Research Area

The area of research is Mitiga International Airport in Libya (13^o16’40”E, N32^o53’40”N, elevation 11 m) (Fig. 2). This airport is located in the municipality of Tripoli, the capital of Libya. The planes operating on these routes are Airbus 320, Airbus 319, Airbus 321, Airbus 332, Boing 738, Boing 734, CRJ9 and ERJ145. The runway has an asphalt surface and has a length of 3,400 m. Near the airport are Mitiga Military Hospital (13^o16’08”E, 32^o54’25”N, elevation 10 m) and Halim Al Saadia Elementary School (13^o16’12”E, 32^o53’29”N, elevation 11 m).

In the period when the noise level was measured, there were an average about 43 landings and take-offs per day. Noise level measurement was carried out in the period of June, July, August 2021, and cases of approximately the same engine power were analyzed. Phase 1 was from June 1 to July 15, 2021, and Phase 2 from July 16 to August 31, 2021. The total number of flights in the observed period is: June 1,363; July 1,364 and August 1,242.

Therefore, measurements were made continuously for 3 months, 24 hours a day (except for the school location). The measurement was carried out in the period from June 1 to August 31, 2021, because it was observed that in that period, although the traffic frequency was not the highest in 2021, it was the highest in previous years. It was not possible to perform longer-term measurements, because this research was conducted only as a part of the student’s doctoral dissertation. As a result, the obtained L_den values are more conservative and higher than the actual ones. The noise level was measured with three sensors of the same type, which were placed at locations M1, M2 and M3 (Fig. 2).

Given that airborne noise has an impact on the health of people living near the airport, as it affects sleep disorders, cardiovascular diseases, including hypertension, altered cognitive abilities in children, preterm birth, hormonal disorders [23, 24, 25], and higher risk of metabolic syndrome [26], the research idea was to determine the noise level in the area that would include these three locations and to implement measures to reduce noise levels in accordance with the measurement results, the effect of which would be determined by a new set of noise measurements, at the same locations.

Measuring of the Noise Level

During three-month period (from June until September in 2021.) noise levels were measured at three locations near the runway threshold 10 (13°16.22’E, 32°53.75’N) (Fig. 2): – M1 (measuring spot 1) - at Mitiga Airport, where the

airport staff is about 1,500 m away from the runway

threshold 10 (13^o16’40”E, N32^o53’40”N), – M2 (measuring spot 2) - in the military hospital

Mitiga, which is at a distance of 1,225 m from the

runway threshold 10 (13^o16’08”E, 32^o54’25”N) and – M3 (measuring spot 3) - Halima Al Saadia Primary

School, which is 550 m from the runway threshold

10 (13^o16’12”E, 32^o53’29”N).

Measurement of noise levels at all three locations was done for the peak load of the airport (the largest number of take-offs and landings of aircraft that generate high noise levels). Also, the measurements were performed in two phases: (1) before the introduction of noise reduction measures and (2) after the introduction of noise reduction measures.

These measures were implemented as a test for this research, which lasted from July 16 to August 31, 2021. The airport authorities have built barriers made of brick and aluminum. All airlines that used the airport in the specified period, as well as ATC (air traffic control), participated in the introduction of the measures.

The impulse precision sound level meter B&K 2209, manufactured by Brüel & Kjær, was used to measure the noise level, and the measurement was performed during three periods, during the day (from 7 h to 19 h), in the evening (from 19 h to 23 h) and at night (from 23 h to 7 h) at the airport and hospital, while at the school location, the measurement was performed only during two periods, during the day and in the evening (the measurement was not done at night, because at the time the school was closed).

Fig. 2.

International airport Mitiga with measuring spots.

One of the goals of this paper is to determine the spatial distribution of 24-hour noise emissions (L_den) generated by air traffic at Mitiga Airport (Tripoli, Libya) using GIS methodology, in order to identify noise pollution levels in relation to the location of Mitiga Military Hospital and Halim Al Saadia Elementary School located near this airport. Due to the fact that noise was measured at three measuring points, during the day, evening and night, 24-hour noise emissions (L_den) had to be calculated using “the following formula [27, 28]:

where: – L_d is the A-weighted long-term average sound level

determined over all the day periods of the year, – L_e is the A-weighted long-term average sound level

determined over all the evening periods of a year,

and – L_n is the A-weighted long-term average sound level

determined over all the night periods of a year” [27,

page 252].

