MAPAN-Journal of Metrology Society of India (June 2024) 39(2):397–415

https://doi.org/10.1007/s12647-023-00692-4

ORIGINAL PAPER

Comparison of Analytical and Machine Learning Models in Traffic

Noise Modeling and Predictions
B. S. Chauhan1,2*, N. Garg1,2, S. Kumar1,2, C. Gautam1,2 and G. Purohit3
1
CSIR-National Physical Laboratory, New Delhi 110012, India
2
Academy of Scientific and Innovative Research (AcSIR), Ghaziabad 201002, India
3
CSIR-Central Electronics Engineering Research Institute, Pilani 333031, India

Received: 04 June 2023 / Accepted: 21 September 2023 / Published online: 14 December 2023

 Metrology Society of India 2023

Abstract: This paper illustrates the applications of analytical models and machine learning methods to predict the
equivalent continuous sound pressure levels (LAeq) along with 10-percentile exceeded sound levels (L10) generated due to
road traffic noise based on rigorous noise monitoring conducted at more than 200 locations in Delhi-NCR. Using the
measured data, regression, back-propagation neural network, and machine learning models were developed, validated, and
tested. The work represents that the developed models are suitable for reliable and accurate predictions of hourly traffic
noise levels. A comparative study reports that the machine learning-based model outperforms the classical analytical
models. Multiple linear regression models and three machine learning techniques, namely decision trees, random forests,
and neural networks, were utilized for developing models that predict the hourly equivalent continuous sound pressure
level (LAeq1h) and 10-percentile exceeded sound pressure level (L10). The developed predicted models have been ascer-
tained to show an accuracy up to ± 3 dB(A). The proposed prediction models in the study can serve as a tool for planning
noise abatement measures and traffic noise forecasts for the Delhi-NCR region. This study is the first rigorous study of its
kind that covers a larger number of areas and zones in Delhi-NCR for assessment and predictions of road traffic noise and
also shows an illustrative example of estimating measurement uncertainty in hourly noise measurements as per ISO
1996-2:2017.

Keywords: Traffic noise level; Noise modeling; Neural networks; Machine learning models

1. Introduction noise pollution close to residences [6, 7]. Some of the

detrimental health consequences of noise pollution are
Road traffic noise plays a very reminiscent role in envi- sleep disruption, speech disruption, annoyance, cardiovas-
ronmental impact assessment (EIA) studies [1]. It has cular damage, hearing problems, etc. [8–10]. Strategic
always been necessary to investigate road traffic noise due noise monitoring is very helpful in assessing the environ-
to the country’s growing urbanization and expansion of the mental noise impact in big metropolitan cities [11]. In
road network. Urbanization brought forth by technological addition to helping with the construction and improvement
advancement has resulted in an issue with growing traffic of road and traffic infrastructure, traffic noise prediction
noise [2]. Increased levels of noise and air pollution as a models are extremely vital [12]. A regression-based
result of heavy traffic congestion, especially in key urban methodology was mostly employed in prior studies from
areas, have been an inevitable result of the ongoing pro- Indian perspective [13–16]. Different noise prediction
duction and purchase of vehicles, especially private ones models have been created in various Indian cities during
[3–5]. Due to the consequences on inhabitants, public the past few decades. A regression equation was formu-
health and town planners have made it imperative to give a lated by Rao et al. 1991 to model LA10 as a function of
thorough assessment, monitoring, and model of traffic traffic density [13]. Traffic signals frequently cause the
majority of the traffic flow in urban areas to be interrupted,
which significantly alters the noise characteristics of urban
*Corresponding author, E-mail: [email protected] roads compared to highways and expressways [14, 15]. For

123
398 B.S.Chauhan et al.

the purpose of calculating Leq values, Agarwal [16] traffic noise prediction models like CRTN, FHWA, and
established the equivalent number of light (LV) and heavy RLS 90 [39]. Industrialized nations have conducted a sig-
vehicles (HV). Noise pollution has been determined to be nificant amount of study in this field, and each nation has
primarily caused by light motor vehicles. The artificial developed and validated their prediction models, such as
neural network, commonly known as ANN is an important CoRTN (in UK), RLS 90 (in Germany), and ASJ-RTN
methodology utilized in research efforts on the machine 2008 (in Japan) [29]. Different vehicular categories having
learning (ML) domain during the past ten years that have widely variable static and dynamic properties make up the
concentrated more on the new techniques of generating homogeneous traffic in Indian cities. Every vehicle type is
ensembles. When compared to multiple linear regression distinct in its own right and cannot be directly compared
models, the ANN technique has demonstrated superior with other kinds because of the significant differences in
performance and accuracy [17–24]. Using the ANN tech- physical dimensions and speeds [40]. As a result, several
nique for road traffic noise assessment in India, Kumar studies report using passenger car units (PCU) to estimate
et al. 2014 forecasted the 10-percentile sound levels (L10) road capacity [40, 41]. An ANN model to predict equiva-
and the equivalent continuous sound levels (Leq) in deci- lent continuous levels and 10-percentile noise levels with
bels (A) [25]. The hourly pass-by traffic volume, average several noise descriptors is commonly used in wide
vehicle speed, and heavy vehicle percentage are the input applications. Many researchers have created traffic noise
parameters taken into account by the ANN approach prediction models using artificial intelligence techniques.
model. The results were also compared to a multiple linear They have implemented artificial neural networks (ANNs)
regression (MLR) model using an ANN technique, yielding in modeling [2, 25, 35, 36]. By using data from the various
more trustworthy outcomes for the modeling of traffic regions where the models have been constructed, genetic
noise [25]. Honking noise is one factor that has not typi- algorithm is also employed to estimate vehicular traffic
cally been taken into account despite being extremely noise [37, 42].
common in Indian road conditions. Numerous researchers In order to determine the applicability of analytical and
have proved that honking is important in the Indian con- machine learning models in traffic noise predictions and
text. For instance, Thakre et al. reported that the enlarge- forecasting the environmental noise levels of a large
ment of the transportation sector in Nagpur resulted in an metropolitan region, the present study used hourly traffic
increase in noise level by 5 dB(A) to 6 dB(A), in honking noise data measured at 200 locations in Delhi-NCR. This is
by 4 to 6 times, and in traffic flow by 1.7 times more than the first rigorous study of its kind that covers a larger
earlier study [26]. Garg et al. 2015 investigated the per- number of areas and zones in Delhi-NCR for assessment
formance of the outputs of the ANN model and the MLR and predictions of road traffic noise. There has not been
model for road traffic noise prediction in Delhi, India. In any prior research on the larger scale of formulation of
their analysis, the ANN model performed superior over the noise prediction models for Delhi-NCR. For various sites
MLR models in terms of reliability and accuracy. How- in Delhi-NCR, the developed models are utilized to fore-
ever, it was found that the ANN has optimized the mea- cast the equivalent continuous sound levels (LAeq) and
sured and predicted data’s root mean square error (RMSE) 10-percentile exceeded sound levels (L10) caused by traffic
and coefficient of determination (R2) [2]. The adaptive noise. Most analytical models in the Indian context are
neuro-fuzzy inference system (ANFIS), which blends ANN based on solo monitoring sites, making it difficult to be
systems with fuzzy interference systems, is another hybrid applied to the other sites as well. The goal of the current
model that has been produced in recent studies (FIS) [27]. study is to create a generalized model using experimental
Numerous research provides a critical evaluation of the data acquired from diverse sites. For each of the con-
most popular models [28–31]. Other types of models structed models, the measured values and the predicted
include those based on artificial intelligence or machine traffic noise levels are compared. The models are validated
learning are dynamic and static, random, consistent, and using correlation coefficient (R), coefficient of determina-
more [32–34]. It was observed in the literature that tion (R2), mean square error (MSE), and mean absolute
numerous researchers used artificial intelligence methods percentage error (MAPE). The novelty in this exhaustive
to predict road traffic noise, including genetic algorithms study lies in the comparison of analytical and machine
and artificial neural networks (ANN) [35–37]. Regression learning models, and the development of an accurate
models and statistical analysis have also been used to machine learning (ML) model with larger datasets covering
construct some models that take honking into account [38], all zones and noisy hotspots of the big metropolitan city of
and other models have been developed by including a horn Delhi along with the National Capital Region (NCR) of
correction factor in reference of level of service in standard country.

