The state-of-the-art in the application of artificial intelligence-based models for traffic noise prediction: a bibliographic overview

Abstract

This paper reviews the application of artificial intelligence (AI)-based models in modeling vehicular road traffic noise. A computerized search method was used to conduct the literature search. Fifty published articles from 2007 to 2023 were reviewed regarding observation time, input data, countries where studies were performed, and modeling techniques. Sixty-three percent of the studies used an observation period of 60 min, and 29% used 15 min. All the reviewed papers considered traffic flow as the major input parameter, followed by average speed, with 95% of the researchers using it as an input parameter. It was found that using AI-based models for traffic noise prediction was popular in countries with no established empirical models. The primary input parameters for the AI-based models are traffic volume and speed. Traffic volume is used either as total traffic volume or classified into sub-categories, and each category is used as an independent input parameter. Although AI-based models have demonstrated reliable performance regarding prediction error and goodness of fit, the accuracy of the AI-based models’ performance should be compared with the results of the empirical models in countries with established models, such as the UK (CoRTN) and the USA (FHWA).

Keywords:

• Artificial intelligence
• noise
• speed
• vehicular traffic
• prediction

Reviewing Editor:

• Sanjay Kumar Shukla
• Edith Cowan University
• Australia

Subjects:

• Civil, Environmental and Geotechnical Engineering
• Transportation Engineering; Environmental Health
• Pavement Engineering

1. Introduction

Traffic noise is the primary source of environmental noise pollution in urban areas and is believed to be higher in areas with high total traffic volume and speed. Vehicular traffic in urban areas has significantly increased over the years due to the high population, leading to increased noise pollution in regions exposed to heavy vehicular traffic. Every year, approximately 10,000 people die in urban areas due to noise pollution (The European Environmental Agency, n.d.). Traffic noise typically results from the interaction of a vehicle’s tires with the road surface, aerodynamic noise generated by airflow through and around the vehicle, and propulsion noise produced by the engine, exhaust, and transmission. On high-speed highways, aerodynamically generated noise is predominant, while tire-pavement interaction prevails on low-speed routes (Sandberg & Ejsmont, 2000). Vehicular traffic noise is generally influenced by various parameters, such as traffic volume, traffic speed, traffic composition, percentage of heavy trucks, road gradient, road surface, and more (Suthanaya, 2015). These parameters can be influenced by other variables, such as speed limits, the type and number of road intersections, drivers’ skills and behavior, and vehicle maintenance responsibilities (Hamoda, 2008). More than 20% of Europeans living in large cities are at risk of suffering ongoing health issues due to nighttime exposure to noise pollution (Alessandro & Schiavoni, 2015). Noise pollution has caused hearing loss in over 120 million people worldwide in their workplaces (Hamoda, 2008). Sleep disturbances, tinnitus, cardiovascular diseases, and cognitive impairments in children are among the major health problems associated with continued exposure to environmental noise pollution (WHO, 2011). Understanding the complex, non-linear relationship of noise, as well as assessing and predicting it, is essential to provide a healthy, noise-free environment (Roadknight et al., 1997).

Monitoring physical traffic noise alongside high-speed roads can be costly, time-consuming, or even impossible (Ahmed & Pradhan, 2019). These challenges have prompted researchers to develop numerous statistical and regression models, which offer reliable and cost-effective methods for more accurate traffic noise estimation. These prediction models help inform decisions and create safer driving conditions. Some classical models used for predicting traffic noise include the Italian CNR model, the French NMPB-Routes-96, and others (Quartieri et al., 2009). Additional models encompass the Nord 2000 model, the ASJ RTN model from Japan, the US FHWA model, and the HARMONOISE model for EU member states (Can & Aumond, 2018; Quartieri et al., 2009). However, these empirical models are known for their high prediction accuracy in countries where they are developed and in countries with similar traffic patterns and driving habits. Nevertheless, classical models tend to be less accurate in regions with significant variations in traffic patterns and characteristics (Federal Highway Administration, n.d.) due to local factors, such as differences in traffic layout, weather patterns, and road design from one country to another. Empirical and regression models encounter generalization issues in their predictions (Hamad et al., 2017). As a response to these issues, several mathematical and artificial intelligence models have been developed to forecast traffic noise in various regions and nations. Due to their adaptability to manage relationships with non-linear characteristics, artificial intelligence-based models, especially ANN (Hamad et al., 2017; Kumar et al., 2014; Mansourkhaki et al., 2018; Tomić et al., 2018) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) (Codur et al., 2017; Sharma et al., 2014; Yasin Çodur & Tortum, 2015), have gained prominence. These models use major input variables like traffic volume, speed, classifications, and distance from the road edge, which vary based on the country or region. Because of their adaptive nature and ability to handle non-linear characteristics, artificial intelligence-based models have proven to be more effective than classical and empirical models in predicting traffic noise. The traffic noise level is typically expressed as either the weighted equivalent noise level L_eq (Can & Aumond, 2018; Federal Highway Administration, n.d.; Hamad et al., 2017; Mansourkhaki et al., 2018; Kumar et al., 2014) or 10th percentile-exceeded sound level L₁₀ (Garg et al., 2015). The purpose of this study is to provide a comprehensive critical analysis of the factors influencing traffic noise and the use of ANFIS and ANN in traffic noise estimation. In this study, ANN and ANFIS models are the primary AI-based models considered, as they are the most widely used models in the literature. Most studies that have used other models, such as Support Vector Regression (SVR), Gaussian Process Regression (GPR), Random Forest (RF), Boosted Regression Trees (BRT), K-Nearest Neighbor (K-NN), and deep learning models, have used the ANN model as the base for comparison, indicating the acceptability and reliability of ANN in predicting traffic noise. This study will make the following contributions to the field:

1. Investigate the reliability of AI-based models in traffic noise modeling in comparison to established empirical models across various countries.

2. Determine the potential input parameters used in traffic noise modeling and their significance in modeling traffic noise. This will aid in utilizing only the most critical parameters in noise modeling.

3. Examine the various methods for measuring parameters in the field and how they are employed as input parameters in AI-based models.

4. Identify potential directions for future research

2. Methodology

A computerized search method was utilized to conduct the literature review. The online resources of the Near East University Grand Library were employed to search for relevant research articles published in the Scopus, Springer, ScienceDirect, Google Scholar, and Web of Science databases using the following keywords: ‘noise’, ‘vehicular’, ‘roadway’, ‘road’, ‘ANN’, ‘ANFIS’, ‘machine learning’, and ‘modeling’. Furthermore, we retrieved additional studies by tracking the references cited in papers we had already found. Ultimately, we identified 50 papers related to modeling road traffic noise using either ANN or ANFIS. Of these, 38 used ANN as the primary model or for comparison with other models, while only six papers were related to modeling road traffic noise using the ANFIS model. Other models employed included Deep Learning, XGB, GPR, GTA, GBR, RF, K-NN, and SMO. The papers were reviewed based on the countries where the studies were conducted, the duration of data collection, and the publication year. In total, 50 articles that employed machine learning to predict traffic noise were found in the literature search. Table 1 summarizes the previous studies conducted from 2007 to 2023 using the two different machine learning approaches. The trend in applying AI-based models has been steadily increasing since 2007, with only one publication in 2007, compared to nine articles in 2021. This indicates a growing interest among researchers in this field of study (see Figure 1).

Figure 1. Articles distribution over the years.

Table 1. Details of studies reviewed where AI was used in the prediction of traffic noise.

