Urban-Scale Acoustic Comfort Map: Fusion of Social Inputs, Noise Levels, and Citizen Comfort in Open GIS

Abstract

With advancements in the Internet of Things (IoT), diverse and high-resolution data sources, such as environmental sensors and user-generated inputs from mobile devices, have become available to model and estimate citizens’ acoustic comfort in urban environments. These IoT-enabled data sources offer scalable insights in real time into both objective parameters (e.g., noise levels and environmental conditions) and subjective perceptions (e.g., personal comfort and soundscape experiences), which were previously challenging to capture comprehensively by using traditional methods. Despite this, there remains a lack of a clear framework explicitly presenting the role of these diverse inputs in determining acoustic comfort. This paper contributes by (1) exploring the relationship between attributes governing the physical aspect of the built environment (sensory data) and the end-users’ characteristics/inputs/sensations (such as their acoustic comfort level) and how these attributes can correlate/connect; (2) developing a CityGML-based framework that leverages semantic 3D city models to integrate and represent both objective sensory data and subjective social inputs, enhancing data-driven decision making at the city level; and (3) introducing a novel approach to crowdsourcing citizen inputs to assess perceived acoustic comfort indicators, which inform predictive modeling efforts. Our solution is based on CityGML’s capacity to store and explain 3D city-related shapes with their semantic characteristics, which are essential for city-level operations such as spatial data mining and thematic queries. To do so, a crowdsourcing method was used, and 20 perceptive indicators were identified from the existing literature to evaluate people’s perceived acoustic attributes and types of sound sources and their relations to the perceived soundscape comfort. Three regression models—K-Nearest Neighbor (KNN), Support Vector Regression (SVR), and XGBoost—were trained on the collected data to predict acoustic comfort at bus stops in Montréal based on physical and psychological attributes of travellers. In the best-performing scenario, which incorporated psychological attributes and measured noise levels, the models achieved a normalized mean squared error (NMSE) as low as 0.0181, a mean absolute error (MAE) of 0.0890, and a root mean square error (RMSE) of 0.1349. These findings highlight the effectiveness of integrating subjective and objective data sources to accurately predict acoustic comfort in urban environments.

1. Introduction

With the rapid development of urban areas, citizens’ expectations for quality of life continue to rise. Comfort is a significant concern in both indoor and outdoor public spaces. While environmental factors in indoor spaces are controllable to a certain extent, the outdoor environment is more complex and influences human comfort in diverse ways. Therefore, studying the factors that affect human comfort outdoors is essential to understanding the physical environmental parameters that influence outdoor comfort and is essential to improving the design of outdoor spaces [1]. Human comfort in outdoor environments has been examined from various perspectives, including thermal conditions [2], acoustic environment [3], building type [4], architectural space [5], urban shading [6], wind conditions [7], and green space structure [8].

Acoustic comfort refers to the perceived state of well-being and satisfaction with the acoustic conditions of an environment [9]. It is a crucial element of the public’s environmental experience in urban settings [10] and plays an essential role in optimizing outdoor space design and urban planning and improving quality of life. Consequently, it directly affects how much people enjoy and utilize urban spaces. While there is a substantial body of research on acoustic comfort, much of it has focused on physical noise parameters, such as noise level and source. Limited attention has been given to integrating social, physiological, and psychological factors into acoustic comfort assessments in urban outdoor environments. For instance, Yang and Kang (2005) emphasized the importance of considering subjective and contextual factors, such as user behavior and demographic characteristics, in assessing urban soundscapes and acoustic comfort [11]. This study seeks to address this gap by exploring the combined influence of these factors, alongside physical parameters, on outdoor acoustic comfort.

Addressing acoustic comfort requires adopting a ‘soundscape’ approach [12]. Acoustic comfort is crucial to the public’s overall urban experience and significantly influences citizens’ enjoyment of urban spaces. The citizen’s perception of an environment’s soundscape represents their level of acoustic comfort in that location. There is a distinction between the acoustic environment (a physical phenomenon) and the soundscape (its perceptual construct) [13]. The acoustic environment of a place is determined by the sound that a person can hear from various sources, which are influenced by the noise sources, the receiver’s location, and the conditions along the propagation path. Each of these factors may vary depending on the time of the day or seasons [14]. In contrast, the soundscape represents a person’s perceptual construct of the acoustic environment in a given place [14]. The soundscape is defined as how people understand, experience, or perceive the acoustic environment of a specific place, with all its physical characteristics [13]. Additionally, the soundscape is influenced by terrain topography and boundary conditions.

Nowadays, noise measurement in cities is conducted by officials using specialized equipment, such as sound level meters or mobile applications like NIOSH Sound Level Meter App [15]. This manual data collection method is expensive and does not work if higher granularity of noise measurements is needed [16]. The simulation of outdoor noise levels has its difficulties, as the calculations require significant computing power due to the large data volume involved [17]. Additionally, outdoor noise measurements are often affected by environmental factors such as wind, temperature variations, and background noise, which can impact the accuracy of readings. Predictions based on engineering models further require detailed information about the area and noise sources and present a limited degree of accuracy, adding complexity to data collection and interpretation.