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 399

2. Materials and Method levels. The traffic flow of all the hourly pass-by vehicles,
average traffic speed, heavy vehicles percentage, and
2.1. Experimental Data Acquisition and Site number of honking events were taken as input variables.
Description The experimental data collected from various locations
were used to develop the models. On the basis of traffic
Traffic noise levels were measured in decibels (dB(A)), density, the prevalence of honking, and the vicinity of the
average vehicle speeds were observed using a laser-based residential population, these variables have been manually
traffic speed gun (Make: Bushnell), and the number of collected utilizing videography at 200 separate specific
moving vehicles was manually recorded. The number of locations. The light, medium, and heavy vehicles and the
honking events was also counted manually. The study number of honks in the measurement period had been
utilized two calibrated sound level analyzers, B&K 2250 recorded. A visual representation of some of the significant
and Norsonic Nor145, both traceable to the national stan- noise monitoring locations in Delhi-NCR during the time
dards of sound pressure level maintained at CSIR-National of the noise measurements is shown in Fig. 2. The strategy
Physical Laboratory, India [43]. According to the standard of selection of monitoring sites was based on selecting all
recommendations stated in ISO 9613-1, the sound level kinds of traffic densities in different areas of the study area
analyzer was installed on a tripod stand and positioned at a so that the 200 selected sites include all kinds of noisy
height of 1.5 m above the ground level. It was also located hotspots, zones i. e. silence, residential, commercial,
around 3–3.5 m away from walls to avoid reflections [44]. industrial, highway and traffic intersections sites in Delhi-
It was ensured during the noise measurements to avoid any NCR. The expanded measurement uncertainty (coverage
kind of reflection from the wall or any obstacle. When the factor, k = 2) in noise monitoring is observed to be ± 2
noise monitoring was performed, meteorological factors dB(A). As specified by ISO 1996-2:2017, Table 2 provides
such as temperature, relative humidity, and wind speed an illustration of uncertainty estimation for a typical noise
were also taken into account. The temperature was measurement under favorable propagation conditions at
observed as (23 ± 8) C and relative humidity was sites with no reflecting surfaces other than the ground
observed as (50 ± 20) % during the measurements. The [48, 49]. The uncertainty parameters as described in ISO
wind speed was observed between 1 to 5 m/s [45]. Short- 1996-2:2017 have been explained to evaluate the uncer-
term, LAeq,1h noise measurements (200 locations) were tainty in LAeq measurements under favorable propagation
undertaken at all zones in Delhi city covering commercial, conditions [48]. Table 2 shows a case study for a locality
residential, industrial, silence zones and additionally where the short-term LAeq level monitored was
highways, intersections, and crowded markets during the 70 dB(A) and the total traffic flow was 1000 vehicles per
period of October 2021 to January 2023 on different days. hour. The expanded uncertainty of ± 2 dB(A) (k = 2) with
It may be noted that designating an area as commercial, a 95% confidence interval was estimated for this case study
residential, industrial, or silent zone is as per the decision as per the Guide to the Expression of Uncertainty in
of competent authority as per the Central Pollution Control Measurement [50]. The sensitivity coefficient is calculated
Board guidelines [46]. However, the present study cate- as per the guidelines recommended in ISO 1996–2:2017
gorizes the various sites based on the author’s perception [48].
about the site characteristics. The locations of the places
where traffic noise monitoring was done are depicted in 2.2. Noise Map Generation
Fig. 1. Desktop application Google Earth Pro version
7.3.4.8248 (64-bit) is used to build the site map [47]. Noise map involves the assessment and visualization of
The present study was performed for 200 sites in Delhi- noise levels in a given study area. It provides a detailed
NCR comprising, 50 silence zone locations, 45 residential representation of the noise distribution and levels across a
zone locations, 48 commercial zone locations, 24 industrial particular region, typically in urban or industrial settings.
zone locations, 10 highway locations, and 23 intersection The purpose of this mapping is to understand and analyze
locations. The silence zone comprises various educational the impact of noise pollution on the environment and
areas, hospitals, religious places, lutein zones/presidential identify the noisy hotspots in Delhi-NCR. This map shall
estates, historical monuments, etc. The commercial zone be very helpful for policymakers, urban planners, and
comprises crowded markets, government and private offi- researchers to make informed decisions regarding noise
ces, shopping malls, ISBT terminals/bus stands, railway control measures, land-use planning, and the implementa-
stations, airports, etc. Table 1 shows the summary of the tion of mitigation strategies. The contours display the dif-
200 noise monitoring sites in Delhi-NCR wherein the noise ferent color-coded zones that represent the levels of noise
monitoring was conducted. The various models were intensity in the study area. Using the ArcGIS v10.8 pro-
developed for predicting the hourly LAeq and L10 sound gram, a noise map showing the traffic noise scenario of 200

123
400 B.S.Chauhan et al.

Fig. 1 Site map of 200 noise monitoring locations in Delhi-NCR

Table 1 Summary of the 200 noise monitoring sites in Delhi-NCR

Zones/sites No. of Maximum traffic flow Minimum traffic flow Sound levels LAeq Sound levels LAeq Percentage of heavy
sites Qmax Qmin max min vehicles (p)

Silence 50 8244 1647 80.5 66.5 10.7

Residential 45 8504 1846 80.4 67.8 10.8
Commercial 48 8783 1689 83.0 68.2 10.5
Industrial 24 8918 4571 80.5 74.9 11.5
Highway 10 8419 6299 82.4 79.0 11.1
Intersection 23 8861 3318 85.5 71.3 12.0

sites in the Delhi-NCR is generated using the inverse dis- prediction is commonly used in several fields, including
tance weighting (IDW) approach depicted in Fig. 3. urban planning, environmental assessment, industrial
design, and transportation planning. The prediction and
forecasting of noise levels facilitates the decision-makers
3. Noise Prediction Models to evaluate the detrimental impacts of noise on human
health, wildlife, and the surrounding environment. This
Noise prediction refers to the process of estimating or perception can then be used to create effective noise mit-
forecasting the noise levels that are expected to be present igation strategies, optimize building or infrastructure
in a given environment or situation. It involves using designs, or evaluate the suitability of certain activities in
various techniques and models to analyze factors such as specific locations. To make accurate noise predictions,
sources of noise, their characteristics, and the propagation factors such as the type and intensity of noise sources, their
of sound waves to determine the expected noise levels at distance from the receiver, the characteristics of the sur-
specific locations or over a certain period of time. Noise rounding environment, and any relevant regulations or

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 401

Table 2 Uncertainty estimation for a typical noise measurement under favorable propagation conditions at sites as per ISO 1996-2:2017 [48, 49]
Parameter Estimate Standard uncertainty, ui dB Sensitivity coefficient, ci Uncertainty contribution ci ui dB

Leq, 1h L0 = 70 dB 0.50 1
1100:1ðL0 Lres Þ
= 1.188 0.59
dslm Normal distribution 0.5 1 0.50
dsou 1000 vehicles pffiffiffiffiffiffiffi
10
¼ 0:3 1 0.14
5000
dmet Favorable 0.5 1 0.5
dloc 0.40 1 0.40
0
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 100:1ðL Lres Þ
dres Lres = 65 dB 2.00 1100:1ðL0 Lres Þ
= 0.188 0.38
P n
uðLm Þ ¼ ðcj uj Þ2 1.00
1

