An Optimized Deep Neural Network Approach for Vehicular Traffic Noise Trend Modeling

Vehicular traffic plays a significant role in terms of economic development; however, it is also a major source of noise pollution. Therefore, it is highly imperative to model traffic noise, especially for expressways due to their high traffic volume and speed, which produce very-high level of traffic noise. Previous traffic prediction models are mostly based on the regression approach and the artificial neural network (ANN), which often fail to describe the trends of noise. In this paper, a deep neural network-based optimization approach is implemented in two ways: i) using different algorithms for training and activation, and ii) integrating with feature selection methods such as correlation-based feature selection (CFS) and wrapper for feature-subset selection (WFS) methods. These methods are integrated to produce traffic noise maps for different time of the day on weekdays, including morning, afternoon, evening, and night. The novelty of this study is the integration of the feature selection method with the deep neural network for vehicular traffic noise modelling. New Klang Valley Expressway (NKVE) in Malaysia was used as a case study due to its increasing heavy and light vehicles, and the motorbike during peak hours, which result in high traffic noise. The results from the models indicate that the WFS-DNN model has the least mean-absolute-deviation (MAD) of 2.28, and the least root-mean-square-error (RMSE) of 3.97. Also, this model shows the best result compared to the other models such as DNN without integration with feature selection methods, CFS-DNN and the ANN networks (MLP and RBF). MAD improvement of 27% - 47% and RMSE improvement of 25% - 38% was achieved compared to other methods. The study provides a generic approach to key parameter selection and dimension reduction with novel trend descriptor which could be useful for future such modelling applications.


I. INTRODUCTION
Population growth and increase in economic activities is directly corelated to the increase in traffic around the world. Along with air pollution, noise pollution is considered as The associate editor coordinating the review of this manuscript and approving it for publication was Lefei Zhang . one of the major issues in the urban environment. Pinto and Mardones [1] inferred that noise in urban areas is highly associated with peoples' activities, particularly due to transportation and industrial activities. This assertion was further emphasized by the U.S. Department of Transportation (1995) [2], that traffic remains the major source of noise in both rural and urban settlements. It is the most annoying type VOLUME 9, 2021 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ of all noise source found in urban centres [3], [4]. Vehicular noise from expressways is considered as one of the key sources of noise in developed cities due to its high traffic volume, high speed, and different classes of vehicles [5]- [10]. Several studies on traffic noise in urban environment show that the noise has a negative impact on physical and mental health of people including annoyance, anxiety, cardiovascular risks etc. [11]- [16]. Therefore, it is worthy to note that traffic noise is a cause for concern in terms of public health and environment coupled with its nuisance effect [17]. Noise pollution was rated by the World Health Organiza¬tion (WHO) as the third most dangerous pollution after air and water pollution [15]. Therefore, the need to develop traffic noise maps and management plans becomes highly imperative in order to abate the aforementioned effects. So, in line with this perspective, the EU Directive 2002/49/EC (2002) Environmental Noise Directive was issued which recommended that the member states should prepare and publish traffic noise maps and management plans every five years [18]. This plan would enable vehicular traffic noise to be properly evaluated toward a better and sustainable environment. Determination of noise level is paramount to the successfully implement of any efficient and sustainable noise action plan that minimizes population exposure to noise. Therefore, it is highly imperative to acquire necessary information regarding level of noise to which the public is exposed to [19], [20], [21]. This would enable countries like Malaysia who a working toward improving environmental noise to meet their 2020 master plan [22].
Therefore, this traffic noise model is proposed to evaluate the noise pollution along the New Klang Valley Expressway (NKVE) of Malaysia. This expressway is exposed to a very high traffic noise due to increased heavy and light vehicles, motorbike etc., especially during the peak hours of the weekdays.

