence confirms that critical bottlenecks are tied to fixed infrastructure features rather than to stochastic demand variations.

The critical bottleneck map shows that the highest-severity locations coincide with the visually identifiable congestion hotspots in the space–time heatmap. This agreement between algorithmic detection and visual inspection provides qualitative validation of the detection framework.

Depending on the parameter combination, the number of critical triggers—those with S-BSI ≥ 35—ranged from 365 (50_L4_K2_m10_dn10) to 14,876 (20_L1_K1_m5_dn5); the corresponding counts for all triggers, irrespective of severity, are those reported in Section 5.3 (832–18,589).

Because no single scalar objective was optimized, the operating combination was chosen by jointly examining three criteria across the results table: (i) a moderate trigger count that avoids both over-detection and excessive filtering, (ii) a mean severity score in the mid-range that preserves discrimination between moderate and severe events, and (iii) strong spatial clustering of triggers onto a small number of recurrent locations. At the study corridor, x30_L3_K2_m10_dn5 combination best balanced these criteria and was therefore adopted for the detailed severity analysis. The triggered bottleneck points belonging to the x30_L3_K2_m10_dn5 combination were filtered to have a Severity Score ≥ 35, and the resulting critical bottleneck points were plotted on a heat map and shown in Figure 14.

The weight structure ensures that the most operationally impactful bottlenecks—those with deep speed drops, long durations, and extensive queues—receive the highest severity scores. This is consistent with an established bottleneck-assessment approach in which intensity (speed drop), duration, and extent are treated as the primary determinants of operational impact, and vehicle-hours of delay are used to rank the most severe bottlenecks [20].

Figure 15 presents an enlarged detail of the corridor-wide speed heatmap shown in Figure 14. The detail is provided so that the critical bottleneck markers produced by the proposed methodology can be examined more clearly. Within the congested (low-speed) regions, the detected bottleneck points fall consistently on the boundary between sustained upstream low speeds and downstream speed recovery. This close spatial agreement between the detected markers and the visually congested cells confirms that the method reliably locates bottlenecks within the high-demand zones of the corridor.

Figure 16 maps the detected critical bottlenecks onto the actual corridor geometry. Each marker corresponds to the midpoint of a segment whose severity score satisfies the critical threshold (S-BSI ≥ 35) for the selected x30_L3_K2_m10_dn5 combination. Projecting the bottlenecks onto the road map links each high-severity location to an identifiable physical feature of the corridor, which supports the cause-related discussion presented in the previous subsection.

A preliminary investigation of the probable causes of the critical bottlenecks was carried out along the corridor in the eastbound direction, without a detailed site-by-site engineering study. The findings indicate that, with respect to critical-bottleneck characteristics, it is more appropriate to divide the corridor into two parts: a western section and an eastern section.

In the western section, critical bottlenecks are associated mainly with on-ramp merges, off-ramp diverges, confluence points, and weaving segments located beneath grade-separated interchanges. In addition, a smaller number of critical bottlenecks were observed near rail-transit stations and pedestrian overpasses.

In the eastern section, critical bottlenecks occur predominantly along rail-transit stations and at the edges of bus stops with high passenger demand. Merges, confluences, and diverges also contribute to bottleneck formation in this section, and several bottlenecks were identified in front of certain commercial establishments with direct mainline access.

Figure 17, Figure 18, Figure 19 and Figure 20 examine four representative critical bottleneck locations in detail. For each location, a site view is shown together with the upstream and downstream speed profiles of the detected bottleneck. In every case, the speed profile displays the characteris-tic bottleneck signature: persistently low speeds on the upstream side and a clear recovery of speed on the downstream side, separated by the detected bottleneck boundary. The An-kara Road–Mahmudiye Street diverge (Figure 17), the Acemler interchange (Figure 18), and the Eski Fakülte interchange (Figure 20) are examples of off-ramp (diverge) bottle-necks, where the local capacity reduction is driven by diverging movements and the asso-ciated geometry. The Arabayatağı light-rail (LRT) station (Figure 19) represents a different mechanism: here the bottleneck is associated with a public-transport stop, which high-lights the importance of stop placement and passenger-transfer design in the formation of recurrent congestion. Taken together, these four cases confirm that the detected critical bottlenecks correspond to identifiable physical and operational features of the corridor rather than to random speed anomalies.

5.5. Grid Search Calibration Results

The grid search evaluated all 128 parameter combinations for eastbound direction. Table A1, Table A2, Table A3 and Table A4 summarize the range of detection outcomes across the entire search space.