Expanded uncertainty 2.00

(coverage prob. 95%)

standards need to be considered. Various mathematical Here, QCar, Q2W, QMCV, Q3W, QBus, and QTruck,
models and simulation techniques, such as sound propa- respectively, represent the overall volume of traffic for
gation models, statistical analysis, and computational cars, two-wheelers (2W), medium commercial vehicles
algorithms, are employed to estimate and predict noise (MCV), three-wheelers (3W), buses, and trucks. The
levels in different scenarios. These predictions can assist in passenger car noise equivalent values were calculated as
SPLT SPLC
making informed decisions and taking appropriate mea- E ¼ 10ð 10 Þ whereas, SPLT is the sound pressure level
sures to manage and control noise pollution effectively. A of heavy vehicles and SPLC is the sound pressure level of a
flowchart of the noise prediction methodology utilized in car [37, 51, 52]. It may be noted that the reference speed
the present paper is shown in Fig. 4. The different noise was assumed to be 50 km/h. Therefore, an adjustment
prediction models developed in the present work are briefly factor may be incorporated in the reference emission level
explained below. as, DLV = A(Veq - 50), where A represents a constant and
Veq represents the number of equivalent vehicles [37, 52].
3.1. Multiple Linear Regression Models (Analytical According to measurement observations of the pass-by
Models) noise levels of different vehicles conducted at various sites,
the empirically determined acoustical equivalence values
Multiple linear regression (MLR) models were developed E2W, EMCV, E3W, EBus, and ETruck for 50 km/h reference
for predicting the equivalent continuous sound pressure speed were: 0.93, 1.58, 2.09, 6.03, and 8.71, respectively,
levels (LAeq) along with 10-percentile exceeded sound as shown in Fig. 5. For comparison, the value of Ecar is
pressure levels (L10). The traffic flow, average speed of assumed to be one as suggested by Garg et al. in 2015 [2].
vehicles, heavy vehicles percentage, and number of honk- The equivalent vehicular speed is evaluated as:
ing events were taken as input variables. The models were Veqv ¼ ðQCar  VCar þ Q2W  E2W  V2W þ QMCV
created using experimental data that was collected from
EMCV  VMCV þ Q3W  E3W  V3W þ QBus
numerous sites. For Eqs. (3) and (4), an analytical regres-
sion model was created, and the parametric constants A, B, EBus  VBus þ QTruck  ETruck  VTruck Þ =Qeqv ;
C, D, and E were assessed. A software program was used to ð2Þ
evaluate these constants (SPSS version v 20). The iteration
The average speed of cars, 2W, MCV, 3W, buses, and
rate was fixed at 500 and the constraints were taken
trucks are shown here as VCar, V2W, VMCV, V3W, VBus, and
assuming zero constraints. Table 3 illustrates the results of
VTruck, respectively, in km/h.
constants estimated in LAeq prediction for the Qeqv objec-
The fundamental model equations are assessed in terms
tive function.
of total traffic flow and equivalent traffic flow denoted by Q
The equivalent traffic flow is evaluated as:
and Qeqv respectively, which are objective functions
Qeqv ¼ QCar þ Q2W  E2W þ QMCV  EMCV þ Q3W  E3W established based on prior investigations [37, 52]
þ QBus  EBus þ QTruck  ETruck ;
ð1Þ

123
402 B.S.Chauhan et al.

Fig. 2 Pictorial view of some noise monitoring locations in Delhi-NCR. a Pari Chowk, Noida, b Delhi–Meerut Highway, c Sarojini Market,
d Naraina, e IGI Airport, f Ashram Chowk, g INA, h Faridabad, i Punjabi Bagh, j Red Fort, k Cyber City, Gurugram, l Delhi–Agra Highway

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 403

Fig. 3 A noise map of traffic noise scenario at 200 locations in Delhi- NCR

LAeq ¼ A þ B  log 10Qeqv þ C  log 10Veqv þ D  p (in km/h). p is the heavy vehicle percentage and n is the honking
þ E  n; events that occurred in measurement time duration. A, B, C, D,
and E are constants that are to be calculated.
ð3Þ
The final MLR model equations are formulated as
and Eqs. (5) and (6) for Qeqv,
LAeq ¼ A þ B  log 10Q þ C  log 10V þ D  p þ E LAeq ¼ 22:3 þ 7:98 log Qeqv þ 12:28 log Veqv þ 0:11
 n;  p þ 0:003  n; R2 ¼ 0:90; ð5Þ
ð4Þ
L10 ¼ 30:3 þ 6:01 log Qeqv þ 13:45 log Veqv þ 0:10
where LAeq is the equivalent sound level and Qeqv and Veqv are  p þ 0:003  n; R2 ¼ 0:82; ð6Þ
equivalent traffic volume and equivalent traffic speed (in km/h),
while Q is total traffic volume and V is average vehicular speed

123
404 B.S.Chauhan et al.

Fig. 4 Flowchart for the

methodology of noise prediction Input experimental traffic data
model

Input parameters

(Q, V, p, n)

Select Prediction Methodology

1. MLR’s
2. ANN
3. Decision Tree
4. Random Forest

Prepare Training, Testing and Validation data

Configure the model

Problem in
Training?

Yes
No

Is M.S.E.
minimum?

No
Yes

Run and validation

Forecasting of output parameters (LAeq and L10 levels)

Table 3 Nonlinear parametric estimation in terms of Qeqv for LAeq prediction

Parametric constant Estimate Standard error 95% CI
Lower bound Upper bound

A 22.304 2.869 16.647 27.962

B 7.984 0.627 6.747 9.221
C 12.284 1.592 9.143 15.424
D 0.110 0.030 0.050 0.169
E 0.003 0.000 0.002 0.004

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 405

(a)
90.0
y = 0.992x + 0.5846
85.0 R² = 0.90

Predicted LAeq dB(A)

80.0

75.0

70.0
Qeqv = 1×Car+0.93×TwoWheeler+1.58×MCV+2.09×Three
Wheeler+6.03×Bus+8.71×Truck 65.0

60.0
60.0 65.0 70.0 75.0 80.0 85.0
Measured LAeq dB(A)

(b)
95.0
y = 1.0249x + 1.7633
90.0
R² = 0.82

Predicted L10 dB(A)

Fig. 5 Passenger car noise equivalent (PCNE) evaluated for acoustic 85.0
equivalence (E) of different vehicular categories in Delhi-NCR
80.0

75.0
A similar technique is involved in estimating the constants 70.0
A, B, C, D, and E for the objective function represented by 65.0
Eq. (4). The constants estimations in LAeq formulation are 60.0
shown in Table 4 for Q. 60.0 65.0 70.0 75.0 80.0 85.0
Measured L10 dB(A)
The final MLR model equations are formulated as
Eqs. (7) and (8) for Q, Fig. 6 (a) Plot between predicted and measured LAeq for MLR model
LAeq ¼ 23:6 þ 7:74 log Q þ 13:27 log V þ 0:13  p in terms of Qeqv. (b) Plot between predicted and measured L10 for
MLR model in terms of Qeqv
þ 0:003  n;
2
R ¼ 0:90;
determination (R2) is observed to be higher in both the
ð7Þ models so developed.
L10 ¼ 31:3 þ 5:77 log Q þ 14:50 log V þ 0:09  p
þ 0:003  n; ð8Þ 3.2. Machine Learning Models
2
R ¼ 0:82;
Machine learning is a subfield of artificial intelligence (AI)
Figure 6a, b shows plots between the forecasted and and computer science that focuses on using data and
measured data for the model developed in terms of the algorithms to simulate how people learn, gradually
passenger car noise equivalent (PCNE) concept. The mean increasing the accuracy of the model. The rapidly
square errors (MSE) between the forecasted and measured expanding discipline of data science includes machine
data obtained from Eqs. (5) and (6) are 1.09 dB(A) for learning as a key element. Algorithms are trained to gen-
LAeq and 1.95 dB(A) for L10, respectively. Figure 7a, b erate classifications or predictions and to find significant
presents the plots between the predicted and measured data insights in a variety of data projects through the use of
for the model formulated in terms of Q. For LAeq and L10, statistical approaches. The empirical noise models for
the mean square error (MSE) for predicted data compared estimating traffic noise are thought to be accurate in pre-
to measured data is 1.09 dB(A) and 1.92 dB(A), dicting the traffic noise of the country for which they are
respectively. It may be noted that coefficient of intended, but they struggle in locations with considerable

Table 4 Nonlinear parametric estimation in terms of Q for LAeq prediction

Parametric constant Estimate Standard error 95% CI
Lower bound Upper bound

A 23.612 2.781 18.128 29.096

B 7.744 0.639 6.485 9.003
C 13.266 1.692 9.929 16.603
D 0.132 0.027 0.079 0.186
E 0.003 0.000 0.002 0.004

123
406 B.S.Chauhan et al.

