Machine Learning-based GIS Model for 2D and 3D Vehicular Noise Modelling in a Data-scarce Environment

Despite the fact that road traffic is inevitable, its annoying sound called noise has continued to draw global attention and possible remedial measures [1,2,3]. Noise pollution has attracted great interest over the decade [4]. Road traffic is the most widespread source of noise in many countries, and it is the most prevalent cause of annoyance and interference [5,6,7]. Traffic noise is generated from engine sound and frictional contact between the ground and the vehicle tires [8]. Noise generated from traffic depends on traffic volume, vehicular speed, vehicular class (heavy duty or light cars), and the type of roads [9,10,11,12,13]. Noise prediction is necessary where a future situation cannot be ascertained [14,15,16]. More so, field measurement of noise pollution seems impractical and difficult since numerous points are required to be measured to ascertain the true position of the noise situation [17,18,19,20].

So, a considerable amount of work has been carried out with the aim of developing models to predict noise levels generated from road traffic. Urban noise maps have been developed in many countries. Most noise maps were produced in a two-dimensional (2D) format by measuring noise levels on a grid of 10 m horizontally and 4 m above the ground. Although the 2D maps are adequate in low-density urban areas where most buildings have only a few stories [21,22,23,24,25,26,27,28]. The accuracy reduces with an increase in the density of the population especially in urban areas where elevation of most buildings is too high. Three-dimensional (3D) noise maps are required for precise characterization of noise levels in major urban cities surrounded by high-rise buildings [29,30,31,32,33,34,35].

However, very little effort has been witnessed in relation to the 3D noise maps predicted in the past with different noise prediction models and tools [31,32,33,34]. Therefore, if 3D noise maps are properly generated and predicted, it would enable planners and architects to locate and orientate their buildings appropriately at the design stage. This would ensure improved noise reduction and a serene environment. It would also improve the quality of life and reduce health risks associated with noise that could arise from long-term exposure to excessive noise levels [36, 37]. The 2D GIS (Geospatial Information Systems) has been widely and successfully used in environmental impact studies to assess the impact of spatial phenomena such as soil pollution, air pollution, and noise, on the environment [38,39,40,41,42,43,44]. However, it can be expected that the proposed 3D approach can offer fundamental improvements when 3D effects are relevant in urban areas.

There are studies in an effort to produce 3D noise maps, for instance, 3D noise map in Paris was created with the French standard NMPB-Route-96 [45] and the geometric data acquired from digital photographs [33, 45]. 3D noise mapping in Hong Kong was reported recently [46]. It was carried out with customized LimA packages using international standard ISO-9613-2 [47] for territory-wide mapping. The UK standard CRTN [48] was applied for project-based 3D mapping. Also, 3D noise mapping using the Dutch standard SMM1 was presented [30, 49]. These models cannot be imported into commercial software for noise mapping as noise propagation algorithms deployed in the commercial software are basically 2.5-dimensional (2.5D).

However, several studies carried out were mainly focused on the production of either 2D noise maps or 3D maps based on a large number of noise samples. This approach proved to be expensive and time-consuming as it required a number of noise samples to produce 3D noise maps with high accuracy. Therefore, the objective of this research is to produce noise impact and improve visualization by using basic 2D and 3D GIS functionalities and to increase the accuracy of noise impact assessment studies by building two models: 2D noise model for roads and 3D noise model for buildings by using fewer noise samples. The samples (2D and 3D) are then combined to produce 3D noise map for the capital of the study area.

This study was carried out in the capital city of Kirkuk, Iraq. This study area was chosen for its features such as industrial and residential buildings characterized by low/high-density population which makes it suitable for noise-related investigations. According to the Iraq Ministry of Planning, Kirkuk is the capital of Iraqi culture that is characterized by high traffic flow. It embodied different types of road networks such as Baghdad Erbil (BE) expressway, primary, and secondary roads all located in the city. These diversities make the site suitable for traffic noise studies. Figure 1 shows the study area.

Figure 1:

The study area.

