Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 5;16:12136. doi: 10.1038/s41598-026-42542-8 Search in PMC Search in PubMed View in NLM Catalog Add to search Optimal deep neural network based road traffic management system for Internet of Things based smart city environment Khaled Abdullah Almejalli Khaled Abdullah Almejalli 1 College of Computing and Informatics, Saudi Electronic University, Riyadh, Saudi Arabia Find articles by Khaled Abdullah Almejalli 1, ✉ Author information Article notes Copyright and License information 1 College of Computing and Informatics, Saudi Electronic University, Riyadh, Saudi Arabia ✉ Corresponding author. Received 2025 Sep 11; Accepted 2026 Feb 26; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13076887 PMID: 41786958 Abstract Internet of Things (IoT) solutions in smart transportation infrastructure deliver a transformative method to enhance operational efficacy and productivity while optimising performance across all scopes. Using IoT abilities, transportation networks might be observed in real-time, permitting data-driven decision making and enhanced connectivity, finally decreasing road congestion and increasing safety integration. A solution of intelligent transportation can improve the traffic flow in urban cities by observing traffic patterns and altering traffic signal times. The goal is to determine and help supportable methods of transport, to increase an Intelligent Transportation System (ITS) that utilises real-time data to improve safety, reduce congestion, and enhance green applications. ITS influences new and evolving technologies to create mobility that is more satisfying and economical in smart cities. At present, deep learning is an effective method to discover hidden visions into ITS without being programmed explicitly by learning from data. In this manuscript, a Hybrid Feature Selection and Deep Neural Network for Decision Support Systems in Road Traffic Management (HFSDNN-DSSRTM) model is proposed for smart IoT-integrated cities. The HFSDNN-DSSRTM model aims to design a decision support system for efficient road traffic management in IoT-enabled smart cities using advanced methodologies. Primarily, the data pre-processing stage is employed at dual levels, such as missing values handling and normalisation methods. For an effective feature selection, the HFSDNN-DSSRTM model employs filter, wrapper, and embedded methods to identify and keep only the most valuable features that contribute to the classification task. Finally, the temporal convolutional network with attention mechanism (TCN-AM) method is used for classification. The comparison analysis of the HFSDNN-DSSRTM approach portrayed a superior accuracy value of 98.75% over existing methods. Keywords: Hybrid feature selection, Deep neural network, Decision support systems, Road traffic management, Internet of Things, Deep learning Subject terms: Engineering, Mathematics and computing Introduction The twenty-first century has seen the swift development of urban populations, resulting in enlarged vehicle ownership and traffic congestion in many cities. Cities are rapidly expanding, a trend fueled by shifts in the global economy and the demands of contemporary living, resulting in a population shift from rural to urban areas 1 . Information and communication technology (ICT) is indispensable in urban development and the sustainability plans of cities. The IoT provides novel as well as advanced mechanisms to develop smart cities and boost the efficacy of city operations and services linked with people 2 . Smart traffic infrastructure is vital for smart cities, tackling pervasive urban congestion that intensifies with urban growth 3 . The traffic management system’s objective is to reduce this type of short-term congestion and to normalise highway traffic flow by controlling the traffic in real-time 4 . For this purpose, traffic control systems swiftly identify and confirm incidents, deploy the correct equipment to the scene, and try to control traffic better at the time of the incident by distracting traffic on other routes, which helps save travel time 5 . Moreover, the decision support system (DSS) aids in achieving the crucial objective of reducing traffic, as demonstrated by its role in a city that successfully lowered traffic congestion to a manageable level 6 . Recently, traffic applications have been developed that employ past and real-time community-generated data from transported vehicles and semantic-based urban data. The widespread adoption of smartphones and ride-hailing platforms has enabled the collection of vast amounts of GPS-tracked transportation data 7 . This data is utilised for investigating the movement of road vehicles in aggregated and restricted temporal resolutions and for different city regions. In addition, urban traffic management is increasingly moving towards information-rich and data-driven methods 8 . Along with the traditional detection-based techniques, traffic data is condensed into several novel types and with improved precision utilising artificial intelligence (AI), which uses the advances that are attained through DL 9 . Furthermore, AI-supported traffic management models focus on managing traffic flow, improving road safety, and reducing congestion, employing real-time data congregated through IoT systems, including cameras, sensors, and connected vehicles. Machine learning (ML), a subfield of AI, can make predictions about the levels of traffic congestion in a specific region of a city 10 . They can effectively model patterns connected to traffic flow and recommend measures that the governments can take to control traffic-related issues. Moreover, combining IoT sensors, cameras, and other devices alongside sophisticated DL models enables the persistent collection and examination of massive traffic data, recognising patterns, predicting congestion, and optimising traffic signal timings 11 . The traffic conditions are addressed for assisting effective road traffic and signal management. It becomes crucial to improve decision quality by incorporating various traffic conditions combined into a management framework. The analysis also concentrates on efficiency to mitigate congestion and improve traffic flow under diverse traffic settings. Figure 1 portrays the general framework of the road traffic management in IoT-integrated smart cities. Fig. 1. Open in a new tab General of road traffic management in IoT-integrated cities. In this manuscript, a Hybrid Feature Selection and Deep Neural Network for Decision Support Systems in Road Traffic Management (HFSDNN-DSSRTM) model is proposed for smart IoT-integrated cities. The HFSDNN-DSSRTM model aims to design a decision support system for efficient road traffic management in IoT-enabled smart cities using advanced methodologies. Primarily, the data pre-processing stage is employed at dual levels, such as missing values handling and normalisation methods. For an effective feature selection, the HFSDNN-DSSRTM model employs filter, wrapper, and embedded methods to identify and keep only the most valuable features that contribute to the classification task. Finally, the temporal convolutional network with attention mechanism (TCN-AM) method