The classification criteria for traffic operation status can be broadly categorized into two types: absolute metrics and relative metrics. Absolute metrics encompass a wide range of fixed values, including traffic volume as outlined in the Freeway Capacity Manual, average travel speed, saturation levels, and the corresponding load factor, which collectively serve as a comprehensive metric. These metrics categorize traffic operating conditions into six classifications, labeled A through F. For instance, China’s Interim Technical Requirements for Road Network Operation Monitoring and Services utilize average travel speed as a metric, dividing traffic operating conditions into five categories: smooth, basic smooth, light congestion, moderate congestion, and severe congestion. Such classifications provide quantitative guidelines for assessing traffic states and facilitate the classification of traffic operation conditions through specific metric indicators. However, it is important to note that traffic flow is influenced by a variety of factors, including road characteristics, traffic patterns, weather conditions, and time of day. Additionally, the inherent uncertainty of traffic flow means that a standardized absolute metric may not accurately represent the actual traffic status on basic road segments under varying spatial and temporal conditions (Manual, 2010; Author Anonymous, 2012; Wang, 2019).

These traditional classification criteria have laid a foundation for traffic state analysis. However, with the development of traffic research, more attention has been shifted to traffic state classification algorithms. Early research focused on California and Bayesian algorithms. Notably, in recent years, researchers have made new improvements to these algorithms. Puangnak uses an improved California Algorithm to detect different types of traffic events, effectively improving search capabilities (Puangnak and Chivapreecha, 2019). This improvement broadens the application scope of the California algorithm. However, it may face difficulties in accurately detecting rare or complex traffic events. Shang utilized the Bayesian algorithm to optimize the handling of traffic data imbalance issues (Shang et al., 2021). Jin employed an improved Bayesian algorithm to enhance the classification of relevant traffic - flow characteristics (Jin et al., 2023). Zhao applied an improved Bayesian algorithm to the analysis and prediction of road safety conditions (Zhao et al., 2024). Ranpura innovated the traditional Bayesian algorithm by combining real-time traffic data to predict the delay time of traffic vehicles (Ranpura et al., 2024).

This research utilizes various traffic parameters as the foundation for analysis, including traffic flow, speed, and occupancy rate. Traffic state classification was achieved by comparing these parameters against established fixed thresholds (Payne and Tignor, 1978; Cook and Cleveland, 1974; Dudek et al., 1974; Collins et al., 1979; Ahmed and Cook, 1982; Martin et al., 2001; Sheu and Ritchie, 1998). Hsiao et al. (1994) were the first to apply the Fuzzy Logic (FL) algorithm for traffic event classification, employing fuzzy rule formulation and membership functions. Hawas (2007) proposed an urban road traffic event detection algorithm that integrated a fuzzy system to establish the membership function for clustering. Bauza et al. (2010) developed a fuzzy classifier-based method to analyze the traffic state of a road segment by examining vehicle networking, thereby facilitating small-scale traffic state exchanges. Liu et al. (2014) utilized Random Forest (RF) techniques to detect traffic events, effectively addressing noise and overfitting issues. Jiang et al. (2020) employed K-means clustering in combination with a Multi-Layer Perceptron (MLP) to detect urban road traffic status, resulting in a more effective real-time monitoring model.

Recently, Machine Learning methods like RF and Neural Networks have been applied. Sharma employed a neural network based on the Convolutional Neural Network (CNN) model, significantly improving the accuracy of vehicle trajectory prediction on highways (Sharma et al., 2023). Dr. P. Hasitha Reddy utilized a Deep CNN model to analyze traffic monitoring data, which enhanced control capabilities in traffic management (Reddy et al., 2024). Park applied Long Short-Term Memory (LSTM) networks to predict vehicle trajectories and traffic volume on urban roads, offering a potentially more accurate and efficient solution compared to traditional methods (Park and Yoon, 2024). Wan Ming implemented the random forest algorithm to detect traffic violations among taxi drivers and accurately predict the severity of these violations (Ming et al., 2023). This approach demonstrated high efficiency and precision, outperforming conventional detection methods. Shaaban employed the SMOTE to analyze traffic accident data, effectively addressing the issue of data imbalance and enhancing the reliability of subsequent data-based analyses (Shaaban et al., 2024).