II. PREVIOUS WORKS
Various scientific models have been proposed in the past to predict emission of traffic noise using regression techniques [22]- [24]. Even though these models were widely used to identify traffic noise in cities with high accuracy, they are still with some limitations [25], [26]. Since 1950s, several traffic noise predictions (TNP) models and methods have been proposed. [7], [27], [28] carried out a comprehensive and critical review of the most frequently used noise level prediction models. The aforementioned studies reported that the TNP models found in literature commonly applied the linear regression approach, which unfortunately does not consider the intrinsic random characteristics of traffic flow and the way the vehicles run. Only traffic volume is taken into consideration in most of these studies.
Several methods in the past used artificial neural networks (ANN) and genetic algorithms [29], [30]. Cammarata et al. [31] first used an ANN model to predict traffic noise level by considering three input parameters: i) vehicle volume estimation was standardized by converting the number of motorcycles, cars, trucks into equivalent number of vehicles, ii) average height of buildings along both sides of the road, and iii) width of the road [32]. The result between the predicted and measured values shows good agreement, with the NN approach yielding better outcome compared with the classical methods. However, the major disadvantage of learning vector quantization (LVQ) neural network approach, as pointed out by the authors, is that its result depends on the training data. This means that network trained for a town layout with a particular road width and building height cannot be used for another city with a different layout [32]. This approach has been further adopted in recent studies [33], [34].
Hamoda [34] predicted the construction noise in the city of Kuwait by applying the neural networks with both the backpropagation and regression analysis. These models were developed based on the phases and equipment types used during construction. The results showed that the general regression network based neural models achieved better accuracy than the backpropagation based networks outcomes. Givargis and Karimi [29] proposed neural, statistical, and mathematical models that predicted the maximum A-weighted noise level (LAmax) for an express train in Tehran-Karaj. A satisfactory result was achieved without any significant differences in the predicted results of the models and the authors suggested that more works are needed to handle sophisticated models.
Genaro et al. [35] proposed a model called Multi-layer Perceptron (MLP) to estimate LAeq (Equivalent Continuous Sound Pressure Level) using street level data sourced from Granada, Spain. The neural network results were compared with the results from other mathematical models. Similar numbers of input parameters (25) used in the neural network model were applied to all the other individual mathematical models investigated. It was observed that better predictions were achieved by the MLP model using the neural network compared to the other mathematical models. Also, when the input parameters were reduced to only 11 by applying principal component analysis, a decline in accuracy was observed. However, the predictions using the neural network still outperformed the other mathematical models.
Mansourkhaki et al. [36] used ANN-MLP and ANN-RBF to estimate LAeq in parts of Tehran, Iran. The variables used in this study include average speed, traffic volume, and percentage of heavy vehicles. The predicted outcomes were then compared with measures such as the mean-squarederror (MSE) and the coefficient of determination (R2). The ANN-MLP network achieved better performance compared to the ANN-RBF model. Torija and Ruiz [37] applied several machine learning approaches with the addition of feature selection process including the sequential minimal optimisation (SMO), multilayer perceptron (MLP), and the Gaussian processes for regression (GPR) to estimate the sound level. It was observed that the feature-subset selection technique, when used with the SMO or GPR algorithms. According to Azeez et al. [38], correlation-based feature selection (CFS) with ANN-MLP showed better performance compared with other methods such as support vector regression (SVR) and liner regression (LR) Models in the prediction of CO emissions from traffic vehicles.
In this study, we develop a novel method for prediction of traffic noise using deep neural network optimisation, where we i) test different algorithms for training and activation, and ii) integrate with feature selection methods such as correlation-based feature selection (CFS) and wrapper for feature-subset selection (WFS) methods. The proposed models are compared with other methods, such as the artificial neural network of the multilayer perceptron (ANN-MLP) and the radial basis function (ANN-RBF) to determine the error (dB) of each model. The performance assessment of the developed models is done based on the mean-absolutedeviation (MAD) and root-mean-square-error (RMSE). The novelty of this study lies in the integration of the feature selection method with the deep neural network, and the output variables of the propped network model that is made in three layers as maximum, minimum and average traffic noise for different time of the day including morning, afternoon, evening, and night. The models were trained and tested with data acquired from the New Klang Valley Expressway (NKVE), Malaysia. Key variables employed in the models are traffic volume, vehicle variety (such as light, heavy vehicles and bus, truck), digital surface elevation (DSM), gradient, density of expressway, temperature, and humidity.