In terms of evaluating acoustical comfort, most research has been limited to indoor or lab settings, and the impact of parameters related to acoustical comfort in outdoor public spaces has not been thoroughly explored. Furthermore, there is no scientific consensus on a standard method for evaluating the soundscape by using acoustic parameters, despite the concept being well defined [18].

The human perception of an acoustic environment depends on (i) the physical features of the built environment and the personal characteristics of the individual exposed to noise, which can, in turn, be grouped into (ii) physiological and (iii) psychological parameters [10]. Physical parameters include dimensions of the surrounding buildings and their materials, vegetation, and physiological parameters, including gender, income, and employment category (employed or unemployed) [19]. These physiological factors can indirectly influence psychological parameters. For example, income may affect an individual’s expectations of their surroundings and tolerance to noise, with higher-income individuals potentially expecting quieter environments. Gender differences can shape perception, as studies suggest that men and women may respond differently to noise, potentially affecting perceived pleasantness and annoyance levels. The employment category may also influence psychological responses, as individuals with more demanding jobs might perceive urban noise differently compared with those who are unemployed or have more flexible schedules. Instances of psychological parameters include pleasantness, expectation, and prior experience [12]. To investigate the relationship between these parameters and acoustic comfort, a data model is needed to store individuals’ physiological characteristics, psychological responses, and information about the built environment. This model can then be connected to measured sound level data, enhancing data retrieval, integrity, and consistency while enabling advanced spatial–temporal analyses.

GISs (geographical information systems) are broadly used to support the collection, storage, and advanced analysis of spatial data at the urban scale (and beyond). This requires a standardized representation of data that includes geometry, appearance, and topology, along with semantic definitions that introduce meaning to various components of the built environment. Additionally, it captures these components’ properties as attributes within a 3D city model. Achieving this requires a modelling approach that facilitates the integration, storage, exchange, and visualization of geospatial information. CityGML (City Geography Markup Language) [20] is an open urban data model providing a standard for representing and exchanging 3D city models. It stores semantically enriched spatial data (in 2D and/or 3D) in standardized formats, enabling data exchange and interoperability. As a result, a CityGML-based semantic 3D city model serves as a hub for integrated comfort-related information.

CityGML can be extended based on the specific application domain. By using CityGML to evaluate citizens’ acoustic comfort, the physical parameters related to acoustic comfort and noise levels can be stored within CityGML’s core and the Noise ADE (an Application Domain Extension for CityGML), respectively. Moreover, physiological and psychological parameters can be stored in the Social ADE, an extension of CityGML recently proposed by the authors to aggregate, organize, and filter end-user information and inputs. This extension integrates distributed citizen data within city digital twins to enhance city models [21]. The ultimate goal of the Social ADE is to enrich the 3D urban models with citizen-contributed data, supporting data-driven decision making in urban infrastructure projects. For the specific use case of acoustic comfort, this extension needs to be updated to incorporate citizens’ feedback as semantic objects within the CityGML model.

While previous studies have explored noise measurement and acoustic comfort in controlled environments, there remains a significant gap in research on outdoor acoustic comfort that integrates objective noise levels with subjective, social parameters in real-world settings. Traditional approaches often neglect the socio-psychological dimensions that shape individuals’ perceptions of noise and their tolerance levels. Our study specifically addresses this gap by introducing a novel GIS-based framework that incorporates psychological and physiological attributes of citizens, utilizing the Social ADE to link social and spatial data seamlessly. This approach allows for a more comprehensive evaluation of acoustic comfort by taking into account not only the physical noise levels but also the nuanced psychological and social factors that influence comfort in urban environments. By integrating these human-centric attributes within a spatial context, our model provides urban planners and policymakers with a deeper understanding of how different demographic and psychological profiles impact perceived comfort across various outdoor spaces.

This paper aims to integrate three categories of parameters essential to evaluating outdoor acoustic comfort: physical environment, and physiological and psychological characteristics. On the one hand, these parameters can be stored and linked automatically in CityGML, and on the other hand, after analyzing and visualizing acoustic comfort at various points of interest, the results are stored back into the initial 3D city model. Such data-rich models can then be used for cross-analysis involving the built environment’s end-users, the services it provides, and its physical characteristics. This, in turn, helps identify necessary changes to better meet user needs. To achieve this, a prediction model is trained, and the results are structured to comply with CityGML, the Noise ADE, and the Social ADE.