The dependent variable of the model is noise levels (time-averaged sound pressure level). It is the equivalent steady level over a given period of time that contains the same amount of noise energy as the actual fluctuating level. Measurements of noise were taken at random sites (selected randomly with the use of the ArcGIS Sampling Design Tools) across the city. The measurements were performed using a CESVA SC102 sound level meter (class 2 ± 0.7 dB). The noise level was measured (dB (A)—Decibel A-weighted) at each site every 20-min and 1.5 m above ground. Measurements were taken in the rush hours of morning (7:00–9:00) and afternoon (12:00–2:00) during summer between March 12 and 15, 2018.

Measurements were taken at various locations such as on and near highways and main roads, back streets (side streets with low noise levels), and different floors of buildings every 20 min during rush hour. Additional data were collected from traffic flows. A total of 26 measurements were performed (Figure 2), where the 26 noise samples for the 3D noise model were split into 18 for training and 8 for testing, and the 9 noise samples for the 2D noise model were split into 6 for training and 3 for testing of the model.

Figure 2:

Training and testing noise samples in the study area.

The purpose of this study is to estimate the traffic noise level for a given location. The noise parameters of the model were determined and selected based on an evaluation taking into account the traffic and weather characteristics of the area under consideration. The noise parameters for the 2D noise model are semitrailer, truck, bus, light vehicle, motorbike, digital surface model (DSM), average speed, and maximum speed, whereas for the 3D noise model are the distance from high light vehicle, medium light vehicle, low light vehicle, high truck, medium truck, low truck, high motorbike, medium motorbike, low motorbike, high semitrailer, medium semitrailer, low semitrailer, high bus, medium bus, low bus, medium average speed, low average speed, high maximum speed, medium maximum speed, low maximum speed, high average speed, and DSM.

The parameters of 2D and 3D noise models were summarized based on statistics as shown in Table 1. Figure 3A shows the type of road network parameter; Figure 3B shows low, medium, and high light vehicle parameters; Figure 3C shows low, medium, and high truck parameters; Figure 3D shows low, medium, and high motorbike parameters; Figure 3E shows low, medium, and high semitrailer parameters; Figure 3F shows low, medium, and high bus parameters; Figure 3G shows low, medium, and high of average speed parameter; Figure 3H shows low, medium, and high of maximum speed parameters; Figure 3I shows the DSM for the study area.

Figure 3:

Different spatial data used in this study: (A) Road networks; (B) Low, medium, and high light vehicles; (C) Low, medium, and high trucks; (D) Low, medium, and high motorbike; (E) Low, medium, and high semitrailer; (F) Low, medium, and high bus; (G) Low, medium, and high average speed; (H) Low, medium, and high maximum speed; (I) DSM. DSM, digital surface model.

Figure 4 presents the methodology adopted in the proposed model, which is a combination of 2D and 3D models of traffic noise based on machine learning (ML). The traffic noise prediction maps were produced depending on the GIS model. In this study, remote sensing data using DSM raster and Quick-bird image were used. The data such as traffic volume and traffic noise were collected using a field survey in the city capital, Kirkuk. Data sets were prepared and maintained in a GIS database and predicted traffic noise maps were generated using GIS. The multilevel prediction models were developed using three ML methods such as artificial neural network (ANN), random forests (RFs), and support vector machine (SVM) and the input parameters for the 2D noise model are light vehicle, truck, motorbike, semitrailer, bus, DSM, average speed, and maximum speed.

Figure 4:

Methodology used in this work.

The input parameters for the 3D noise model are distance from high light vehicle, distance from medium light vehicle, distance from low light vehicle, distance from high truck, distance from medium truck, distance from low truck, distance from high motorbike, distance from medium motorbike, distance from low motorbike, distance from high semitrailer, distance from medium semitrailer, distance from low semitrailer, distance from high bus, distance from medium bus, distance from low bus, distance from high average speed, distance from medium average speed, distance from low average speed, distance from high maximum speed, distance from medium maximum speed, distance from low maximum speed, and DSM.