The grid search results confirm that detection outcomes are strongly dependent on the parameter configuration. Combinations with the strictest thresholds (x = 50, L = 4, K = 2, m = 10, dn = 10) produce fewer triggers, while combinations with the most relaxed settings (x = 20, L = 1, K = 1, m = 5 dn = 5) produce a large number of triggers. This trade-off between sensitivity and specificity is a fundamental feature of threshold-based detection algorithms. An evaluation of an established bottleneck identification method [23] on a Portland freeway corridor demonstrated that parameter choices produce systematically different trade-offs between detection rate and false-alarm rate, confirming that optimal threshold settings are corridor-specific [33].

5.6. Performance Metrics

The 8606 critical triggers produced by the selected combination correspond to repeated space–time activations of a far smaller set of physical locations. After applying the spatial-aggregation rule of Section 3.5.4 (merging contiguous critical segments and attributing each cluster to its most downstream segment) and grouping recurrent activations of the same segment, these triggers collapse into 32 distinct critical bottleneck locations, which form the basis of the validation below. The spatial-aggregation rule is applied algorithmically at each time step, whereas the grouping of recurrent activations of the same physical location across time, together with the subsequent field verification, was carried out as a manual post-processing step outside the detection code.

In this study, a total of 32 critical bottleneck locations were detected. To validate the accuracy of these detections, preliminary research and data validation were conducted to identify the probable causes for each bottleneck.

Out of the 32 detected locations, 31 were successfully verified with identifiable traffic flow disruptions. Therefore, these 31 locations are classified as True Positives (TPs). However, for one specific location—Node 19, located near the KGM campus exit—no physical or operational cause could be identified for the bottleneck detected during the evening peak hour. Consequently, this single detection was classified as a False Positive (FP). Since the field validation was conducted only on the positively detected locations, the primary performance metric applicable for this study is Precision (Positive Predictive Value). Precision measures the reliability of the model’s positive predictions and is calculated using the following formula:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(20)

Given the validation results (TP = 32 and FP = 1), the precision score of the proposed detection method is calculated as follows: 0.969. The results demonstrate that the model achieved a precision rate of 96.9%. This high precision score indicates that the proposed methodology is highly reliable in identifying critical bottlenecks, with a remarkably low false discovery rate (only 1 out of 32 detections).

To assess whether the detected bottlenecks correspond to real congestion-generating locations rather than artefacts of the speed data, an independent corroboration step was carried out for all 32 critical bottlenecks identified by the proposed framework. Each detected location was cross-checked using two independent sources: (i) the typical traffic layer of Google Maps (Google LLC; accessed on 11 June 2026), which provides historical congestion patterns derived from large-scale probe data independent of the dataset used in this study, and (ii) on-site field inspections combined with examination of the road geometry and adjacent land use. For each location, a plausible physical cause of recurrent congestion was sought.

For 31 of the 32 detected bottlenecks (97%), a clear physical congestion-generating feature was identified at the detected location. These features fall into three groups: public-transport stop layouts (bus stops and light-rail transit stations, commercials) accounting for 15 locations (48.4%); ramp- and weaving-related geometries (on-ramps, off-ramps, on–off ramps, and weaving sections) accounting for 15 locations (48.4%); and one signalized control section (3.2%). Only a single location (No. 19, near the exit of the KGM campus) could not be associated with an obvious physical cause and requires further investigation.

The strong spatial coincidence between the detected bottlenecks and known traffic-generating features—observed both in the independent Google Maps typical traffic patterns and during field inspections—indicates that the framework reliably identifies genuine recurrent bottlenecks. This agreement provides face validity for the proposed Speed-only Bottleneck Severity Index and confirms that high-ranked locations reflect real operational deficiencies rather than data noise. This spatial coincidence establishes the face validity (physical plausibility) of the detections rather than an independent confirmation of each bottleneck; a full precision–recall evaluation against independent traffic measurements is left for future work.