Although traffic state classification has evolved from early threshold-based algorithms (e.g., California and Bayesian algorithms) to advanced machine learning methods (e.g., Fuzzy Logic and Random Forest), significant challenges remain. Existing studies predominantly focus on balanced datasets, neglecting the real-world prevalence of data imbalance. For instance, fuzzy logic often relies on subjective membership functions, while standard Random Forest models struggle to accurately classify minority states (e.g., severe congestion) in imbalanced or noisy data environments. Furthermore, while recent deep learning approaches improve prediction, they often overlook sensor noise characteristics.

To address these gaps, this study proposes a novel framework denoted as FCM-RF-SMOTE. This approach utilizes SUMO simulation to generate realistic trajectory data with radar-like noise characteristics. By integrating Fuzzy C-Means (FCM) clustering, Random Forest (RF), and the Synthetic Minority Over-sampling Technique (SMOTE), this study aims to develop a robust model for freeway traffic state classification.

The primary contributions of this study are summarized as follows. A standardized classification system is established. By introducing the SMOTE algorithm, the representation of the minority class (severe congestion) is significantly improved, increasing its proportion from 3.67% to 19.83%. Consequently, the overall classification accuracy enhanced from 77.67% to 97.80%. The fuzzy clustering feature of FCM is utilized to define traffic states objectively, improving upon traditional hard-threshold methods and better reflecting the continuity of traffic flow. The framework is validated using SUMO simulation with Gaussian noise (σ = 0.1 m) and a 10 Hz sampling rate, demonstrating the algorithm’s robustness in scenarios approximating real-world millimeter-wave radar detection.

In this study, we employ the Simulation of Urban Mobility (SUMO) platform to model vehicle dynamics on a mainline freeway. The SUMO simulation parameters were calibrated to emulate basic characteristics of millimeter-wave radar. Gaussian positional noise (σ = 0.1 m) was injected via SUMO ‘noise’ module, and detectors were configured with a 10 Hz sampling rate to approximate radar measurement intervals. However, this simplified model does not account for multi-target tracking or signal attenuation effects (e.g., rain fade).

The simulation is conducted over a duration of 10 h, utilizing a time step of 1 s and a random seed value of 42. The input traffic flow on the mainline roadway varies between 2000 and 7,000 vehicles per hour. The composition of the vehicle fleet is characterized by 90% small cars and 10% trucks, as detailed in Table 1. Furthermore, to enhance the realism of the simulation, a single accident is introduced at a random point during the simulation period to reflect actual traffic conditions.

The approximate simulation of the characteristics of radar sensors, including noise and multi-target tracking, provides a realistic dataset for model training. The classification performance achieved in this study (overall accuracy >97%) demonstrates the framework’s robustness in handling sensor-like noise, which is a common challenge in real-world deployments. This validates the simulation’s capability to emulate physical sensor conditions effectively, supporting the feasibility of applying the framework in practical traffic monitoring systems.

It is important to note that while real-world validation is ideal, obtaining high-fidelity trajectory data that includes specific ‘severe congestion’ events (such as accidents) is extremely difficult and dangerous to instrument in physical environments. Simulation allows us to generate these ‘minority class’ events safely and consistently, which is essential for validating the effectiveness of the SMOTE algorithm in handling imbalanced data.

The data obtained from the roadside radar video detector encompasses various parameters, including time and coordinate points, vehicle type, instantaneous vehicle speed, and license plate information, among others (see Table 2).