S/N	Author	Study area	Temporal scale	AI technique	Input variables
1	Cammarata et al., 1995	Italy	10 min	ANN	C, MC, HV, RW, H
2	Parbat & Nagarnaik, 2007	India	1 h	ANN	C, MC, HV, H, RW
3	Parbat & Nagarnaik, 2008	India	1 h	ANN	B/HV, C, TC, MC, W, D
4	Genaro et al., 2010	Spain	1 h	ANN	Tm, En, CW, St, Q, Cantecedent, Cdescendant, MCantecedent, MCdescendant, HVantecedent, HVdescendant, Vsiren, traffic condition, V, RS, LAS, LDS, PV, ST, D
5	Givargis & Karimi, 2010	Iran	1 h	ANN	Q, P, G, V, a (165–180)
6	Cirianni & Leonardi, 2011	Italy	1 h	ANFIS	Q, HV, V, and D
7	Cirianni & Leonardi, 2012	Italy	1 h	GRNN	Q, HV, MV, C, B, MC, V, 0.5RW
8	Al-Mutairi et al., 2012	Kuwait	5 min	ANN	V, equivalent C, P, RW, and H
9	Arora & Mosahari, 2012	India	1 h	ANN	Q, V, HV
10	Kumar et al., 2014	India	1 h	ANN	Q (Log Q), V (Log V), P
11	Nedic et al., 2014	Serbia	1 h	ANN	C, MV HV, HV, B, V
12	Žilionienė et al., 2014	Italy	1 h	ANN	Operational V (V85), DSS, Q
13	Sharma et al., 2014	India	1 h	ANFIS	Q, V, HK
14	Cirianni & Leonardi, 2015	Italy	1 h	ANN	P, Q, V, W
15	Garg et al., 2015	India	15–60 min	ANN	MC, TC, C, B, HV, V
16	Khouban et al., 2015	Iran	1 h	ANN	Q, G, RW, V, H
17	Torija & Ruiz, 2015	Spain	5 min	GPR, MLP, SMO
18	Garg et al., 2016	India	15–60	ANN	Lday and Lnight
19	Singh et al., 2016	India	1 h	ANN	Q, P, V
20	Codur et al., 2017	Turkey	1 h	ANN, ANFIS	Q, HV, V
21	Hamad et al., 2017	UAE	15 min	ANN	Q, V, D, T
22	Huang et al., 2017	China	1 h	ANN	HV, MV, C, V, D, angle, building height
23	Freitas et al., 2018	Portugal	1 h	ANN	PV, TX, V, and PD
24	Kumar et al., 2018	India	1 h	ANN	Q, P, V
25	Mansourkhaki et al., 2018	Iran	15 min	ANN	Q, V, C, MV, HV, MC, G, DBF, BRF
26	Tomić et al., 2018	Serbia	15 min	ANN	C, MV, HV, B, MC, D
27	Sharma et al., 2018	India	1 h	ANFIS	Equivalent Q, equivalent V, HK
28	Khalil et al., 2019	UAE	15 min	ANN	D, C, T, V, HV
29	Ahmed & Pradhan, 2019	Malaysia	15 min	ANN	MC, Q, C/Q, P, HWD, C, DSM
30	Bravo-Moncayo et al., 2019	Ecuador	1 h	ANN	Q, PV, V
31	Chen et al., 2020	Europe	1 h	ANN	Q, V, G
32	Nourani et al., 2020a	Cyprus	15 min	Ensemble,	C, B, HV, MC, P, V
33	Nourani et al., 2020b	Cyprus	15 min	Emotional ANN	C, B, HV, MC, P, V
34	Yin et al., 2020	USA	1 h	ANN, XGB, RF,	Q, Road network features, meteorological condition, LU
35	Ahmed et al., 2021	Malaysia	15 min	ANN, DNN	MC, Q, C/Q, P, HWD, C, DSM
36	Ahmed et al., 2021	Malaysia	15 min	ANN, RF	MC, Q, C/Q, P, HWD, C, DSM, time, T, Hm
37	AlKheder & Almutairi, 2021	Kuwait	15 min	ANFIS	LV, HV, V, RW, H, PC, Troad, Tair
30	Gilani & Mir, 2021	India	1 h	GTA	V, Q, RW, HV, HK
39	Chen et al., 2021	China	20 min	ANN	Q, P, V, G
40	Kim et al., 2021	China	1 h	ANN	MC, C, C, MV, HV& B, V
41	Singh et al., 2021	India	1 h	ANFIS	Q, V, P
42	Ranpise et al., 2021	India	15 min	ANN	Q, V, RW
43	Zhang et al., 2021	Czech	9 h	RNN	Vehicle ID, type, location, speed, accc/dec
44	Debnath et al., 2022	India	1 h	ANN	Q, V, P, G, RW
45	Fallah-Shorshani et al., 2022	USA	1 h	XGB	Q, Road network features, meteorological condition, LU
46	Singh et al., 2022	India	1 h		Q, P, V
47	Umar et al., 2022	Cyprus	15 min	SVR, GPR, ANN	C, B, HV, MV, MC, P, V, HK
48	Umar et al., 2023	Cyprus	15 min	Hybrid	C, B, HV, MV, MC, P, V, HK
49	Wang et al., 2023	China		Deep learning
50	Zhang et al., 2023	China	1 h	RF, K-NN, DT, GB, AR, SVR	Traffic composition, street form variables, LU variables, meteorological parameters

a: angle of view; Q: traffic volume; HV: heavy vehicles; C: cars; MC: motorcycles; P: percentage of HV; V: average speed; RW: road width; T: temperature; MV: medium vehicles; B: buses; G: road gradient; H: average height of the buildings facing the road; PV: pavement type; TX: road texture; BRF: building reflection factor; BF: building façade; DSS: distance of the vehicle from the sound source; HK: honking; D: distance of the receiver from the road edge; L: number of lanes; TC: tricycle; PD: pavement distress; HWD: highway density; DSM: digital surface model; RS: road slope; ST: street type; Tm: time of the day; En: environment; Hm: humidity; LU: land use.

Figure 2 displays the distribution of studies based on the countries where they were conducted. Only 16 countries utilized artificial intelligence to predict traffic noise. In countries lacking a well-established empirical traffic noise model, such as the FHWA model, Calculation of Road Traffic Noise (CoRTN) model, German RLS 90 model, ASJ-RTN model 2008, etc., it was observed that the use of AI in traffic noise prediction attracted less attention from researchers. Traffic noise measurement proved to be difficult and time-consuming. Consequently, the use of AI-based techniques serves as an alternative approach in countries with different traffic compositions and characteristics compared to countries with established empirical models. Comparing the results of AI-based models with those of empirical models showed that AI-based models performed better. It was noted that no studies were conducted in certain parts of the world, such as Africa, despite the absence of any established empirical model. Therefore, more research should be conducted in these regions to address this potential research gap.

Figure 2. Number of published papers concerning the countries where the study areas are located.

Regarding the observation periods employed in the studies, all of them utilized observation periods of either 5, 15, 20, or 60 min. Of these, 63% of the studies used a 60-min observation period, while 29% used a 15-min observation period. It’s worth noting that most studies with a 15-min observation period had larger datasets compared to those with a 1-h observation period. The percentage distribution based on the duration of each dataset is presented in Figure 3.

Figure 3. Data observation time.