In this way, an experiment was conducted in this study where the participants were asked to measure the noise level and complete questionnaires regarding their sensations. Utilizing smartphones as distributed sensors among citizens has become a common approach in previous studies, such as the CITI-SENSE EU project [22], although professional sound level meters are generally used for higher accuracy. Various mobile apps enable noise level measurement by using cellphone microphones, such as Decibel X: dB [23], SPLnFFT (Sound Pressure Level and Fast Fourier Transform) Noise Meter [24], and Decibel Meter Pro [25]. This paper used the ‘NIOSH (National Institute for Occupational Safety and Health) Sound Level Meter App’ mobile application [15] to measure noise levels. The app is free, easy-to-use and logs the user’s location during measurements. Its accuracy is within 2 dBA, which is slightly lower than that of professional sound level meters but remains acceptable for the purposes of this study. Since the application is only available on iPhones, all devices used in this study were similar in hardware and software specifications to minimize potential measurement discrepancies. It should be noted, however, that the accuracy of the app is outside the scope of this study, as the methodology can be adapted to any alternative application running on mobile devices. The noise level is reported based on an A-weighted, equivalent continuous sound pressure level (LAeq).

Several bus stops in Montréal, Canada, were chosen as the sites for evaluating acoustic comfort. Bus stops were selected for several reasons: (i) they are used daily by travellers, with the same person often visiting the same stop at different times of the day and on different days of the week; (ii) they are frequented by a diverse range of users with varying profiles; (iii) citizens typically spend a considerable amount of time waiting there; (iv) bus stops in congested areas were specifically chosen, as they offer a greater potential for long-term exposure.

On the other hand, assuming that traffic is the primary noise source, noise levels were simulated by using CadnaA software version 2023 MR 1 (32 bit) (Licensee: SLR Consulting, Guelph, Canada) [26]. This software predicts noise levels from various sources, including road traffic, industry, and railways. CadnaA evaluates noise levels in both outdoor and indoor environments, considering all relevant noise sources, as well as absorption and reflection effects. Finally, by using the actual noise levels, the simulated noise levels, and participants’ physiological and psychological parameters, the three regression models of KNN (K-Nearest Neighbor), SVR (Support Vector Machine), and XGBoost (Extreme Gradient Boosting) were trained to predict acoustic comfort. These models were selected due to their proven effectiveness in handling small-to-medium-sized datasets and their interpretability in relation to key influencing factors on acoustic comfort. Unlike more complex models, such as artificial neural networks (ANNs) or recurrent neural networks (RNNs), KNN, SVR, and XGBoost provide easily interpretable results that allow us to understand the influence of individual predictors, which aligns well with the study’s objectives.

The remainder of this paper is organized as follows: Section 2 presents the architecture of the proposed methodology, detailing the systematic approach for data acquisition and model development essential to this study. Section 3 discusses the implementation of the reference architectures, including the data collection campaign (Section 3.1), the process of feeding the GIS with relevant data (Section 3.2), and the training of predictive models (Section 3.3). Section 4 focuses on integrating the collected data within the GIS framework, establishing a cohesive platform for analysis. Section 5 delves into model development, including data preparation and preprocessing (Section 5.1), the evaluation of the prediction models’ performance (Section 5.2), and their subsequent deployment (Section 5.3). Finally, Section 6 provides the concluding remarks, summarizing the study’s key findings and contributions.

2. Materials and Methods

Figure 1 presents the architecture of the proposed methodology. The designed experiment begins with the first step, ‘data collection’ where the participants are asked to choose three or more bus stops from nine pre-selected bus stops located in downtown Montréal, QC, Canada and measure noise levels at each chosen bus stop. For this purpose, they were asked to install a smartphone application, NIOSH version 1.2.5, to measure the noise levels. They then completed a survey and answered some questions about their perceptions. Moreover, traffic data for the selected locations were downloaded, and all the collected data were integrated by using CityGML version 3.0. In the next step, ‘data integration’, all the collected data were stored and linked in 3DcityDB 4.4.1. The expected noise level at the studied locations was calculated by using CadnaA. In step 3, ‘model training’, participants’ acoustic comfort levels were modelled as functions of their social characteristics and the sensory data collected. Finally, in the last step, ‘deployment’, the trained model was applied to new locations in Montréal that were not used for model training.

The training of the comfort model was performed to provide an outdoor acoustic comfort estimator system. In this regard, three different categories of variables were considered as influencing factors in calculating acoustic comfort: physical, physiological, and psychological variables (Figure 2a). Our selection of variables was informed by prior research [27] and our available dataset. As actual physiological and psychological parameters were inaccessible, we opted for proxy variables as a suitable alternative. These variables are equally reliable and do not compromise the integrity of the study. Here, the acoustic comfort prediction model is described as a transformation that maps the three categories of variables to the acoustic comfort level. This transformation is denoted by f(.).