5.7. Corridor Result Summary

The two dominant groups together account for approximately 95% of all detected points; This suggests that critical speed drops along the corridor are primarily due to local flow disturbances rather than homogeneous congestion. In Group A, frequent slowdowns and stops for boarding and alighting, combined with pedestrian activity and lane congestion near stations, lead to recurring speed reductions. In Group B, merging and diverging movements, lane change turbulence, and limited passage lengths reduce speeds in the intersection impact zones. Of the remaining two points, the signalized intersection at the end of the corridor at the city exit in Kestel was included in the study area as a control point to test the accuracy of the methodology. The results show that S-BSI successfully identified this point as a critical bottleneck. No significant bottleneck indication was found at the other point identified as a critical bottleneck by S-BSI, near the KGM campus exit. Further field research is needed at this point, but this point was considered an incorrect identification in the study. These results demonstrate that the proposed S-BSI consistently captures bottlenecks in infrastructure characteristics where geometric or operational friction is expected, and supports its validity as a speed-based severity measure.

Although the case study focuses on a single corridor, the D-200 expressway spans 31 km and is divided into 747 individual segments. These segments are highly heterogeneous, encompassing a wide range of geometric and operational conditions, including on-ramps, off-ramps, merge and diverge areas, weaving sections, and signalized at-grade intersections, as well as varying free-flow speeds and demand levels. The framework was therefore not validated on a single uniform facility but across 747 distinct spatial units exhibiting diverse congestion-generating mechanisms. This intra-corridor heterogeneity provides a substantial and varied test bed, demonstrating that the method performs consistently across different roadway configurations rather than under a single specific condition.

While these results demonstrate strong within-corridor generalizability across heterogeneous segments, validation of cross-city transferability is left for future work.

6. Conclusions and Future Work

6.1. Summary of Findings

Applied to the Bursa D-200 corridor, the framework produced several consistent findings. The spatially aware moving-window algorithm detected bottleneck activation events whose locations match the congestion visible in the spatiotemporal speed maps, confirming that the detection logic captures genuine capacity constraints rather than incidental slow-downs. Detections concentrate at a small number of locations—a few segments account for most events and most critical (S-BSI ≥ 35) events—and recur with high daily regularity, indicating demand-driven congestion suited to proactive management. The five-component S-BSI clearly separates moderate from severe bottlenecks, enabling prioritization, and queue-propagation tracking reproduces the growth–plateau–dissipation pattern of first-order traffic-flow theory. The critical bottlenecks coincide with identifiable physical features—off-ramp diverges, light-rail stations, and high-demand bus stops—and the signalized intersection included as a reference point was correctly detected, providing qualitative validation. In the 128-combination grid search, the congestion-threshold parameter (x) is the most influential, while temporal persistence (m) and the downstream-verification settings provide secondary control, so that the most severe detections remain stable across reasonable parameter choices. Because the detected critical bottlenecks are recurrent and spatially stable, they provide candidate locations for ramp-metering deployment.

6.2. Limitations and Future Research Directions

6.2.1. Limitations

Several limitations apply: (1) the provider does not disclose probe penetration, so speed estimates may be less reliable off-peak and on low-volume segments; (2) without volume or density data, all flow-related quantities—including queue extent and growth—are approximate speed-based proxies; (3) the framework was validated on a single corridor; (4) a fixed corridor-wide FFS (82 km/h) ignores local speed-limit, geometric, and weather variations; (5) multi-day averaged speed matrices smooth day-to-day variability and may underestimate worst-day severity; (6) results depend on the chosen threshold and persistence values, with moderate-severity events more sensitive than severe ones; (7) bottleneck locations are identified but not automatically attributed to geometric causes; and (8) the field corroboration confirms physical plausibility (true positives) but is not a full precision–recall evaluation. The averaged space–time matrix should be interpreted as the typical recurrent pattern of the corridor rather than as any single day; by construction it suppresses non-recurrent disturbances but also attenuates day-specific peaks, so the reported severities are conservative with respect to worst-day conditions.

6.2.2. Future Research Directions

Future work will (1) validate the framework on multiple corridors; (2) compare its output with concurrent loop-detector or radar data to quantify accuracy, including false-positive and false-negative rates; (3) adapt it to streaming FCD for real-time detection; (4) combine the physics-based criteria with machine-learning models, including convolutional networks applied to spatiotemporal speed images; (5) integrate traffic counts or FCD-derived volume estimates for flow-based analysis; (6) add weather and incident data to separate recurrent from non-recurrent bottlenecks; (7) extend the method to network-level analysis of bottleneck interactions and congestion propagation; and (8) couple detection outputs with ramp-metering control algorithms (e.g., ALINEA) for closed-loop operation.

6.3. Contributions and Practical Implications

6.3.1. Contributions

This research makes several methodological and practical contributions to the traffic-engineering literature.

