Modelling and Prediction of Road Transportation Noise Pollution in Some Capital Cities in Eastern Nigeria by Use of Artificial Neural Network

The study attempts to model and predict road transportation noise pollution in five capital cities in Eastern Nigeria. The capital cities are Calabar, Uyo, Umuahia, Owerri and Port Harcourt. Feed-forward neural network (FNN) with negative back-propagation algorithm was used to do this. The software used was NeuroXL. The ability of this software to handle multiple non-linear relationships makes it ideall y suited for this work. The input data used were total road traffic volume, road traffic mix, road traffic noise pollution response data, and distances from road centre-line to measurement points. The output data used was A-weighted energy mean sound level (L A eq). Models based on this negative back- propagation neural network were trained, validated and tested using data collected. The performance of the model was tested by an error measure, root mean square error (RMSE). RMSE is low as expected, ranges from 1.007 - 1.814, showing that the model is good for the prediction of road traffic noise data. The correlation between observed and predicted noise levels (LAeq) was also obtained, and ranges between +0.757 to +0.974, showing that there is no significant difference between observed and predicted noise levels, thereby, proving the model accurate and reliable.


Introduction
Nigerian urban dwellers are excessively exposed to severe environmental/city noise pollution. The most disturbing city noise source, as generally established in the developing and developed urban communities being road transportation, as noise from it causes a lot of socio-psychological and physiological problems such as annoyance, sleeplessness, hearing loss, communication disturbances, speech intelligibility, cardiovascular disorders and other health problems [1 -9]. The heterogeneous nature of urban environments, coupled with the characteristics of road transportation noise, their spatial, temporal and spectral variability, makes the matter of modeling and prediction of road transportation pollution a very complex and non-linear problem, to which the application of artificial neural networks becomes imperative. Artificial neural networks (ANNs) are widely used in road transportation noise modeling and prediction as a preference to more conventional statistical techniques, because ANNs are non-linear, relatively insensitive to noise data, perform reasonably well when limited data are available, and provide flexibility, accuracy and fault tolerance in changing environments [9][10][11][12][13][14][15][16][17].

Materials and methods 2.1 Measurement sites
One hundred (100) measurement sites were randomly selected from the five (5) Nigerian capital cities surveyed. Fifty (50) sites were chosen from road transportation high noise pollution zones, where heavy road transportation volume and dense traffic mix (composition) are experienced, on daily basis to serve as study group, while 50 sites were from low noise zones to serve as control group.

Materials for data collection 2.2.1 Materials for acoustic data collection
A precision sound level meter, Bruel and Kjaer (B & K), type 732 was used to assess road transportation noise levels at each measurement sites. Other materials used included measuring tape (to measure distance from the road centre line to the measurement points); stop watch/clock (to take sampling/measurement times); tally sheets (to record motor vehicle volume and motor vehicle mix during measurement/sampling times); and tripod stand (to support the sound level meter).

Materials for psycho-social data collection
Subjective (psycho-social) responses of respondents exposed to intense road transportation noise were obtained by use of questionnaire items. The questionnaire was designed after Fields [18] with some variations to suit the objectives of this study. The questionnaire contains a number of noise response questions to help elicit the needed social noise data from respondents on road transportation noise-induced health problems such as sleeplessness, annoyance, hearing loss, auditory communication disturbance, and others. Information on effects of road transportation noise pollution on various health challenges has six (6) rating options: Extremely severe disturbance (ESD) with response rating of 6; Very severe disturbance (VSD) with rating of 5; Severe disturbance (SD) with rating of 4; Moderate disturbance (MD) with rating of 3; Little disturbance (LD) with rating of 2; and No disturbance (ND) with response rating of 1. The questionnaire also contains information on some demographic/socio-economic variables such as: sex (male and female), age (15 years and above), marital status (single, married, divorced), educational level (primary, secondary, tertiary schools), occupation (student, civil/public servants, business/trader, artisan, jobless), occupational status (junior, senior, executive), income level (low, medium, high), among others.

Methods of data collection 2.3.1 Methods for acoustical data collection
A precision sound level meter was used to collect the road transportation noise levels in line with ISO 1996 -1 and ISO 1996 -2 standards [19,20]. All measurements were done when motor vehicles (motorcycles/tricycles, cars/jeep, buses and trucks/trailers, etc) were moving past the measurement points. Readings of noise levels, background noise levels (BNLs) and A-weighted energy mean noise levels (LAeq) at each measurement point were taken every fifteen (15) minutes (sampling time or time rate) for a period of about 15 hours (7am -10pm) daytime period, and 9 hours (10pm -7am) nighttime period. Sound level meter (SLM) was held on a tripod stand with a microphone directly pointing toward noise source about 1.5 -2.0m high from the ground, and 3.5m from reflecting surfaces. The distance between measurement point and road Centre line was 10 -15m. Measurement sites were randomly selected to reflect roads with high and low transportation noise pollution levels, also away from airports, factories, construction sites and any other sources of heavy and intense noise other than motor vehicles. This was to prevent or reduce undue influence of these sources to road transportation noise levels. Total road traffic volume and road traffic mix (composition) were also recorded at each measurement sites. Tables 6-10 show observed and predicted LAeq data and mean road transportation volume per hour during recording time at daytime and nighttime periods in the surveyed capital cities.