3. Bibliographic reviews on the application of AI-based models in traffic noise modelling

3.1. Artificial neural network (ANN)

Due to its capacity to simulate intricate non-linear interactions, several researchers have developed Artificial Neural Network (ANN) models for estimating traffic noise. The optimal model structure for all these models is typically obtained through a trial-and-error technique (Federal Highway Administration, n.d.; Hamad et al., 2017; Kumar et al., 2014; Mansourkhaki et al., 2018). The Levenberg-Marquardt (L-M) algorithm and the sigmoid function were the most commonly used training algorithms and transfer functions in ANN-based traffic noise models. The effectiveness of the ANN model largely depends on its architecture, including the number of hidden layers and the neurons within each layer (Mansourkhaki et al., 2018). Irrespective of the input parameters used, ANN models consistently demonstrated superior prediction accuracy when compared to classical and analytical models. For instance, Parbat and Nagarnaik (2007) was the first paper to employ artificial intelligence for traffic noise prediction. The study was conducted in India using parameters, such as C, MC, HV, H, and RW as inputs. The ANN models accurately predicted traffic noise compared to multilinear regression models (MLR) in terms of mean absolute error (MAE) and coefficient of correlation (R). Since this pioneering study, the use of AI techniques in traffic noise modeling has steadily increased, highlighting researchers’ growing interest in this field. In another study, Parbat and Nagarnaik (2008) compared the performance of the ANN model with MLR in predicting traffic noise in Maharashtra, India, using data obtained from 16 observation points. The ANN demonstrated lower error and a higher R-value. Similarly, Genaro et al. (2010) used ANN to model equivalent traffic noise levels in Spain, and the ANN exhibited low prediction error. Furthermore, Al-Mutairi et al. (2012) employed 620 datasets from four distinct roadways from 2007 to 2008 to assess the performance of MLR, backpropagation neural networks (NN), and generalized regression NN for the prediction of traffic noise. Backpropagation NN demonstrated superior performance. Nevertheless, training the generalized regression NN with a genetic algorithm enhanced its performance over the backpropagation NN. sing 95 data sets from 88 locations, Arora and Mosahari (2012) utilized the ANN model to estimate traffic noise along the 90-km NH2 road in India, employing Q, P, and V as input parameters. The ANN model showed strong predictive power. Given the non-linear relationship between noise-generating elements and noise levels, Kumar et al. (2011) analyzed several ANN models created using various variables in different nations around the world. The results consistently demonstrated the superior performance of ANN models over conventional and statistical models. Additionally, Žilionienė et al. (2014) used five sensors along the length of a road to record noise data for 6 months, 24 h a day, and stored it in a GIS database. A 20 km long freeway’s traffic noise was modeled using ANN, with operational speed (V85), the DSS, and Q as the input variables. When compared to nine widely used conventional models, ANN was found to be more accurate and exhibited fewer residuals when predicting traffic noise. In Punjab city, India, Kumar et al. (2014) trained numerous multi-layer FFNNs using the L-M algorithm to forecast a 10-percent exceeded sound level (L10) and equivalent continuous sound level (Leq). The model’s inputs included parameters, such as P, the log (Q), and V for 1 h on various. The t-test results at a 5% level of significance showed that the ANN model could make predictions more accurately than the regression model. Furthermore, the performance of the ANN model for estimating highway traffic noise was compared with various statistical models by (Nedic et al., 2014), and the results supported the ANN's superiority over the other models used. Cirianni and Leonardi (2015) developed an NN model using data from 14 survey sites along a stretch of unbroken road, reasonably far from stop signs and intersections, to estimate the noise level in the Italian city of Villa S. Giovanni. The model’s inputs included parameters, such as Q, P, V, and DSS. The study demonstrated that ANN could accurately predict traffic noise even when using a small database. The study suggests the inclusion of more variables, such as ground type, vehicle classification, PV, and reflecting surface. Lastly, Khouban et al. (2015) proposed an expert system built on ANNs to model traffic noise. They used Mean Squared Error (MSE) and the coefficient of determination (R2) as statistical performance metrics to evaluate the model’s efficiency. Additionally, Singh et al. (2016) explored the capability of ANN to forecast Leq and L10 caused by traffic noise at different sites in Delhi In a study conducted in Patiala, India, the effectiveness of decision trees, random forests, MLRs, and ANNs for estimating traffic noise was assessed and compared, with the outcome demonstrating that Random Forest RF is more reliable and accurate in predicting traffic noise.

