Advanced Noise Indicator Mapping Relying on a City Microphone Network

Citation: Van Renterghem, T.; Le Bescond, V.; Dekoninck, L.; Botteldooren, D. Advanced Noise Indicator Mapping Relying on a City Microphone Network. Sensors 2023, 23, 5865. https://doi.org/ 10.3390/s23135865 Academic Editor: Hector Eduardo Roman Received: 25 May 2023 Revised: 16 June 2023 Accepted: 21 June 2023 Published: 24 June 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). sensors Article Advanced Noise Indicator Mapping Relying on a City Microphone Network Timothy Van Renterghem 1,* , Valentin Le Bescond 2, Luc Dekoninck 1and Dick Botteldooren 1 1WAVES Research Group, Department of Information Technology, Ghent University, Technologiepark 126, B 9052 Gent-Zwijnaarde, Belgium 2Joint Research Unit in Environmental Acoustics (UMRAE), Centre for Studies on Risks, Mobility, Land Planning and the Environment (CEREMA) and University Gustave Eiffel, F-44344 Bouguenais, France *Correspondence: timothy.vanrenter[email protected] Abstract: In this work, a methodology is presented for city-wide road traffic noise indicator mapping. The need for direct access to traffic data is bypassed by relying on street categorization and a city microphone network. The starting point for the deterministic modeling is a previously developed but simplified dynamic traffic model, the latter necessary to predict statistical and dynamic noise indicators and to estimate the number of noise events. The sound propagation module combines aspects of the CNOSSOS and QSIDE models. In the next step, a machine learning technique—an artificial neural network in this work—is used to weigh the outcomes of the deterministic predictions of various traffic parameter scenarios (linked to street categories) to approach the measured indicators from the microphone network. Application to the city of Barcelona showed that the differences between predictions and measurements typically lie within 2–3 dB, which should be positioned relative to the 3 dB variation in street-side measurements when microphone positioning relative to the façade is not fixed. The number of events is predicted with 30% accuracy. Indicators can be predicted as averages over day, evening and night periods, but also at an hourly scale; shorter time periods do not seem to negatively affect modeling accuracy. The current methodology opens the way to include a broad set of noise indicators in city-wide environmental noise impact assessment. Keywords: noise monitoring networks; microphones; road traffic noise; environmental noise mapping ; noise indicators 1. Introduction Road traffic is commonly the main source of exposure to environmental noise in European cities [ 1 ]. A basic step in road traffic noise mapping is gaining access to traffic parameters such as traffic intensity, vehicle speed, acceleration and traffic composition on each road segment in the network [ 2 ]. However, most traffic models focus on major roads only to perform congestion analysis during rush hours [ 3 ]. On smaller roads, in contrast, traffic data are most often lacking. Although this might be in line with the Environmental Noise Directive [ 4 ] in Europe, stipulating that noise maps should only be produced from 55 dB(A) Lden on, this is nevertheless problematic in view of city-wide noise mapping. When assessing human sleep disturbance due to noise, exposure mapping becomes even more challenging and should go down to levels as low as 40 dB(A) Lnight. Knowledge of less exposed zones is relevant as well since these zones should be of primary interest for future residential developments and are potentially restorative places in a city. Furthermore, only mapping exposure in part of a city could introduce bias in environmental justice studies, an important concern nowadays when making sustainable cities (see, e.g., [5]). An interesting line of research showed that street categorization in a city is able to estimate street-side exposure levels reasonably well [ 6 – 11 ], possibly accompanied with limited sets of snapshot measurements, where efforts can be minimized by suited sampling Sensors 2023,23, 5865. https://doi.org/10.3390/s23135865 https://www.mdpi.com/journal/sensors Sensors 2023,23, 5865 2 of 19 strategies [ 6 , 12 ]. Similarly, roadside noise measurements were shown to be able to adequately predict the underlying road traffic parameters such as vehicle speed, traffic intensity and the share of heavy vehicles [ 13 ]. In the open GIS initiative Open Street Map (OSM), every road in a city is present and assigned a specific category. This opens possibilities for full city noise mapping, including low(er) exposure zones. Linking noise exposure maps to human health effects is currently not very successful. For an important noise policy indicator such as self-reported noise annoyance, less than 30% of the observed variance found in a surveyed population is actually captured [ 14 ]. A possible reason for this low predictive power is that current noise mapping initiatives focus on long-term equivalent sound pressure levels only. This undermines a noise map as an efficient urban sound planning instrument. Only recently, the shortcomings of the commonly used energetically equivalent levels have been officially acknowledged [15]. The way people perceive environmental noise is much more complex than what can be quantified with these standard energetically averaged sound pressure levels. A wider set of noise indicators and psycho-acoustical indicators have long been used in other contexts, e.g., in product design [ 16 ] and soundscape studies [ 17 , 18 ]. Essentially, people are very good listeners, and even subtle changes in the spectro-temporal content of a sound might impact the perception and reaction to it. Noise indicators of potential interest are statistical sound pressure levels, the number of events and indicators describing the dynamic nature of urban