Urban advanced noise indicator mapping relying on street categorization and measurements Timothy Van Renterghem 1 , Wout Van Hauwermeiren 2 , Luc Dekoninck 3 , Dick Botteldooren 4 Ghent University Department of Information Technology, WAVES Research Group Technologiepark 126, B 9052 Gent-Zwijnaarde, Belgium Valentin Le Bescond 5 UMRAE, Univ Gustave Eiffel, IFSTTAR, CEREMA, F-44344 Bouguenais, France

ABSTRACT In order to more accurately estimate health outcomes related to environmental noise exposure, indicators beyond long-term equivalent sound pressure levels might be needed (such as statistical levels, number of events, psycho-acoustical indices, etc). In urban noise mapping, predicting these more advanced noise indicators is especially challenging. In the current work, an open source noise mapping code (NoiseModelling) is combined with a simplified dynamic traffic estimation model. However, in most cities, traffic data availability is poor, especially in low traffic streets. To overcome this issue, the noise mapping procedure developed here assumes no direct access to traffic information and fully relies on Open Street Map (OSM) street categorization. These street categorizations were then assigned sets of plausible traffic compositions, counts and speeds; various scenarios were explicitly simulated. In a next step, these traffic scenarios were weighted to best fit a set of 29 noise indicators on 23 measurement stations deployed in the city of Barcelona, during various periods of the day. It was shown that this procedure leads to adequate assessments of a wide range of noise indicators.

1. INTRODUCTION

In order to more accurately estimate health outcomes due to environmental noise exposure such as sleep disturbance and noise annoyance, indicators beyond long-term equivalent sound pressure levels might be needed. Possible candidates are statistical levels, number of events, and all kinds of psycho-acoustical indices. In urban road traffic noise mapping, predicting these less common noise indicators is especially challenging and procedures to do so are currently lacking. Besides the need for more detailed information on road traffic dynamics, they also vary strongly with location, making the data gathering process problematic when aiming at the full city scale.

1 timothy.vanrenterghem@ugent.be 2 wout.vanhauwermeiren@ugent.be 3 luc.dekoninck@ugent.be 4 dick.botteldooren@ugent.be 5 valentin.lebescond@univ-eiffel.fr

Note that data availability is typically poorest in low traffic streets, often neglected in traditional noise mapping aiming primarily at the highly exposed zones. In order to find the noise indicators with the highest predictive power, also less exposed environments should be included. Only when this information is added, objective noise indicators and noise impact (e.g. from city-wide questionnaires) can be accurately linked. In this work, a methodology is developed to predict a wide range of noise indicators, with the potential to cover a full city. The procedure uses deterministic modeling, relies on a city microphone network and uses a machine learning technique. 2. DETERMINISTIC MODELLING

2.1. Dynamic traffic model Simplified but realistic vehicle movements are modeled based on hourly averaged vehicle counts and speeds. Vehicles are launched on a road segment with a fixed speed. Upon reaching the end of that segment, the vehicle is removed from the simulation, meaning there is no vehicle transfer from one segment to another. The inter-vehicle times respect a Poisson distribution and vehicle speeds follow a normal distribution. At the end of the simulated hour, at each road segment, the (static) vehicle counts and average speeds are respected. More information on this traffic modeling procedure can be found in Ref. [1]. The current procedure assumes no direct access to traffic information whatsoever and fully relies on Open Street Map street categorization. Street categorization was shown to be a basic predictor of street noise levels before [2]. Here, each street category was assigned a set of plausible traffic parameters. Various scenarios were consequently explicitly assumed to capture temporal variations in the correspondence between street category and these traffic parameters.

2.2. Traffic noise emission model Vehicle category (light, medium or heavy vehicles, or motorized two-wheelers), number of vehicles per hour, and average speed are used as input for the CNOSSOS [3] road traffic acoustic emission model. These traffic related inputs come from the dynamic traffic modeling procedure described in Section 2.1.