is used for classification. An extensive experiment of the HFSDNN-DSSRTM methodology is carried out under the Smart Traffic Management dataset. The key contributions are listed below. Initially, missing value handling and normalisation are performed at dual levels to improve data quality and ensure more accurate and reliable traffic classification. Also, an integrated filter, wrapper, and embedded methods are used for effectively detecting relevant traffic data features, thus improving the performance and mitigating computational overhead for traffic management. The TCN-AM method is employed for capturing short- and long-term temporal dependencies in traffic data, thus enhancing the efficiency and accuracy of dynamic traffic pattern prediction. The HFSDNN-DSSRTM model incorporates TCN-AM, temporal dependencies and attention-based feature weighting to improve classification performance compared to conventional traffic prediction models significantly. Previous studies on road traffic management in IoT and smart cities This section reviews the existing traffic management techniques in the smart city environment. In 12 , a Smart Traffic Management System (STMS) recommended in this study utilises an enhanced YOLO-driven DL technique for current vehicle recognition and dynamic flow of traffic optimisation. For evaluating traffic film, accurately identifying various vehicle types, and calculating the density of traffic, the system integrates AI and CV. This examination allows an intelligent controlling system to dynamically modify traffic signals and reroute automobiles to reduce congestion. In 13 , Conjecture Interaction Optimiser (CIO) method utilising terminal-communication assistance is proposed. This method explains the road’s physical state and the traffic density to set up effective routing. Moreover, this approach heightens instantaneous vehicle interaction for analysing the vehicles live and route refinement, and unnecessary interactions for minimum delay. Khan and Thakur 14 described a self-evolving real-time technique that depends on the real-time flow of traffic and observation. Integrating image pre-processing with AI-powered, self-evolving ML to control road clearance at junctions is a modern method with high potential. The recommended system utilises the YOLOv3 technique and single-image processing employing an NN for determining road clearance at the signal. Mrudula et al. 15 developed a system by utilising the IoT for real-time data collection and ML techniques for accurate forecasting. Feature selection is performed using Particle Swarm Optimisation (PSO) to enhance prediction accuracy, and models like K Nearest Neighbour (KNN), Multi-Layer Perceptron (MLP), and Bayes Network are employed for classification. In 16 , an IoT-based robotic (IoRT) method is improved with an advanced system that combines cameras and IoT sensor nodes for gathering actual traffic information. The key contributions of this paper are the implementation of two DL methods, such as Inception-V3 and LeNet-5, for traffic sign recognition. In 17 , a DL technique intended for current traffic control in advanced cities is proposed. The method utilises a Recurrent Neural Network (RNN) and a Convolutional Neural Network (CNN) for pre-processing, decision-making, feature extraction, and data collection. The significant aid of this study is its complete strategy to handle the geographical and temporal components of traffic information, which finally enhances decision-making and flow of traffic. Real-time traffic camera pictures are obtained throughout the data collection stage. Johny and Sharma 18 presented a new ML-driven traffic congestion management approach that integrates Euclidean distance trackers alongside the YOLO object detection structure. As urban issues are affected by the complexities of intensifying traffic, the need for advanced platforms able to perform congestion analytics and real-time vehicle surveillance is emphasised. The recommended solution overtakes conventional limitations by utilising ML for precisely detecting and tracking automobiles in a city environment. The authors 19 introduced an IoT-driven Intelligent Traffic Signal System (ITSS), which involves a programmable microcontroller and induction loops for determining the density of traffic. Interconnection in the unified control component sets the traffic signal timer and harmonises with the density of traffic for seamless movement of vehicles with reduced delay. Sun et al. 20 utilized Discrete Cosine Transform (DCT) for frequency-domain feature extraction, Squeeze-and-Excitation Networks (SENet) for channel-wise feature recalibration, Temporal Convolutional Network (TCN) for capturing temporal dependencies, and iTransformer for modeling complex multivariate inter-correlations. Prasad et al. 21 utilized a Two-Tier Optimization Strategy for Robust Adversarial Attack Mitigation (TTOS-RAAM) methodology by integrating a Min–Max Scaler for data normalization, Coati–Grey Wolf Optimization (CGWO) for optimal feature selection, Conditional Variational Autoencoder (CVAE) for adversarial attack detection, and Improved Chaos African Vulture Optimization (ICAVO) for parameter tuning. Yang 22 predicted and identified highway traffic events by utilizing real-time multi-dimensional data collected through IoT devices and sensor networks. A nonlinear dynamic system integrated with complexity theory and Graph Neural Network (GNN) techniques is also utilized for processing spatiotemporal traffic data. Hernandez-Jaimes et al. 23 improved anomaly detection in Internet of Medical Things (IoMT) networks by implementing an attention-driven Deep Neural Network (DNN) model. The unsupervised One-Class Support Vector Machine (OCSVM) using just nine generic features and protocol presence information is also employed, Furthermore, natural language processing (NLP) is used for capturing intrinsic traffic patterns. Bansal and Bali 24 utilized an Attention-based Graph Convolutional Network (AGCN) methodology for capturing both spatial and temporal dependencies in traffic data. The model also utilizes GCN for spatial pattern extraction, standard convolutions for temporal feature modeling, and an attention-based mechanism. Pawar et al. 25 utilized a Hybrid Adaptive and Attention Deep Learning Network (HAADLNet) methodology by integrating Multi-Scale Capsule Network (MCapsNet) for hierarchical feature extraction, Temporal Convolution Network (TCN) is also used for capturing temporal dependencies, and Iteration-based Random Variable for Preschool Education Optimization Algorithm (IRV-PEOA) for tuning. Singh and Kashyap 26 integrated a Neural Network (NN)-based Network Intrusion Detection System (NIDS) with Mutual Information (MI)-driven feature selection. Khan and Qin 27 improved Traffic State Estimation (TSE) by integrating a Graph Attention Temporal Convolutional Network (GAT-TCN) approach with advanced Kalman Filter variants, including Extended KF (EKF), Unscented KF (UKF), and Sliding Window KF (SWKF). Potharaju et al. 28 detected and mitigated False Data Injection Attacks (FDIAs) in Industrial IoT (IIoT) systems by employing a weighted voting ensemble that integrates Random Forest (RF), Extreme Gradient