The acceptability of ANN models for predicting traffic noise makes them a benchmark for validating the efficiency of other machine learning models, such as SVR (Zhang et al., 2023), GPR (Torija & Ruiz, 2015; Umar et al., 2022), EANN (Nourani et al., 2020b), linear-nonlinear hybrid (Umar et al., 2023), ensemble models (Nourani et al., 2020a), XGB (Yin et al., 2020), RF (Ahmed et al., 2021; Zhang et al., 2023). Evaluation based on R2 values shows an improvement over the ANN model of up to 11.73, 12, 14.62, and 19% for GPR, EANN, hybrids, and ensemble models, respectively. The enhanced performance of these advanced models over the ANN model results from additional pre/post-processing techniques incorporated into the models. For example, EANN achieves accuracy by integrating neural emotions (confidence and anxiety) into the classical ANN model, which are adjusted during the model’s calibration. Optimal results are achieved when the confidence level is high and the anxiety level is low (Nourani et al., 2020b). On the other hand, SVR models were generally found to be less efficient than the ANN model in most comparative studies (Nourani et al., 2020a; Umar et al., 2022). The GPR model gets its high prediction ability from its flexibility to provide uncertainty representation (Cai et al., 2017). The GPR model derives its high prediction ability from its flexibility in providing uncertainty representation (Cai et al., 2017). Similarly, ensemble models, such as BRT and neural ensemble models also outperformed the ANN model. This is attributed to the ensemble power of BRT, which involves creating multiple individual trees before ultimately ensembling them into a single model (Zhang et al., 2021, 2023). It’s important to note that while these models provided higher accuracy than the ANN model, the ANN model consistently delivers satisfactory results.

3.2. Adaptive neuro-fuzzy inference system (ANFIS)

ANFIS was employed to predict vehicle noise levels in Messina and Villa S. Giovanni, Italy, using a dataset consisting of 176 observations. When compared to the conventional regression models used in the literature, ANFIS exhibited superior performance in forecasting traffic noise (Singh et al., 2022). In another application, ANFIS was utilized to evaluate the traffic noise generated by various traffic characteristics in Nagpur, India, using V, Q, and HK as the model’s input variables (Umar et al., 2022). A study conducted in Erzurum, Turkey, aimed to predict traffic noise in urban areas using both ANN and ANFIS. The study’s results indicated that the ANFIS model outperformed the ANN model in its ability to anticipate traffic noise in an urban setting. The R² values for the ANN and ANFIS models were estimated at 0.81 and 0.91, respectively (Umar et al., 2023). The model was also found to provide higher prediction accuracy compared to FHWA, CoRTN, and RM models in a study conducted by (Wang et al., 2023). Zhang et al. (2023) employed a dataset of 480 records, which included L_eq, the average speed of heavy and light vehicles, C, H, RW, pavement quality, H, air temperature, and highway temperature to estimate traffic noise levels on the Kuwaiti ring road using the ANFIS model. The RMSE value for the model’s traffic noise prediction was 0.0022. Additionally, Al-Mutairi et al. (2012) conducted a comparison of the effectiveness of ANFIS, FFNN, generalized linear models, RF, decision trees, and SVM for predicting traffic noise. The ANFIS model demonstrated the lowest error when modeling traffic noise during the validation stage.

In ANFIS-based models for traffic noise prediction, the best performance is typically achieved through a hybrid learning algorithm that combines backpropagation and feed-forward algorithms (Arora & Mosahari, 2012; Huang et al., 2017; Kumar et al., 2011). Studies have indicated that ANFIS exhibits a higher prediction ability than ANN (Codur et al., 2017). However, its application in noise prediction has not been fully realized. Upon reviewing the literature, it is apparent that only six researchers have employed ANFIS in noise prediction studies, as detailed in Table 1. In a comparative study, ANFIS was found to model road traffic noise significantly better than ANN (Codur et al., 2017). Just like in ANN models, the main input variables in ANFIS models include V, Q, HV, and HK, with Leq being the only output parameter. Regardless of the number and type of input parameters, ANFIS was observed to effectively model traffic noise even with limited data, demonstrating its superiority over classical and regression models (Cirianni & Leonardi, 2011). In comparison to other AI-based models, the results from the bibliographic review revealed that ANFIS outperforms RF, DT, and SVM.

4. Factors contributing to traffic noise

Selecting the appropriate input parameters is crucial in any data-driven model. This is because irrelevant variables tend to increase the model’s complexity and decrease its performance accuracy. The study conducted a comprehensive review of potential input parameters to identify the most significant variables for use in developing AI-based models. According to the literature survey, primary input variables employed in the development of statistical and regression models for predicting traffic noise include traffic volume, traffic composition, speed, distance from the noise source, reflective surface, temperature, building facade, gradient, honking, land use, building height, vegetation, obstruction, pavement type, gradient, and acceleration/deceleration (Žilionienė et al., 2014). Additional factors that may impact the previously mentioned criteria include driving style, driver skills, vehicle maintenance responsibilities, speed limits, and road geometry (Kumar et al., 2014). Figure 4 summarizes the factors contributing to traffic noise based on the reviewed studies.

Figure 4. Variables affecting traffic noise (Kumar et al., 2014).