Mihajlov and Prascevic [27] used noise indicators L_d, L_e and L_n to present daily, evening and night noise indicators (L_day, L_evening and L_night).

The Guideline Development Group (GDG) of the World Health Organization (WHO) has suggested that the noise levels produced by aircraft should be reduced to levels below 45 dB L_den and 40 dB L_night during night time as it can have negative effect on health and sleep [29]. The GDG has also expressed its confidence in the fact that exposure to noise levels below 45 dB L_den can lead to an increased risk of annoyance [29]. We noticed that these threshold values are about 10 dB lower than the threshold values in the EU member states, so we increased the WHO limit values for L_den and L_night by 10 dB and conducted this research with a limit value of L_den = 55 dB and L_night = 50 dB.

Noise at the school location was measured only during the day and evening, because the school does not work at night, so when calculating the L_den value for a given location for L_night, a limit value of 50 dB was taken into account. The pupils are not affected by noise during the night, since the school is closed, but the people living near the school are. Since the aim of our research was to assess the noise impact on pupils who are attending school near Mitiga airport, not the people living nearby, we did not measure noise during the night at the measuring spot M3, but we assumed that it was within the limit value (50 dB), and the sensitivity analysis have shown that the results for L_den would not differ much in case that the values for L_night are equal to 0 dB. Besides, in reality outdoor noise value at night is never 0 dB. This is why we used the limit value for L_night = 50 dB. Although, if the noise was measured during the night period in the measuring spot M3, the values would certainly be higher than 50 dB, as well as the calculated L_den values.

It was assumed that the noise at the M3 measuring point at night would be greater than 50 dB because it is close to the runway, and the noise level measured at the M2 measuring point is the farthest from the runway before and after the introduction of noise reduction measures, was greater than 50 dB. Also, it is a well-known fact that the noise is decreasing by increasing the distance. The sensitivity analysis was done separately for the measurement results before and after the introduction of noise decrease measures. The value for the M3 measuring point at night was increased by 10 and 20% to get the value for L_night as closest to the mean noise value measured at the measuring point M2 at night which according to the results in Table 1 was 58.04 dB.The results have shown that in this case values for L_night would be 55 dB (increased by 10%), and 60 dB (increased by 20%), and the calculated mean values for L_den would be 70.78 dB and 71.80 dB, respectively. The same analysis was done for the measurements after the application of noise decrease measures. However, since the mean noise level at measuring point M2 at night (Table 2) in this case was 51.34 dB, it was necessary to increase noise at the M3 measuring point at night by 2.5 and 5%, thus the new values for L_night would be 51.25 dB (increased for 2.5%) and 52.5 dB (increased for 5%). New calculated mean L_den values for the M3 measuring point were 58.64 dB and 59.45 dB, respectively. As can be seen, the sensitivity analysis has shown that by increasing the L_night values, the L_den values will also be higher. However, the influence of increased L_night values on mean L_den values at the M3 measuring point will not be so big, as can be seen by comparing with the results for calculated mean L_den values at the M3 measuring point (Tables 1, 2).

Geographic Information System

Fig. 2 shows a map with three locations where noise was measured (measuring points M1, M2 and M3). This map was created in GIS, using the QGIS software package (version 3.6). Since GIS has a very powerful spatial analysis capabilities [30, 31], it was used to create layers that contain all the necessary information for segmenting grids based on the boundaries of the research area. The research area consisted of 4x4 grids covering an area of 5.968 square kilometers. It has also been considered dividing the area into different numbers of grids but the results were not significantly different. In order to generate noise levels per grid, the measured noise values (L_day, L_evening and L_night) were used to calculate the 24-hour noise (L_den) measured at three locations and these results were then used in zonal statistics in GIS. The spatial distribution of noise emission was calculated using the IDW interpolation method, and the previously calculated L_den values for all three measuring points. IDW interpolation method [32] was used in this study to show the spatial distribution [33] of noise emission. The IDW method generates maps that offer valuable information for monitoring purposes, even when only a small number of monitoring units are available [34]. This is one of the simplest methods based on the assumption that the value at an uncompressed point can be approximated as a weighted average of values within points within certain distances or from a given number of the nearest points (usually 10 to 30) [35]. The weights are usually inversely proportional to the distance [27], which leads to estimation in an unspecified location. Therefore, the interpolation is calculated as follows [36]:

where m represents the number of the nearest points, r location, and p represents the parameter [36].