2.3. Sound propagation model : CNOSSOS To calculate sound propagation from any relevant noise source towards a receiver, the CNOSSOS [3] sound propagation model was used, as implemented in the open access NoiseModelling framework [4][5]. CNOSSOS is the recommended model for noise mapping in the European Union, and is essentially an update of the ISO9613-2 model with somewhat improved diffraction formulae. The model captures the basic physics of sound propagation in the outdoor environment, and is suited for strategic noise mapping of large zones. 2. MICROPHONE NETWORK

Historical data from 23 measurement stations, deployed in the city of Barcelona, was used in this work. Depending on the location, there was data for several weeks to about a full year, with a basic temporal resolution of 1 s, at the spectral detail of 1/3 octave bands. In a next step, the data for 29 acoustical indicators were aggregated to day-evening-night periods using all available data at a specific sensor. These indicators are summarized in Table 1. The Intermittency ratio is defined as discussed in Ref. [6].

Table 1: Sound indicators considered in this work.

Equivalent

Statistical

Number of events above

Number of events above a

Sound dynamics indicators

Intermittency

Sound pressure

sound pressure

ratio

x dBA

specific indicator

levels

levels

Leq LA01 EN55 ENLA10 LA10-LA90 intRatio

LAeq LA05 EN60 ENLA50 LC10-LC90

LCeq LA10 EN65 ENLA50+3 sigmaAS

LA50 EN70 ENLA50+10 sigmaCS

LA90 EN75 ENLA50+15

LA95 EN80 ENLA50+20

LA99 ENLAeq+10

ENLAeq+15

Note that the position of the microphones relative to the facade might change from location to location, and this information is not available. When at a few meters from the facade, a rather constant level increase (relative to the absence of the facade) of 3 dBA can be expected for typical road traffic noise (see e.g. Ref. [7]). When flush mounted on a rigid plate against the facade, this level might increase to about 6 dBA. During the fine tuning of the model with the measurements, similar distances between microphone and facade as in the measurement stations dataset are then assumed. 3. FINE-TUNING THE DETERMINISTIC PREDICTIONS WITH THE MEASUREMENTS

In a final step, the deterministic noise indicator predictions for the different street type-traffic parameter pairs need to be combined. There are various candidate techniques for this fine tuning on the measurements, ranging from simple regression modeling to machine learning techniques. In this work, the artificial neural network (ANN) methodology was used as a multi-input multi-output model. The inputs were 29 noise indicators calculated by 11 different traffic scenarios (=319 predictors) at 23 locations. The output of the model were 29 indicators at 23 measurement locations. Models were developed for the day, evening and night period, separately. For the ANN, 15 hidden layers were used, with a random assignment of data to the training, validation and test set for 70%, 15%, and 15%, respectively. To check the influence of such a division, the network was built 50 times, and the means of the predictions were used, allowing to determine confidence intervals on the predictions. Figure 1 shows the predictions at the 23 measurement locations on which the model has been trained, including the uncertainty evaluation, for each of the noise indicators. Overall, the model is able to capture well the variation in indicator values at the different sensor locations. The same information is represented in Figure 2, but ordered per location.

Figure 1: The black lines show the long-term measurements during day time for the Barcelona network. The data is grouped per indicator; where 1=locID 8, 2=locID 16, ... , 23=locID 227 on the horizontal axis. The red lines are the predicted values, as the means of 50 neural network constructions. The errorbars show the 95% confidence intervals on these means, assuming a normal distribution of the model results. Note that the different indicators have different units.