Boosting (XGBoost, XGB), NN, and Logistic Regression (LR). The ensemble assigns dynamic weights based on F1-score. Vatambeti et al. 29 utilized a hybrid DL ensemble of Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Bidirectional Gated Recurrent Unit (BiGRU) networks. Moreover, EfficientNet is used for feature extraction, hyperparameters are optimized via the Eurygasters Optimization Algorithm (EOA), and Tournament-Selected Glowworm Swarm Optimization (TSGSO) enhances ensemble performance, while Fuzzy Logic (FL) classifies traffic severity. A summary of the literature review on road traffic management in smart cities using IoT is established in Table 1 . Table 1. Outline of related works. References Goals Techniques Dataset Outcomes Talaat et al. 12 To recognise diverse vehicle types and analyse traffic density, the method integrates AI and CV. It permits an intellectual control method to adaptively adjust traffic signals and redirect vehicles to lessen congestion. To effectually modify various traffic situations, decreasing congestion and enhancing the traffic flow in real-time Improved YOLOv11 and DL Real-World Traffic Datasets mAP of 92.4% Al Asmari et al. 13 To enhance vehicle communication by examining the vehicles’ active and discarded communications for route refinement and minimal delay. Combining multi-instance learning with the process of communication labelling allows the routing classification by discriminating between erroneous and consistent data CIO IAMCV Dataset Throughput of 11.67% Khan and Thakur 14 To identify objects from videos. Then, the centroid object identification model is employed to observe the movement of every vehicle in the presented model ML, YOLOv3 Traffic Dataset Accuracy of 88.43% Mrudula et al. 15 To present a model for predicting traffic flow utilising IoT and the process of selecting features. The attainment of real-time sensor information is accomplished by the employment of IoT gadgets, which are utilised on cars or roadways PSO, KNN, MLP, and Bayes Network UCI Traffic Dataset Accuracy of 96% Kheder and Mohammed 16 To project an IoRT-based method for increasing mobility and incorporating cameras and IoT sensor nodes by gathering traffic details, which is analysed to employ several image processing and dual presented models before being wirelessly transmitted to the cloud and reachable for commuters and drivers Modified LeNet5 and InceptionV3 GTSRB, EGTSRB, and LISA Datasets Accuracy of 99.12%, 99.78% and 98.6% Akash 17 To provide a sufficient and comprehensive approach intended particularly for managing traffic in a smart city. It encompasses the intricate temporal and geographical features of traffic data with a well-structured model and is capable of precisely recording and simulating the dynamics of traffic conditions CNN and RNN NA Accuracy of 99.2% Johny and Sharma 18 To project a thorough method that incorporates learning models by tackling the complex issues of urban traffic congestion YOLO and CNN Large Dataset F1 scores ranging from 98.36% to 99.01% Rai et al. 19 To examine vehicular traffic networks and present an ITSS method intended for this kind of network. Notably, the amount of traffic in major cities is expected to continue to rise all over the world, and technological advancements and their uses in traffic management will become significant IoT Network NA NA Sun et al. 20 For precisely predicting traffic flow to alleviate congestion DCT + SENet + TCN, iTransformer, TCN Parallel UK Motorway Dataset R 2 : + 0.01–6.42%; MAE: − 3.07–62.25%; RMSE: − 3.41–64.43%; MAPE: − 1.83–39.27%; MASE: − 35.35–64.59% Prasad et al. 21 For detecting and reducing adversarial attacks in IoT networks by utilizing a robust two-tier optimization strategy TTOS-RAAM, Min–Max Scaler, CGWO, CVAE, ICAVO RT-IoT2022 Accuracy of 99.91% Yang 22 To predict and detect abnormal highway traffic events utilizing real-time IoT data Nonlinear Dynamic System Modeling, Complexity Analysis, GNN Expressway Traffic Section Data Traffic Flow: 13.4 vehicles/min; Speed: 28.0 km/h; Congestion: 76.0%; Prediction Accuracy: 96.7% Hernandez-Jaimes et al. 23 To improve anomaly detection in IoMT networks by utilizing an attention-driven DNN Attention-driven Extraction, DNN, OCSVM, Protocol Presence Characterization CICIoMT2024, MQTT-IoT-IDS2020 Precision of 84.43–92.14%, Recall of 98.73–99.17%, F1-Score of 91.02–95.53% Bansal and Bali 24 For anticipating traffic density on roads for supporting urban traffic management and route planning Attention-based Spatial–Temporal Mechanism, GCN, Standard Temporal Convolutions, Output Fusion California Transportation Agencies Performance Measurement System (CalTrans PeMS) Prediction Accuracy of + 4% Pawar et al. 25 To predict traffic congestion in IoT-based ITS using an adaptive DL model HAADLNet,MCapsNet, TCN, IRV-PEOA Standard Traffic Datasets High Accuracy and Congestion Management Singh and Kashyap 26 To improve anomaly detection in IoT networks by employing NN-based intrusion detection with MI-driven feature selection NN, NIDS, MI IoT-Botnet 2020 Accuracy of + 0.57–2.6%, False alarm rate (FAR) of − 0.23–7.98% Khan and Qin 27 For improving traffic state estimation and safety management in rural arterial networks by utilizing a hybrid intelligence framework GAT-TCN, EKF, UKF, SWKF Real-World Rural Toll Corridor Reduced Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) Potharaju et al. 28 For detecting and preventing FDIAs in IioT systems using an optimized ensemble learning framework Weighted voting ensemble, EF, XGBoost, XGB, NN, LR UKMNCT_IIoT_FDIA Accuracy of 99.71%, Precision of 99.72%, Recall of 99.72%, F1-Score of 99.72% Vatambeti et al. 29 For predicting urban traffic flow accurately and classify congestion levels using a hybrid DL approach EfficientNet, Ensemble LSTM, BiLSTM, BiGRU, EOA, TSGSO, FL Traffic Data from Four Urban Intersections Improved Prediction Accuracy and Processing Time Open in a new tab Though the existing studies are efficient, they exhibit low performance as some models depend on static assumptions and also lack robustness. Various approaches illustrate high accomplishment but restricted scalability and interpretability in large scale smart city deployments. Several studies also concentrate on isolated tasks such as prediction or detection. Furthermore, research gap can be seen due to the reliability on labelled and balanced datasets. Moreover, real-world applicability is restricted by various models and integrating multi-source heterogeneous data is insufficiently explored. Moreover, it is crucial to examined the comparison analysis under inconsistent and standard datasets. Thus, a research gap is highlighted through adaptive, interpretable, and data efficient models. Proposed methodology This manuscript designs an HFSDNN-DSSRTM model in smart IoT-integrated cities. The HFSDNN-DSSRTM model focuses on creating a decision support system for effective road traffic management in IoT-based smart cities, utilising an advanced technique. To perform this, the HFSDNN-DSSRTM model involves input pre-processing, dimensionality reduction, and classification processes. Figure 2 depicts the entire process of the HFSDNN-DSSRTM method. Fig. 2. Open in a new tab Entire process of the HFSDNN-DSSRTM method. Data pre-processing techniques At first, the data pre-processing stage is utilised by dual levels, such as missing values handling and normalisation techniques. This model is essential to guarantee input consistency and quality 30 . Dealing with missing values through mean imputation: all missing entries in the feature were substituted by the average of its experimental rates. Provided that fewer than 1% of all data points are missing, a comprehensive case study might face insignificant data loss; then it uses global mean imputation for preserving the complete size of the samples. 