4.1. Speed

Speed plays a significant role in generating traffic noise due to its contribution to the aerodynamically generated noise resulting from vehicle movement. All of the reviewed studies have incorporated speed as an input parameter in their models. Speed, as an input parameter in AI-based models, is employed in three different forms: as the average speed of all vehicles (Hamad et al., 2017; Khalil et al., 2019; Žilionienė et al., 2014), as the average speed for each vehicle category (Garg et al., 2015), as operational speed (Žilionienė et al., 2014) and as the logarithmic value of the average speed (Kumar et al., 2014). The inclusion of speed in traffic noise modeling significantly contributes to the modeling of traffic noise generation (Cirianni & Leonardi, 2015; Nedic et al., 2014). Researchers (Cirianni & Leonardi, 2015; Khouban et al., 2015) discovered a strong correlation between vehicle speed and annoyance for all types of road surfaces. Controlling speed within a limit of 50 km/h reduces the level of traffic-induced noise (Singh et al., 2016). The increased speed of heavy vehicles on smooth road surfaces leads to greater exposure compared to the speed of light vehicles (Nourani et al., 2020b). Traffic noise exhibits a linear relationship with speed, and increasing the speed from 25 to 35 km/h results in a noise increase of 4–5 dB (Nourani et al., 2020a). On average, speed was ranked fourth in terms of its contribution to traffic noise (Hamad et al., 2017). Regardless of the form in which speed was used in ANN, ANFIS, or any other machine learning models, its inclusion consistently improved the efficiency of the traffic noise model. The speed of vehicles can be obtained using speed guns (Garg et al., 2015), traffic counters (Hamad et al., 2017), radar devices, instrumented vehicles, moving observers (Kumar et al., 2014), or by analyzing video recorded by a camera (Mansourkhaki et al., 2018).

4.2. Traffic volume, traffic composition, and proportion of heavy vehicles

Traffic volume is a crucial input variable that significantly contributes to traffic noise. It is widely recognized that noise pollution increases with traffic volume when the road geometry remains the same (Cirianni & Leonardi, 2011; Sharma et al., 2014; Zhang et al., 2021). A 50% reduction in the total vehicle flow or a 50% decrease in heavy vehicle flow can lead to a reduction of approximately 3 dB in noise level (Codur et al., 2017). Total traffic volume can be used as an input or converted into equivalent road traffic flow and employed as an input variable, as demonstrated by (Huang et al., 2017; Yasin Çodur & Tortum, 2015). It can also be categorized into sub-categories, such as 2-wheelers, 3-wheelers, cars, medium commercial vehicles, buses, and trucks, with each category used as a separate input for the model (Cirianni & Leonardi, 2015; Federal Highway Administration, n.d.; Kumar et al., 2014). The classification of noise emissions into six sub-categories was found to enhance the performance of the developed traffic noise model (Sharma et al., 2018). Kumar et al. (2014) transformed the total volume (in passenger car equivalents) into logarithmic form before inputting it into the model. Additionally, the study utilized the ratio of trucks to traffic volume as a distinct input. The presence of heavy vehicles in traffic was identified as a major source of traffic noise compared to light vehicles (AlKheder & Almutairi, 2021; Cirianni & Leonardi, 2011; Nourani et al., 2020a). In a related study by (Singh et al., 2021) the number of heavy vehicles surprisingly ranked as the least influential factor (5th) contributing to traffic noise, whereas the number of light vehicles ranked 2nd in a research conducted in Dubai (Hamad et al., 2017). In their study, Rastogi et al. (2008) found that light vehicles had a greater impact on traffic noise than other traffic classes. The volume of motorcycles was determined to be the most significant predictor of traffic noise (Rao & Tripathy, 2018). Another study indicated that medium-weight vehicles were the dominant vehicles resulting in higher noise levels (Covaciu et al., 2015; Rao & Tripathy, 2018). Traffic volume can be measured using an automatic counter (Hamad et al., 2017), analysis of video recordings (Mansourkhaki et al., 2018), or counted manually (Freitas et al., 2012; Kuşkapan et al., 2022).

4.3. Distance from the noise source

Distance is the primary factor influencing traffic noise when compared to traffic volume, speed, roadway temperature, and the proportion of heavy vehicles (Hamad et al., 2017). Residents living near noisy roadways frequently experience higher annoyance rates (Sharma et al., 2014). Since measuring the exact distance of a passing vehicle to the noise-measuring device on a multilane road is challenging, either the distance from the roadway edge to the device is used (Cammarata et al., 1995; Federal Highway Administration, n.d.) or the distance from the road centerline (Çolakkad & Yücel, 2018). Regardless of the method chosen to represent the distance, the distance from the noise source is an essential factor in modeling traffic noise. The impact of horizontal distance on noise attenuation can only be observed for distances of no more than 60 m (Ece et al., 2018).

4.4. Honking

Studies have demonstrated that honking significantly impacts traffic noise levels (Vijay et al., 2015; Zannin & Ferraz, 2016). In a study aimed at evaluating the effect of honking on traffic noise, it was observed that honking has an effect in heterogeneous traffic but no effect in homogeneous traffic (Zannin & Ferraz, 2016). Honking was found to increase noise levels by approximately 2–5 dB and was considered the most significant source of traffic noise after speed and volume (Nourani et al., 2020a). Depending on the traffic conditions, honking may increase noise levels by 0.5–13 dBA (Ma et al., 2006). Including the honking equivalent for different vehicles is believed to enhance the performance of AI-based models (Sharma et al., 2018). The number of honks in a specified time can be measured using videography (Al-Mutairi et al., 2011) or counted manually (Sharma et al., 2018).