Methods for Psycho-social data collection
Subjective (Psycho-social) responses of respondents exposed to intense road transportation noise pollution were obtained by use of road transportation noise pollution survey questionnaire (RTNPSQ) and analysed and evaluated. Persons who have literacy skills (reading and writing skills in English), who reside at the place for atleast three (3) years as at the time the survey took place, and who were upto 15 years and above by age, were given copies of the questionnaire to complete objectively and return to the researcher. These precautions were taken to help reduce information bias on the part of the respondents. Two thousand and five hundred (2,500) persons were given copies of the questionnaire at road transportation high noise pollution sites, to serve as experimental group, while another 2,500 persons were given some copies of questionnaire at low noise pollution sites, to serve as control group. In all, the response rates at high and low noise pollution sites were 93.5% and 94.8% respectively.

Artificial neural network training process
Every neural network has input, hidden and output layers (nodes). Feed-forward neural network (FNN) and many other networks learn using back-propagation algorithm. The input data used in this study include total road traffic volume, road traffic composition (mix), distance from measurement point to road centre-line; and respondents' road traffic noise pollution-induced response data. The input data were divided into two sets -training (learning) data set and checking (testing) data set. Data points for road traffic high noise pollution sites were 486, 472, 454, 464 and 461 in Calabar, Uyo, Umuahia, Owerri and Port Harcourt cities respectively, while for road transportation low noise sites were 465, 480, 478, 476 and 471 in Calabar, Uyo, Umuahia, Owerri and Port Harcourt cities respectively. Table 11 shows summary of ANN training and checking data used for the study. Data points used for training ANN at high noise sites in Calabar, Uyo, Umuahia, Owerri and Port Harcourt cities were 301, 295, 297, 295 and 284 respectively, while at low noise sites were 296, 304, 295, 299 and 288 respectively. Also data points used for checking the validation of ANN at high noise sites in Calabar, Uyo, Umuahia, Owerri and Port Harcourt cities were 185, 177, 157, 169 and 177 respectively, while at low noise sites were 169, 176, 183, 177 and 183 respectively. With back-propagation, the input data were fed into the input layer to the hidden layer. Within the hidden layer they got summed, then processed by a non-linear function (usually either zero-based log sigmoid function or the hyperbolic tangent). The data were then finally multiplied by interconnection weights, then processed within the output layer to produce the neural network output. The output of the neural network was compared to the desired output, and the model error was computed. This error was then fed back (back-propagated) to the neural network and used to adjust the weights such that the model error decreased with each iteration, and the neural model got closer and closer in accuracy until the desired output was obtained, when the network no longer seemed to be learning, or an acceptable model error was reached. Fig. 2 shows a diagram demonstrating ANN training process [21], while table 12 shows summary of initial ANN training parameters. Table 13 shows the validation parameters of the ANN model.

Data analysis/reductions
The following noise measure or descriptor was used: Energy mean A-weighted sound pressure level (LAeq): This is mathematically expressed in Eqn. 1.

Results
The findings of this study are summarized in tables 6 -10, 13 and Figs. 3 -7. Tables 6 -10 show observed (measured) and predicted (calculated) noise levels (LAeq) and mean road traffic volume and traffic mix at daytime and nighttime periods in the surveyed Nigerian cities. Table 13 shows the calculated validation parameters of the ANN model. Figs. 3a -7a show correlation curves and R 2 -values between observed and predicted LAeq, while Figs. 3b -7b show ANN performance curves for checking (testing) data for road traffic high and low noise pollution sites, indicating respondents' noise reactions against observed and predicted LAeq at surveyed Nigerian cities.

Discussion of Results
From Tables 6 -10 the observed and predicted LAeq appear to be correlating well. They are found to be high, beyond the recommended World Health Organization's standard [24]. The LAeq ranged from 87.1 -98.5 dB(A) (observed)in Calabar city high noise sites. Similar trends were observed in other surveyed cities. Such levels of noise are high enough to cause human annoyance, discomfort, sleeplessness, hearing loss, communication disturbances, among other physiological and psycho-social health disorders [9]. The mean road traffic volume per hour (VPH) is much as observed in tables 6 -10 at high noise sites. It was shown that noise level is a function of traffic volume. Percentage of heavy duty vehicles ranged from 9.1 -20.3%. This magnitude of motor vehicles is alarming [9,24].  +0.950 showing that there is no significant difference between observed and predicted LAeq, further proving that the ANN model is accurate [9].
In order to certify the good results obtained with the developed ANN based prediction model, correlation values between observed and predicted LAeq are shown in Figs. 3a -7a while Figs. 3b -7b display the ANN model performance curves of observed and predicted values of the output variables (LAeq) for all data used for the checking (testing) phase based on noise impact responses from respondents. From the results obtained, the proposed ANN based model has achieved prediction with a reasonably low RMSE, and has shown a great capacity for generalization. The neural network is capable of predicting, with considerable precision and accuracy, the sound pressure level (LAeq) and even temporal and spectral composition of the different types of situations presented to the network [26].

Conclusion
Due to their well-known characteristics, the use of artificial neural networks to approach a complex problem of modelling and prediction of urban noise seemed highly recommended [9,17]. Based on the results discussed in this paper this hypothesis is certified. The developed ANN based prediction model is capable of predicting, with great accuracy, road traffic noise levels as well as their temporal and spectral compositions in cities. In this study, the model developed is not only able to learn and predict those data presented during the training phase, but also is able, with great success, to predict noise data used for the testing phase, which inform about its great capacity of generalization. This goes to show that the model will not only be very useful for cities surveyed under this study, but also for other cities which have similar noise situations and characteristics [9,26].

Acknowledgements
The authors are very grateful to all those who helped in data collection/collation.