After applying the IDW interpolation method and creating spatial distribution maps [37] of noise emission and obtained average L_den values per grid, using zonal statistics as table tool in GIS, deviations from the limit value in each grid were calculated, assuming that the limit value for L_den is 55 dB.

Artificial Neural Networks

ANN technology was used to assess the reliability of noise level prediction [17]. This technique was used within the SPSS software package (version 20) and on that occasion the MLP structure of the neural network was used as one of the most commonly used ANNs. MLPs include input layer, hidden layer and output layer [38], and in this study they were used to predict variations in noise levels. Each MLP had one output (L_day, L_evening, L_night, or L_den) and three inputs (engine power, take-off and landing runway distance, and air traffic volume) that were used to identify differences between noise measurement locations. The training, validation and testing phases play a crucial role in the field of machine learning [39]. During the training phase, the corpus was organized in a manner that aligns with the algorithms’ computational requirements. This allows the model to learn from their trial and error experiences. To ensure comprehensive training, the dataset was divided into three segments, so when defining MLPs, it was determined that 70% of the sample would be random selection for training, i.e., calibration model, 20% for testing and 10% for model validation. Ratio 70:20:10 was also used by other authors like Zhang et al. [40], Kowsher et al. [39] and Karpathy et al. [41]. As Kowsher et al. [39] stated the training segment, comprising 70% of the dataset, was utilized to train the models. The testing phase involved using the remaining 20% of the dataset as a dedicated testing set to assess the models’ performance [39]. Lastly, 10% was set aside for validation, aiding the models in evaluating their own performance and facilitating iterative improvements during the training process [39]. A training dataset for an ANN can have any number of hidden layers, which can also have any number of neurons [42]. Until the sum of the square error of the training template reached the minimum, the training is performed. The MLP analysis in this research was done based on 198 samples (for training, testing and holdout) for all outputs except for L_night (132 samples) because the measurement was not done during the night period. To assess how well the ANN models are performing in terms of precision, the criterion used is the root mean square error (RMSE). The RMSE is calculated using the following equation [16]:

where O_i is the observed value and P_i is the computational value. The n parameter represent the number of data.

Multiple Linear Regression

MLR method was also used in predicting noise pollution levels at Mitiga International airport in Tripoli, in order to compare the results achieved by ANN method with another traditional model like MLR.

The equation for multiple linear regression for k variables used for calculating noise pollution levels in the research was [43]:

where the dependent variable i.e. noise pollution levels (L_day, L_evening, L_night and L_den) is y; coefficients of the regression evaluation are β₁, β₂ and β_k while β₀ is the y-intercept (value of y when all other parameters are set to 0); x_i values are independent variables (runway distance (RD), air traffic volume (ATV) and engine power (EP)); e_i is model error.

Application of GIS

After all measurements were performed both before and after the introduction of corrective measures to reduce noise levels, at all three locations, the data were entered into the GIS database for further analysis. First, based on the entered data on measured values and the application of the IDW interpolation method (Equation 2), maps of noise emission distribution were created before and after the introduction of corrective measures (Fig. 3).

It is noted that the noise level was significantly higher before the introduction of corrective measures (Fig. 3a), with the lowest noise emission values (from 59.02 dB to 70.30 dB) recorded in the military hospital (M2), while in the school area (M3) recorded values were between 63.28 dB and 75.29 dB. The highest values of noise emission were measured in the area of the airport (M1) and ranged from 76.07 dB to 87.90 dB. After the introduction of corrective measures (Fig. 3b), the noise emission values in the hospital area (M2) ranged from 55.72 dB to 59.65 dB, in the school area (M3) between 56.96 dB and 58.94 dB. The highest noise level was still in the area of the airport (M1) and ranged between 59.99 dB and 63.76 dB.