sound. Currently, city-wide mapping of such noise indicators is very scarce. A few measurement-based initiatives can be found, where walkers equipped with microphones scan a particular city quarter [ 19 – 21 ]. The use of these more advanced indicators to better predict the impacts of environmental noise is currently underexplored. In this work, a methodology is described to predict both equivalent sound pressure levels and a wide range of other noise indicators, by means of deterministic noise modeling, where the accuracy of the predictions is improved by fitting to long-term measurements of a city noise monitoring network in a final step. The state-of-the-art deterministic noise modeling procedure, facilitating the calculation of dynamic noise indicators, is described in brief in Section 2. The proposed methodology is illustrated for the city of Barcelona (Spain) in Section 3, where a city-wide microphone network has been operational for more than a decade. 2. Deterministic Noise Mapping Procedure 2.1. Linking Traffic Data and Open Street Map Road Categorization Street categorization data were directly used from Open Street Map (OSM). Each street category was assigned a set of plausible traffic parameters (more precisely traffic intensity, vehicle speed and share of heavy vehicles). This assignment starts from existing (highway) traffic count databases, and it is ensured that the expected logics such as a lower traffic intensity, lower vehicle speed and a lower share of heavy vehicles on minor streets compared to major streets are present. Depending on the deterministically predicted noise indicators corresponding to a given scenario, additional scenarios were manually added. In total, 15 scenarios were used (see Appendix Afor an overview of the parameter settings). In a final step, the calculated outcomes for a wide set of noise indicators are weighted to minimize the difference with measurement from the microphone network, as will be discussed in Section 3.2. 2.2. Dynamic Traffic Model Simplified vehicle movements are modeled based on the hourly averaged number of vehicles and their speeds. Vehicles are launched on a road segment at a fixed speed. When 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 of the different cars 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 simplified micro-simulation traffic procedure can be found in [22]. A time step of 1 s was considered in this work. Sensors 2023,23, 5865 3 of 19 2.3. Traffic Noise Emission Model Vehicle category, number of vehicles per hour, and average speed are used as input for the CNOSSOS [ 23 ] road traffic acoustic emission model. These traffic-related inputs come from the dynamic traffic modeling procedure described in Section 2.2. 2.4. Sound Propagation Model The sound propagation modeling procedure combines the CNOSSOS sound propagation model [ 23 ] with aspects from the QSIDE urban sound propagation model [ 24 ]. Vehicles close to a receiver (within a radius of 500 m) are treated differently from those further away (between 500 m and at maximum 2000 m). At close distance, and when a direct line-of-sight propagation path is possible between a source and a receiver, geometrical divergence, ground effect and atmospheric absorption are included following the CNOSSOS sound propagation model, implemented in the open access NoiseModelling framework [ 25 , 26 ]. Only in the absence of a line-of-sight propagation path, diffractions around horizontal edges and reflections on vertical objects are accounted for. The maximum reflection order is 2, and the maximum source-reflection distance is 50 m (which are standard settings; see, e.g., [ 27 ]). A standard noise mapping receiver height of 4 m is used. Scattering on atmospheric turbulence is added to the attenuation factors to avoid levels becoming unrealistically low, especially behind objects. The QSIDE engineering scattering approach [ 28 ] was used, providing an easy-to-evaluate expression adapted to the urban environment, accounting for sound frequency, propagation distance, street canyon geometry and turbulence strength. Although the model could include local street canyon geometry in detail, standard building heights and widths were used to avoid time-demanding retrieval from geographical input data. Turbulence structure parameters C v2 and C T2 (see Table 1) depend on whether the propagation occurs in rural/suburban settings or in the dense urban fabric, and during the daytime or at night. These turbulence parameters are based on longterm observations over flat rural zones with dispersed smaller cities [ 29 ]; in the dense urban environment, turbulence strength is doubled in a simplified approach. Table 1. Overview of the sound propagation models and parameter settings. Close-by Traffic Far Traffic Radius around receiver (in m) <500 ≥500 and <2000 Traffic (noise emission) modeling Simplified dynamic traffic modeling following [22], at a 1 s time interval. Aggregated traffic at discrete emission points. Number of emission points minimized by NoiseModelling [26] If a direct line-of-sight path is possible CNOSSOS sound propagation model [23] without reflections on vertical objects, without diffractions, and in a non-refracting atmosphere. Only obstructed sound paths are present CNOSSOS sound propagation model [23] including reflections on vertical objects (reflection order 2, maximum source-reflection distance 50 m) and including diffractions on horizontal edges. Downward refraction (“favorable conditions”) is assumed with 50% occurrence in any direction. Turbulent scattering model [28] Rural/suburban Dense urban fabric Distance to façade (m) Not applicable 5 City canyon width (m) Not applicable 15 Building height (m) 8 20 Day Night Day Night Cv2(m4/3/s2)0.4 0.2 0.8 0.4 CT2(K2/m2/3)0.7 0.04 1.4 0.08 In order to capture dynamic noise indicators and noise events, considering individual nearby vehicles is essential. This is not the case anymore for road traffic further away Sensors 2023,23, 5865 4 of 19 contributing mainly to the background noise at a receiver. This allows bundling the acoustic energy of cars in a limited number of emission points as optimized by the NoiseModelling framework. For the propagation simulations, an approach similar to that for nearby traffic is followed, so depending on whether a line-of-sight path is possible or not. The CNOSSOS favorable sound propagation approach (i.e., downward refraction) is only considered in the absence of line-of-sight paths. In the CNOSSOS simplified curved ray approach, the difference between refraction/no refraction is mainly relevant in the case of propagation over objects. The probability of favorable sound propagation is then set to 50% in any propagation direction. 