A. STUDY AREA
The study area is located in Kuala Lumpur, Malaysia as shown in Fig. 1. The NKVE is a heavy traffic route that passes through high-density areas including Kuala Lumpur, Subang, Petaling Jaya, Damansara, Klang and Sungai Buloh. Geographically, it is bounded on longitude 101 • 27 30 E to 101 • 36 E and latitude 03 • 03 30 N to 03 • 07 N. The NKVE highway is a stretch of about 25 kilometres running from Bukit Raja near Klang town to Jalan Duta, Subang Jaya. The average temperature of the area under consideration is between 80 • F to 83 • F and a wind speed of 5 to 8 mph. The humidity varies between morning and the afternoon, recording an average of 92% -96% and 66% -72%, respectively. Recent population indices show that the area harbours over 400,000 people, which is expected to increase by about 0.32% yearly. The area was selected for this research due to its connection to two major settlements -Subang Jaya and Klang regions. It also serves different land uses such as hospital, schools, public service offices, religious infrastructure such as mosques and temples, housing and residential condominium, commercial buildings, and industrial developments.

B. DATASET
In traffic noise calculation and modeling, three basic data types are required [10]. Firstly, traffic flow information such as traffic volume and the proportion of vehicle types that ply the road. These data can be acquired directly in the field using either by manual or automated recording. The second data type is the actual noise coming from the vehicles. These data is usually collected in the field with the aid of noise level meters. The noise data should be collected using suitable devices fitted with advanced filtering technology [23], [24]. This ensures that only vehicular noise is captured, while other types of noise are discarded. There are many types of noise level meters available in the market, such as Sound Level Meter and Datalogger (Class 2) -CENTER323, Digital Sound Level Meter -CEL-240 and others. Finally, the information regarding the road characteristics is important for noise modeling.
In this study, noise levels were measured with Sound Level Meter TES-52 in terms of minimum, maximum, and average values continuously at 15-min intervals using type A filter (dB (A), 0.1dB resolution). The installation of the noise meters was carried out at 10 cm from signs or poles which are separated from noise barriers by at least 2 m allowance. Garmin Global Positioning System (GPS) GPS 60 was used to acquire the geographic coordinates of each sampling location. The noise level measurement was taken four times every day during weekdays. This comprise of morning hours (6.30am-8.30am), afternoon (11.30am-1.30pm), evening (6.30pm -8.30pm), and night (11pm-12midnight) each day. The traffic volume data were segregated into five classes including light vehicle, heavy vehicle, motorbike, truck, and bus. Meanwhile, the predicted traffic noise maps for the study area are based on GIS modelling. In this study, remote sensing data using light detection and ranging (LiDAR) point clouds, and Worldview-3 images were used (TABLE 1). The LiDAR data were captured by using an airborne system on March 8, 2015. The camera had a spatial resolution of 10 cm, and the laser scanner had a scanning angle of 60 • with a camera angle of ±30 • . The posting density of the LiDAR data was 3-4 pts/m 2 (average point spacing = 0.41 m). The minimum and maximum elevations  were 36 and 69 m, respectively. The Worldview-3 image with eight bands of panchromatic spatial resolution, multispectral, short-wave infrared, and Clouds, Aerosols, Vapors, Ice, and Snow (CAVIS) resolution at 0.31, 1.24. Fig. 2 shows the methodology adopted in this proposed traffic noise model which is based on deep neural network (DNN). The dataset was prepared and managed in GIS database, as well as the predicted traffic noise maps were achieved using GIS. On other hand, optimisation using DNN model was performed using grid search and integration with feature selection methods including correlation-based feature selection (CFS), and wrapper for feature-subset selection (WFS). In order to comprehend the behavior of the proposed DNN model, sensitivity analysis and contribution of various factors were studied. The results of the proposed model were finally validated using data collected from the field. Furthermore, the proposed model was compared with other machine learning models, such as ANN-MLP and ANN-RBF. The comparison of these models is based on performance measures of MAD and RMSE. In addition, we randomly divided the noise data samples into training (70% ≈ 116) and testing (30% ≈ 64) datasets.