Lea rea Lat Laos. ato cy 0 20| 0 bal 70] 1 c 70] a ol oo ol oo rr a a ) o 0 oF WW © oO 0 @% ” aso Aso tase ‘eo eNsS ‘eo Neo ol oo oo a al so 0 «0 «o «0 Cr rs a ) rr) 0 20 0 0 2 ‘00 Nes N70 ol Nao so, ENATO ENLASO ‘00 100 oo al ‘) 20 ol ol " ol ° 0 m 0 0 2 . 0 2» 70 0 % 0 2 ENLASO+S. ENLASO+10 a__ENLASO820 ENLAGg* 10 ENLAGgH 5 9 ao 4 Ey 10] 2 1 ol dl al 0 2 oO 0 2 oo 2 0 20 0 20 ‘Sigmaas Sigmacs Leto..cs0 IntRatio al 6 1s al A oo 2m oF 0 2 Cr )

Figure 2: The black lines show the long-term measurements during day time for the Barcelona network. The data is grouped per location; with 1=Leq, 2=LAeq, ... 29=IntRatio (see Figure 1 for follow-up) on the horizontal axis. The red lines are the predicted values, as the means of 50 neural network constructions. The errorbars show the 95% confidence intervals on these means, assuming a normal distribution of the model results. Note that the different indicators have different units.

tooo =8 toold = 16 told = 28 toold = 34 told = 35 tool = 37 100 100 100 100 100 ry 0 * oo a «0 20 ay ° ol ol a rr rr a a a rr) 0 3% 0 0 00,__eeld= 48 told = 61 0010s = 88 tociD = 91 tocid = 99 tc0;_—_leetO = 108 c 100] oo oo oo ao “ oo 0 ” a ol ° ol ol dl a ° 0 2 w» % 0 2 % ‘oO 0 am 0 0 0 2 % o 1 mm % 0 10 2 focid = 105 00,22 = 110 00,__leelo = 114 focid = 120 focid = 130 00;__leelo = 142 © oo oo “ a 0 «0 ao oo ao 2» 20 ° OD ol ol él ol ow 2 w» % 0 » 0 w am wo % 0 a » 0 1 2 w% 0 1 2 focid = 162 focio = 167 focid = 218 focid = 220 focio = 227 ao o8 88 g_8 = o 8 8 ad o8 88 o8 ss 7 a 0 % 0 mo % o W m wo 0 1 2 %™

The absolute error analysis in Figure 3 shows that equivalent sound pressure levels are predicted within a few decibels. Predictions of statistical noise levels typically deviate a few decibels more from the measured values. The number of events above a fixed level ENx seems a less stable parameter, in contrast to e.g. ENLA50+x and ENLAeq+x. Also the indicators linked to the dynamic nature of the sound fields lead to limited absolute errors. The ANN model applied at locations not considered during the model development show plausible variations from one location to another, together with plausible value ranges, as shown with the examples in Figure 4 and Figure 5. Note that at these locations, the deterministic noise modeling is performed with the same models and parameter settings as before, after which the developed ANN was applied to fine-tune these modeling results.

Figure 3: Absolute error between the (averaged) neural network predictions and the measurements. Note that different units are used (decibels and number of events) but the coloring uses a single scale.

Figure 4: Final predictions at about 250 locations not considered during the ANN model development for the LAeq indicator (in dB), during the daytime period. The errorbars show the 95% confidence intervals on the predicted means, assuming a normal distribution of the model results.

“a L L L L L

Figure 5: Final predictions at about 250 locations not considered during the ANN model development for the LA10-LA90 indicator (in dB), during the daytime period. The errorbars show the 95% confidence intervals on the predicted means, assuming a normal distribution of the model results.