The input variables of f(.) are categorized into three types: physical variables ($a_1$, $a_2$,…, $a_n$), physiological variables ($b_1$, $b_2$,…, $b_m$), and psychological variables ($c_1$, $c_2$,…, $c_l$), where n, m, and l represent the number of physical, physiological, and psychological variables, respectively. For the physical variables, CadnaA software can simulate noise levels by using data such as traffic volume, speed limit, gradient, and surface material of the street (g(.)). Hence, considering that the effect of physical variables is known and independent of other variables, the function f(.) can be reformulated as shown below:

$$f\left(a_1,\dots ,a_n,b_1,\dots ,b_m,c_1,\dots ,c_l\right)=h\left(g\left(a_1,\dots ,a_n\right),b_1,\dots ,b_m,c_1,\dots ,c_l\right)$$

(1)

where

$$\widehat{h}(.,\Theta )\cong h(.)$$

(2)

and h(.) represents the result of partial decomposition of the function f(.). Based on the above equation, estimating f(.) is equivalent to determining a parametric function structure $\widehat{h}(.,\Theta )$, as an approximation of h(.), and estimating the vector of parameters, denoted by $\theta$. To train the model and optimize the parameters $\theta$, we use an error minimization approach. Specifically, we minimize the mean square error (MSE) between the actual acoustic comfort levels reported in the survey (${Comfort}_{actual,i}$) and the predicted comfort levels generated by the model (${Comfort}_{predicted,i}$). This error function $E\left(\theta \right)$ is defined as

$$E\left(\theta \right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N{({Comfort}_{actual,i}-{Comfort}_{predicted,i})}^2$$

(3)

where $N$ is the number of data points. This MSE metric quantifies the discrepancy between the observed and predicted values, guiding the model training process to find the optimal parameter vector $\theta$ that minimizes the error.

In Figure 2b, the input data and actual output data are needed to estimate the function $\widehat{h}(.,\Theta )$. For the input data, the physiological and psychological variables were collected from the survey, while physical variables were measured by the cellphone microphone by using the smartphone application ‘NIOSH’. For the actual and estimated output data, the collected acoustic comfort levels from the surveys and the output of the function $\widehat{h}(.,\Theta )$ are considered the actual and estimated values of the acoustic comfort level, respectively. By subtracting these values, the error in estimating acoustic comfort levels is calculated and sent to the parameter estimation algorithm. Here, three different regression models—KNN, SVR, and XGBoost—are used as the function $\widehat{h}(.,\Theta )$ and corresponding learning algorithms for parameter estimation.

3. Implementing the Reference Architectures

3.1. Data Collection Campaign

For data collection, a campaign was conducted to evaluate acoustic comfort levels at selected bus stops. The bus stops were located within 1 km of Concordia University Campus in downtown Montréal, Canada. These locations were intentionally chosen to cover streets with varying characteristics such as speed limit, gradient, and the number of people using the bus stations.

In the initial step of this campaign, participants registered by completing a registration form that gathered personal information, including age, gender, level of education, and income. It is important to note that participants had the option to choose “not to say” for the questions related to gender, age group, and income. While no participants opted for “not to say” for the gender or age group, one participant selected this option for income and was therefore excluded from the analyses. Then, the participants visited the pre-selected bus stops and completed a questionnaire on their perceptions, covering aspects such as their expectations, convenience, loudness, pleasantness, appropriateness, and their acoustic comfort level. (Notably, we reviewed the existing literature and carefully selected questions likely to capture individuals’ acoustic comfort in urban settings. In this study, the acoustic comfort level was measured by using a 5-level Likert scale, where level 1 indicates the lowest and level 5 shows the highest.) Participants also documented environmental parameters, including survey location, survey time, noise source, and noise type. The questionnaire is included in Supplementary Materials.

After completing the survey, participants were instructed to use the NIOSH mobile application on their iPhones to measure the noise level for approximately two minutes. As the app was only available on iPhones, all measurements were performed by using the iPhone microphone, which maintains a similar level of accuracy (within ±2 dB) [15]. While the NIOSH app’s accuracy is somewhat lower than that of professional sound level meters, it offers an acceptable level of precision given the study’s context and the need for convenience and accessibility across multiple participants. We selected the NIOSH app because of its ease of use, accessibility, and relatively consistent performance, making it a pragmatic choice for large-scale data collection where professional equipment was not feasible for all users [28].

Following each measurement, participants saved the sound level and submitted the results along with the completed questionnaire. This process enabled the storage the noise level, physiological and psychological variables, and corresponding acoustic comfort level in the database. Information on physical characteristics, such as vehicle count, speed limit, and slope, was obtained from the ‘Partenariat Données Québec’ website [15].

3.2. Feeding the GIS

The Social ADE was recently developed and is used to store the physiological and psychological parameters of citizens and link this information to the other aspects of the built environment [21]. Similarly, the transportation package in the core of CityGML and the Noise ADE can help to store other parameters, such as traffic data and measured noise levels, respectively. Through these ADEs, all the collected parameters relevant to calculating acoustic comfort can be stored and interconnected.