(a) 90.0
complex nonlinear processes, have been utilized to model
y = 0.9186x + 8.7636 vehicular traffic noise in order to overcome the limits of the
85.0 R² = 0.90
classical models. It is well known that different approaches
Predicted LAeq dB(A)

80.0 to a given problem may provide various outcomes

75.0
depending on the circumstances. Three different machine
learning approaches have been used (presented in Table 5),
70.0
in order to develop models for the forecasting of sound
65.0 pressure levels LAeq and L10. The performance of the
60.0 developed models has been examined using different cri-
60.0 65.0 70.0 75.0 80.0 85.0 90.0
teria of Pearson correlation (R), root mean square error
Measured LAeq dB(A)
(RMSE), coefficient of determination (R2), mean absolute
(b)
95.0 percentage error (MAPE), etc. using these soft computing
90.0
y = 0.8114x + 16.473 methods. The traffic noise prediction models were devel-
R² = 0.82
Predicted L10 dB(A)

oped using three machine learning approaches, namely

85.0
artificial neural network (ANN), decision trees (DT), and
80.0
random forest (RF). The above-mentioned machine learn-
75.0
ing (ML) approaches utilized in the present paper are
70.0 explained below in brief.
65.0

60.0 3.2.1. Artificial Neural Network (ANN)

60.0 65.0 70.0 75.0 80.0 85.0 90.0 95.0
Measured L10 dB(A)
ANN is one of the soft computing methods in spirit by the
Fig. 7 (a) Plot between predicted and measured LAeq for MLR model human brain. It works well as a stand-in for performing
in terms of Q. (b) Plot between predicted and measured L10 for MLR multiple regression analysis or solving the least squares
model in terms of Q problem. There are several other types of neural networks,
but the back-propagation one is a more basic and popular
variations in the composition of the traffic. The classical kind of it [53]. The multilayer perceptron performs well by
models may have certain limitations in the prediction of the addition of one hidden layer or more hidden layers in
traffic noise because of the complexity of traffic noise. compare of the layer containing a single neuron. It is the
Artificial intelligence (AI) techniques, which are recog- more reliable approach in learning and has a generic input–
nized to be more reliable instruments in dealing with output relationship. Thus, a neural network is made up of a

Table 5 Error analysis of the developed traffic noise prediction models

Parameter Minimum Maximum error Mean square error Root mean square error Mean absolute Coefficient of
error (dB(A)) in (dB(A)) (MSE) in dB(A) (RMSE) in dB(A) percentage error determination (R2)
(MAPE %)

Analytical model, MLR (equivalent traffic flow)

LAeq - 3.86 3.09 1.09 1.05 1.05 0.90
L10 - 5.99 4.11 1.95 1.39 1.35 0.82
Analytical model, MLR (total traffic flow)
LAeq - 3.83 3.04 1.09 1.05 1.04 0.90
L10 - 5.97 3.60 1.92 1.39 1.36 0.82
ANN model
LAeq - 2.31 4.03 0.80 0.89 0.86 0.93
L10 - 3.53 5.27 1.40 1.18 1.13 0.87
Decision trees model
LAeq - 1.89 3.16 0.49 0.70 0.68 0.95
L10 - 2.39 3.92 0.87 0.93 0.93 0.92
Random forest model
LAeq - 1.93 3.55 0.32 0.56 0.49 0.97
L10 - 1.85 4.94 0.49 0.70 0.62 0.95

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 407

Table 6 Neural network architecture for training phase

Parameter ANN model structure Training algorithm Activation function MSE in dB(A) Pearson correlation, R
Training Testing Training Testing

LAeq 11:16:1 Trainlm Logsig 0.62 1.36 0.98 0.96

L10 11:16:1 Trainlm Logsig 0.68 1.81 0.97 0.88

1 first one is ‘‘steepest descent approach’’ and the second one

Training Testing
Correlation Coefficient, R

is ‘‘Gauss–Newton approach.’’ This retains the steepest

descent method’s stability and the speed advantage of the
Gauss–Newton algorithm. In order to minimize the dis-
crepancy between the predicted and target (output) values,
ANN examines the input-target data from the training
dataset and updates the weighted link values accordingly.
Through numerous training cycles commonly known as
0
4 6 8 10 12 14 16 18 20 22 24 ‘‘epochs,’’ the forecasting error is reduced until the desired
No. of Neurons target performance is attained. The MATLAB neural net-
work toolbox version R2021a was utilized for developing
Fig. 8 Correlation coefficient, R between predicted and measured
noise data for ANN structure (11-N-1), where N is number of hidden the desired model. The data is separated into train phase,
layers test phase, and validation phase datasets in order to assess
the model’s validity. Although a network has no more than
4 one hidden layer in general, it might be challenging to
Mean Squared Error, dB(A)

Training Testing
establish the optimal number [57]. The network ran
3 through many trials in the hidden layer by using different
numbers of neurons. The number of neurons for the hidden
2 layer is optimally placed to achieve the performance cri-
terion for MSE and R between the measured and predicted
1 data. Utilizing the dataset of test phase, the created and
validated network may be used to make accurate
0
4 6 8 10 12 14 16 18 20 22 24
forecasting.
No. of Neurons
3.2.1.1. Training of ANN Model The process of feeding
Fig. 9 Mean square error (in dB(A)) between predicted and measured input and output datasets to a back-propagation neural
data for ANN structure (11-N-1), where N is number of hidden layers
network is insistent, requiring the network to change its
weights until the performance requirement is met. An input
number of layers of artificial neurons that are connected to
vector with traffic parameters and an output that represents
one another containing a ratio of 1: C 1:1 for the input
the noise levels make each training pair. Data from the first
layer, hidden layer, and output layer, respectively. Artifi-
input vector is received by the network, which then suit-
cial neurons serve as the fundamental processing units of
ably processes it through the activation of hidden neurons
neural networks. The output for that neuron is created by
to produce a target through a unit in the output layer. Back-
multiplying the inputs by adjusted weights and then pass-
propagation of error is used to optimize weights with biases
ing the result through a transfer function [54]. The sigmoid
for all network units during the learning session. The
(logistic) function is the transfer function that is most fre-
training proceeds until the network either concenter and
quently utilized. The properties of ANN are appropriate for
achieves its aim for the lowest error between the projected
system-recognizing applications where a set of examples
value and the proposed intended target. The other option is
may be used to represent the system behavior with the
choosing the maximum number of cycles (epochs). Eleven
condition that there is no programmable connection
input parameters were taken into account, including two-
between the system’s input and output [55]. The most
wheelers (2W), three-wheelers (3W), cars, MCVs, buses,
widely used training technique is the back-propagation
trucks, the mean speed of heavy, medium, and light vehi-
algorithm, which is basically a gradient steepest descent
cles along with percentage heavies, and the number of
approach [56]. The Levenberg–Marquardt (LM) algorithm
honking events per hour. Only LAeq and L10 were taken into
is used in this method that combines two algorithms, the
account as output parameters for the two networks created

123
408 B.S.Chauhan et al.

Fig. 10 ANN model structure

with 11 inputs and one output
developed for LAeq and L10
levels

Car

2W

3W

MCV

Bus

Truck Output layer

(LAeq and L10)
Speed LMV

Speed MCV

Speed HV

Honking per hour

Percentage heavies

Input layerr

Hidden layer

for each of these parameters. The 200 observations in the parameter by maintaining all inputs as same as earlier.
collection of noise measurements are split into the training Thus, it was observed that 16 neurons in hidden layer
part (70%), the testing part (15%), and the validation part provide the best-optimized results. At 16 neurons, the
(15%). A large number of simulations were performed to results come out as an MSE of 0.62 dB(A) in the train
find the optimal configuration for the network architec- phase and 1.36 dB(A) in the test phase for variable LAeq, on
ture’s characteristics. The single hidden layer’s neurons the other hand, an MSE of 0.68 dB(A) in train phase and
were varied in a range of 4 to 24, and the MSE and R be- 1.81 dB(A) in test phase is reported for variable L10. The
tween the observed and forecasted data were examined at correlation R is obtained to be 0.98 in train phase and 0.96
each iteration. The network variables used to train the in test phase for the LAeq variable. Among all other training
network are listed in Table 6. An 11:16:1 structure (neu- functions, the ‘‘trainlm’’ function is shown to be the most
rons in input layer:hidden layer:output layer, respectively) applicable. The finalized network structure is presented in
demonstrated the most accurate forecast for the dataset of Fig. 10 as 11:16:1. The plots between observed and fore-
traffic noise. Figures 8 and 9 show the R and MSE between casted data for the ANN model show superiority over the
the measured and predicted data for various numbers of MLR models. These observations corroborate the superi-
neurons for the LAeq parameter for the neural network ority of neural networks over regression methods repre-
approach. It is clear that both MSE and R, are optimum for sented in earlier studies [58, 59]. Figure 11a, b presents the
training and testing datasets with 16 neurons in the hidden plots between observed and predicted data for LAeq and L10
layer. These trials were same processed for L10 output parameters. The value of R2 between the ANN predicted