(1) Speed is calculated using Equation 1

V = ∑ 1 i v i i (1)

(2) The speed deviation for an individual vehicle is defined in Equation 2. The average speed deviation is calculated as shown in Equation 3.

v d e v i a t i o n i = v i + 1 − v i (2) V d e v i a t i o n = ∑ 1 i v d e v i a t i o n i i (3)

(3) Headway distance is determined by Equation 4. The average headway is calculated using Equation 5.

h i = x i − x i + 1 (4) H = ∑ 1 i h i i (5)

(4) Headway time is defined in Equation 6. The average headway time is derived using Equation 7.

h t i = x i − x i + 1 v i + 1 (6) H T = ∑ 1 i h t i i (7)

(5) Vehicle density is calculated according to Equation 8.

D = i c a r + i t r u c k * 1.5 200 × 5 (8)

Within each 1-s interval, vehicles are organized by lane and direction. Due to high-frequency fluctuations, traffic parameters are aggregated into 1-min intervals to enhance stability. Table 3 presents the processed dataset parameters: front and rear vehicle positions ( x i , x i + 1 ) and speeds ( v i , v i + 1 ); average interval speed ( v ); average speed deviation ( V d e v i a t i o n ); average headway distance ( H ); average headway time ( H T ); and vehicle density ( D , calculated as equivalent passenger cars per kilometer). These aggregated metrics serve as the foundational feature vectors for the subsequent classification model.

In contemporary traffic state analysis, addressing data complexities is of paramount importance. The Synthetic Minority Over-Sampling Technique (SMOTE) plays a pivotal role in the preprocessing stage. In traffic datasets, class imbalance is a prevalent issue, characterized by the underrepresentation of certain traffic states. SMOTE addresses this concern by generating synthetic samples for minority classes. By incorporating these synthetic samples into the dataset, SMOTE effectively rectifies the class distribution, ensuring that subsequent classification algorithms are not biased towards majority classes.

Following this, the Fuzzy C-means algorithm is applied to categorize traffic states. In the context of traffic analysis, it employs a set of traffic-status metrics to partition data into distinct traffic-state classes. Metrics such as vehicle speed, traffic flow rate, and occupancy are typically considered. By iteratively updating the membership values of each data point across different clusters and adjusting the cluster centers, the algorithm generates a traffic-status dataset for each class.

Finally, the Random Forest (RF) algorithm is implemented within the traffic-state classification decision module. RF, a combinatorial classifier, operates as a non-parametric classification algorithm. By analyzing the four classes of traffic-state data generated by the Fuzzy C-means algorithm, a traffic-state classifier is developed. This classifier can accurately assess the real-time operational status of road traffic, even in the presence of noisy data and missing values. In order to assess the temporal variability of traffic state changes, detector deployment intervals of 100 m, 200 m, 500 m, and 1,000 m were utilized. In summary, this paper constructs a traffic state discriminative model based on FCM-RF. The process of the FCM-RF traffic state classification model is shown in Figure 3.

The historical traffic data were clustered and analyzed to generate a dataset comprising four distinct types of traffic states. Sixty percent of the data were designated as the training set, while the remaining forty percent were utilized as test samples. The composition of the data samples is presented in Table 5.

The Random Forest (RF) algorithm is executed using the scikit-learn library in Python, and it computes the confusion matrix, which is presented in Table 6.

As shown in Table 7, the accuracy of traffic status classification exceeds 95% when traffic conditions are smooth and stable, indicating the feasibility of the algorithm. However, the limited sample size for congestion and blockage states results in an unbalanced data distribution, which adversely affects the algorithm’s discriminatory accuracy. To address this issue, the Synthetic Minority Over-sampling Technique (SMOTE) is employed to ensure a balanced sample size across all categories.

The unbalanced data samples from each state adversely affect the performance of the classification algorithm, the distribution of traffic conditions is shown in Figure 5. To address this issue, the Synthetic Minority Over-sampling Technique (SMOTE) is employed to generate synthetic data, which is then incorporated into the original dataset. This approach aims to balance the data distribution and enhance the accuracy of the classification algorithm.