Methodologically, this study makes several contributions. It establishes a formalized bottleneck definition based on the spatial relationship between upstream congestion and downstream speed recovery, grounded in traffic-flow theory and distinguishing true capacity constraints from uniform congestion [32,45]. It introduces an adaptive downstream criterion that references the boundary segment, so it adjusts to congestion intensity, stays independent of the upstream extent L, and avoids the false negatives of fixed-threshold methods [23,33]. Its central contribution is the Speed-only Bottleneck Severity Index (S-BSI), a fully specified 0–100 measure that ranks active bottlenecks using speed alone. It advances the literature in four ways: it is data-parsimonious (needing no flow, density, occupancy, or image inputs, unlike volume- or reliability-based indices); integrated (combining detection and five-component severity scoring in one framework); dynamics-aware (capturing queue length and growth from a speed-based, first-order-consistent shockwave approximation, which static composite measures cannot); and purpose-built and transferable (its fixed, corridor-adaptive 0–100 scale supports comparison across settings and corridors and the prioritization of ramp-metering locations). Taken together, these features make the S-BSI a low-data, dynamics-aware, and operationally oriented complement to the more data-intensive severity methods currently available, extending severity assessment into the temporal domain of queue evolution while relying strictly on speed [23]. The framework also operates at 1-min resolution, enabling detection of short-duration events, and calibrates all parameters through a transparent 128-combination grid search rather than ad-hoc choices. Practically, it links severity ranking to ramp-metering feasibility assessment, is implemented in Python for standard commercial FCD with GIS-compatible outputs, and demonstrates transferability to a developing-country setting where high probe penetration cannot be assumed.

6.3.2. Practical Implications

For practitioners, the framework offers a cost-effective complement or alternative to fixed sensors for corridors lacking detection infrastructure; a severity-based ranking that prioritizes ramp-metering deployment toward high-severity, high-recurrence locations to maximize return on investment; a consistent basis for tracking bottleneck severity over time and evaluating interventions before and after; and a practical analysis path for developing regions, where commercial FCD is increasingly available without the capital cost of fixed-sensor networks.

6.4. Concluding Remarks

This study presents a complete, theoretically grounded, and field-corroborated methodology that converts widely available commercial speed data into actionable bottleneck intelligence.

By combining spatially aware moving-window detection, temporal persistence filtering, and a five-component severity index, it moves beyond simple speed thresholds and reliably identifies bottlenecks whose location, timing, and severity match the observed speed fields. Its central innovation, the five-component S-BSI—spanning spatial extent, temporal persistence, speed-deficit magnitude, queue length, and queue growth rate—provides a single within-corridor transferable severity scale derived from speed alone yet still captures queue dynamics through a speed-based, first-order-consistent shockwave approximation. Because it requires only speed, the same procedure applies wherever commercial probe data exist—an advantage that will grow as connected and autonomous vehicles expand probe coverage. By lowering the data and infrastructure requirements for rigorous bottleneck analysis, the framework makes evidence-based congestion management accessible to a much wider range of corridors, particularly in developing metropolitan areas where fixed-sensor networks remain sparse. These qualities make the proposed approach a practical and broadly applicable addition to the traffic-engineering toolkit. From a sustainability perspective, ranking the most severe recurrent bottlenecks lets agencies concentrate limited resources where congestion relief yields the largest fuel and emission reductions; quantifying these savings and pursuing the corridor-level, real-time, and network-level extensions outlined above are left for future work. Similar assessment approaches developed for advanced traffic management systems show that removing forced deceleration and queuing produces substantial reductions in CO₂ and NO_x emissions [22]; the S-BSI ranking therefore provides a low-cost basis for selecting the interventions with the highest expected environmental return.

This study has several limitations that should be noted. First, the framework relies only on speed data; without volume or density measurements, queue-related quantities are approximate proxies. Second, the analysis was validated on a single corridor (D-200, Bursa), so the results may not directly transfer to other cities or road types. Third, a constant free-flow speed (82 km/h) was used across the corridor, independent of local geometric or weather conditions. Finally, the field corroboration confirms physical plausibility but is not a full precision–recall evaluation.

Future work will address these points in several ways. The framework must be tested on multiple corridors and compared with loop-detector or radar data to measure accuracy, including false-positive and false-negative rates. It can also be adapted to streaming FCD for real-time detection and combined with machine-learning models for improved performance. In addition, integrating volume estimates, weather, and incident data can support a more complete, flow-based analysis and strengthen the transferability of the method.