A. NOISE PARAMETERS
The main purpose of our model is to estimate the traffic noise level at a particular location and during a specific period of time. In this study, the dependent parameter, that is the highway noise descriptor, is the equivalent continuous noise level per 15 minutes (L_(eq,15)) for the morning, afternoon, evening and night. The noise parameters were pre-determined and selected based on literature review which consider traffic and weather characteristics of the area under consideration. The noise parameters inputted into the model are light vehicle, heavy vehicle, motorbike, truck and lorry, bus, digital surface elevation (DSM), time (i.e. morning, afternoon, evening and night), gradient, density of road, temperature, and humidity. Whilst, the maximum, minimum and average traffic noise are the outputs of the model. Summary statistics of these parameters are shown in TABLE 2.

B. DEEP NEURAL NETWORKS
Deep learning exhibits special feature that enable it to reduce local optima problems in non-convex objective function [40]. Three approaches have contributed to the success recorded with deep learning methods which are better learning algorithm, large number of hidden units and better parameter initialization technique [41]. Furthermore, deep architecture seems to be appropriate for higher-level abstractions [42]. Some features of deep learning are helpful across domains which makes it well-suited for transfer learning.
DNN is a machine learning approach that relies on biologically inspired statistical learning models. It is a perception based on a multilayer approach that consists of an interconnection of simple nodes or neurons. It is a nonlinear model represented by inputs and outputs values. The neurons are series of structured nodes systematically connected together form the layers which are randomly connected to the successive layers [43]. DNN is theoretically structured into three layers. The layers are input, hidden, and output layers, which form a complete process sufficient to yield results [43]. The nodes are allocated some numeric weights throughout the input and output processes and are transformed through a simple activation function [44].
The major attraction to the DNN model is the ability of learning. Paul Werbos (1974) developed back propagation and has soon become the commonest learning algorithm employed in ANN. This approach was later rediscovered by researchers (Priddy and Keller, 2005). The DNN algorithm is designed based on error minimization principle through iteration and gradient design as shown in (1). This concept has been successfully used in remote sensing applications. However, their applications are not without challenges, such as high computational complexity coupled with overfitting [45].
where, d j and o M j refers to output and current responses at the node ''j'' of the output layer, respectively, while ''L'' means the number of nodes found in the output layer. This approach is deployed in an iterative manner in which corrections are made to the parameter weights through computation and addition to the previous outputs as shown in Eqn. 2.
where, w i,j is a weight parameter for node i and j, is a positive constant that regulates adjustment to be made refer to as learning rate, α is a momentum factor with values between 0 and 1, while ''t'' represents the iteration number. Also, α parameter can be referred to as stabilizing or smoothing factor due to its ability to smoothen the changes between the weights [46].

C. OPTIMIZATION PROCEDURE 1) OPTIMIZATION ALGORITHMS THROUGH SEARCH SPACE
The performance of DNN is based on its structure and the hyperparameters used in developing the model. In this research, several hyperparameters are combined and tested to obtain the sub-optimal network model to calculate the vehicular traffic noise. TABLE 3 shows the structure and hyperparameters used to evaluate the model and their domain within the search space.
The former employed the dot product between the inputs and weight parameters with monotonic activation functions,  [47]. When training networks, optimization score, or an objective function is minimized based on the training dataset, the optimizer usually has gradient momentum parameters and learning rate is given. Furthermore, various activation functions including relu (rectified linear unit), logistic, identity, elu (Exponential Linear Unit), and sigmoid can be applied. The hyperparameters in our models were selected through systematic grid search and executed within the Scikit-Learn environment for 500 epochs. Even though, this approach involves cost of high computation, more realistic results are obtained through systematically tuning of the hyperparameters. Many models were constructed and tested with various combinations of parameters.

2) OPTIMIZATION METHODS FOR FEATURE SELECTION
In this section, two methods of integration with DNN model are explained with correlation-based feature are-subset selection (CFS) and wrapper for feature-subset selection (WFS). Also, the best method of integration with DNN model was selected based on the lower value of MAD and RMSE.