The distribution maps shown in Fig. 3 indicate that the area around Mitiga Airport and the population living in that location are still exposed to noise above 55 dB. According to Souza & Zannin [44] exposure to noise above 55 dB can reflect the health of the population living nearby [45]. Also, Baloye and Palamuleni [46] stated that noise sensitivity zones with noise level between 50 dB and 60 dB are considered risky, between 60 dB and 65 dB moderately risky, between 65 dB and 70 dB highly risky, between 70 dB and 75 dB dangerous, between 75 dB and 80 dB highly dangerous, while noise levels above 80 dB are extremely dangerous.

After creating maps of spatial distribution of noise emissions, the investigated area was divided into 16 grids and zonal statistics as table tool in GIS was applied for calculating deviations from the limit value for each grid. Then these deviations were entered into the GIS database and maps were created showing the deviations from the limit values per grid. Deviations from the limit values were calculated for L_den before and after the introduction of corrective measures (Fig. 4).

As can be seen in Fig. 4a), before the introduction of corrective measures less than 20% deviations was on the following grids: 672 (19.94%), 771 (18.43%) and 772 (16.91%), between 20% and 25% deviations was on the grids: 471 (22.05%), 472 (21.96%), 473 (22.55%), 474 (24.32%), 571 (21.65%), 572 (21.88%), 573 (23.22%), 671 (20.02%) ) and 673 (24.54%), while over 25% of the deviations were on the grids: 574 (26.28%), 674 (30.21%) and 774 (31.96%). Regarding deviations after the introduction of corrective measures (Fig. 4b), less than 6% of deviations were on the following grids: 471 (5.48%), 472 (5.19%), 473 (5.58%), 571 (5.58%), 572 (5.37%), 671 (5.68%), 672 (5.73%), 771 (5.53%) and 772 (5.22%), between 6% and 9% deviations were on the grids: 474 (6.65%), 573 (6.10%), 574 7.70%), 673 (7.33%) and 773 (7.45%), while more than 9% of the deviations were still on the grids: 674 (9.66%) and 774 (10.50%).

All deviations from the limit value of noise levels used in the research, before and after the introduction of measures to reduce noise levels caused by air traffic are shown in Table 3.

As can be seen from Table 3, after the application of certain measures to reduce noise levels caused by air traffic, a significant reduction in noise emissions was achieved, but it is also noticeable that it is necessary to introduce additional measures to keep noise levels within limits, i.e. below 55 dB.

For this reason, the following additional measures have been proposed, which can be applied in the research area: – planting trees in the immediate vicinity of the school

and military hospital, – implementation of the so-called green roofs on both

buildings, – thicker facade of buildings and replacement of

windows and doors (sound insulation), etc. –

Application of ANNs

In order to assess the reliability of predicting the emission of the examined parameters (L_day, L_evening, L_night and L_den), i.e., to achieve the second goal of the research, the method of ANNs, i.e., MLP was used. Input, i.e. independent variables of this model included: engine power (%), runway distance (m) and air traffic volume, while the dependent variables included noise emission parameters (L_day, L_evening, L_night and L_den).

The results of ANN analysis for all four outputs (L_day, L_evening, L_night, or L_den) are presented in Table 4.

The results in Table 4 show that in all cases, ie. for all four outputs (L_day, L_evening, L_night, or L_den) RMSE values are lower than 1 which implies a higher coefficient of determination (Figs 5-8). RMSE is a performance metric that provides information about the short-term efficiency of a model. It measures the difference between predicted values and observed values, with a lower RMSE indicating a more accurate evaluation. On the other hand, the coefficient of determination (R²) measures the variance explained by the model, reflecting the reduction in variance when using the model. R² ranges from 0 to 1, with a value close to 1 indicating a model with strong predictive ability and a value close to 0 indicating a model that is not effective in analyzing the data. These performance metrics, RMSE and R², are reliable measures of the overall predictive accuracy of a model [47]. Accuracy metrics are commonly used in ANNs to assess the quality of predictions. Beside Mean square error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and symmetric mean absolute percentage error (SMAPE), RMSE is widely used accuracy metrics in various fields such as weather forecasting, medical, and engineering [48]. RMSE, specifically, measures the average magnitude of error between predicted and actual values. It can be visualized as the average vertical distance from the actual value to the corresponding predicted value on a fitted line. In simple terms, RMSE is the square root of MSE. Similar to MSE, the range of RMSE is from 0 to positive infinity, and a smaller RMSE value indicates a higher accuracy of the prediction model. However, unlike MSE, the units of RMSE are the same as the original units, making it more easily interpretable [48].