The location of noise samples and the output or dependent parameter is the equivalent continuous sound pressure level every 20 min (L_eq,20 dB(A)) during the rush hours of weekday morning and afternoon. Furthermore, the results of the proposed model and other models were validated based on correlation (R), correlation coefficient (R²), and root mean square error (RMSE) algorithms. Finally, the prediction of 2D and 3D noise maps for the study area based on the best model achieves less RMSE.

The performance of the three models was determined by calculating three performance measures: correlation (R), correlation coefficient (R²), and RMSE, which can give an estimated Leq value. Four performance measures were used to assess the predictive power of the model. These performance measures measure the accuracy of a model’s predictions by comparing the actual parameter values a_i, the predicted values b_i, and the number of sample data points n with, for example, the mean of all observations indicate and the mean of all predicted values (b_i). This can be useful when comparing different models.

First, Eq. (7) shows the correlation between the two data sets to calculate the linear relationship. The value of R is between −1 and +1. Then, using Eq. (8), the coefficient of determination was calculated and the resulting value for R² is between −1 and +1. Finally, the RMSE is calculated using Eq. (9) to assess the average performance of the model across different test samples [66]. (7) R=∑i=1nai−a¯bi−b¯∑i=1nai−a¯2∑i=1nbi−b¯2 R = {{\sum\nolimits_{i = 1}^n {\left( {{a_i} - \bar a} \right)\left( {{b_i} - \bar b} \right)} } \over {\sqrt {{{\sum\nolimits_{i = 1}^n {\left( {{a_i} - \bar a} \right)} }^2}{{\sum\nolimits_{i = 1}^n {\left( {{b_i} - \bar b} \right)} }^2}} }} (8) R2=R2 {R^2} = {\left( R \right)^2} (9) RMSE=∑i=1nbi−ai2n RMSE = \sqrt {{{\sum\nolimits_{i = 1}^n {{{\left( {\left| {bi - ai} \right|} \right)}^2}} } \over n}}

Vehicular traffic noise is one of the major sources of environmental pollution in urban areas, where road networks and traffic flow are present. Traffic noise prediction models and spatial models are used to assess the impacts of vehicular emissions on human health and the environment. In this study, noise impact and improved visualization of the noise impact are proposed using 2D and 3D GIS functionalities. This approach is aimed at improving the accuracy of noise impact assessment studies using fewer noise samples. The two models such as 2D noise model for roads and 3D noise model for buildings are developed and combined to produce the final 3D noise map for the capital of the study area, which is considered the novelty of this research. The proposed 2D model noise prediction achieved 0.003 for RMSE, 1.00 for correlation (R), and correlation coefficient (R²) of 1. For the testing model, 7.14, 0.87, and 0.75 were recorded for RMSE, R, and R² of the traffic noise prediction, respectively. The proposed 3D noise prediction model shows 0.058 for RMSE and 1.00 for correlation (R), correlation coefficient (R²). For the testing model, 4.46, 0.82, and 0.68 were recorded for RMSE, R, and R² of the traffic noise prediction, respectively. The result shows that the best architecture in the 2D ANN noise prediction model was achieved by a network of 8 input parameters, 18 hidden layers, BFGS algorithm of trained network, identity algorithm of hidden and output activation, and 0.3 of gradient momentum and learning rate. While the best architecture in the 3D ANN noise prediction model was achieved by a network of 22 input parameters, 11 hidden layers, RBFT algorithm of trained of network, logistic algorithm of hidden and output activation, and 0.3 of gradient momentum and learning rate. Moreover, it was observed that the best model is ANN, and RF is better than SVM. This was evident by the result that RF is less than RMSE of SVM, whereas RMSE of RF model was (1.82, 6.00) and (9.83, 4.50) for training and testing 2D and 3D model, respectively, while RMSE of SVM model was (3.60, 6.16) and (10.34, 4.75) for training and testing 2D and 3D model, respectively. The GIS modeling was applied to improve visualization of 2D and 3D noise maps and these maps were the average traffic noise level of the study area for weekdays in the morning and afternoon.