Environmental Justice in the 15-Minute City: Assessing Air Pollution Exposure Inequalities Through Machine Learning and Spatial Network Analysis

Highlights

What are the main findings?

• The study uncovers significant socioeconomic and racial disparities in air pollution exposure within 15-minute walking ranges, with lower-income and Black communities facing higher PM2.5 levels.
• Traditional exposure assessments often underestimate disparities by failing to account for residents’ daily mobility patterns and activity spaces.

What is the implication of the main finding?

• The findings emphasize the need for dynamic, accessibility-based environmental justice assessments that reflect real-world mobility and exposure patterns.
• The study calls for tailored urban planning strategies to address localized pollution inequalities and ensure equitable outcomes in 15-minute city initiatives.

Abstract

The intersection of environmental justice and urban accessibility presents a critical challenge in sustainable city planning. While the “15-minute city” concept has emerged as a prominent framework for promoting walkable neighborhoods, its implications for environmental exposure inequalities remain underexplored. This study introduces an innovative methodology for assessing air pollution exposure disparities within the context of 15-minute activity zones in New York City. By integrating street-level PM2.5 predictions with spatial network analysis, this research evaluates exposure patterns that more accurately reflect residents’ daily mobility experiences. The results reveal significant socioeconomic and racial disparities in air pollution exposure, with lower-income areas and Black communities experiencing consistently higher PM2.5 levels within their 15-minute walking ranges. A borough-level analysis further underscores the influence of localized urban development patterns and demographic distributions on environmental justice outcomes. A comparative analysis demonstrates that traditional census tract-based approaches may underestimate these disparities by failing to account for actual pedestrian mobility patterns. These findings highlight the necessity of integrating high-resolution environmental justice assessments into urban planning initiatives to foster more equitable and sustainable urban development.

1. Introduction

Urban air pollution continues to be a critical environmental and public health challenge, with growing evidence highlighting its disproportionate impact on various socioeconomic and racial groups [1]. While considerable progress has been made in understanding these disparities at larger geographic scales, there remains a lack of research into how pollution exposure impacts residents in their daily activity spaces [2]. This gap in knowledge is especially important as cities globally adopt new urban planning paradigms that emphasize walkability and local accessibility. Understanding how these planning strategies intersect with pollution exposure at the micro-scale is crucial for developing equitable and effective public health interventions.

The “15-minute city” concept, pioneered by Carlos Moreno [3], has become a significant framework in urban planning, promoting neighborhoods where residents can access essential services within a 15-minute walk or bike ride. This model has gained widespread global recognition, with cities such as Paris and Melbourne implementing variations of this approach to improve urban livability, sustainability, and resilience [3,4,5,6]. While the framework aims to reduce car dependency and foster social cohesion, it also provides a valuable lens for analyzing environmental exposure patterns within residents’ daily activity spaces [7,8,9].

However, the transformation toward 15-minute cities faces multiple environmental challenges that affect walkability beyond mere distance metrics. While air pollution—particularly fine particulate matter (PM2.5)—represents a significant health concern for pedestrians, other environmental factors, such as noise pollution from traffic and construction, and visual pollution from excessive signage and poor urban design, also impact the quality and desirability of walking experiences. These multiple forms of pollution can interact to create cumulative environmental burdens that disproportionately affect certain neighborhoods and demographic groups, potentially undermining the equitable implementation of the 15-minute city concept. Despite the importance of these multiple dimensions of pollution, this study focuses primarily on air pollution due to its well-documented health impacts and the methodological challenges in its street-level assessment.

Accurately assessing exposure to air pollution (PM2.5) poses significant technical and methodological challenges. Traditional air quality monitoring networks, which rely on sparsely distributed fixed stations, often lack the spatial resolution required for detailed street-level analysis [10]. This limitation has historically constrained environmental justice research to broader geographic scales, potentially obscuring critical variations in exposure patterns that occur at the micro-scale—precisely where individuals walk, commute, and spend their daily lives [11,12]. As a result, there is a pressing need for more granular approaches to capture the nuanced dynamics of pollution exposure within the context of urban mobility and daily activity spaces.

Recent advancements in machine learning, coupled with the growing availability of urban environmental data, have opened new avenues for addressing the research gap in micro-scale exposure analysis. These technological innovations have facilitated the development of high-resolution air pollution models capable of estimating pollutant concentrations at the street level [10,13,14,15]. However, their potential application to detailed exposure assessments—particularly within the context of accessible, walkable neighborhoods—remains underexplored. This represents a critical research gap, as integrating these advanced modeling approaches with urban mobility data could provide deeper insights into the ways in which pollution exposure varies across micro-scale activity spaces, thereby informing more equitable and targeted urban planning and public health interventions.