Fig. 3.

L_den distribution maps: a) noise before corrective measures (L_den_NB) and b) noise after corrective measures (L_den_NA).

Fig. 4.

L_den cut of scheme: a) noise before corrective measures (L_den_NB_cut off) and b) noise after corrective measures (L_den_NA_cut off).

The assessment of the reliability of the prediction using MLP showed that changes in noise emission during the day (L_day) in the study area resulted in a coefficient of determination of 0.996 (Fig. 5a). As can be seen on Fig. 5a, the distinction between the actual values and the predicted values was in range of -2.04 to +1.32 dB. Significance analysis (Fig. 5b) showed that air traffic volume has the greatest impact on predicting L_day variations.

The assessment of the reliability of the prediction of noise levels in the evening (L_evening) resulted in a coefficient of determination of 0.997 (Fig. 6a) and that the difference between the actual values and the predicted values was in range of -1.55 to +1.40 dB, while the significance analysis (Fig. 6b) showed that air traffic volume has the greatest impact on predicting L_evening variations.

The assessment of the reliability of the prediction of noise levels at night (L_night) resulted in a coefficient of determination of 0.997 (Fig. 7a), and that the distinction between the actual values and the predicted values was in range of -1.34 to +1.24 dB, while analysis of importance (Fig. 7b), as with other examined parameters showed that air traffic volume has the greatest influence on predicting L_night variations.

Finally, a 24-hour noise emission prediction reliability assessment (L_den) was performed, which showed that the coefficient of determination was 0.998 (Fig. 8a), and that the difference between the observed and predicted values was in rang of -1.56 to +0.92 dB. The air traffic volume had the greatest impact on noise pollution, as Hamida et al., [49] also stated in their study. This was confirmed by a significance analysis (Fig. 8b).

Therefore, it can be concluded that the air traffic volume and the engine power have the greatest impact, while the distance from the runway has the least impact on the variations of the analyzed noise emission parameters. Also, as Mansourkhaki et al. [17] stated that the high coefficients of determination for all dependent variables (in this case L_day, L_evening, L_night and L_den) confirms the reliability of the ANNs in predicting noise pollution caused by air traffic.

Considering that ANN analysis results have indicated lower values of RMSE (Table 4) it can be concluded that the model used in this part of research is good for prediction of dependent variables (L_day, L_evening, L_night and L_den).

Application of MLR

Finally, MLR analysis was used for determining coefficients of determination also for all four dependent variables (L_day, L_evening, L_night and L_den). The independent variables used in this part of the research were the same as the ones used in ANNs (runway distance (RD), air traffic volume (ATV) and engine power (EP)).

After applying MLR method (Equation 4) for predicting dependent variables, the developed noise prediction model is given in Equations (5-8).

According to Table 5, the correlation between independent variables (runway distance (RD), air traffic volume (ATV) and engine power (EP)) and dependent variables (L_day, L_evening, L_night and L_den) are 0.952, 0.951, 0.949 and 0.952, respectively. The R Squared for the regression models for dependant variables (L_day, L_evening, L_night and L_den) are 0.906, 0.904, 0.901 and 0.906, respectively. This results indicate that 90.6% of the L_day values, 90.4% of the L_evening values, 90.1% of the L_night values and 90.6% of the L_den values can be explained by the independent variables. RMSE values for L_day, L_evening, L_night and L_den are 3.521, 3.534, 3.390 and 3.366, respectively. Also the level of significance was less than 0.05 in all four cases, which means that all independent variable coefficients are statistically significantly different from zero.

Fig. 5.

Graphical representation of a) actual compared to predicted L_day emission values and b) importance analysis.

Fig. 6.

Graphic representation of a) actual compared to predicted L_evening emission values and b) importance analysis.

Fig. 7.

Graphic representation of a) actual compared to predicted L_night emission values and b) importance analysis.

Fig. 8.

Graphic representation of a) actual compared to predicted L_den emission values and b) importance analysis.

If we take into consideration the results of the conducted ANN analysis and MLR (the values of R² and RMSE) it can be concluded that ANN model used in the research offers better results, as Yusof et al. [50] and Kuznetsov et al. [16] also stated in their research.