4.5. Pavement type and condition

Pavement types and conditions are directly related to traffic noise as they influence the noise generated from tire/pavement interaction. Concrete pavement produces higher noise at the source than asphalt pavement by approximately 5 dB (Fiedler & Zannin, 2015). In a study evaluating the interior noise of electric vehicles, cement roads generated more noise than asphalt roads due to higher surface roughness (Peng et al., 2019). The sound pressure level increases significantly with water on asphalt roads for all vehicle type (Rodríguez-Molares et al., 2011). Pavement distresses, including no distress, alligator cracking, and ravelling, were confirmed to increase traffic noise (Freitas et al., 2018). Pavement surface roughness and texture impact pavement noise reduction and should therefore be considered essential factors (Mishra et al., 2010). Cobblestone pavements result in more annoying traffic noises than dense asphalt and open asphalt rubber (Cirianni & Leonardi, 2015). The use of porous road surfaces reduces population exposure to noise on high-speed roads (Nourani et al., 2020b). The sound levels of measured and predicted traffic are decreased by the sound absorption provided by porous pavements (Agarwal & Swami, 2011).

4.6. Building façade material, density, and reflection factor

Including building density and the refraction factor were found to reduce the average mean square error and enhance the performance of the ANN model (Mansourkhaki et al., 2018). Improving the insulation of building facade materials helps reduce the noise pollution that reaches the building (Zhang et al., 2021). In a comparative study, the presence of multi-story buildings along the main roads of Wuhan city, combined with an even distribution of roads in the city, results in lower traffic noise levels compared to Greater Manchester (Gündoğdu et al., 2005). Prediction models developed to forecast traffic noise using road gradient, traffic volume, maximum legal noise emission levels for each vehicle type, and the ratio of the highest building height to the road were found to provide accurate predictions (Žilionienė et al., 2014). Vacant spaces were found to have a negative impact on noise level increases, while built-up areas contributed positively to noise level increases (Gökdag, 2012).

4.7. Temperature

Temperature is another factor that affects traffic noise levels. In a study conducted in Tarsus-Adana-Gaziantep, noise distribution was found to be higher in summer than in winter (Nemes et al., 2018). The inclusion of temperature in ANN models enhances the performance of traffic noise models in Sharjah, Dubai (Hamad et al., 2017). The maximum and minimum average noise levels were recorded during summer mornings and winter evenings, respectively, in a study carried out in Taiwan (Vij et al., 2016).

4.8. Road gradient

It was found that road gradient had a significantly greater impact on traffic noise compared to the existing model (Givargis & Karimi, 2010). It was believed to enhance the performance of the ANN model in predicting traffic noise levels (Mansourkhaki et al., 2018) as well as the genetic algorithm model (Žilionienė et al., 2014). Road gradients above 5% have a significant effect on noise attenuation values (Ece et al., 2018).

5. Comparing machine learning models with classical models

For validation, researchers typically compare the performance of the developed machine learning models with that of classical models, either using error metrics, such as mean absolute error, root mean square error, etc., or goodness-of-fit measures like R² or the correlation coefficient. For instance, Mansourkhaki et al. (2018) compared the efficiency of the ANN model with the IRAN model, RLS90, CORTN, and C.N.R model for the prediction of vehicular traffic noise. The ANN model outperformed the classical models. Nedic et al. (2014) also compared the performance of six noise empirical models (Burgess, Griffiths, Fagotti, RLS90, C.S.T.B, and C.N.R) with that of an ANN model. The ANN model outperformed all the classical models in terms of R², ME, MAE, and MAPE. The result was statistically found to be accurate. Hamad et al. (2017) also compared the performance of ANN model with that of Basic Statistical Traffic Noise model (BSTN) and the Ontario Ministry of Transport Road Traffic Noise model (ORNAMENT) and the result proved the superiority of the ANN model. Fallah-Shorshani et al. (2022) evaluated the efficiency of XGB with some classical models (Harmonoise, Howloud-TNM, and CadnaA) using traffic and noise data obtained from USA. The CadnaA and XGB provided the optimum results over the other classical model. The high accuracy of the CadnaA model in this study was due to its calibration with US data. Comparing the accuracy of the machine learning models with CNR and BURGESS shows higher precision and accuracy in the machine learning models (Nourani et al., 2020a, 2020b; Umar et al., 2022). The ability of machine learning models to identify patterns in the data and learn from previous data makes them suitable for modeling complex problems like traffic noise. Machine learning models provide higher accuracy since they are calibrated with data from the case study, whereas classical models are calibrated with data from different locations, likely with different traffic patterns and characteristics.