3. The Barcelona Microphone Sensor Network 3.1. Measurements and Data Handling The Barcelona microphone measurement network is unique in its kind due to its size (roughly 250 monitoring points spread over the city) and since it has been operational for more than a decade. The network contains both fixed sensors and sensors that are repositioned period-wise to maximize the zone monitored. Together with the fact that individual sensors are prone to accidental failure, the dataset is rather discontinuous in nature. Nevertheless, at some fixed sensors, continuous sound pressure level measurements over several years are present. The sensors are opportunistically positioned, e.g., directly attached to window sills or near balconies. This means that the extent to which façade reflections impact the sound pressure level measurements is not fixed (see Section 4). Microphones are always facing the streets and are representative of the most exposed building side. To limit the impact of changes in the traffic network infrastructure and its management (such as limiting traffic in specific streets, changing the direction of circulation, ban of heavy traffic), a 3-year period was selected, which was a compromise between keeping this period as short as possible and having a sufficient amount of data while keeping as many sensor locations as possible for processing. Nevertheless, changes in the traffic network cannot be fully avoided within this time frame, and if this was the case, the measurements were then the average between the two different traffic situations. Note that convergence must still be reached at such locations (see next paragraph) for a microphone position to be used. The processing of the measurement sensors was performed as follows. A basic time period of 15 min was chosen for all indicators. Previous research [ 12 ] showed that this is a suitable time frame in road traffic noise-dominated urban environments. Shorter periods could lead to difficulties in stabilizing the noise indicators, giving too much emphasis on momentary variations. When extending to longer periods, the temporal variations in the sonic environment might not be captured sufficiently. For a sensor location to be considered in further analysis, at least 3 weeks of data (not necessarily continuous) should be available. Weekends were excluded to avoid uncommon traffic situations. As a simplified convergence criterion, the difference between taking 80% of the data and all available data (in a chronological way) should be less than 1 dB when (linearly) averaging a noise indicator that uses a decibel scale. For event-based indicators, this criterion is set to five events, and for the intermittency ratio set to 5% (see further). If this condition is not met, this sensor location is disregarded at least for a specific time period. Removing sensor data during the day period, e.g., does not necessarily mean that the sensor location is also disregarded during the evening and night periods. The measurement network contains sensors with two levels of detail. Most microphone stations report total A-weighted sound pressure levels with a basic integration period of 1 min. These data were available at 93 sensors in the current study (see Section 3.3), during the period 2020–2021–2022. Secondly, measurement stations logging 1/3-octave bands with a basic integration period of 1 s were used during the period 2016–2017–2018. These more detailed data were available at 23 stations and allowed calculating more advanced noise indicators as discussed in Section 3.4. Sensors 2023,23, 5865 5 of 19 3.2. Machine Learning Fitting Procedure As an example supervised machine learning fitting algorithm, an artificial neural network was used, as implemented in Matlab [ 30 ]. A standard split into training, validation and test sets using 70%, 15% and 15% of the data, respectively, was chosen. The training algorithm “Levenberg–Marquardt backpropagation” was used, which is recommended as a fast, first-choice procedure [ 30 ]. To prevent overfitting, only five neurons were used, with a single hidden layer [ 31 ]. Given the random split into training, validation and test datasets, models were repeatedly constructed, allowing the use of averaged model predictions and giving an indication of confidence intervals on repeated predictions. In the case of predicting A-weighted equivalent sound pressure levels (see Section 3.3), the input consists of 93 locations × 15 traffic scenarios; there are 93 (locations) × 3 (day, evening and nightly averaged) or 93 (locations) × 24 (hourly averaged) outputs. In case more advanced noise indicators were included (see Section 3.4), 23 (locations) × 15 (traffic scenario) × 29 (indicators) inputs were used to predict 23 (locations) × 29 (indicators) × 3 (day, evening and nightly averaged) outputs. This work does not aim at finding the most accurate or fastest machine learning approach for this specific application, but rather showcases what can be achieved with a standard and well-established supervised machine learning fitting approach. Similarly, further optimization of the neural