3.3. Model Training

After storing the comprehensive dataset in CityGML, a hypothesis was formulated to investigate the correlation between acoustic comfort and social parameters. First, the correlation between physical parameters and noise level (LAeq) was studied to test this hypothesis. Then, the correlation between the noise level and the acoustic comfort level was calculated. Pearson correlation was applied in both experiments, and variables with a higher correlation level with the comfort level were selected as influencing features for predicting the comfort level. Furthermore, to determine which physiological and psychological attributes (all categorical variables) impact the acoustic comfort level, a Linear Mixed Model (LMM) [29,30] was used. Similarly, statistically significant variables were considered influencing features in predicting comfort.

After identifying the top influential physical, physiological, and psychological features, we attempted to predict the comfort level. One new method for predicting outdoor noise levels is Land Use Regression (LUR), which explores the statistical relationship between land use characteristics (including land use type, traffic density, and building height) and measured noise levels. However, in cases involving non-linear relationships, random forest algorithms have proven superior to LUR for modelling noise in five cities in Canada [31]. Hence, we tested various regression models and ultimately selected the three best-performing models with the lowest error.

Moreover, various parameters were adjusted to achieve the lowest prediction error. The first applied regression model was KNN, using 3 as the number of neighbors. The second regression model, SVR, included parameters such as C (regularization parameter) set to 1.0 and $\epsilon$ (insensitivity parameter) set to 0.1. The third regression model was XGBoost. To assess model performance on the test data, mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and Max error are reported. MSE provides an absolute measure of the goodness-of-fit, while RMSE is the square root of MSE. Unlike MSE, MAE calculates the average of the absolute prediction error. Also, to evaluate model performance in deployment, simulated noise levels were used instead of measured noise levels, and prediction error results with simulated noise levels were also calculated and reported.

4. Results

During the data collection process, the noise level and the participants’ overall impressions were recorded across nine pre-selected points of interest (POIs). Each participant independently chose three POIs to measure, resulting in an uneven distribution of measurements across locations.

Table 1. Number of measurements at different points of interests (POIs).

POIs	Bus Stops	Numbers of Measurements
POI 0	Guy/Maisonneuve	51
POI 1	Côte-des-Neiges/Docteur-Penfield (int. n-e)	20
POI 2	Côte-des-Neiges/Summerhill	32
POI 3	Atwater/Maisonneuve	7
POI 4	Atwater/Sainte-Catherine	20
POI 5	Montagne/Sainte-Catherine	43
POI 6	Mansfield/Sainte-Catherine	4
POI 7	Peel/Sainte-Catherine	42
POI 8	Saint-Mathieu/Sherbrooke	43

displays the number of measurements recorded at each POI during the experiment. Some bus stops had only a few single-digit measurements, while others had between 40 and 50 measurements. At the conclusion of the two-week data collection period, a total of 738 samples were gathered, encompassing noise level readings and responses to the corresponding questionnaire. However, the number of samples per location was not crucial for the analysis, as each sample was treated as an independent data point in the regression model. This allowed us to use each data point individually, ensuring that no single POI disproportionately influenced the results.

Figure 3a represents the distribution of the measured noise levels across different POIs. As seen, POI 0 (Guy/Maisonneuve) and POI 2 (Côte-des-Neiges/Summerhill) exhibited a wider range of noise levels (the difference between maximum and minimum) compared with the other POIs. This is due to additional noise sources at these locations, such as public activity and construction noise, affecting POI 0 and POI 2, respectively. Note that in this paper, noise levels are reported based on LAeq, which is the A-weighted equivalent continuous sound level representing a constant noise level that would produce the same amount of sound energy over a specific period [32]. Figure 3b provides the geographic location of each POI, along with a colour map indicating the average noise level at each site. The circle sizes in the figure reflect the number of measurements per POI, while darker colours signify higher noise levels.

Figure 4 represents patterns observed in the hourly measured sound levels across two different POIs. In busier locations, such as Guy/Maisonneuve, the measured noise level remained relatively stable without significant fluctuations. Conversely, at other sites, like Atwater/Saint Catrine, two notable peaks occurred at around 10 a.m. and 7 p.m.

Figure 5 represents the noise level measurements based on the weekday vs. weekend (a) and morning vs. evening (b) timeframes, with morning being defined as 7 a.m. to 12 p.m. and evening as 12 p.m. to 6 p.m. As shown, there was no significant difference in the average noise levels between these two groups. However, the variation in noise levels was greater during weekday evenings compared with weekday mornings. In future studies, incorporating heatmaps and temporal analysis will be explored to capture variations in noise levels and comfort ratings across different times of the day and peak hours. This enhancement will require a larger sample size and a data collection period exceeding two weeks to ensure adequate data coverage for meaningful temporal insights.

After measuring the noise levels in the studied areas, the noise levels were simulated in those areas based on the physical attributes (traffic data and the built environment) by using CadnaA (Computer-Aided Noise Abatement) [26], which can communicate with CityGML through import and export interfaces. The parameters used to estimate LAeq include vehicles per hour, speed limit, heavy-vehicle percentage, gradient, surface type, building height, temperature, and wind speed.

Table 2. Comparison of simulation results and measurements.