3) CORRELATION-BASED FEATURE-SUBSET SELECTION (CFS)
One of the most famous methods used for feature selection based on the correlation function is the CFS model. The algorithm is designed based on subgroups selection. It must VOLUME 9, 2021 contain features that strongly correlate to a specific class. This means that all features with low correlation with the class would are neglected. Besides, repeating features are identified due to their exceptional relationship with any one of the other features. The feature will be obliged if its level of prediction within the classes in the territory of the instance space is not as expected by different features. Equation (3) presents the CFS feature subset assessment function.
where, Ms represent the heuristic ''merit'' containing k features and feature subset s, r cf represents the mean of the feature-class correlation (f 2 s), and r ff means the average of the feature-feature inter-correlation.

4) WRAPPER FOR FEATURE-SUBSET SELECTION (WFS)
An induction algorithm, along with a set of training data are presented in the supervised machine learning approach. The induction algorithm acts as a black box while selecting the feature subset in the wrapper approach. The input variables were chosen based on the DNN model. Because its parameters are selected based on each of the regression algorithms. Thereafter, searches were carried out in feature-selection algorithm to find an optimal subset using the induction algorithm itself which is an aspect of the evaluation function. The feature-subset-selection algorithm is considered as a wrapper around the induction algorithm [48]. The WFS approach assesses the attribute sets with the aid of the learning scheme. The procedure requires in the WFS approach are as follows: the induction algorithm is executed on the dataset and divided into internal training and holdout set. However, a different set of features is excluded from the data. The feature subset that has the highest assessment is chosen as the final dataset to be used to run the induction algorithm [48]. Finally, a crossvalidation method was used to determine the accuracy of the learning scheme for a set of attributes.

D. THE GIS MODEL
The GIS model was spatially designed as a representation of predicted traffic noise level discharged to atmosphere from highway traffic of the study area. The model was proposed based on the implementation of the final proposed model. The statistical model parameters were converted to a geodatabase by mapping the sample attributes with their corresponding locations obtained via GPS. The model's parameters were transformed to raster format through inverse distance weighting interpolation (IDW) for noise predictor information [48]. IDW technique was selected due to its ability to provide a higher degree of correlation compared with the Kriging and Spline method [49]. On the other hand, a higher distortion was observed in the interpolated results obtained from Kriging and Spline results compared with the IDW values. The model parameters were combined in GIS based on the overlying analysis with the proposed model. This was spatially overlaid on (5 * 5) m 2 high-resolution grid, to predict the road traffic noise in the unsampled areas. An overall grid value was calculated using the intersected parameter values which represent the variation and distribution in traffic noise levels in the study area.

E. MODEL EVALUATION
The effectiveness and potential of the developed models were ascertained by calculating Mean-Absolute-Deviation (MAD) and Root-Mean-Square-Error (RMSE), in the knowledge that this would give estimates of L eq15 minutes . To evaluate the predictive performance of the models, two performance measures were used. These performance measures indicate the accuracy of predictions of the model by comparing the actual value of the parameters (a i ), predicted value (b i ), number of sample data points (n) and others such as an average of all observed values (a) and average of all predicted values (b) which could be useful when comparing different models.
Firstly, the MAD was calculated using (4). Determining MAD enables researchers to note the relationship between two continuous variables. Next, (5) was used to calculate the RMSE for evaluation of the average performance of the model across different test samples. The best validation result was obtained with a network of 11 input parameters and two stage 23-7 hidden layers. Also, it shows that the network is best trained with the Levenberg-Marquardt algorithm, while identity algorithm indicated the best output for hidden and output activation layers. Furthermore, the best gradient momentum and learning rates obtained are 0.9 and 0.5, respectively. All the hyperparameters associated with the DNN model were used for the noise prediction while fine-tuning within their search space. The DNN training model achieved 3.4 and 5.2 for MAD and RMSE, respectively. While, during testing the DNN model achieved 3.61 and 5.57 for MAD and RMSE of the traffic noise prediction respectively for the study area. The output of the DNN model is defined by maximum, minimum and average equivalent continuous noise level (dB) L eq,15 . Fig. 4 (a) shows the impact of the number of the hidden units on MAD and RMSE. Where, we observed that RMSE with MAD is increasing gradually with an increase in hidden number units. The DNN model training stage results shown in Fig. 4 (b), where four separate algorithms were tested. The best results were obtained by the Levenberg-Marquardt during the training stage, with 3.4 and 5.2 for MAD and RMSE, respectively. Likewise, the performance of hidden and output activation methods for the DNN model is shown in Fig. 4 (c) where the best results were obtained by the Identity algorithm with MAD of 3.4 and RMSE of 5.2.
Regarding the learning rate, the best value was found to be 0.5. The MAD and RMSE values were significantly decreased at the learning rate between 0.1 -0.5. On the other hand, it was observed that an increase in momentum of the optimization algorithm improved the MAD and RMSE of noise prediction. The DNN model has been enhanced from momentum value 0.6 to 0.9. Momentum is vital if local minima stuck is to be avoided. In general, large values of momentum enable fast convergence, while small values cannot always avoid local minima, which slows down training of a system. Fig. 5 shows the best gradient momentum with learning rate of 0.9 and 0.5, respectively with MAD of 3.4 and RMSE of 5.2.