network settings is also beyond the scope of the current work. 3.3. Predicting A-Weighted Equivalent Sound Pressure Levels Using the basic L Aeq,1min values, an integration is performed to 15 min. In the next step, L Aeq,15min data are linearly averaged over day (7:00–19:00), evening (19:00–23:00) and night (23:00–7:00) periods, thus providing a typical value in each period, and form the basis for the artificial neural network predictions. In a second set of predictions, hourly averaged LAeq,15min data are used as well. Figures 1–3depict the 15 deterministic predictions at each sensor location that formed the basis for the weighting by the artificial neural network, together with the measured values (i.e., the ground truth), the mean predicted values and the 90th and 10th percentiles based on repeated model constructions. On the horizontal axis, the location ID number is used, which is an arbitrary number but easily allows assessing changes from location to location, both in measurements and predictions. Figures are shown for the daily, evening and nightly averaged L Aeq,15min . Note that sensors with an insufficient number of data points or sensors not leading to converged indicators (see Section 3.1) were obviously not used during the construction of the machine learning model. Once the model was constructed, predictions with the model were performed at all 93 sensor locations. Clearly, only locations with both measurements and predictions were considered in the subsequent accuracy analysis. In Figures 4–6, the measured data are plotted on the street map of Barcelona, complying with the selection criteria (see Section 3.1), the (mean) predictions at (all) sensor locations, and the difference between measurements and predictions where possible (as root-meansquare error, RMSE). As an example, daytime data only are shown. At most locations, prediction errors are limited, although a few points give rise to larger errors. Sensors 2023,23, 5865 6 of 19 Sensors 2023, 23, x FOR PEER REVIEW 6 of 19 locations. Clearly, only locations with both measurements and predictions were considered in the subsequent accuracy analysis. Figure 1. Deterministic predictions of LAeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged LAeq,15min during the daytime. Figure 2. Deterministic predictions of LAeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged LAeq,15min during the evening. Figure 1. Deterministic predictions of L Aeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged L Aeq,15min during the daytime. Sensors 2023, 23, x FOR PEER REVIEW 6 of 19 locations. Clearly, only locations with both measurements and predictions were considered in the subsequent accuracy analysis. Figure 1. Deterministic predictions of LAeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged LAeq,15min during the daytime. Figure 2. Deterministic predictions of LAeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged LAeq,15min during the evening. Figure 2. Deterministic predictions of L Aeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged LAeq,15min during the evening. Sensors 2023,23, 5865 7 of 19 Sensors 2023, 23, x FOR PEER REVIEW 7 of 19 Figure 3. Deterministic predictions of L Aeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90 th and 10 th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged L Aeq,15min during the night. In Figures 4–6, the measured data are plotted on the street map of Barcelona, complying with the selection criteria (see Section 3.1), the (mean) predictions at (all) sensor locations, and the difference between measurements and predictions where possible (as root-mean-square error, RMSE). As an example, daytime data only are shown. At most locations, prediction errors are limited, although a few points give rise to larger errors. Figure 4. Linearly averaged L Aeq,15min from measurements during daytime. Only sensor locations with converged measurements and data falling within the pre-selected timeframe are shown. Figure 3. Deterministic predictions of L Aeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90th and 10th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged LAeq,15min during the night. Sensors 2023, 23, x FOR PEER REVIEW 7 of 19 Figure 3. Deterministic predictions of L Aeq,15min for each single traffic scenario, by the artificial neural network (showing the mean prediction and the 90 th and 10 th percentiles, based on repeated model constructions), together with the converged measurements, at each of the 93 locations where a sensor node is/was operational. Data shown here are the linearly averaged L Aeq,15min during the night. In Figures 4–6, the measured data are plotted on the street map of Barcelona, complying with the selection criteria (see Section 3.1), the (mean) predictions at (all) sensor locations, and the difference between measurements and predictions where possible (as root-mean-square error, RMSE). As an example, daytime data only are shown. At most locations, prediction errors are limited, although a few points give rise to larger errors. Figure 4. Linearly averaged L Aeq,15min from measurements during daytime. Only sensor locations with converged measurements and data falling within the pre-selected timeframe are shown. Figure 4. Linearly averaged L Aeq,15min from measurements during daytime. Only sensor locations with converged measurements and data falling within the pre-selected timeframe are shown. Sensors 2023,23, 5865 8 of 19 Sensors 2023, 23, x FOR PEER REVIEW 8 of 19 Figure 5. Mean L Aeq,15min predictions during daytime, at 93 spots where a sensor node is/was operational. Figure 6. Root-mean-square error (RMSE) between measured and (mean) predicted L Aeq,15min during daytime. Only sensor locations with converged measurements and data falling within the pre-selected timeframe were used for this analysis. The histograms in Figure 7 depict the actual differences between measurements