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 409

(a) of the tree, is defined by the MaxDepth variable distance

90.0 depth or length of the tree. The minimal number of data
y = 0.8811x + 9.0038
85.0 R² = 0.93 that a leaf node may hold is shown by the MinBucket.
Predicted LAeq dB(A)

Typically, one-third of the MinSplit value is the default

80.0
value. Here is a general overview of how the decision trees
75.0 algorithm works:
70.0
3.2.2.1. Feature Selection The algorithm begins by
65.0 selecting the most relevant feature from the available
dataset. Various measures like information gain, Gini
60.0
60.0 65.0 70.0 75.0 80.0 85.0 90.0 index, entropy, etc. are used to determine the feature that
Measured LAeq dB(A) best separates the data [60].
(b)
90.0 3.2.2.2. Node Creation Once a feature is chosen, a
y = 0.8711x + 9.9706
R² = 0.87 decision node is created, representing that feature’s test
85.0
condition. Subsets of the dataset are then created based on
Predicted L10 dB(A)

80.0 the potential values for that characteristic.

75.0
3.2.2.3. Recursive Splitting The algorithm repeats the
70.0 process recursively for each subset, considering one feature
65.0 at a time, and creating additional decision nodes. This
splitting process continues until the process reaches max-
60.0
60.0 65.0 70.0 75.0 80.0 85.0 90.0 95.0 imum depth or remains left with a single glass subset and
Measured L10 dB(A) stops.
Fig. 11 (a) Plot between predicted and measured equivalent contin-
uous sound level, LAeq for developed ANN model. (b) Plot between
3.2.2.4. Leaf Node Creation When the stopping criterion
predicted and measured 10 percentile exceeded sound level, L10 for is met, a leaf node is created for each subset. The leaf node
developed ANN model represents the predicted class or value for the given subset
[61].
and measured data was reported to be higher than analyt-
ical models as 0.93 for LAeq and 0.87 for L10 parameter. 3.2.2.5. Prediction Once the decision tree is built, it can
be used to make predictions on the latest unforeseen data.
3.2.2. Decision Trees (DT) Starting from the root node, each instance traverses down
the tree following the decision rules until it reaches a leaf
‘‘Decision trees’’ is a popular machine learning (ML) node, which provides the final prediction.
algorithm used for classification and regression tasks. In Decision trees have several advantages, such as inter-
this, each internal node depicts a characteristic or property, pretability, ease of understanding, and the ability to handle
each branch works as a decision rule based on that feature, both numerical and categorical features. However, they can
and each leaf node indicates the result or prediction. Its be prone to overfitting, especially when the tree becomes
structure is like a flowchart. The DT algorithm builds the too complex or when the training data contains noise.
trees by recursively dividing the data based on the selected
features, aiming to maximize information gain or minimize 3.2.3. Random Forest (RF)
impurity measures at each node. This technique is used in
machine learning applications when data analysis is ‘‘Random forest’’ is an ensemble learning method that uses
required. They have a form akin to a schematic diagram several decision trees to produce predictions. It is a DT
since they are based on stochastic data structures and are algorithm expansion that is prominent in machine learning
utilized for classification tasks. There is a root node at the and usually beneficial in classification and regression
top of the tree, and its branches stand for the tests and its applications. The main idea behind random forests is to
leaves represent the test results. Sub-nodes are created by enhance the reliability and accuracy of predictions by
dividing nodes in a decision tree. The MinSplit variable aggregating the predictions of multiple individual decision
displays the bare minimum number of node members that trees. Random forest uses a technique called ‘‘bootstrap
must exist before splitting. The maximum distance from aggregating’’ or ‘‘bagging.’’ It involves creating multiple
root node to leaf node, which indicates the depth or length

123
410 B.S.Chauhan et al.

(a) (a)
90.0 90.0
y = 0.9475x + 5.1152 y = 0.9474x + 5.1009
R² = 0.95 R² = 0.97
85.0 85.0

Predicted LAeq dB(A)

Predicted LAeq dB(A)

80.0 80.0

75.0 75.0

70.0 70.0

65.0 65.0

60.0 60.0
60.0 65.0 70.0 75.0 80.0 85.0 90.0 60.0 65.0 70.0 75.0 80.0 85.0 90.0

Measured LAeq dB(A) Measured LAeq dB(A)

(b) (b)
95.0
95.0 y = 0.8856x + 9.9709
y = 0.9399x + 5.6514 R² = 0.95
R² = 0.92 90.0
90.0

Predicted L10 dB(A)

Predicted L10 dB(A)

85.0
85.0
80.0
80.0
75.0
75.0
70.0
70.0
65.0
65.0 65.0 70.0 75.0 80.0 85.0 90.0
65.0 70.0 75.0 80.0 85.0 90.0 Measured L10 dB(A)
Measured L10 dB(A)

Fig. 13 (a) Plot between predicted and measured equivalent contin-

Fig. 12 (a) Plot between predicted and measured equivalent contin- uous sound level, LAeq for developed Random Forest model. (b) Plot
uous sound level, LAeq for developed decision tree model. (b) Plot between predicted and measured 10 percentile exceeded sound level,
between predicted and measured 10 percentile exceeded sound level, L10 for developed Random Forest model
L10 for developed Decision Trees model

randomly sampling with replacement. Each subset, known

subsets of the original training data by randomly sampling as a bootstrap sample, has the same size as the original
with replacement. Once all the decision trees are built, dataset but contains some repeated instances and may
predictions are made by aggregating the predictions of each exclude others.
tree. Random forest is created by a group of decision trees.
They made random decision forests to overcome the 3.2.3.2. Tree Construction For each bootstrap sample, a
overfitting problem in decision trees which is a drawback decision tree is constructed using the same process as the
in the decision tree method. The observations work as input standard decision tree algorithm. However, there is a slight
for each decision tree and major result is utilized as the modification. At each node of the decision tree, instead of
extreme output. This is a more accurate prediction method considering all available features, a random subset of fea-
due to its error cancelation property. As the tree is being tures is considered for splitting. This randomness helps to
formed, Breiman (2001) proposed the theory of random decorate the individual trees and promote diversity.
sampling of parameters at each node [62]. Additionally, he
looked at the above-mentioned bagging concept for taking 3.2.3.3. Ensemble Prediction Once all the decision trees
samples [63]. For each tree in this approach, a random are built, predictions are made by aggregating the predic-
sample selection will be made for the training dataset. tions of each tree. In classification tasks, the most common
Thus, the model becomes more resistant to noise and its prediction (mode) of the individual trees is chosen as the
deviations as a result. The variable ‘‘ntree’’ indicates the final prediction. In regression tasks, the forecasting of the
total number of trees constructed in RF model. Here is how individual trees is averaged to obtain the final outcomes.
the random forest algorithm works: By aggregating multiple decision trees and using ran-
domness in feature selection, random forests tend to be
3.2.3.1. Data Sampling Random forest uses a technique more robust against overfitting compared to individual
called ‘‘bootstrap aggregating’’ or ‘‘bagging.’’ It involves decision trees. Random forests generally have higher
creating multiple subsets of the original training data by

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 411

Table 7 t-test results for measured and predicted traffic noise levels for all the models
Statistical variables MLR (Qeqv) predicted MLR (Q) predicted ANN predicted Decision trees’ Random forest
values values values predicted values predicted values
LAeq L10 LAeq L10 LAeq L10 LAeq L10 LAeq L10