2) INTEGRATION THE FEATURE SELECTION (CFS AND WFS) WITH DNN MODEL
Based on the results shown in TABLE 4, the noise predictors have different impact levels when used in conjunction with each feature selection method (CFS and WFS) for prediction of the maximum, minimum and average traffic noise level in our selected study area. When the CFS method was used, it was found that the noise predictors such as motorbike, bus and humidity are significant at 100% confidence level, which is imperative to use in the DNN model prediction.
In addition, there are other noise predictors that can yield good prediction, such as heavy vehicle, DSM and temperature VOLUME 9, 2021 parameters. While the noise predictors such as light vehicle, truck and lorry, time, gradient and density of road are not important and used in the model. The CFS method was excluded from the DNN model due to its low correlation with traffic noise predictors, which makes it unsuitable for our DNN prediction model. Based on the WFS method, we found that the noise predictors, such as time and humidity, are significant at 100% confidence level, with substantial use in the DNN model. Also, there are other important parameters used in the DNN model prediction, such as light vehicle, heavy vehicle, motorbike, truck and lorry, bus, DSM and gradient. While, the noise predictors were not significant for DNN prediction model, such as density of road and temperature.
The statistical results indicate that the CFS method was able to establish that the parameters such as light vehicle, truck and lorry, time and gradient are not significant for DNN prediction model. On the other side, the WFS method found those parameters significant, especially the time, truck and lorry parameters for DNN model.
Finally, feature selection methods (CFS and WFS) were integrated with the DNN model and trained. It was found that the training and testing of the WFS-DNN model has the least MAD and RMSE values. Fig. 6 shows the proposed deep neural network architecture and TABLE 5 describes the hyperparameters of each model which consist of input, number of hidden layer and the output of the model.

B. COMPARISON WFS-DNN MODEL WITH OTHER MODELS
The proposed model was compared with two ANN variations -ANN-MLP, and ANN-RBF models. The WFS-DNN supersedes the performance of the other models through as shown in TABLE 6. This can be seen in   In addition, we found that time and humidity of noise predictors are significant at 100% confidence level, as well as the density of road and temperature. Whereas in the other models, the most important parameters were found to be light vehicle, heavy vehicle, motorbike, truck and lorry, bus, DSM and gradient. Also, results revealed that the density of road and temperature of noise predictors were not significant using the other prediction model.

C. VEHICULAR TRAFFIC NOISE PREDICTION MAPS
Noise distribution maps for the study area were generated using the proposed noise model in GIS. The model produces continuous noise level as the output, accounting for the field conditions and other factors such as topography, weather and other noise predictors. In this research, the noise and traffic volume were measured at different periods, morning, afternoon, evening and night of weekdays. Fig. 8, Fig. 9 and Fig. 10 show the maps of each period. However, this section presents only the recommended maps for planning purposes.   It was discovered that the road is characterized by high traffic noise level in the morning, afternoon and night hours. The following figures show the generated noise distribution maps (maximum, minimum and average traffic noise level) of the study area for a weekday in the morning, afternoon, evening and night.