and predictions, showing that the zero error class is most populated in all time periods considered. During the night, the distribution is still symmetrical, but the spread seems somewhat larger. The RMSEs are 2.0 dB(A) during the daytime, 2.1 dB(A) during the evening and 3.3 dB(A) during the night. Figure 5. Mean L Aeq,15min predictions during daytime, at 93 spots where a sensor node is/was operational. Sensors 2023, 23, x FOR PEER REVIEW 8 of 19 Figure 5. Mean L Aeq,15min predictions during daytime, at 93 spots where a sensor node is/was operational. Figure 6. Root-mean-square error (RMSE) between measured and (mean) predicted L Aeq,15min during daytime. Only sensor locations with converged measurements and data falling within the pre-selected timeframe were used for this analysis. The histograms in Figure 7 depict the actual differences between measurements and predictions, showing that the zero error class is most populated in all time periods considered. During the night, the distribution is still symmetrical, but the spread seems somewhat larger. The RMSEs are 2.0 dB(A) during the daytime, 2.1 dB(A) during the evening and 3.3 dB(A) during the night. Figure 6. Root-mean-square error (RMSE) between measured and (mean) predicted L Aeq,15min during daytime. Only sensor locations with converged measurements and data falling within the pre-selected timeframe were used for this analysis. The histograms in Figure 7depict the actual differences between measurements and predictions, showing that the zero error class is most populated in all time periods considered. During the night, the distribution is still symmetrical, but the spread seems somewhat larger. The RMSEs are 2.0 dB(A) during the daytime, 2.1 dB(A) during the evening and 3.3 dB(A) during the night. Results for hourly predictions are depicted in Figure 8, shown as temporal patterns over 24 h periods at each sensor location. The measured temporal patterns are shown as well. This figure does not allow the comparison of measurements and predictions at any individual sensor location, but it nicely shows that the bulk of the temporal patterns are well predicted. Both locations with a rather flat pattern and those with stronger level drops during the night Sensors 2023,23, 5865 9 of 19 hours can be distinguished, both in the measurements and predictions. Directly related to Figure 8, Figure 9shows the hourly RMSEs. Minimum values are found around noon, near 2 dB(A), and increase slightly above 3 dB(A) between 3 and 4 o’clock at night. Sensors 2023, 23, x FOR PEER REVIEW 9 of 19 Figure 7. Histograms showing the difference between the (mean) predicted and measured LAeq,15min, linearly averaged over daytime, evening and night hours. Results for hourly predictions are depicted in Figure 8, shown as temporal patterns over 24 h periods at each sensor location. The measured temporal patterns are shown as well. This figure does not allow the comparison of measurements and predictions at any individual sensor location, but it nicely shows that the bulk of the temporal patterns are well predicted. Both locations with a rather flat pattern and those with stronger level drops during the night hours can be distinguished, both in the measurements and predictions. Directly related to Figure 8, Figure 9 shows the hourly RMSEs. Minimum values are found around noon, near 2 dB(A), and increase slightly above 3 dB(A) between 3 and 4 o’clock at night. Figure 8. Hourly temporal patterns of LAeq,15min at all 93 measurement locations. The measurements are shown together with the mean predictions based on repeated model construction. Data shown here are the linearly averaged LAeq,15min during a specific hour. Interrupted lines indicate hours where measurements are not converged due to an insufficient amount of data. Figure 7. Histograms showing the difference between the (mean) predicted and measured L Aeq,15min , linearly averaged over daytime, evening and night hours. Sensors 2023, 23, x FOR PEER REVIEW 9 of 19 Figure 7. Histograms showing the difference between the (mean) predicted and measured LAeq,15min, linearly averaged over daytime, evening and night hours. Results for hourly predictions are depicted in Figure 8, shown as temporal patterns over 24 h periods at each sensor location. The measured temporal patterns are shown as well. This figure does not allow the comparison of measurements and predictions at any individual sensor location, but it nicely shows that the bulk of the temporal patterns are well predicted. Both locations with a rather flat pattern and those with stronger level drops during the night hours can be distinguished, both in the measurements and predictions. Directly related to Figure 8, Figure 9 shows the hourly RMSEs. Minimum values are found around noon, near 2 dB(A), and increase slightly above 3 dB(A) between 3 and 4 o’clock at night. Figure 8. Hourly temporal patterns of LAeq,15min at all 93 measurement locations. The measurements are shown together with the mean predictions based on repeated model construction. Data shown here are the linearly averaged LAeq,15min during a specific hour. Interrupted lines indicate hours where measurements are not converged due to an insufficient amount of data. Figure 8. Hourly temporal patterns of L Aeq,15min at all 93 measurement locations. The measurements are shown together with the mean predictions based on repeated model construction. Data shown here are the linearly averaged L Aeq,15min during a specific hour. Interrupted lines indicate hours where measurements are not converged due to an insufficient amount of data. Sensors 2023,23, 5865 16 of 19 obtained, in excess of the deviation amongst reference microphones themselves [ 49 ]. More recent developments and experience with MEMS microphones [ 50 – 52 ] could further boost the deployment of city-wide microphone networks. The implicit traffic data retrieval in the current work could further benefit from including computer vision technologies [ 53 ]. In [54], e.g., camera images were directly used for noise mapping using machine learning. A relevant question is whether the trained fitting network could be transferable to other cities. It is expected that this is unlikely, since the link between street categories and traffic parameters could be strongly locally dependent. In addition, as discussed before, the network might not only weight the traffic scenarios, but also traffic and propagationrelated aspects are corrected for. Examples are the typical street widths and number of lanes corresponding to a specific street category in a city, traffic management policy, and preferred road surfaces and their maintenance. 