As illustrated in the figure, the distribution of each traffic state is markedly uneven, with the sample size for the stable state significantly exceeding that of the blocked and congested states. This issue of data imbalance leads to erroneous learning by the random forest algorithm, resulting in a low accuracy rate for the blocked and congested states, which subsequently undermines the effectiveness of traffic state classification. Consequently, the application of the SMOTE algorithm is necessary to address the imbalance in the data.

From Table 8, it can be observed that after clustering the traffic flow dataset into 1-min intervals, there are only 164 instances of severe congestion and 614 instances of congested state. By employing the SMOTE algorithm to synthetically augment the blocked data, the quantity of severe congestion data has been increased sevenfold, while the quantity of congested state data has been doubled. The distribution of data categories for each state following the application of the SMOTE algorithm is presented in Table 9 and illustrated in Figure 6.

The original dataset has been enhanced with data on congested and severely congested states to ensure that the proportions of the four traffic states are balanced, while also preserving a larger sample size for the stable state. After applying the SMOTE algorithm, the distribution of the four traffic states has become more balanced, with severe congestion now accounting for 19.83% of the total. This adjustment effectively addresses the class imbalance issue present in the original dataset, where severe congestion only made up 3.67%. As a result, the dataset processed using the SMOTE algorithm can now serve as a more representative sample for traffic state classification.

Using the balanced dataset generated by SMOTE (as detailed in Table 9), the Random Forest model was retrained using the same 60:40 training-test split ratio. The Random Forest (RF) algorithm is implemented using the scikit-learn library in Python, and it computes the confusion matrix, as illustrated in Table 10.

As illustrated in Table 9, the accuracy of the four categories of traffic state classification exceeds 97%, indicating superior algorithm performance. The application of SMOTE (Synthetic Minority Over-sampling Technique) to address imbalanced data has enhanced the discriminative capability for congestion and blockage states. As illustrated in Table 11, the classification accuracy exceeds 97% in simulation, indicating potential for real-world applications pending field validation.

The comparative analysis between the baseline and the proposed framework demonstrates the critical role of data balancing. While the baseline RF model struggled with minority classes due to data imbalance (0% precision for severe congestion), the integration of SMOTE significantly enhanced the model’s sensitivity, boosting the classification accuracy for severe congestion to 97.80%. These results demonstrate the framework’s robustness, which is further validated by the rigorous design of the simulation environment. The emulation methodology is informed by established research in vehicular sensing and communications, notably the contributions of Sommer (Sommer et al., 2011) et al., who elucidated the effects of shadowing and signal degradation on vehicular communications, and Stiller (Stiller et al., 2025), who conducted a review of sensor fusion techniques aimed at enhancing the reliability of traffic monitoring. Furthermore, insights from Li and Yoon (Li and Yoon, 2023) regarding radar-camera fusion have further guided the design of the simulation, effectively replicating sensor imperfections and multi-target scenarios that are typical in real-world applications.

The use of SUMO simulation allowed for the injection of Gaussian noise (σ = 0.1 m) and limited sampling rates (10 Hz), approximating the characteristics of commercial millimeter-wave radars. The model’s high performance under these noisy conditions suggests a degree of robustness suitable for deployment in hardware-constrained environments.

Despite these promising results, several limitations warrant future investigation. Firstly, while the simulation mimics basic sensor noise, it does not fully replicate complex real-world factors such as multi-path interference, occlusions in multi-lane scenarios, or signal attenuation caused by adverse weather (e.g., rain or fog). Secondly, the current framework assumes homogeneous freeway traffic; its applicability to urban environments with complex interactions (e.g., intersections) remains to be tested. Thirdly, relying on a single sensor type may limit reliability.

Future work will focus on validating the model with physical sensor data, extending the framework to urban scenarios, and exploring multi-source data integration (e.g., fusing radar with camera or GPS data) to further enhance system robustness.