1 For standardising feature scales, Min‐Max normalisation was used, converting all features in the interval of ( as demonstrated: 2 whereas and represent the maximum and minimum values of . The model stops bias from different sizes amongst features and improves the convergence of the model. Feature reduction models For an effective feature selection, the HFSDNN-DSSRTM model utilises filter, wrapper, and embedded methods to recognise and keep only the most beneficial features that contribute to the classification task within the training set and the final features are fixed prior to analysis for preventing data leakage 31 . The hybrid models MI, Fisher Score (FS), Sequential Forward Selection (SFS), Recursive Feature Elimination (RFE), and RF Importance (RFI) are chosen for eliminating irrelevant features, while the feature subsets are evaluated by the SFS and RFE models. Furthermore, RFI effectively captures feature significance during model training. The hybrid model also ensures a more robust and efficient feature selection compared to using any single method alone. With the continued development of data in scale, complexity, and diversity, decreasing data dimensions has now become essential for preserving effective performance. Features are frequently separated into four dissimilar groups: unnecessary, inapplicable, strongly associated with the target, or poorly applicable but non‐redundant. FS concentrates on constructing an efficient method with decreased computational efficiency by highlighting the most relevant and unique characteristics for prediction and classification. Additionally, voting strategy is utilized for combining the features. With FS, a sub-group of attributes is established, with the features gained utilising statistical techniques applied for creating the method. Filter models for FS use statistical conditions for assessing and ranking features based on their link to the targeted variable, regardless of the learning method. Filter techniques act as the pre-processing stage for ranking features according to their importance, which is then chosen and discovered by the estimator, utilising different variable sorting methods. Filter techniques are applied as the pre-processing stage, irrelevant to the classifier, for reducing data dimensionality by choosing just the essential attributes. Examples of general filter techniques consist of MI that determines the dependency among variables; Pearson Correlation (PC) that estimates linear relationship among mathematical variables; Chi‐Square that evaluates the individuality among categorical variables; Information Gain (IG) that measures feature significance according to entropy decline; FS that evaluates the variance ratioamong and inside classes; and ANOVA that examines changes in means through groups. The PC model is applied for numerical values, while the Chi‐Square model is used for categorical values. MI MI is a statistical measure applied to calculate the dependence level between 2 random variables. In contrast to other filtering techniques, it might further capture nonlinear dependences. Ranking conditions according to information theory measure the interconnection of dual variables, using MI mathematically characterised as , as presented in Eq. ( 3 ). 3 whereas the joint function of the likelihood distribution of and was characterised by , whereas and designate the marginal likelihood distributions of and , correspondingly. Additionally, MI is redeveloped utilising entropy, as presented in Eqs. ( 4 – 7 ). 4 5 6 7 whereas refers to marginal entropy, characterises the conditional entropy, and denote the combined entropy of as well as . calculates the vagueness related to the random variable , however, computes the residual uncertainties in after is renowned. This expression gives emphasis to the type of MI as the amount of data that some variable offers regarding the other, taking either linear or nonlinear dependency. FS FS is among the frequently used methods to identify extremely discriminatory feature subsets. The most informative characteristics are chosen once the ranking of every feature is measured, as demonstrated: 8 9 Here, denotes class count and represent the Mean and Std regarding attribute , and and represent the Std and Mean of class with respect to attributes respectively. The most discriminatory features are selected from the top ‐ranked attributes with the maximum score counts. Wrapper techniques are progressive FS methods that evaluate the performance of the ML method by systematically assessing dissimilar feature subsets for determining the best incorporation of features. These models depend on the trained method’s validation error for deciding which features to choose, with attributes to iteratively add or eliminate according to their influence on the precision of the model. It includes estimating each possible feature subset for identifying the optimal incorporation of attributes. For larger data, it isn’t viable in terms of time or computational efficiency. Wrapper techniques frequently result in overfitting. Here, two wrapper techniques are used: SFS and RFE. SFS The SFS method gradually adds features to the process, targeted to improve the performance index while reducing the selected feature counts. The performance index, frequently dependent on metrics like classification precision or accuracy, directs the FS procedure, guaranteeing the top features are chosen to improve the method. SFS begins the procedure with a null feature set and gradually combines particular features at all steps, picking out the feature that exploits the goal function. This procedure iteratively persists, with every novel feature assessed for its role in the performance index till the pre-determined feature count is chosen. Still, SFS is considered a simple technique because it fails to explain dependencies among features, possibly resulting in sub-optimal feature subsets. RFE RFE is an iterative approach applied to remove features till the chosen feature count remains. This model recursively eliminates features one after another based on the scoring. During all iterations, the attribute with the low score is eliminated, and the method is reconstructed utilising the residual features. It is reiterated till just the significant qualities are maintained. Then, it constructs a method for the features that are persistent. The features that considerably impact the prediction of the targeted feature according to the accuracy of the model. Embedded techniques in ML concurrently process FS and the model training. In the training, they detect the