5. Conclusions The proposed advanced noise indicator mapping procedure, using a set of deterministic predictions combined with data from a city microphone measurement network, has been shown to be an approach with high potential. Both equivalent sound pressure levels and more advanced noise indicators expressed in decibel units lead to RMSEs between 2 and 3 dB. These deviations should be positioned relative to the 3 dB variation in street-side urban road traffic noise exposure measurements when the microphone positioning relative to the façade is not fixed. The current work further shows that city-wide noise mapping without access to direct traffic data is feasible on the condition that a microphone network is available, and at the same time, systematic inaccuracies occurring at any stage during the deterministic modeling process might be implicitly corrected for, at least to some extent. Continued research and more case studies are needed to see whether the current concept can grow to a mature urban traffic noise mapping methodology. Author Contributions: Conceptualization, D.B. and T.V.R.; methodology, D.B., T.V.R., L.D. and V.L.B.; software, V.L.B., T.V.R., D.B. and L.D.; validation, T.V.R.; formal analysis, T.V.R.; investigation, V.L.B., D.B., T.V.R. and L.D.; resources, D.B. and T.V.R.; data curation, T.V.R.; writing—original draft preparation, T.V.R.; writing—review and editing, T.V.R., V.L.B., D.B. and L.D.; visualization, T.V.R.; supervision, T.V.R. and D.B.; project administration, D.B.; funding acquisition, D.B. All authors have read and agreed to the published version of the manuscript. Funding: 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. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: This study made use of third-party data and software that are partly open. The new data created in this work is available upon request. Acknowledgments: We are grateful to Javier Casado Novas and Julia Camps Farres for enabling access to the historical data from Barcelona’s microphone network. Conflicts of Interest: The authors declare no conflict of interest. Appendix A In Table A1, the link between the Open Street Map categories and the traffic intensity, vehicle speed and share of heavy vehicles is shown, for the 15 scenarios that were used in this work. Sensors 2023,23, 5865 17 of 19 Table A1. Traffic parameters assigned to the Open Street Map street categories that were explicitly calculated with the deterministic noise mapping methodology. Vehicle intensities (VIs) are expressed in cars per hour, the share of heavy vehicles (SHV) in %. Period Road Type Vehicle Speed (km/h) VI (SHV) Scenario 1 VI (SHV) Scenario 2 VI (SHV) Scenario 3 VI (SHV) Scenario 4 VI (SHV) Scenario 5 VI (SHV) Scenario 6 Day motorway 130 20,400 (15%) 10,200 (15%) 5100 (15%) 20,400 (15%) 20,400 (20%) 20,400 (15%) trunk 110 8400 (15%) 4200 (15%) 2100 (15%) 8400 (15%) 33,600 (20%) 16,800 (15%) primary 80 4800 (0%) 2400 (0%) 1200 (0%) 4800 (0%) 19,200 (5%) 9600 (0%) secondary 80 3300 (0%) 3300 (0%) 3300 (0%) 1750 (0%) 26,400 (5%) 6600 (0%) tertiary 50 350 (0%) 350 (0%) 350 (0%) 175 (0%) 8400 (0%) 2100 (0%) residential 30 175 (0%) 175 (0%) 175 (0%) 85 (0%) 350 (0%) 1400 (0%) service 30 80 (0%) 80 (0%) 80 (0%) 42 (0%) 175 (0%) 175 (0%) Evening motorway 130 20,400 (11%) 10,200 (11%) 5100 (11%) 20,400 (11%) 20,400 (16%) 20,400 (11%) trunk 110 1600 (11%) 800 (11%) 400 (11%) 1600 (11%) 12,800 (16%) 3200 (11%) primary 80 1000 (0%) 500 (0%) 250 (0%) 1000 (0%) 8000 (5%) 2000 (0%) secondary 80 600 (0%) 600 (0%) 600 (0%) 300 (0%) 9600 (5%) 1200 (0%) tertiary 50 100 (0%) 100 (0%) 100 (0%) 50 (0%) 2400 (0%) 600 (0%) residential 30 50 (0%) 50 (0%) 50 (0%) 25 (0%) 100 (0%) 400 (0%) service 30 25 (0%) 25 (0%) 25 (0%) 12 (0%) 50 (0%) 50 (0%) Night motorway 130 20,400 (32%) 10,200 (32%) 5100 (32%) 20,400 (32%) 20,400 (37%) 20,400 (32%) trunk 110 800 (32%) 400 (32%) 200 (32%) 800 (32%) 6400 (37%) 1600 (32%) primary 80 640 (0%) 320 (0%) 160 (0%) 640 (0%) 5120 (5%) 1280 (0%) secondary 80 360 (0%) 180 (0%) 180 (0%) 160 (0%) 5760 (0.5%) 720 (0%) tertiary 50 50 (0%) 50 (0%) 50 (0%) 25 (0%) 1200 (0%) 300 (0%) residential 30 25 (0%) 25 (0%) 25 (0%) 12 (0%) 50 (0%) 50 (0%) service 30 12 (0%) 12 (0%) 12 (0%) 6 (0%) 25 (0%) 100 (0%) VI (SHV) Scenario 7 VI (SHV) Scenario 8 VI (SHV) Scenario 9 VI (SHV) Scenario 10 VI (SHV) Scenario 11 VI (SHV) Scenario 12 VI (SHV) Scenario 13 VI (SHV) Scenario 14 VI (SHV) Scenario 15 33,300 (12.9%) 28,404 (16.2%) 34,315 (11.8%) 35,481 (7.7%) 20,400 (15%) 20,400 (20%) 20,400 (15%) 20,400 (15%) 20,400 (15%) 26,928 (5.4%) 21,012 (7.3%) 26,794 (4.4%) 27,705 (2.3%) 8400 (15%) 8400 (20%) 16,800 (15%) 8400 (15%) 33,600 (15%) 26,928 (5.4%) 21,012 (7.3%) 26,794 (4.4%) 27,705 (2.3%) 4800 (0%) 4800 (5%) 9600 (0%) 4800 (0%) 19,200 (0%) 18,192 (10.7%) 14,652 (13.6%) 18,061 (9.8%) 19,562 (5.7%) 6600 (0%) 6600 (5%) 13,200 (0%) 13,200 (0%) 26,400 (0%) 8928 (6.2%) 7476 (7.6%) 9050 (5.1%) 9717 (2.6%) 2100 (0%) 2100 (5%) 2100 (0%) 4200 (0%) 8400 (0%) 3216 (3.5%) 2400 (4.4%) 3062 (3.1%) 3404 (1.6%) 350 (0%) 350 (5%) 350 (0%) 700 (0%) 1400 (0%) 1098 (2.7%) 768 (3.6%) 1059 (2.4%) 1110 (1.4%) 175 (0%) 175 (5%) 175 (0%) 350 (0%) 700 (0%) 20,400 (11%) 