POIs	Bus Stops	Simulated LAeq (dBA)	Measured LAeq (dBA)	Difference Between Simulated and Measured LAeq (dBA)
POI 0	Guy/Maisonneuve	69.5	67.3	3.26%
POI 1	Côte-des-Neiges/Docteur-Penfield (int. n-e)	71.3	67.7	5.31%
POI 2	Côte-des-Neiges/Summerhill	72.4	69.6	4.02%
POI 3	Atwater/Maisonneuve	70.8	67.1	5.15%
POI 4	Atwater/Sainte-Catherine	70.5	68.3	3.22%
POI 5	Montagne/Sainte-Catherine	71.7	70.3	1.99%
POI 6	Mansfield/Sainte-Catherine	71.6	70.9	0.98%
POI 7	Peel/Sainte-Catherine	72.7	69	5.36%
POI 8	Saint-Mathieu/Sherbrooke	74.7	72.8	2.6%

compares the measured and simulated noise levels. For most POIs, the difference was under 3 dBA. Reasons considered for these differences include the following:

• CadnaA used 2020 data to predict noise levels for 2022.
• The measured noise levels included all possible noise sources, such as traffic, construction, and human activity, whereas the simulating software only considered traffic noise.
• CadnaA simulation does not account for road congestion effects.
• CadnaA simulation does not factor in the impact of traffic signals or other traffic control measures on flow.

Figure 6 presents the CadnaA simulation results as a noise map. In this simulation, the receivers at each selected POI were assumed to be stationary to ensure consistency across measurement points.

During the experiment, the results were collected and stored in the CityGML by using a 3dCitydB database. In this phase, when the participants filled out the registration form, their data were stored in the Social ADE::User package (Figure 7). Also, when participants completed the questionnaire about their feelings, specifically their acoustic comfort level, their responses were stored in Social ADE::User::UserPreference. Other information from the questionnaires, such as the participant’s expectations and experience, was stored in ADE::User::UserInspiration and ADE::User::UserMood, which are proposed in the new version of the Social ADE. Data about the built environment of the POIs and traffic data were stored in Transportation::Road [33] (Figure 8). Moreover, the measured noise level and the correspondent data were stored by using the Noise ADE [34] (Figure 9).

5. Model Development

5.1. Data Preparation and Preprocessing

Once all the collected data were stored in 3dCitydB, the correlations between physical variables and the measured noise level were analyzed. The analysis revealed that the numbers of cars, buses, and medium buses did not have a significant correlation with noise level. The p-value was used as an indicator of the strength of evidence against the null hypothesis (no correlation). A low p-value, generally below 0.05, suggests statistical significance for the observed correlation, reinforcing the presence of a meaningful relationship between variables.

The correlation between noise level and acoustic comfort level was calculated as −0.4302, with a p-value of 0.041, indicating a ‘moderate negative’ correlation between the two variables [35]. Two distinct statistical methods were employed to evaluate the correlation between categorical and ordinal features on the acoustic comfort level. The Linear Mixed Model (LMM) was utilized for categorical features, and the outcome of these analyses facilitated the selection of relevant features. Some variables were defined based on the survey that the participants filled out (

Table 3. Questions in the survey and their related variables.

Variables	Questions in the Survey to Measure Categorical Variables (Possible Answered Values)
Convenience Level	If you were asked to stay at this location for one hour from now, would you find the acoustic environment too noisy or acceptable?
Expectation Level	Do you prefer/expect the environment to be quieter, or it is OK? (Quieter/It is OK)
Appropriateness	How appropriate is the sound to the surrounding? (Rank between 1 to 5)
Being Temporary/Permanent	Is the main source of the noise temporary or permanent?
Being Pleasant	From the aspect of noise, on a scale of 1 to 5, how pleasant is it here?
Income Group	Which of the following income groups do you belong to?
Loudness	How loud is it here? (Rank between 1 to 5)
Noise sources	What is the main source of the noise?
Being Tired	Which of the following best explains your emotion when you were waiting at this location?
Being Confused	Which of the following best explains your emotion when you were waiting at this location?
Having Deadline	Have you had a deadline or exam recently, or will you have one in the near future? (Yes/No)
Age group	Which of the following age groups do you belong to?
Being Bored	Which of the following best explains your emotion when you were waiting at this location?
Education Category	Which of the following categories best describes you: (Undergraduate student/Graduate student)
Being Happy	Which of the following best explains your emotion when you were waiting at this location?

). Moreover,

Table 4. Effect of categorical variables on acoustic comfort level: Linear Mixed Model results.