related features to heighten the model performance, whereas removing those that give fewer or are unnecessary. In embedded techniques, ML techniques are essential for FS, as they explain interdependency amongst features and the shifting relationships between inputs and outputs. Embedded techniques provide numerous benefits, including reducing overfitting and improving either the method or the FS procedure concurrently. RFI RF is a strong EL method that improves either the dependability or precision of a technique by incorporating the predictions from various decision trees, all trained on dissimilar data subsets, to create a stronger prediction. It utilises the bootstrap aggregation (bagging) model, where multiple instances of the data are made with substitution, and all cases are applied for training the individual tree. By aggregating the outputs of these error-prone yet sound methods, RF decreases prediction variability and improves strength. This model decreases overfitting in comparison with conventional decision trees and offers precise predictions by utilising numerous different trees. RF selects variables randomly from variables. 10 Here, characterises the complete tree counts, refers to the detection completed through tree for the input , and , the last detection completed utilising the RF model, is measured as the prediction’s average from every tree. Feature significance is a measure that defines the growing tree splitting point utilising RF methods. Therefore, features with a higher correlation display small or no steps of significance. RFI uses Out‐of‐Bag (OOB) instances as the basis for every tree for assessing and recording prediction performance. Classification using TCN-AM method Finally, the TCN-AM technique is employed for the classification process 32 . This model effectively captures both short- and long-term temporal dependencies in traffic data. The TCN ensures that the temporal order is maintained, while the AM highlights the most relevant time steps, enhancing the accuracy and interpretability. Furthermore, fixed length sliding window is used for efficiently capturing the patterns, and also, each window contains consecutive time steps that are used as input sample. This also presents faster training, better parallelisation, and more stable gradient propagation, compared to conventional sequence models, making it highly appropriate for traffic classification and dynamic pattern prediction. Figure 3 signifies the architecture of the TCN-AM technique. Fig. 3. Open in a new tab TCN-AM architecture. TCNs are convolution‐based methods designed for sequence modelling, targeted at seizing either short‐term or longer‐term dependences within time series data. Unlike RNNs like LSTM and GRU, TCNs apply dilated convolutions, which permit the method to have a larger receptive field without increasing the parameter counts considerably. Dilated Convolutions Provided the sequence of input , while refers to sequence length, the output of a convolution with filter dimension , rate of dilation , and weights at time step is provided by: 11 Now, the rate of dilation controls the space among the elements in the sequence of input applied by the convolution, permitting the receptive area to grow exponentially as the layer count improves. This allows the TCN to seize longer‐range dependences without enhancing the depth of the network. 2. Residual Connections To avoid problems of vanishing gradient and increase the information flow over deep TCN structures, use residual connections. The residual connection at layer is described as: 12 whereas refers to dilated convolution output at layer , and denote input to the preceding layer. It permits the system to directly pass input information through layers, facilitating the learning of either short‐term or long‐term dependencies. The AM permits the method to focus on the most significant time-steps in the sequence of input, increasing its ability to generate accurate predictions. After processing the input over the TCN layers, use attention to the output of the last TCN layer. Assume the output of the TCN at every time step is signified as . The AM calculates a context vector which is a weighted sum of the hidden states at every time step. The attention weights are calculated as demonstrated: 13 14 15 Now, and refer to learnable parameters, denotes the unnormalised score of attention, and denotes the normalised attention weight for every time step . The context vector takes the most significant information from the complete sequence and is applied as input for the last prediction layers. Table 2 depicts the hyperparameters involved in the TCN-AM model. Table 2. Hyperparameters involved in the TCN-AM model. Hyperparameter Value/Setting Description TCN_LAYER_NO 3 Temporal Conv. layer number FILTER/KERNEL_SIZE 3 Conv. kernel size FILTERS_PER_LAYER 64 Feature maps in every layer DILATION_FACTOR 1, 2, 4 The receptive field is controlled for the temporal context ATTEN_TYPE Sefl-attention The crucial time step is focused DROP_RATE 0.5 Averts overfitting BATCH_SIZE 5 Samples per training batch LEAR_RATE 0.01 Optimiser step size OPTMIZER Adam Optimisation algorithm EPOCHS 200 Maximum training iterations Open in a new tab Evaluation metrics Many standard performance metrics applied usually for ML are estimated for assessing the model’s efficacy in recognising and for measuring the ML precision of the models. ( ) This computes the overall correctness by measuring the proportion of correctly predicted instances out of all predictions. 16 ( ) This calculates the precision of positive prediction. 17 Recall (R) It computes the fraction of real positive class instances properly recognised by the method. 18 ‐ score A harmonic mean of and , delivering a balance among them to assess the complete model performance. 19 AUC score It is computed through the ROC curve, signifying the relationship between true positive rate (TPR) and false positive rate (FPR). ROC-AUC is utilised for binary classification and determines how well a model differentiates between negative and positive target classes. Particularly, if the significance of positive and negative classes is equal for us, the ROC-AUC score is a valuable performance measure. Results and discussions The performance assessment of the HFSDNN-DSSRTM model is examined under the Smart Traffic Management dataset 33 . The oversampling of minority classes and class-weighted loss functions are utilised for addressing data imbalance and to ensure fair model training. Also, dropout, early stopping, and cross-validation are utilised for mitigating overfitting. Thus, the robustness and generalisation are improved on unseen traffic data. The model is simulated using Python 3.6.5 on a PC with an i5-8600k, 250GB SSD, GeForce 1050Ti 4GB, 16GB RAM, and 1TB HDD. Parameters include a learning rate of 0.01, ReLU activation, 200 epochs, 0.5 dropout, and a batch size of 5. The utilised dataset is created for examining and managing urban traffic utilising DL models. It comprises real-time traffic metrics like location ID, timestamp, average vehicle speed, traffic volume, and numbers of diverse vehicle types (cars, trucks, bikes). Environmental factors, including temperature, weather