4. DISCUSSION AND CONCLUSIONS

The procedure followed shows to be capable of producing city noise maps for a wide variety of complex acoustic indicators, going beyond equivalent sound pressure levels. A major and common practical issue, namely the lack of spatially and temporally detailed traffic data, including information on minor roads, is efficiently solved here by relying on the open street map street categorization and a simplified procedure for simulating dynamic traffic. Most noise indicators will typically exhibit a daily pattern. This means that the traffic parameters suited for a specific street type might change over time. Therefore, it is mandatory to feed the artificial neural network model with a variety of plausible correspondences between street type and traffic parameters, from which the most suited combinations are then implicitly chosen for each period of the day. In the current work, the processing is restricted to day, evening and night period, leading to separate models. Further analysis might be needed to check potential over-fitting on the data, given the rather limited locations with measurements. Nevertheless, the predictions of the noise indicators at 250 facade locations, other than the measurement stations used to build the ANN model, show plausible value ranges and plausible variations from location to location, which gives at least some confidence in the followed procedure. The uncertainty assessment on the measurements further show which indicators are more easily predicted. The current measurement-based tuning of coefficients will probably only work well in the city for which it is designed. In this work, the city of Barcelona was used given the availability of long-term environmental noise measurements at several locations with a sufficient spectro-temporal detail. To further check the feasibility of the methodology, microphones should be positioned in less exposed streets as well. This is especially interesting when studying micro-environments (e.g. a quiet facade [8]). It can be reasonably expected that there will be a dominant traffic parameter set while training the ANN, but other and often complex links between the various noise indicators, and also their behavior at various locations, might help the predictions. The ANN is capable of unraveling and using such (hidden) information if present. This might hold as well for possible links between the indicators during the day, evening or night period. However, such information was not used since separate models were constructed for each period. In addition, simplifications and inaccuracies arising from the deterministic modeling process might be corrected for in the final fitting. As an example, in urban noise mapping, the number of reflections in street canyons is most often set to a small number to speed up calculations. Research showed that a much larger number is actually needed [9]. As another example, the local vehicle fleet might not fully correspond to the standard sound emission model and can also be corrected in this way.

5. ACKNOWLEDGEMENTS

We acknowledge the funding from the European Union’s Horizon 2020 research and innovation programme for the project “Equal life”, as part of the European Human Exposome Network, under grant agreement No 87474. We are also grateful to Mrs. Julia Camps Farres and Dr. Maria Forester for giving access to the historical data from Barcelona’s microphone network. 6. REFERENCES

1. B. De Coensel, A. L. Brown, D. Tomerini. A road traffic noise pattern simulation model that

includes distributions of vehicle sound power levels. Applied Acoustics 111, 170-178 (2016). 2. J. Barrigón Morillas, V. Gómez Escobar, J. Méndez Sierra, R. Vílchez-Gómez, J. Vaquero

Martínez, J. Trujillo Carmona. A categorization method applied to the study of urban road traffic noise. Journal of the Acoustical Society of America 117, 2844–2852 (2005). 3. S. Kephalopoulos, M. Paviotti, F. Anfosso-Lédée. Common noise assessment methods in

Europe (CNOSSOS-EU). Publications office of the European Union, 180 p (2012). 4. P. Aumond, N. Fortin, A. Can. Overview of the NoiseModelling open-source software version 3

and its applications. Proceedings NOISE-CON Congress (261, 4, 2005-2011). Institute of Noise Control Engineering (2020). 5. E. Bocher, G. Guillaume, J. Picaut, G. Petit, N. Fortin. NoiseModelling: An Open Source GIS

Based Tool to Produce Environmental Noise Maps. ISPRS International Journal of Geo- Information 8, 130 (2019). 6. J-M. Wunderli, R. Pieren, M. Habermacher, D. Vienneau, C. Cajochen, N. Probst-Hensch, M.

Röösli, M. Brink. Intermittency ratio: A metric reflecting short-term temporal variations of transportation noise exposure. Journal of Exposure Science and Environmental Epidemiology, 26, 575–585 (2016). 7. G. Memoli, M. Paviotti, S. Kephalopoulos, G. Licitra. Testing the acoustical corrections for

reflections on a facade. Applied Acoustics 69, 479–495 (2008). 8. E. Öhrström, A. Skånberg, H. Svensson, A. Gidlöf-Gunnarsson. Effects of road traffic noise and

the benefit of access to quietness. Journal of Sound and Vibration 295, 40-59 (2006). 9. E. Salomons, H. Polinder, W. Lohman, H. Zhou, H. Borst, H. Miedema. Engineering modeling

of traffic noise in shielded areas in cities. Journal of the Acoustical Society of America 126, 2340-2349 (2009).