Variables	Coef.	Std. Err.	z	p > |z|
Gender	−16.756	95138839.2	0	1
Age group	−3.100	38296914.932	0	1
Income Group	527.491	16869572.260	0	1
Convenience Level	−0.205	0.085	−2.421	0.015
Expectation Level	0.045	0.134	0.336	0.737
Being Temporary/Permanent	0.297	0.149	1.997	0.016
Noise sources	0.127	0.154	0.823	0.41
Being Tired	−0.056	0.59	−0.095	0.925
Being Confused	−0.072	0.622	−0.116	0.908
Having Deadline	0.071	0.206	0.342	0.732
Being Bored	0.084	0.589	0.142	0.887
Education Category	0.259	0.622	0.416	0.677
Being Happy	0.008	0.6	0.013	0.99
Appropriateness	0.252	0.062	4.096	0
Being Pleasant	0.197	0.082	2.417	0.016
Loudness	−0.205	0.085	−2.421	0.015

presents the outcomes of the categorical feature analysis, highlighting the selection of convenience and temporary/permanent nature (with a p-value of less than 0.005) as significant factors. It is also seen that some attributes, such as the age group, did not correlate significantly with acoustic comfort. The absence of a significant correlation between these variables and acoustic comfort could indicate that these variables cannot add much information to machine learning.

As seen in Table 4, the physiological parameters (Age, Gender, and Income) have a p-value of 1, indicating they provide limited information for inference and hypothesis testing. In practice, such high p-values suggest that the prediction model under evaluation may require additional refinement. Hence, a wrapper method for evaluation using forward selection [36] was applied for feature selection for these three features. Through this method, Income was identified as the most relevant physiological parameter and was included in the prediction models.

5.2. Prediction Models’ Performance

After identifying the most relevant features—Convenience, Appropriate, Being Temporary/Permanent, Being Pleasant, and Income—three regression models (KNN, SVR, and XGBoost) were trained to predict the acoustic comfort level.

Table 5. Comparison of regression results with selected features (Convenience, Loudness, Ap-propriate, Temporary/Permanent, Pleasant, and Income) and measured noise level, simulated noise level, and physical attributes as inputs.

		Using LAeq	Using Simulation LAeq	Using Physical Attributes
Machine Learning Model	Normalized Error	Selected Features, LAeq	Selected Features, LAeq	Selected Features, Car, Bus, Camions, and Medium Bus
KNN	NMSE	0.0181	0.0277	0.0307
	MAE	0.0890	0.1178	0.1198
	RMSE	0.1349	0.1666	0.1752
	Max error	0.4166	0.5000	0.6666
SVR	MSE	0.0146	0.0268	0.0358
	MAE	0.0979	0.1210	0.1300
	RMSE	0.1211	0.1639	0.1894
	Max error	0.3746	0.5640	0.9001
XGBoost	MSE	0.0205	0.0250	0.0396
	MAE	0.1030	0.1141	0.1098
	RMSE	0.1434	0.1583	0.1990
	Max Error	0.3923	0.4787	0.9999

compares the errors from these models across three different scenarios. In the first scenario, psychological attributes and measured noise levels were used as inputs, with the acoustic comfort level as the output.

In the second scenario, similar to the first, the simulated noise level was used as input instead of the measured noise level. Across all three regression models, the errors from the first and the second scenarios—where the measured and simulated noise levels were used as the input, respectively—were lower than those from the third scenario, in which the physical attributes alone were used as input.

5.3. Deployment

To create an acoustical comfort level map for the 256 bus stops in Montréal, data from the city’s noise map [37] were utilized alongside generated physiological and psychological attributes for 13,000 hypothetical travelers. By randomly assigning these hypothetical individuals to the bus stops, the noise map data and the generated traveler attributes were input to the pre-trained models from the previous section. As a result, the comfort levels of all hypothetical individuals were estimated for each bus stop. Here, two attributes, ‘gender’ and ‘having a deadline’, are selected as examples to show the acoustic comfort levels across various locations in Montréal (Figure 10 and Figure 11).

Figure 10 shows a difference in comfort levels between male and female commuters at approximately 25% of bus stops, suggesting that individuals of different genders may have varying sensitivity to noise, leading to distinct acoustic experiences in the same environment. This difference may be linked to differences in hearing thresholds, hormonal influences, and cognitive responses to sound stimuli. However, identifying the root cause of this and any other similar pattern determined by the analytics requires further investigation and is beyond the scope of this study. Still, in Montréal, individuals across all genders generally experience moderate levels of acoustic comfort at most bus stops, with fewer instances of extreme discomfort. These findings align with previous research; for example, a 2016 study in Vitoria-Gasteiz reported no significant gender differences in acoustic comfort in outdoor public spaces [38]. Such outcomes may be influenced by the city’s diverse population and a range of activities contributing to a lively, fluctuating soundscape, often resulting in noise levels that remain within a moderate comfort range. This dynamic acoustic environment may help maintain a balance between high and low levels of acoustic comfort.