conditions, and humidity, are also involved, with indicators for accidents and the present traffic signal status. It comprises 2000 instances and 100% under four signal strategies as displayed in Table 3 . The no. of features is 12, but only 11 are chosen. Table 3. Dataset used. Signal strategy Instances Percentage (%) Critical congestion 1514 75.7 Heavy congestion 309 15.5 Moderate congestion 139 6.9 Low congestion 38 1.9 Total 2000 100 Open in a new tab Table 4 provides a decision support system for class labels. For critical congestion, the vehicle count is 650, the signal strategy is Traffic Volume > 300, and the decision support system is enabling the alternate route and increasing the green light duration by 60 s. Besides, for low congestion, the vehicle count is 50, the signal strategy is Traffic Volume < 70, and the decision support system is shortening the green light or activating the pedestrian cycle. Table 4. Decision support system for various class labels. Vehicle count Class labels Signal strategy Decision support system 650 Critical congestion Traffic Volume > 300 Enable the alternate route and increase the green light duration by 60 s 290 Heavy congestion Traffic Volume > 150 Increase green by 30 s 135 Moderate Congestion Traffic Volume > 70 Keep normal timings 50 Low Congestion Traffic Volume < 70 Shorten the green light or activate the pedestrian cycle Open in a new tab Figure 4 displays four different visualisations of vehicle count data to deliver insights into distribution patterns. Figure 4 a demonstrates the car count cumulative distribution, in which the progressively increasing curve specifies that car counts are evenly spread without sudden jumps. Figure 4 b represents the truck count distribution as a histogram with an overlaid curve, displaying frequency variations on various truck count ranges. Figure 4 c exemplifies the bike count density plot, where the smooth, filled curve indicates that bike counts are comparatively evenly distributed between 10 and 50, with less extreme values. Lastly, Fig. 4 d gives a total vehicle count violin plot, merging density distribution and summary statistics, presenting the overall spread and concentration of total vehicle counts. Fig. 4. Open in a new tab Count distribution of HFSDNN-DSSRTM model. Figure 5 describes the traffic volume distribution across five various location IDs. Every box signifies the interquartile range (IQR) of vehicle counts, with the horizontal line in the box indicating the median. Locations 1, 3, and 5 illustrate relatively greater median traffic volumes when compared with Locations 2 and 4. In general, traffic patterns differ across locations. The plot emphasises that Locations 1, 3, and 5 usually manage heavier traffic loads, but Locations 2 and 4 are inclined to have relatively lesser central traffic volumes. Fig. 5. Open in a new tab Traffic volume distribution by location. Figure 6 shows the classifier outcome of the HFSDNN-DSSRTM model at 80:20. Figure 6 a–d depicts the confusion matrices with exact recognition of every class. Figure 6 b points out the PR investigation, denoting maximal performance on each class. Figure 6 c embodies the ROC examination, exhibiting effective outcomes with the highest ROC values for individual classes. Fig. 6. Open in a new tab 80:20 ( a , d ) Confusion matrices and ( b , c ) PR and ROC curves. Table 5 and Fig. 7 depict the classifier outcome of the HFSDNN-DSSRTM method at 80:20. On 80%TRPHE, the HFSDNN-DSSRTM model got , , , , and of 98.41%, 93.06%, 83.21%, 86.98%, and 90.49%, respectively. Likewise, at 20%TSPHE, the HFSDNN-DSSRTM model got , , , , and of 98.75%, 95.13%, 96.10%, 95.41%, and 97.21%. Table 5. Classifier outcome of HFSDNN-DSSRTM model under 80:20. Class labels TRPHE (80%) Critical congestion 97.75 97.64 99.42 98.52 96.02 Heavy congestion 98.19 94.24 93.85 94.05 96.41 Moderate congestion 98.94 95.37 89.57 92.38 94.61 Low congestion 98.75 85.00 50.00 62.96 74.90 Average 98.41 93.06 83.21 86.98 90.49 TSPHE (20%) Critical congestion 98.25 98.39 99.35 98.87 96.99 Heavy congestion 98.25 100.00 89.23 94.31 94.62 Moderate congestion 98.50 82.14 95.83 88.46 97.25 Low congestion 100.00 100.00 100.00 100.00 100.00 Average 98.75 95.13 96.10 95.41 97.21 Open in a new tab Fig. 7. Open in a new tab Average values of the HFSDNN-DSSRTM approach with 80% and 20% Figure 8 presents the training (TRAN) and validation (VALD) accuracy of the HFSDNN-DSSRTM model at 80:20 over 200 epochs. Both curves steadily surge and gradually converge, indicating that the model is learning efficiently. The VALD accuracy persistently stays slightly greater than the TRAN accuracy, indicating that the model is not over-fitting and is generalising better to unseen data. The fluctuations in accuracy are because of the task complexity, but overall upward tendencies prove stronger performance and stability of the model. Fig. 8. Open in a new tab curve of HFSDNN-DSSRTM approach with 80% and 20% Figure 9 presents the TRAN and VALD loss of the HFSDNN-DSSRTM model at 80:20 over 200 epochs. Both curves display a persistent downward trend, denoting that the model is efficaciously minimising error during learning. The VALD loss remains slightly lower than the training loss across most epochs, implying better generality and no signs of over-fitting. Even though some fluctuations are noted, it is becoming progressively reliable and stable. Fig. 9. Open in a new tab Loss curve of HFSDNN-DSSRTM approach at 80% and 20% Figure 10 portrays the classifier outcome of the HFSDNN-DSSRTM model at 70:30. Figure 10 a–d illustrates the confusion matrices with precise detection of every class. Figure 10 b exhibits the PR inspection, denoting maximal performance on every class. Ultimately, Fig. 10 c shows the ROC evaluation, proving efficacious outcomes with the most excellent ROC values for distinct classes. Fig. 10. Open in a new tab 70%:30% ( a , d ) Confusion matrices and ( b , c ) PR and ROC curves. Table 6 and Fig. 11 signify the classifier outcome of the HFSDNN-DSSRTM model at 70:30. On 70%TRPHE, the HFSDNN-DSSRTM approach got , , , , and of 97.68%, 93.08%, 78.07%, 83.09%, and 87.16%, respectively. Similarly, at 30%TSPHE, the HFSDNN-DSSRTM approach got , , , , and of 98.17%, 96.07%, 82.11%, 87.43%, and 89.54%, respectively. Table 6. Classifier outcome of HFSDNN-DSSRTM approach with 70% and 30% Class labels TRPHE (70%) Critical congestion 95.93 95.80 98.96 97.35 92.74 Heavy congestion 98.43 96.98 92.34 94.61 95.92 Moderate congestion 97.64 87.23 79.61 83.25 89.34 Low congestion 98.71 92.31 41.38 57.14 70.65 Average 97.68 93.08 78.07 83.09 87.16 TSPHE (30%) Critical congestion 97.00 96.59 99.56 98.05 94.26 Heavy congestion 97.83 96.77 90.00 93.26 94.70 Moderate congestion 98.50 90.91 83.33 86.96 91.40 Low congestion 99.33 100.00 55.56 71.43 77.78 Average 98.17 96.07 82.11 87.43 89.54 Open in a new tab Fig. 11. Open in a new tab Average values of the HFSDNN-DSSRTM approach with 70% and 30% Figure 12 describes the TRAN and VALD accuracy of the HFSDNN-DSSRTM model at 70:30 over 200 epochs. Both curves steadily rise and gradually converge, which denotes that the model is effectively learning. The VALD accuracy persistently stays slightly superior