6. Conclusions

The following conclusions can be drawn from the study: (1) The most significant factors contributing to increased traffic noise are traffic characteristics, which include volume, composition, and speed. The vehicle category is believed to be a major factor contributing to traffic noise. However, it remains controversial as each category (heavy vehicles, medium vehicles, light vehicles, and motorcycles) has been reported to be a major factor contributing to traffic noise. (2) The significance of input parameters depends on the traffic characteristics of the study area. Further research is needed to determine the most important category and the extent to which each category influences traffic noise. (3) The use of machine learning models in modeling traffic noise has also been reviewed, and the results demonstrate that machine learning approaches predict traffic noise models better than analytical and classical models for different data sets and input variables. ANN was found to be the most commonly used and accepted model for traffic noise prediction. However, the application of advanced models, such as ANFIS, EANN, nonlinear hybrid models, and ensembles proved to enhance the performance of the ANN model due to additional pre- and post-processing techniques in these models. (4) L_eq and L₁₀ are the two major noise descriptors used to represent noise levels in urban areas.

7. Recommendations

Recent studies have begun to explore the potential of other AI-based techniques, such as Gaussian process regression, graph neural networks, and expert systems (Jia et al., 2005; Konbattulwar et al., 2016; Rajakumara & Mahalinge Gowda, 2006), for modeling traffic noise. However, further research needs to be conducted to assess their reliability and applicability in modeling traffic noise. Future research should also include additional independent variables, such as pavement type, street aspect ratio, vegetation, locality type, the presence of industries, building facade materials, acceleration and deceleration effects, and barriers. Another aspect that future research should consider is a microscopic approach to predicting traffic noise, where engine noise and rolling noise are modeled separately by equipping a test vehicle to measure the different noises individually.

Author contributions

All authors equally contributed to the study.

Acknowledgements

The authors wish to acknowledge the support of Prince Sultan University, Riyadh Saudi Arabia. This study was also supported by Prince Sattam bin Abdulaziz University project number (PSAU/2023/R/1444).

Disclosure statement

No potential conflict of interest was reported by the author(s).

This study receives no funding.

Data availability statement

No data has been generated during the study.

Additional information

Notes on contributors

Ibrahim Khalil Umar

Ibrahim Khalil Umar obtained his PhD from Near East University Cyprus in 2022. He is currently a Lecturer at the Kano State Polytechnic, Nigeria. His area of research interest includes Machine learning, noise pollution, road safety, traffic induced pollutions.

Musa Adamu

Musa Adamu obtained his PhD from Universiti Teknologi Petronas Malaysia. He is currently a Researcher at Prince Sultan University, Saudi Arabia. His area of research interest includes Artificial intelligence and machine learning, Natural Fiber reinforced concrete, Sustainable Concrete utilizing Supplementary cementitious materials, Rubberized concrete, Modelling and Optimization of concrete’s properties using response surface methodology.

Nour Mostafa

Nour Mostafa received a PhD degree from the Queen’s University Belfast School of Electronics, Electrical Engineering and Computer Science, UK, in 2013. He previously worked as a software developer with Liberty Information Technology, USA/UK, and is currently an associate professor of computer science in the College of Engineering and Technology, American University of the Middle East. His current research interests include grid computing, large database management, artificial intelligence, machine learning, distributed computing, cloud, fog, and IoT computing. He has authored and co-authored many refereed journal articles, conference papers and book chapters. He is an active reviewer for many reputed international and IEEE journals and letters. Also, he has been selected as an International Steering Committee Member of many conferences, and he has joined the editorial board of international journals.

Sadi I. Haruna

Sadi I. Haruna obtained his PhD in Civil Engineering from Tianjin University in 2023. He is currently a lecturer at Bayero University Kano, and a Researcher at Prince Sultan University Saudi Arabia. His area of research interest includes dynamic analysis, Artificial intelligent, Sustainability of Construction Materials; Polymer Materials, Reuse, Durability of Concrete Materials, Materials Characterization Mineral and chemical admixtures for concrete, Structural Rehabilitation with Composite Systems and PU materials.

Mukhtar Fatihu Hamza

Mukhtar Fatihu Hamza obtained his PhD (Control, Automation and Robotics) from University of Malaya 2017. He is currently an assistant professor in the department of mechanical engineering, Prince Sattam bin Abdulaziz University, Al Kharj Saudi Arabia. His area of research interest includes Artificial general intelligence and intelligent controllers, Lower limb exoskeleton robot stability for power argument, Cloud-based CNC machine tools; and cloud-based status monitoring of CNC machine tool.

Omar Shabbir Ahmed

Omar Shabbir Ahmed is a master’s student at King Saud University in Saudi Arabia. He joined Prince Sultan University (PSU) in 2020 as a lab engineer in the Engineering Management department. He is a member of the American Society of Civil Engineers (ASCE) and the Structure and Materials Research Lab. Omar has participated in numerous research projects and published several papers in the areas of construction management, concrete design, and materials structure.

Marc Azab

Marc Azab is currently an Assistant Professor with the Department of Civil Engineering, American University of the Middle East, Kuwait. He received the Ph.D. degree in materials and fracture mechanics from Université Grenoble Alpes, in 2016. His extensive experience in teaching and research, combined with his passion for innovative solutions to complex problems, has earned him a well-deserved reputation as a respected authority in civil engineering and related disciplines. He has published more than 30 articles in prestigious journals, contributing significantly to the advancement of knowledge in his field. His research interests include fracture mechanics, construction materials, and sustainability.