Based on the data in Figure 11, acoustic comfort in Montréal is generally moderate for most individuals, with the presence of a deadline having minimal impact on perceived comfort. However, comfort levels do vary across different locations within the city. Specifically, about 25% of the locations studied displayed either high or low comfort levels, with 14 out of 100 locations showing high comfort and 19 out of 100 l reporting low comfort. These variations may stem from localized factors, such as differences in noise sources, urban design, traffic patterns, and other context-specific elements unique to each area. To make informed urban planning decisions that enhance acoustic comfort, it is essential to leverage comprehensive data—including CityGML information—and integrate them with demographic insights. By combining these data sources, urban planners can gain a deeper understanding of the city’s acoustic dynamics, which would allow them to create more effective strategies for improving acoustic comfort across various neighborhoods.

The results confirm that the proposed method for estimating comfort levels can enhance GISs, offering urban planners a dynamic and scalable tool to address the multifactorial nature of acoustic comfort. This tool recognizes the diverse experiences of citizens, accounting for variables such as age, gender, environmental factors, and architectural designs. The integration of this method provides a promising approach to optimize urban environments and better cater to the acoustic preferences and sensitivities of different demographic groups. Ultimately, it contributes to the creation of more inclusive and comfortable urban spaces.

6. Conclusions

Providing safe, sustainable, and comfortable urban spaces is essential to meeting citizens’ expectations and enhancing their quality of life. This study categorized the factors contributing to citizens’ comfort into three main groups: physical, physiological, and psychological. The primary objective was to examine the linkage between the physical aspects of the built environment and noise sources and the physiological and psychological parameters of individuals to predict outdoor acoustic comfort. An experiment was conducted to collect data to train a prediction model for acoustic comfort levels. In this experiment, all collected data from the questionnaires and the traffic data from the study locations were stored in 3DCityDB. Pearson correlation analysis and the Linear Mixed Model (LMM) were used to analyze numerical and categorical variables, respectively, to identify factors influencing outdoor acoustic comfort. The selected influencing factors (Convenience, Appropriate, Temporary/Permanent, and Pleasant) were then used to train acoustic comfort prediction models, achieving an acceptable error level in predictions.

The key contributions of this research are as follows:

• Integration of diverse data types: This study integrated actual noise measurements, simulated noise data, and social parameters of hypothetical travelers to comprehensively analyze the real-world impact of urban noise on citizens.
• Integrated model for acoustic comfort: A CityGML-based data model was developed, combining physical, physiological, and psychological parameters relevant to acoustic comfort. This facilitated the systematic storage, retrieval, and spatial–temporal analysis of acoustic comfort data, filling a gap in prior research.
• Identification of social and psychological influences: This study demonstrated that acoustic comfort is influenced by not only physical and environmental factors, such as traffic, but also social and psychological factors, including age, gender, and stressors. This highlights the importance of these parameters in managing urban comfort.
• Predictive modeling for acoustic comfort: Predictive models—KNN, SVR, and XGBoost—were developed to forecast acoustic comfort levels by using integrated data. These interpretable models provided insights into the relative influence of each factor, making them valuable for urban planning applications.
• Case study at urban bus stops: A case study at selected bus stops in Montréal demonstrated the application of the proposed model. This highlighted how the approach could assess acoustic comfort in real-world settings and inform actionable infrastructure improvements.

Beyond these contributions, this study provided digital tools and methods to facilitate large-scale experiments on outdoor comfort. Such experiments have traditionally been expensive and hence limited to small groups of participants. The proposed GIS-based method, however, can take advantage of crowdsourcing at large to create more inclusive comfort models.

This study has certain limitations that may affect the results. Specifically, we had to rely on proxies to represent the physiological and psychological characteristics of the subjects. Given the available data, we used ‘income’ and ‘education’ as proxies for physiological parameters, and ‘convenience’ and ‘appropriate’ to represent psychological parameters. However, we acknowledge that these proxies may not be entirely representative of the subject’s characteristics. Also, the numbers of participants, points of interest (POIs), and measurements were all limited. Additionally, the accuracy of the measurement devices (such as the microphone on mobile phones) is only reliable within the frequency range of human speech and therefore may not accurately measure environmental noise. Finally, the software used (NIOSH v.1.2.5) may have limitations that could affect the accuracy of the results. Additionally, the duration of each measurement was only 2 min, as opposed to the usual granularity of physical data such as traffic volume, which are typically available for at least 1 h.

Future studies can help resolve these limitations and also deploy the data schema and proposed methodology to scale data collection (on both subjective and objective aspects of the built environment) through crowdsourcing. Firstly, additional details shall be included in the data collection campaigns to better represent social, psychological, and physiological features of service users. Such campaigns can be readily scaled, using the technology developed in this work, to take the experimental research on urban-scale comfort to the next level. Additionally, features of the physical built environment (geometry, material, furniture, vegetation, etc.) and further data of the service layer (i.e., mobility-related features) can be integrated into a single GIS, using the schema presented here. Cross-analyses of social, physical, and service layers can then help to uncover causalities among infrastructure attributes and end-users’ comfort. Adopting the social input-infused GIS can be also taken beyond the transportation infrastructure and noise comfort function to include aspects such as thermal comfort, safety, and level of service (among others) by storing all layers of data in 3dCitydB as the CityGML database.