20,400 (11%) 20,400 (11%) 20,400 (11%) 20,400 (11%) 20,400 (16%) 20,400 (11%) 20,400 (11%) 20,400 (11%) 3200 (11%) 3200 (11%) 3200 (11%) 3200 (11%) 1600 (11%) 1600 (16%) 3200 (11%) 1600 (11%) 12,800 (11%) 2000 (0%) 2000 (0%) 2000 (0%) 2000 (0%) 1000 (0%) 1000 (5%) 2000 (0%) 1000 (0%) 8000 (0%) 1200 (0%) 2400 (0%) 2400 (0%) 2400 (0%) 1200 (0%) 1200 (5%) 2400 (0%) 4800 (0%) 9600 (0%) 600 (0%) 600 (0%) 600 (0%) 600 (0%) 600 (0%) 600 (5%) 600 (0%) 1200 (0%) 2400 (0%) 400 (0%) 100 (0%) 100 (0%) 100 (0%) 100 (0%) 100 (5%) 100 (0%) 200 (0%) 400 (0%) 50 (0%) 50 (0%) 50 (0%) 50 (0%) 50 (0%) 50 (5%) 50 (0%) 100 (0%) 200 (0%) 20,400 (32%) 20,400 (32%) 20,400 (32%) 20,400 (32%) 20,400 (32%) 20,400 (37%) 20,400 (32%) 20,400 (32%) 20,400 (32%) 1600 (32%) 1600 (32%) 1600 (32%) 1600 (32%) 800 (32%) 800 (37%) 1600 (32%) 800 (32%) 6400 (32%) 1280 (0%) 1280 (0%) 1280 (0%) 1280 (0%) 640 (0%) 640 (5%) 1280 (0%) 640 (0%) 5120 (0%) 1440 (0%) 1440 (0%) 1440 (0%) 1440 (0%) 720 (0%) 720 (5%) 1440 (0%) 2880 (0%) 5760 (0%) 300 (0%) 300 (0%) 300 (0%) 300 (0%) 300 (0%) 300 (5%) 300 (0%) 600 (0%) 1200 (0%) 50 (0%) 50 (0%) 50 (0%) 50 (0%) 50 (0%) 50 (5%) 50 (0%) 100 (0%) 200 (0%) 25 (0%) 25 (0%) 25 (0%) 25 (0%) 25 (0%) 25 (5%) 25 (0%) 50 (0%) 100 (0%) References 1. European Environmental Agency. Environmental Noise in Europe 2020; EEA report No 22/2019; Publications Office of the European Union: Copenhagen, Denmark, 2020. 2. Licitra, G. Noise Mapping in the EU: Models and Procedures; CRC Press: Boca Raton, FL, USA; Taylor and Francis Group: Germantown, NY, USA, 2013. 3. Kessels, F. EURO Advanced Tutorials on Operational Research. In Traffic Flow Modelling: Introduction to Traffic Flow Theory Through a Genealogy of Models; Springer: Berlin/Heidelberg, Germany, 2018. 4. END. Directive 2002/49/EC of the European Parliament and of the Council of 25 June 2002 Relating to the Assessment and Management of Environmental Noise; European Commission: Brussels, Belgium, 2002. 5. Rickenbacker, H.; Brown, F.; Bilec, M. Creating environmental consciousness in underserved communities: Implementation and outcomes of community-based environmental justice and air pollution research. Sust. Cities Soc. 2019,47, 101473. [CrossRef] 6. Barrigón Morillas, J.M.; Gómez Escobar, V.; Méndez Sierra, J.; Vílchez-Gómez, R.; Vaquero Martínez, J.; Trujillo Carmona, J. A categorization method applied to the study of urban road traffic noise. J. Acoust. Soc. Am. 2005,117, 2844–2852. [CrossRef] Sensors 2023,23, 5865 18 of 19 7. Rey Gozalo, G.; Barrigón Morillas, J.M.; Gómez Escobar, V. Urban streets functionality as a tool for urban pollution management. Sci. Total Environ. 2013,461–462, 453–461. [CrossRef] [PubMed] 8. Zambon, G.; Benocci, R.; Brambilla, G. Statistical Road Classification Applied to Stratified Spatial Sampling of Road Traffic Noise in Urban Areas. Int. J. Environ. Res. 2016,10, 411–420. 9. Zambon, G.; Benocci, R.; Brambilla, G. Cluster categorization of urban roads to optimize their noise monitoring. Environ. Mon. Assess. 2016,188, 26. [CrossRef] [PubMed] 10. Barrigón Morillas, J.M.; Montes González, D.; Gómez Escobar, V.; Rey Gozalo, G.; Vílchez-Gómez, R. A proposal for producing calculated noise mapping defining the sound power levels of roads by street stratification. Environ. Pollut. 2021 ,270, 116080. [CrossRef] [PubMed] 11. Staab, J.; Schady, A.; Weigand, M.; Lakes, T.; Taubenböck, H. Predicting traffic noise using land-use regression—A scalable approach. J. Exp. Sci. Environ. Epidem. 2022,32, 232–243. [CrossRef] 12. Can, A.; Van Renterghem, T.; Rademaker, M.; Dauwe, S.; Thomas, P.; De Baets, B.; Botteldooren, D. Sampling approaches to predict urban street noise levels using fixed and temporary microphones. J. Environ. Monit. 2011,13, 2710–2719. [CrossRef] 13. Can, A.; Dekoninck, L.; Rademaker, M.; Van Renterghem, T.; De Baets, B.; Botteldooren, D. Noise measurements as proxies for traffic parameters in monitoring networks. Sci. Total Environ. 2011,410, 198–204. [CrossRef] 14. Brink, M. A Review of Explained Variance in Exposure-Annoyance Relationships in Noise Annoyance Surveys. In Proceedings of the International Commission on Biological Effects of Noise (ICBEN), Nara, Japan, 18–22 June 2014. 15. WHO. Environmental Noise Guidelines for the European Region; WHO Regional Office for Europe: Geneva, Switzerland, 2018. 16. Spence, C.; Zampini, M. Auditory contributions to multisensory product perception. Act. Acust. Acust. 2006,92, 1009–1025. 17. Kang, J.; Aletta, F.; Gjestland, T.; Brown, L.; Botteldooren, D.; Schulte-Fortkamp, B.; Lercher, P.; van Kamp, I.; Genuit, K.; Fiebig, A.; et al. Ten questions on the soundscapes of the built environment. Build. Environ. 2016,108, 284–294. [CrossRef] 18. Lionello, M.; Aletta, F.; Kang, J. A systematic review of prediction models for the experience of urban soundscapes. Appl. Acoust. 2020,170, 107479. [CrossRef] 19. Can, A.; Gauvreau, B. Describing and classifying urban sound environments with a relevant set of physical indicators. J. Acoust. Soc. Am. 2015,137, 208–218. [CrossRef] 20. Aumond, P.; Can, A.; De Coensel, B.; Botteldooren, D.; Ribeiro, C.; Lavandier, C. Modeling Soundscape Pleasantness Using perceptual Assessments and Acoustic Measurements Along Paths in Urban Context. Act. Acust. Acust. 2017 ,103, 430–443. [CrossRef] 21. Van Renterghem, T.; Thomas, P.; Dekoninck, L.; Botteldooren, D. Getting insight in the performance of noise interventions by mobile sound level measurements. Appl. Acoust. 2022,185, 108385. [CrossRef] 22. De Coensel, B.; Brown, A.L.; Tomerini, D. A road traffic noise pattern simulation model that includes distributions of vehicle sound power levels. Appl. Acoust. 2016,111, 170–178. [CrossRef] 23. Kephalopoulos, S.; Paviotti, M.; Anfosso-Lédée, F. Common Noise Assessment Methods in Europe (CNOSSOS-EU); Publications Office of the European Union: Luxembourg, 2012; 180p. 24. Wei, W.; Botteldooren, D.; Van Renterghem, T.; Hornikx, M.; Forssén, J.; Salomons, E.; Ögren, M. Urban background noise mapping: The general model. Act. Acust. Acust. 