Mean absolute error 0.79 1.07 0.79 1.07 0.66 0.89 0.51 0.74 0.37 0.49
Standard deviation 1.05 1.39 1.05 1.38 0.89 1.17 0.70 0.93 0.57 0.70
Correlation coefficient 0.95 0.91 0.95 0.91 0.96 0.93 0.98 0.96 0.99 0.98
df 199 199 199 199 199 199 199 199 199 199
t-stat 0.44 1.89 0.71 1.65 1.74 1.95 0.46 0.60 0.11 0.51
P(T B t) one-tailed 0.33 0.03 0.24 0.05 0.04 0.03 0.32 0.27 0.46 0.30
t-critical one-tailed 1.65 1.65 1.65 1.65 1.65 1.65 1.65 1.65 1.65 1.65
P(T B t) two-tailed 0.66 0.06 0.48 0.10 0.08 0.05 0.64 0.54 0.91 0.61
t-critical two-tailed 1.97 1.97 1.97 1.97 1.97 1.97 1.97 1.97 1.97 1.97

accuracy compared to single decision trees, especially predicted data and observed data for the developed models
when dealing with complex and high-dimensional data. are presented in Table 5 for 200 datasets including training,
testing, and validation observations. It is clear that random
forest outperforms all the models, including analytical and
4. Results and Discussions machine learning models. The MLR model’s performance
is nearly identical for both equivalent and total traffic flow,
4.1. Comparison of All Models as indicated in Table 5. The MSE in the developed ANN
model for LAeq is 0.8 dB(A) and 1.4 dB(A) for L10. In
The suitability of models for forecasting traffic noise in the contrast to MLR models, which are better in acquiring the
Delhi-NCR region was investigated through a comparison linear pattern of data, ANN models are more versatile in
of all developed noise models. The error analysis of the their ability to acquire noise and acute values [64, 65].

Fig. 14 a Error histograms for (a)

predicted LAeq sound levels for
independent datasets (14 Sites, 30
N = 25) by various developed
25
models for the model validation. MLR (Qeq)
No. of Sites (in %)

b Error histograms for predicted

20 MLR (Q)
L10 sound levels for independent
ANN
datasets (14 Sites, N = 25) by 15 DT
various developed models for
RF
the model validation 10

0
-5.0 -4.5 -4.0 -3.5 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Error range for LAeq levels in dB(A)

(b)
30

25
No. of Sites (in %)

MLR (Qeq)
20
MLR (Q)
ANN
15
DT
10 RF

0
-5.0 -4.5 -4.0 -3.5 -3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Error range for L10 levels in dB(A)

123
412 B.S.Chauhan et al.

Therefore, ANN gave better results in compare of devel- design is problem-dependent since it depends on the large
oped MLR models. Figures 12a, b and 13a, b present the dimension, diversity, and elevated nonlinearity of the
plots between predicted and measured data for LAeq and L10 issues to be modeled [71]. An essential component of
levels for decision trees and random forest models analytical models is their straightforward structure, which
respectively. The coefficients of determination, R2 are makes it possible for non-specialists working in govern-
observed to be 0.95 for LAeq and 0.92 for L10 parameter for ment and other public institutions to use them with ease.
DT and 0.97 for LAeq and 0.95 for L10 parameter for RF However, the intermittent nature of neural networks and
respectively. The comparative analysis of all the models is the flexibility of frequent selection of input and output
shown in Table 5. It can be seen that the highest value of parameters in machine learning approach with their
the R2 was obtained between the random forest forecasted capacity to model nonlinear interactions make them con-
and measured data for both the parameters among all the venient for use in the scientific environment [72]. As a
developed models. Thus, it is evident that developed ran- result, the analytical and neural network models predicted
dom forest model gives the most acute results among all promising results; however, for more accurate results, we
the other models in these criteria and can be utilized for have utilized machine learning algorithms with greater
highly decisive traffic noise predictions and forecasting. reliability in predicting road traffic noise levels in Delhi-
All the developed models are also validated with the NCR.
outcomes of paired t-tests performed at a significance level
(a = 0.05) for the forecasted and observed sound data. As 4.3. Validation of the Models
indicated in Table 7, this test compares the test statistic (t-
stat) with the t critical value, and if the t-stat value comes With one hidden layer and 100 epochs, neural networks
in range of ± t critical value for a two-tailed test, it have 16 hidden layer neurons. The decision trees’ MinS-
demonstrates that there is no remarkable difference plit, MaxDepth, and Random State values were taken 20,
between two samples (accept null hypothesis). In decision 20, and 0, respectively. Regression random forest is
tree and random forest models, it is seen that the t-stat employed in the random forest. 500 trees have been con-
values are very low, considerably below the critical levels, sidered for the algorithm in random forest. The ‘‘Bagging’’
and lying in a non-rejection zone, indicating that the concept is used for sampling step. The criteria of Pearson
forecasted and measured data fit the models well correlation (R), coefficient of determination (R2), mean
[25, 66, 67]. square error (MSE), and mean absolute percentage error
(MAPE) were tested to compare the outcomes produced by
4.2. Implication of Developed Models various developed models. For checking the accuracy and
validation of the models, the independent datasets have
The current research investigates the significant potential been taken for the selected 14 sites for road traffic noise in
of machine learning techniques in developing a road traffic Delhi-NCR comprising all types of zones and traffic
noise prediction model in a large Delhi-NCR region. No characteristics. A total of 25 traffic datasets for one hour
such prior study has been conducted on the development of for the 14 sites have been measured apart from the 200
machine learning models for Delhi-NCR on a broad scale, noise-monitored datasets and applied in the developed
but machine learning approaches have been effectively models to check their accuracy. It has been observed that
implemented for other Indian cities like Patiala [68] and all the sites predicted the sound levels within the accuracy
Yavatmal [41]. The present developed models should not of ± 3 dB(A) as shown in Fig. 14a, b. Thus, it is evident
be considered site-specific models as they are generalized that developed noise models can be used for big
models based on field monitoring from multiple sites. A metropolitan cities in forecasting the traffic noise scenario.
large database collected from numerous sites with fluctu- It is concluded that random forest outperforms all the
ating traffic densities (low, medium, and high) will be other models. The value of MSE is 0.32 dB(A) for LAeq
helpful in establishing general empirical formulations for and 0.49 dB(A) for L10 levels for the random forest which
calculating the road traffic noise in Delhi-NCR because is the lowest and the value of R2 is 0.97 for LAeq and 0.95
these MLR models are adequate for places having all sorts for L10 levels which is the highest among all other models.
of traffic scenarios. The current analysis corroborates ear- The results of this study match with the study of vehicular
lier research that vehicular traffic plays a major role in traffic noise prediction using soft computing approach by
determining noise levels of road traffic [69]. With the aim D. Singh et al. 2016 [73]. The scatter plots of predictions
of minimizing noise pollution, the current model might versus measured equivalent levels of various developed
thus be very beneficial for urban executives and controllers models showed the general efficacy of the proposed models
in the monitor and managing road traffic noise in the Delhi- as all the model’s results are gathered almost closer to the
NCR region [43, 70]. The complexity of neural network bisector linear line in the graphs. The random forest plot