to the TRAN accuracy, signifying that the model is not over-fitting and is generalising better to unseen data. The fluctuations in accuracy are due to the complexity of the task; however, the overall upward trend reveals strong performance and stability of the model. Fig. 12. Open in a new tab curve of HFSDNN-DSSRTM method with 70% and 30% Figure 13 describes the TRAN and VALD loss of the HFSDNN-DSSRTM model at 70:30 over 200 epochs. Both curves exhibit persistent downward tendencies, denoting that the model is effectively minimising error during learning. The VALD loss remains slightly lower than the training loss across most epochs, indicating good generalisation and no signs of overfitting. Even though some fluctuations are observed, it is becoming increasingly stable and reliable. Fig. 13. Open in a new tab Loss curve of the HFSDNN-DSSRTM method at 70% and 30% Table 7 and Fig. 14 portray a comparative examination of HFSDNN-DSSRTM methodology with existing models under various metrics 16 , 34 . Under , the HFSDNN-DSSRTM model got a maximum of 98.75%. In contrast, the Improved LeNet-5 CNN, Viola-Jones, transfer learning (TL)-VGG16, Improved CNN, MDC + LSTM, OKM-CNN, and MBiMGCGRU models got a minimum of 95.02%, 90.02%, 97.18%, 96.02%, 97.15%, 98.11%, and 97.14%, respectively. Table 7. Comparative analysis of HFSDNN-DSSRTM model with existing approaches. Models Improved LeNet-5 CNN 95.02 90.01 94.00 91.54 Viola-Jones 90.02 94.59 94.28 93.54 TL-VGG16 97.18 92.62 90.80 93.40 Improved CNN 96.02 91.85 94.56 93.90 MDC + LSTM 97.15 94.99 92.14 92.99 OKM-CNN 98.11 90.62 92.88 93.86 MBiMGCGRU 97.14 91.13 94.15 93.73 HFSDNN-DSSRTM 98.75 95.13 96.10 95.41 Open in a new tab Fig. 14. Open in a new tab Comparative analysis of HFSDNN-DSSRTM model with existing approaches. Furthermore, on , the FTDL-PESD model got a maximum of 95.41% whereas the Improved LeNet-5 CNN, Viola-Jones, TL-VGG16, Improved CNN, MDC + LSTM, OKM-CNN, and MBiMGCGRU methodologies got a minimum of 91.54%, 93.54%, 93.40%, 93.90%, 92.99%, 93.86%, and 93.73%, respectively. These results stated that the HFSDNN-DSSRTM model has resulted in better performance over other existing models. Table 8 depicts the ablation study analysis of the HFSDNN-DSSRTM method. The baseline TCN with Filter model attained an of 94.250%, of 90.410%, of 91.670%, and of 91.200%. Furthermore, the TCN-AM + Filter model achieved an of 94.950%, of 91.160%, of 92.190%, and of 91.730%. Moreover, by replacing filter features with wrapper features additionally illustrated an of 95.650%, of 91.930%, of 92.800%, and of 92.350%. Additionally, the TCN-AM + Wrapper method reached an of 96.250%, of 92.470%, of 93.440%, and of 92.970%. Also, the TCN + Embedded model attained an . of 96.810%, of 93.190%, of 94.050%, and of 93.590%, while TCN-AM + Embedded model attained an of 97.420%, of 93.950%, of 94.620%, and of 94.150%. Finally, the HFSDNN-DSSRTM model attained an of 98.150%, of 94.630%, of 95.350%, and of 94.890%. Table 8. Ablation study analysis of the HFSDNN-DSSRTM method. Models TCN + Filter (Without AM and wrapper and embedded features) 94.250 90.410 91.670 91.200 TCN-AM + Filter (With AM without wrapper and embedded features) 94.950 91.160 92.190 91.730 TCN + Wrapper (Without AM and filter and embedded features) 95.650 91.930 92.800 92.350 TCN-AM + Wrapper (With AM without filter and embedded features) 96.250 92.470 93.440 92.970 TCN + Embedded (Without AM and wrapper and filter features) 96.810 93.190 94.050 93.590 TCN-AM + Embedded (With AM without wrapper and filter features) 97.420 93.950 94.620 94.150 HFSDNN-DSSRTM (TCN-AM with Hybrid feature selection) 98.150 94.630 95.350 94.890 Open in a new tab Table 9 indicates the analysis in terms of Floating-Point Operations (FLOPs), Graphics Processing Unit (GPU), and inference time 35 . The HFSDNN-DSSRTM model needed only 5.76 G FLOPs, 875 M GPU memory, an inference time of 0.79 ms, and latency of 0.49. In contrast, YOLOv5m operates at 48.30 G FLOPs with 4195 M GPU usage and 3.93 ms inference time, and latency of 2.54, while YOLOv6s mitigates complexity to 44.20 G FLOPs and 2094 M GPU with 2.90 ms inference time, and latency of 1.34. YOLOv7 and RT DETR L illustrate significantly higher computational costs, requiring 103.50 G and 110.00 G FLOPs with inference times of 2.96 ms and 4.09 ms, and latency of 3.91 and 1.72, respectively. Although YOLOv8n utilizes relatively low FLOPs of 8.20 G, it suffers from higher inference of 6.74 ms, and latency of 2.19, and YOLOv8s balances performance with 28.70 G FLOPs and 2.10 ms inference time, and latency of 4.61. Thus, the HFSDNN-DSSRTM model attained superior computational and memory requirements. Table 9. Analysis of the HFSDNN-DSSRTM method in terms of FLOPs, GPU, and inference time. Models FLOPs (G) GPU (M) Inference Time (ms) Latency (sec) YOLOv5m 48.30 4195 3.93 2.54 YOLOv6s 44.20 2094 2.90 1.34 YOLOv7 103.50 3502 2.96 3.91 RT-DETR-L 110.00 3118 4.09 1.72 YOLOv8n 8.20 3682 6.74 2.19 YOLOv8s 28.70 4778 2.10 4.61 HFSDNN-DSSRTM 5.76 875 0.79 0.49 Open in a new tab Conclusion In this study, the HFSDNN-DSSRTM model is presented for a smart IoT-integrated city environment. The HFSDNN-DSSRTM model aimed to create a decision support system for effective road traffic management in IoT-based smart cities, utilising an advanced method. At the initial stage, the data pre-processing phase is utilised by dual levels, such as missing values handling and normalisation techniques. For an effective feature selection, the HFSDNN-DSSRTM approach employs filter, wrapper, and embedded methods to recognise and keep only the most beneficial features that contribute to the classification task. Finally, the TCN-AM method is utilised for classification. The comparison analysis of the HFSDNN-DSSRTM approach portrayed a superior accuracy value of 98.75% over existing methods. The limitations include the dependence on a single dataset. The sensor noise or incomplete data streams may be affected in real-time deployment. Moreover, external factors such as weather, road construction, and special events were not explicitly modelled. Future studies may explore multi-source data integration, adaptive learning for changing traffic patterns, and incorporating predictive analytics to support proactive traffic management strategies. Author contributions All contributions are made by Khaled Abdullah Almejalli. Funding No funding. Data availability The data that support the findings of this study are openly available in the Kaggle repository at https://www.kaggle.com/datasets/smmmmmmmmmmmm/smart-traffic-management-dataset/data , reference number 33 . Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Musa, A. A. et al. Sustainable traffic management for smart cities using internet-of-things-oriented intelligent transportation systems (ITS): Challenges and recommendations. Sustainability 15 (13), 9859 (2023). [ Google Scholar ] 2. Hilmani, A., Maizate, A. & Hassouni, L. Automated real‐time intelligent traffic control system for smart cities using wireless sensor networks. Wirel. Commun. Mob. Comput. 2020 (1), 8841893 (2020). [ Google Scholar ] 3. Sudha, S., Shoba, B., Rajkumar, A. & Said, B. Bipolar Triangular Neutrosophic Chromatic Numbers with the Application of traffic light system. Int. J. Neutrosophic Sci. (IJNS) , 22 (2) (2023). 