VI. CONCLUSION
Vehicular emissions such as traffic noise are considered as one key source of environmental pollution affecting urban areas. Plethora of predictive and spatial models have been developed to estimate the impacts of vehicular noise on the environment and public health. In this study, we developed a new DNN based model integrating the feature section methods (CFS and WFS) with GIS mapping. The proposed model accurately predicts the vehicular noise with lowest MAD and RMSE of 2.23, 2.28 and 2.85, 3.97 for training and testing, respectively. The default model parameters were 11 parameters. After the implementation of the CFS and WFS models, the input parameters were reduced to 6 and 9 parameters each for the CFS-DNN and WSF-DNN models receptively. The WFS-DNN model was observed to be the best model and outperformed the other models such as DNN without integration with feature section methods, CFS-DNN and the ANN based networks (MLP and RBF). Moreover, the model found that the noise predictors such as the time and humidity are significant at 100% confidence level. According to the noise prediction maps, it was observed that the high traffic noise level in the morning, afternoon and night hours. The proposed noise distribution maps were displayed with the maximum, the minimum and the average traffic noise level of the study area for a weekday in the morning, afternoon, evening and night. Although the developed models produced an excellent result, however, it is also important to test the transferability model in different geographical sites to check its efficacy and robustness. AHMED ABDULKAREEM AHMED received the master's degree in remote sensing and GIS from Universiti Putra Malaysia (UPM). He is currently pursuing the Ph.D. degree in advanced modeling and geospatial information systems (CAMGIS) with the Faculty of Engineering and IT, University of Technology Sydney (UTS). His scientific background is civil engineering. His research interests include remote sensing and GIS, traffic noise modeling and environmental modeling, and soft-computing applications. VOLUME 9, 2021 BISWAJEET PRADHAN (Senior Member, IEEE) received the Habilitation degree in remote sensing from Dresden University of Technology, Germany, in 2011. From 2015 to 2021, he has worked as the Ambassador Scientist of the Alexander Humboldt Foundation, Germany. He is currently the Director of the Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS), Faculty of Engineering and IT. He is also a Distinguished Professor with the University of Technology Sydney. He is also an internationally established scientist in the fields of geospatial information systems (GIS), remote sensing and image processing, complex modeling/geo-computing, machine learning and softcomputing applications, natural hazards, and environmental modeling. Out of his more than 650 articles, more than 500 have been published in science citation index (SCI/SCIE) technical journals. He has authored eight books and 13 book chapters. He has widely travelled abroad, visiting more than 52 countries to present his research findings. He was a recipient of Alexander von Humboldt Fellowship from Germany. He was also a recipient of Alexander von Humboldt Research Fellowship from Germany. He received 55 awards in recognition of his excellence in teaching, service, and research, since 2006. From 2016 to 2020, he was listed as the World's Most Highly Cited Researcher by Clarivate Analytics Report as one of the world's most influential mind. From 2018 to 2020, he was awarded as the World Class Professor by the Ministry of Research, Technology and Higher Education, Indonesia. He is also an associate editor and an editorial member of more than eight ISI journals.
SUBRATA CHAKRABORTY (Senior Member, IEEE) received the Ph.D. degree in decision support systems from Monash University, Australia. He is a Senior Lecturer with the Faculty of Engineering and IT, School of Information, Systems and Modelling, University of Technology Sydney (UTS), Australia. He is also a core member of the Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS) at UTS. Previously, he worked as an Academic with the University of Southern Queensland, Charles Sturt University, Queensland University of Technology, and Monash University. His current research interests include optimization models, data analytics, machine learning, and image processing with decision support applications in diverse domains, including business, agriculture, transport, health, and education. He is a Certified Professional Senior Member of ACS.
ABDULLAH ALAMRI received the B.S. degree in geology from King Saud University, in 1981, the M.Sc. degree in applied geophysics from the University of South Florida, Tampa, in 1985, and the Ph.D. degree in earthquake seismology from the University of Minnesota, USA, in 1990. He is currently a Professor of earthquake seismology and the Director of the Seismic Studies Center at King Saud University (KSU). His recent projects also involve applications of EM and MT in deep groundwater exploration of empty quarter and geothermal prospecting of volcanic harrats in the Arabian shield. He has published more than