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 413

shows that the results are relatively stable, which supports forest model outperforms all the models. These obser-
the model’s viability for usage in such applications on vations corroborate the earlier findings presented by
noisy data with significant input from road traffic. The Singh et al. 2021 [68].
predicted noise levels nearly above 80 dB(A) seem to show • The prediction and forecasting accuracy of the models
a general underestimation of the model due to having is ascertained by an independent noise monitoring
higher honking events at some particular sites. dataset acquired from 14 sites. It is evident that these
models predict traffic noise levels with an accuracy
of ± 3 dB(A).
5. Conclusions • The developed models can be used for traffic planning
and environmental impact assessment studies for
This paper presents an extensive research study with the metropolitan and smart cities to reduce traffic noise
objective of noise monitoring, prediction, and forecasting pollution.
of hourly traffic noise levels at more than 200 sites in the
The present work can also be expanded to other cities of
Delhi-NCR. Such a rigorous study has not been reported so
the country for developing the validated traffic noise
far in the past two decades. The paper presents the devel-
models. The classification of honking and non-honking
opment and comparison of analytical models and machine
scenarios can also be added as an important objective of
learning approaches for noise modeling, prediction, and
this work. In order to expand the potential applications to
forecasting of road traffic noise levels in the Delhi-NCR
additional scenarios, future advancements of this study will
region. The analytical and machine learning models pre-
put more of an emphasis on calibrating and validating the
sented in this paper serve to be a reliable tool for noise
proposed models using data from various other sites in
predictions and forecasting in the given situation when
Delhi-NCR with varied traffic patterns. Furthermore, the
there is no validated road traffic noise prediction model
classification of honking and non-honking scenarios,
available for metropolitan cities like Delhi. The models
acceleration–deceleration regimes, and different traffic
considered the input parameters with different traffic flows,
characteristics will be a possible development of the
vehicular speed, percentage heavies, and the number of
methodology presented.
honking events. The models developed were examined for
their suitability using various statistical tests. Some Acknowledgements The authors are very thankful to the Director,
important conclusions from the study are drawn as: CSIR-National Physical Laboratory, New Delhi. The corresponding
author is thankful to AcSIR (Academy of Scientific and Innovative
• The study presents simple, easy-to-use analytical Research) and DST (Department of Science and Technology) for
models based on large noise-monitored data of 200 providing support and fellowship to carry out his Doctoral Disserta-
locations in Delhi-NCR covering all the zones includ- tion work at CSIR-NPL, New Delhi. The author expresses his grati-
tude to the Metrology Society of India (MSI) for providing financial
ing mixed zones that shall facilitate the prediction and support in conducting various International and National workshops
forecasting of hourly traffic noise levels. Additionally, on Noise Pollution Monitoring, Building Acoustics, Noise Mapping
the honking component is also considered in the model and Control, and Noise awareness over the last two years that pro-
as honking always accentuates the hourly ambient noise vided a large database for the present study. The author is also
thankful to Acoustics and Vibration Standard’s (CSIR-NPL) team
levels. These models can be utilized for long-term (16- members Mr. Mahender, Mr. Ayush, Mr. Abhishek, Mr. Ashish, Mr.
h or 24-h) predictions as well. Also, the analytical Kamesh, and Mr. Gautam (CSIR-CEERI) for their valuable support
models in conjunction with the GIS framework and for field measurements and technical help.
sound propagation algorithms can be used for devel-
oping noise maps of various sites.
• The study also presents an empirical model based on References
the concept of acoustic equivalence of vehicles
[1] A. Gilpin, Environmental impact assessment: cutting edge for
(PCNE). This model is observed to have similar the 21st century. Cambridge University Press (1995).
accuracy in comparison with the model developed [2] N. Garg, S.K. Mangal, P.K. Saini, P. Dhiman and S. Maji,
based on total traffic flow. Comparison of ANN and analytical models in traffic noise
• The analytical models presented in the study can be modeling and predictions. Acoust. Aust., 43 (2015) 179–189.
[3] H. Moshammer, J. Panholzer, L. Ulbing, E. Udvarhelyi, B.
widely used in comparison with the machine learning Ebenbauer and S. Peter, Acute effects of air pollution and noise
models which are data-dependent. from road traffic in a panel of young healthy adults. Int.
• Apart from the MLR approach, the study shows the J. Environ. Res. Public Health, 16(5) (2019) 788.
application of machine learning algorithms like deci- [4] J. Gieseke and G.J. Gerbrandy, Report on the inquiry into
emission measurements in the automotive sector; Committee of
sion trees (DT), random forest (RF), and artificial inquiry into emission measurements in the automotive sector.
neural networks (ANN). It is observed that random European Parliament, Brussel (2017).

123
414 B.S.Chauhan et al.

[5] A. Pascale, P. Fernandes, C. Guarnaccia and M. Coelho, A study scenario of vehicles and honking. Environ. Sci. Pollut. Res., 27
on vehicle noise emission modelling: correlation with air pol- (2020) 38311–38320.
lutant emissions, impact of kinematic variables and critical [27] D. Singh, R. Upadhyay, H.S. Pannu and D. Leray, Development
hotspots. Sci. Total Environ., 787 (2021) 147647. of an adaptive neuro-fuzzy inference system based vehicular
[6] W. Babisch, The noise/stress concept, risk assessment, and traffic noise prediction model. J. Ambient Intell. Humaniz.
research needs. Noise Health, 4(16) (2002) 1. Comput., 12 (2021) 2685–2701.
[7] E. Murphy and E.A. King, Environmental noise pollution: Noise [28] C. Steele, A critical review of some traffic noise prediction
mapping, public health, and policy. Elsevier (2022). models. Appl. Acoust., 62(3) (2001) 271–287.
[8] W. Babisch, Stress hormones in the research on cardiovascular [29] N. Garg and S. Maji, A critical review of principal traffic noise
effects of noise. Noise Health, 5(18) (2003) 1. models: strategies and implications. Environ. Impact Assess.
[9] World Health Organization, Environmental noise guidelines for Rev., 1(46) (2014) 68–81.
the European Region. World Health Organization Regional [30] J. Quartieri, N.E. Mastorakis, G. Iannone, C. Guarnaccia, S.
Office for Europe UN City, Copenhagen Ø (2018). D’Ambrosio, A. Troisi, T.L. L. Lenza, ‘‘A review of traffic noise
[10] D. Banerjee, Research on road traffic noise and human health in predictive models.’’ Recent Advances in Applied and Theoret-
India: Review of literature from 1991 to current. Noise Health, ical Mechanics, 5th WSEAS International Conference on
14(58) (2012) 113. Applied and Theoretical Mechanics (MECHANICS’09) Puerto
[11] B.S. Chauhan, S. Kumar, N. Garg and C. Gautam, Evaluation De La Cruz, Tenerife, Canary Islands, Spain December, 2009.
and analysis of environmental noise levels in NCT of Delhi, [31] C. Guarnaccia, Advanced tools for traffic noise modeling and
India. MAPAN-J. Metrol. Soc India, 25 (2023) 1–21. prediction. WSEAS Trans. Syst., 12(2) (2013) 121–130.
[12] D. Singh, S.P. Nigam, V.P. Agarwal and M. Kumar, Modelling [32] K. Kumar, M. Parida and V.K. Katiyar, Road traffic noise pre-
and analysis of urban traffic noise system using algebraic graph diction with neural networks—a review. Int. J. Optim. Control, 2
theoretic approach. Acoust. Aust., 44 (2016) 249–261. (2012) 29–37.
[13] P.R. Rao and M.S. Rao, Prediction of LA10T traffic noise levels [33] J. Quinlan, Induction of decision trees. Mach. Learn., 1 (1986)
in the city of Visakhapatnam, India. Appl. Acoust., 34(2) (1991) 81–106.
101–110. [34] E. Eliseeva, A.E. Hubbard, and I.B. Tager, An application of
[14] R.S. Nirjar, S.S. Jain, M. Parida, V.S. Katiyar, N. Mittal, A study machine learning methods to the derivation of exposure-re-
of transport-related noise pollution in Delhi. J. Inst. Eng. India sponse curves for respiratory outcomes (2013).
Environ. Eng. Div. 84(1) (2003). [35] G. Cammarata, S. Cavalieri and A. Fichera, A neural network
[15] H.N. Rajakumara and R.M. Mahalinge Gowda, Road traffic architecture for noise prediction. Neural Netw., 8(6) (1995)
noise prediction model under interrupted traffic flow condition. 963–973.
Environ. Model. Assess., 14 (2009) 251–257. [36] S. Givargis and H. Karimi, A basic neural traffic noise prediction
[16] S. Agarwal and B.L. Swami, Comprehensive approach for the model for Tehran’s roads. J. Environ. Manag., 91(12) (2010)
development of traffic noise prediction model for Jaipur city. 2529–2534.
Environ. Monit. Assess., 172 (2011) 113–120. [37] S. Rahmani, S.M. Mousavi and M.J. Kamali, Modeling of road
[17] A.C. Tan and D. Gilbert, Ensemble machine learning on gene traffic noise with the use of genetic algorithm. Appl. Soft
expression data for cancer classification (2003). Comput., 11(1) (2011) 1008–1013.
[18] Z. Erdem, R. Polikar, F. Gurgen, N. Yumusak, Ensemble of [38] A. Sharma, G.L. Bodhe and G. Schimak, Development of a
SVMs for incremental learning. In Multiple Classifier Systems: traffic noise prediction model for an urban environment. Noise
6th International Workshop, MCS 2005, Seaside, CA, USA, Health, 16(68) (2014) 63.
June 13–15, 2005. Proceedings 6 2005 (pp. 246–256). Springer [39] R. Kalaiselvi and A. Ramachandraiah, Honking noise correc-
Berlin Heidelberg. tions for traffic noise prediction models in heterogeneous traffic
[19] R. Polikar, Ensemble learning. Ensemble machine learning: conditions like India. Appl. Acoust., 111 (2016) 25–38.
methods and applications. 2012:1–34. [40] P.K. Paul, and P.K. Sarkar, Determination of dynamic PCUs of
[20] Z.H. Zhou, Ensemble methods: foundations and algorithms. different types of passenger vehicles on urban roads: A case
CRC Press (2012). study, Delhi urban area. Indian Highw. 41(4) (2013).
[21] J. Heinermann and O. Kramer, Machine learning ensembles for [41] G. Tiwari, J. Fazio and S. Gaurav, Traffic planning for non-
wind power prediction. Renew. Energy, 1(89) (2016) 671–679. homogeneous traffic. Sadhana, 32 (2007) 309–328.
[22] B. Krawczyk, L.L. Minku, J. Gama, J. Stefanowski and M. [42] Ö. Gündoğdu, M. Gökdağ and F. Yüksel, A traffic noise pre-
Woźniak, Ensemble learning for data stream analysis: a survey. diction method based on vehicle composition using genetic
Inf. Fusion, 1(37) (2017) 132–156. algorithms. Appl. Acoust., 66(7) (2005) 799–809.
[23] C.H. Fox, F.H. Huettmann, G.K. Harvey, K.H. Morgan, J. [43] N. Garg and O. Sharma, Measurement accuracy of secondary
Robinson, R. Williams and P.C. Paquet, Predictions from standards of sound pressure in comparison to primary standards.
machine learning ensembles: marine bird distribution and den- MAPAN-J. Metrol. Soc India, 27 (2012) 219–229.
sity on Canada’s Pacific coast. Mar. Ecol. Progress Ser., 27(566) [44] ISO 9613-1:1993. Acoustics—attenuation of sound during
(2017) 199–216. propagation outdoors—part 1: calculation of the absorption of
[24] B.T. Pham, D.T. Bui, I. Prakash and M.B. Dholakia, Hybrid sound by the atmosphere.
integration of Multilayer Perceptron Neural Networks and [45] ISO 9613-2:1996. Acoustics—attenuation of sound during
machine learning ensembles for landslide susceptibility assess- propagation outdoors—part 2: general method of calculation.
ment in the Himalayan area (India) using GIS. Catena, 149 [46] Central Pollution Control Board. Annual report, (2011–12).
(2017) 52–63. p. 94–96.
[25] P. Kumar, S.P. Nigam and N. Kumar, Vehicular traffic noise [47] Google Earth Pro 7.3.4.8248 (64-bit) desktop app.
modeling using artificial neural network approach. Transp. Res. https://www.google.com/intl/en_in/earth/versions/.
Part C Emerg. Technol., 1(40) (2014) 111–122. [48] ISO 1996-2:2017, Acoustics—description, measurement and
[26] C. Thakre, V. Laxmi, R. Vijay, D.J. Killedar and R. Kumar, assessment of environmental noise, Part 2: Determination of
Traffic noise prediction model of an Indian road: an increased environmental noise levels.