4. Dui, H., Zhang, S., Liu, M., Dong, X. & Bai, G. IoT-enabled real-time traffic monitoring and control management for intelligent transportation systems. IEEE Internet Things J. 11 (9), 15842–15854 (2024). [ Google Scholar ] 5. Elassy, M., Al-Hattab, M., Takruri, M. & Badawi, S. Intelligent transportation systems for sustainable smart cities. Transp. Eng. 16 , 100252 (2024). [ Google Scholar ] 6. Saleem, M. et al. Smart cities: Fusion-based intelligent traffic congestion control system for vehicular networks using machine learning techniques. Egypt. Inform. J. 23 (3), 417–426 (2022). [ Google Scholar ] 7. Jagatheesaperumal, S. K., Bibri, S. E., Huang, J., Rajapandian, J. & Parthiban, B. Artificial intelligence of things for smart cities: Advanced solutions for enhancing transportation safety. Comput. Urban Sci. 4 (1), 10 (2024). [ Google Scholar ] 8. Ait Ouallane, A., Bahnasse, A., Bakali, A. & Talea, M. Overview of road traffic management solutions based on IoT and AI. Proc. Comput. Sci. 198 , 518–523 (2022). [ Google Scholar ] 9. Moumen, I., Abouchabaka, J. & Rafalia, N. Enhancing urban mobility: Integration of IoT road traffic data and artificial intelligence in smart city environment. Indones. J. Electr. Eng. Comput. Sci. 32 (2), 985–993 (2023). [ Google Scholar ] 10. Dikshit, S., Atiq, A., Shahid, M., Dwivedi, V. & Thusu, A. The use of artificial intelligence to optimise the routing of vehicles and reduce traffic congestion in urban areas. EAI Endorsed Trans. Energy Web 10 , 1–13 (2023). [ Google Scholar ] 11. Visan, M., Negrea, S. L. & Mone, F. Towards intelligent public transport systems in smart cities; collaborative decisions to be made. Proc. Comput. Sci. 199 , 1221–1228 (2022). [ Google Scholar ] 12. Talaat, F. M., El-Balka, R. M., Sweidan, S., Gamel, S. A. & Al-Zoghby, A. M. Smart traffic management system using YOLOv11 for real-time vehicle detection and dynamic flow optimisation in smart cities. Neural Comput. Appl. 37 (24), 19957–19974 (2025). [ Google Scholar ] 13. Al Asmari, A.F., Almutairi, A., Alanazi, F., Alqubaysi, T. & Armghan, A. Conjecture interaction optimisation model for intelligent transportation systems in smart cities using reciprocated multi-instance learning for road traffic management. IEEE Access (2025). 14. Khan, H. & Thakur, J. S. Smart traffic control: Machine learning for dynamic road traffic management in urban environments. Multimed. Tools Appl. 84 (12), 10321–10345 (2025). [ Google Scholar ] 15. Mrudula, S. T. et al. Internet of things and optimised knn based intelligent transportation system for traffic flow prediction in smart cities. Meas. Sens. 35 , 101297 (2024). [ Google Scholar ] 16. Kheder, M. Q. & Mohammed, A. A. Real-time traffic monitoring system using IoT-aided robotics and deep learning techniques. Kuwait J. Sci. 51 (1), 100153 (2024). [ Google Scholar ] 17. Akash, F.M.M. Employing deep learning for real-time data collection and decision-making for traffic management in smart cities. 18. Johny, T. & Sharma, A. Smarter roads smarter cities: Machine learning integration for dynamic traffic management. In Proceeding of SRSC Conf erence 101–112 (2023). 19. Rai, S. C. et al. ITSS: An intelligent traffic signaling system based on an IoT infrastructure. Electronics 12 (5), 1177 (2023). [ Google Scholar ] 20. Sun, J. et al. DSTIT-TCN: Research on traffic flow prediction method based on big data multi-scale information fusion. Cluster Comput. 29 (2), 107 (2026). [ Google Scholar ] 21. Prasad, K. S. et al. A two-tier optimization strategy for feature selection in robust adversarial attack mitigation on Internet of Things network security. Sci. Rep. 15 (1), 2235 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Yang, R. Research on prediction model of expressway traffic abnormal events based on nonlinear Internet of Things. Secur. Priv. 9 (1), e70136 (2026). [ Google Scholar ] 23. Hernandez-Jaimes, M. L., Martinez-Cruz, A., Ramírez-Gutiérrez, K. A. & Morales-Reyes, A. Network traffic inspection to enhance anomaly detection in the Internet of Things using attention-driven deep learning. Integration 103 , 102398 (2025). [ Google Scholar ] 24. Bansal, N. & Bali, R. S. Graph convolutional network and attention for traffic prediction-A deep learning approach. Recent Adv. Comput. Sci. Commun. 19 (2), e26662558351567 (2026). [ Google Scholar ] 25. Pawar, A.S., Shirode, U.R., Vhatkar, K.N., Sontakke, P.V. & Sarwade, J.M. Development of hybrid adaptive and attention network with enhanced optimization for traffic congestion control in the IoT environment. In: Iranian Journal of Science and Technology, Transactions of Electrical Engineering 1–30 (2025). 26. Singh, R. & Kashyap, R. A hybrid approach for analyzing and securing anomalies in network traffic data using Internet of Things and deep learning. In Internet of Things Security 129–146 (Elsevier, 2026). [ Google Scholar ] 27. Khan, T. A. & Qin, Y. Graph-based deep learning and multi-source data to provide safety-actionable insights for rural traffic management. Vehicles 7 (4), 151 (2025). [ Google Scholar ] 28. Potharaju, S., Tambe, S. N., Tirandasu, R. K., Kumar, D. A. & Kantipudi, M. P. Enhancing cybersecurity in industrial Internet of Things systems using ensemble learning against false data injection attacks. J. Curr. Sci. Technol. 16 (1), 151–151 (2026). [ Google Scholar ] 29. Vatambeti, R. et al. Enhancing urban traffic congestion prediction through EfficientNet and optimized ensemble learning models. Sci. Rep. 15 (1), 40224 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Ouhssini, M. et al. Transparent DDoS defense by combining Kolmogorov-Arnold networks and XAI for real-time protection in cloud environments. Telemat. Inform. Rep. 19 , 100242 (2025). [ Google Scholar ] 31. Süpürtülü, M., Hatipoğlu, A. & Yılmaz, E. An analytical benchmark of feature selection techniques for industrial fault classification leveraging time-domain features. Appl. Sci. 15 (3), 1457 (2025). [ Google Scholar ] 32. Biswas, A.K., Bhuiyan, M.S.A., Mir, M.N.H., Rahman, A., Mridha, M.F., Islam, M.R. and Watanobe, Y. A dual output temporal convolutional network with attention architecture for stock price prediction and risk assessment. IEEE Access ( 2025). 33. https://www.kaggle.com/datasets/smmmmmmmmmmmm/smart-traffic-management-dataset/data [09-09-2025] 34. Saini, K. & Sharma, S. Smart road traffic monitoring: Unveiling the synergy of IoT and AI for enhanced urban mobility. ACM Comput. Surv. 57 (11), 1–45 (2025). [ Google Scholar ] 35. Gao, G., Deng, D., Shi, G., Guo, Y. & Ning, Z. Research on methodology of intelligent traffic accident detection based on enhanced YOLOv8 algorithm. In Proceedings of the 2024 International Conference on Generative Artificial Intelligence and Information Security 340–348 (2024). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data that support the findings of this study are openly available in the Kaggle repository at https://www.kaggle.com/datasets/smmmmmmmmmmmm/smart-traffic-management-dataset/data , reference number 33 . Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (4.1 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top