2014,100, 1098–1111. [CrossRef] 25. Aumond, P.; Fortin, N.; Can, A. Overview of the NoiseModelling Open-Source Software Version 3 and its Applications. In Proceedings of the NOISE-CON Congress (261, 4, 2005–2011), Seoul, Republic of Korea, 12 October 2012. 26. Bocher, E.; Guillaume, G.; Picaut, J.; Petit, G.; Fortin, N. NoiseModelling: An Open Source GIS Based Tool to Produce Environmental Noise Maps. ISPRS Int. J. Geo. Inform. 2019,8, 130. [CrossRef] 27. Le Bescond, V.; Can, A.; Aumond, P.; Gastineau, P. Open-source modeling chain for the dynamic assessment of road traffic noise exposure. Transp. Res. Part D Transp. Environ. 2021,94, 102793. [CrossRef] 28. Forssén, J.; Hornikx, M.; Botteldooren, D.; Wei, W.; Van Renterghem, T.; Ögren, M. A model of sound scattering by atmospheric turbulence for use in noise mapping calculations. Act. Acust. Acust. 2014,100, 810–815. [CrossRef] 29. Van Renterghem, T.; Horoshenkov, K.; Parry, J.; Williams, D. Statistical analysis of sound level predictions in refracting and turbulent atmospheres. Appl. Acoust. 2022,185, 108426. [CrossRef] 30. Matlab. The MathWorks Inc., version: 9.13.0 (R2022b); The MathWorks Inc.: Natick, MA, USA, 2022. Available online: https: //www.mathworks.com (accessed on 1 December 2022). 31. Hagan, M.; Demuth, H.; Beale, M.; De Jesus, O. Neural Network Design, 2nd ed.; Martin Hagan: Stillwater, OK, USA, 2014. 32. Wunderli, J.-M.; Pieren, R.; Habermacher, M.; Vienneau, D.; Cajochen, C.; Probst-Hensch, N.; Röösli, M.; Brink, M. Intermittency ratio: A metric reflecting short-term temporal variations of transportation noise exposure. J. Exp. Sci. Environ. Epidemiol. 2016 , 26, 575–585. [CrossRef] [PubMed] 33. Hall, F.L.; Papakyriakou, M.J.; Quirt, J.D. Comparison of outdoor microphone locations for measuring sound insulation of building facades. J. Sound Vib. 1984,92, 559–567. [CrossRef] 34. Memoli, G.; Paviotti, M.; Kephalopoulos, S.; Licitra, G. Testing the acoustical corrections for reflections on a facade. Appl. Acoust. 2008,69, 479–495. [CrossRef] 35. Mateus, M.; Carrilho, J.D.; Da Silva, M.G. An experimental analysis of the correction factors adopted on environmental noise measurements performed with window mounted microphones. Appl. Acoust. 2015,87, 212–218. [CrossRef] Sensors 2023,23, 5865 19 of 19 36. Barrigón Morillas, J.M.; Montes González, D.; Rey Gozalo, G. A review of the measurement procedure of the ISO 1996 standard. Relationship with the European Noise Directive. Sci. Total Environ. 2016,565, 595–606. [CrossRef] [PubMed] 37. Heutschi, K. A simple method to evaluate the increase of traffic noise emission level due to buildings for a long straight street. Appl. Acoust. 1995,44, 259–274. [CrossRef] 38. Montes González, D.; Barrigón Morillas, J.M.; Rey Gozalo, G. The influence of microphone location on the results of urban noise measurements. Appl. Acoust. 2015,90, 64–73. [CrossRef] 39. ISO 9613-2; Acoustics-Attenuation of Sound Propagation Outdoors, Part 2: General Method of Calculation. International Organization for Standardization: Geneva, Switzerland, 1996; revised in 2017. 40. Salomons, E.; Polinder, H.; Lohman, W.; Zhou, H.; Borst, H.; Miedema, H. Engineering modeling of traffic noise in shielded areas in cities. J. Acoust. Soc. Am. 2009,126, 2340–2349. [CrossRef] 41. Thomas, P.; Van Renterghem, T.; De Boeck, E.; Dragonetti, L.; Botteldooren, D. Reverberation-based urban street sound level prediction. J. Acoust. Soc. Am. 2013,133, 3929–3939. [CrossRef] 42. Jonasson, H. Acoustical Source Modelling of Road Vehicles. Act. Acust. Acust. 2007,93, 173–184. 43. Hadden, W.; Pierce, A. Sound diffraction around screens and wedges for arbitrary point source locations. J. Acoust. Soc. Am. 1981 , 69, 1266–1276. [CrossRef] 44. Öhrström, E.; Skånberg, A.; Svensson, H.; Gidlöf-Gunnarsson, A. Effects of road traffic noise and the benefit of access to quietness. J. Sound Vib. 2006,295, 40–59. [CrossRef] 45. Forssén, J.; Hornikx, M. Statistics of A-weighted road traffic noise levels in shielded urban areas. Act. Acust. Acust. 2006 , 92, 998–1008. 46. Farres, J.C. Barcelona Noise Monitoring Network. In Proceedings of the Euronoise 2015, Maastricht, The Netherlands, 31 May–3 June 2015; pp. 2315–2320. 47. Mydlarz, C.; Sharma, M.; Lockerman, Y.; Steers, B.; Silva, C.; Bello, J.P. The Life of a New York City Noise Sensor Network. Sensors 2019,19, 1415. [CrossRef] 48. Mietlicki, F.; Mietlicki, C.; Sineau, M. An Innovative Approach for Long Term Environmental Noise Measurement: RUMEUR Network in the Paris Region. In Proceedings of the Euronoise 2015, Maastricht, The Netherlands, 31 May–3 June 2015; pp. 2315–2320 . 49. Van Renterghem, T.; Thomas, P.; Dominguez, F.; Dauwe, S.; Touhafi, A.; Dhoedt, B.; Botteldooren, D. On the ability of consumer electronics microphones for environmental noise monitoring. J. Environ. Mon. 2011,13, 544–552. [CrossRef] 50. Mydlarz, C.; Salamon, J.; Bello, J.P. The implementation of low-cost urban acoustic monitoring devices. Appl. Acoust. 2017 , 117, 207–218. [CrossRef] 51. Quintero, G.; Balastegui, A.; Romeu, J. A low-cost noise measurement device for noise mapping based on mobile sampling. Measurement 2019,148, 106894. [CrossRef] 52. Yang, D.; Zhao, J. Acoustic Wake-Up Technology for Microsystems: A Review. Micromachines 2023,14, 129. [CrossRef] 53. Buch, N.; Velastin, S.; Orwell, J. A Review of Computer Vision Techniques for the Analysis of Urban Traffic. IEEE Trans. Intell. Transport. Syst. 2011,12, 920–939. [CrossRef] 54. Fredianelli, L.; Carpita, S.; Bernardini, M.; Del Pizzo, L.; Brocchi, F.; Bianco, F.; Licitra, G. Traffic Flow Detection Using Camera Images and Machine Learning Methods in ITS for Noise Map and Action Plan Optimization. Sensors 2022 ,22, 1929. [CrossRef] [PubMed] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.