123
Comparison of Analytical and Machine Learning Models in Traffic Noise Modeling and Predictions 415

[49] N. Garg, Environmental Noise Control: The Indian perspectives [64] P. Goyal, A.T. Chan and N. Jaiswal, Statistical models for the
in an International Context, 2022, Springer Nature, Switzerland, prediction of respirable suspended particulate matter in urban
p: 1–592. https://doi.org/10.1007/978-3-030-87828-3. cities. Atmos. Environ., 40(11) (2006) 2068–2077.
[50] JCGM 100, Evaluation of measurement data—Guide to the [65] G.P. Zhang, Time series forecasting using a hybrid ARIMA and
expression of uncertainty in measurement, 2008. neural network model. Neurocomputing, 50 (2003) 159–175.
[51] N. Garg, S. Mangal, P. Dhiman, A multiple regression model for [66] P. Pamanikabud and P. Vivitjinda, Noise prediction for high-
urban traffic noise in Delhi. In: Proceedings of the International ways in Thailand. Transp. Res. Part D Transp. Environ., 7(6)
Conference on Advances in Engineering and Technology 2014 (2002) 441–449.
(pp. 344–349). [67] D.C. Montgomery and C.R. George, Applied statistics and
[52] B. Li, S. Tao, R.W. Dawson, J. Cao and K. Lam, A GIS-based probability for engineers. Wiley (2010).
road traffic noise prediction model. Appl. Acoust., 63(6) (2002) [68] D. Singh, A.B. Francavilla, S. Mancini and C. Guarnaccia,
679–691. Application of machine learning to include honking effect in
[53] D.E. Rumelhart, E.H. Geoffrey and J.W. Ronald, Learning vehicular traffic noise prediction. Appl. Sci., 11(13) (2021)
representations by back-propagating errors. Nature, 323(6088) 6030.
(1986) 533–536. [69] K. Kumar, V.K. Jain and D.N. Rao, A predictive model of noise
[54] A. Ghaffari, H. Abdollahi, M.R. Khoshayand, I.S. Bozchalooi, for Delhi. J. Acoust. Soc. Am., 103(3) (1998) 1677–1679.
A. Dadgar and M.R. Tehrani, Performance comparison of neural [70] S. S. Jain, M. Parida, N. Mittal, Urban transport environment
network training algorithms in modeling of bimodal drug interaction—defining a national level action plan. In: CODATU
delivery. Int. J. Pharm., 327(1–2) (2006) 126–138. XI: world congress: towards more attractive urban transporta-
[55] N. Genaro, A. Torija and A.R. Ridao, A neural network-based tion, 2004.
model for urban noise prediction. J. Acoust. Soc. Am., 128(4) [71] K. Kumar, M. Parida, V.K. Katiyar, Artificial neural network
(2010) 1738–1746. modeling for road traffic noise prediction. In: Third international
[56] G. Zhang, B.E. Patuwo and M.Y. Hu, Forecasting with artificial conference on computing communication and network tech-
neural networks: the state of the art. Int. J. Forecast., 14(1) nologies (ICCCNT), Coimbatore (2012).
(1998) 35–62. [72] A. Nucara, M. Pietrafesa, G. Scaccianoce, G. Staltari, A com-
[57] D.R. Hush and B.G. Horne, Progress in supervised neural net- parison between analytical models and artificial neural networks
works. IEEE Signal Process. Mag., 10(1) (1993) 8–39. in the evaluation of traffic noise levels, In: Proceedings 17th
[58] M. Cai, Y. Yin and M. Xie, Prediction of hourly air pollutant International Congress on Acoustics, ICA Rome, pp. 208–209
concentrations near urban arterials using artificial neural net- (2002).
work approach. Transp. Res. Part D Transp. Environ., 14(1) [73] D. Singh, S.P. Nigam, V.P. Agrawal and M. Kumar, Vehicular
(2009) 32–41. traffic noise prediction using soft computing approach. J. Envi-
[59] D. Srinivasan, A.C. Liew and C.S. Chang, A neural network ron. Manag., 183(1) (2016) 59–66. https://doi.org/10.1016/j.
short-term load forecaster. Electr. Power Syst. Res., 28 (2008) jenvman.2016.08.053.
227–323.
[60] N. Patel, and S. Upadhyay, Study of various decision tree Publisher’s Note Springer Nature remains neutral with regard to
pruning methods with their empirical comparison in WEKA. Int. jurisdictional claims in published maps and institutional affiliations.
J. Comput. Appl. 60(12) (2012).
[61] Y.Y. Song and L.U. Ying, Decision tree methods: applications
Springer Nature or its licensor (e.g. a society or other partner) holds
for classification and prediction. Shanghai Arch. Psychiatry,
exclusive rights to this article under a publishing agreement with the
27(2) (2015) 130.
author(s) or other rightsholder(s); author self-archiving of the
[62] L. Breiman, Random forests. Mach. Learn., 45 (2001) 5–32.
accepted manuscript version of this article is solely governed by the
[63] L. Breiman, Bagging predictors. Mach. Learn., 24 (1996)
terms of such publishing agreement and applicable law.
123–140.

123