This study addresses these methodological challenges by introducing a novel framework that integrates street-level PM2.5 predictions with spatial network analysis to explore pollution exposure patterns within the 15-minute city concept. The framework is designed to achieve three primary objectives:

• Develop and validate a robust machine learning methodology for estimating high-resolution, street-level air pollution concentrations.
• Design a graph network to examine exposure patterns within 15-minute walking ranges, capturing variations in pollution levels at the hyper-local scale (daily activity zones).
• Investigate the relationship between these exposure patterns and demographic characteristics, with a specific focus on income and racial disparities.

Using New York City (NYC) as a case study, this study demonstrates that this integrated framework uncovers previously unrecognized patterns of exposure inequality within residents’ daily activity spaces. By bridging the gap between high-resolution pollution assessment and accessibility-based neighborhood analysis, this research advances both the environmental justice literature and urban planning practice. The findings offer critical insights for urban planners and policymakers to foster more equitable and sustainable urban development.

2. Literature Review

2.1. The 15-Minute City and Urban Environmental Quality

The 15-minute city concept, which gained prominence during the COVID-19 pandemic, represents a transformative shift in urban planning paradigms [3]. The model advocates for the development of polycentric urban environments where residents can access essential services—such as healthcare, education, commerce, and recreational facilities—within a 15-minute walk or bicycle ride [16]. Originally popularized in Paris, the concept has been embraced by cities worldwide, each adapting the framework to its unique urban contexts [3]. Empirical studies have highlighted the multifaceted benefits of this approach, including significant reductions in transportation-related carbon emissions, strengthened social cohesion, and increased economic resilience at the neighborhood level [17]. These findings underscore the potential of the 15-minute city model to address pressing urban challenges while promoting sustainable and inclusive urban development.

However, the implications of the 15-minute city framework for urban environmental quality assessment remain insufficiently explored. While Moreno et al. [3] emphasized that the viability of walkable neighborhoods hinges on favorable environmental conditions (e.g., air quality), this aspect has not been thoroughly examined. This gap is particularly critical, as emerging evidence shows that increased physical activity in areas with poor air quality may exacerbate adverse health outcomes, potentially undermining the model’s intended benefits [5]. Additionally, studies have highlighted that the spatial distribution of walkable amenities often mirrors historical patterns of environmental inequities, such as disparities in green space access and pollution exposure [18]. These findings raise critical questions about the equitable implementation of the 15-minute city framework, particularly in diverse urban contexts, where pre-existing environmental injustices may persist. Addressing these challenges is essential to ensure that the 15-minute city model not only promotes accessibility and sustainability but also advances environmental justice and public health.

2.2. Air Pollution Monitoring

Assessing urban air pollution at the street level poses significant methodological challenges that have historically limited both research advancements and policy applications. Conventional air quality monitoring networks, despite offering high-temporal resolution data, are constrained by low spatial density, typically deploying only one station per 100–300 km² in urban areas [10]. This spatial inadequacy has spurred the development of alternative approaches to assess air pollution, such as mobile monitoring campaigns [19,20] and the deployment of low-cost sensor networks [21,22]. However, these methods face their own challenges, including inconsistent temporal coverage, data quality issues, and accessibility barriers.

Compounding these challenges is the unequal distribution of air quality monitoring infrastructure, which mirrors broader inequities in urban resource allocation [23,24]. Studies have revealed that wealthier neighborhoods often benefit from dense networks of air quality sensors, while marginalized communities experience significant monitoring gaps—a type of smart city inequality with profound implications [25,26]. This disparity not only conceals the disproportionate pollution burden on vulnerable populations but also perpetuates a cycle where resource allocation decisions based on incomplete data reinforce existing environmental injustices. The ramifications of this imbalance extend into public advocacy and policy-making: communities with access to robust air quality data are better equipped to demand and secure interventions, while those without such data struggle to advocate effectively, even though they often endure higher pollution levels.

To address these disparities, a concerted effort is needed to democratize air quality monitoring and data accessibility. Expanding low-cost sensor networks in marginalized areas, supporting community-driven monitoring initiatives, and leveraging open data platforms can help close monitoring gaps. By ensuring equitable access to air quality data, cities can more effectively identify and mitigate pollution hotspots in marginalized communities, fostering more inclusive and effective urban air quality management strategies. This aligns with the broader need for inclusive smart city governance, as highlighted by Kolotouchkina et al. [23], who emphasized the importance of addressing digital inequalities and ensuring that vulnerable groups and low-income populations are not left behind in the digital urban transition.

2.3. Machine Learning Applications in Air Pollution Prediction

The application of machine learning to air quality prediction has undergone substantial advancements in recent years, transitioning from traditional statistical methods to advanced deep learning architectures. Neural networks, particularly Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), have shown exceptional promise in modeling the intricate dependencies that govern urban air pollution dynamics [27,28,29,30]. Research by Singh et al. and Wang et al. [31,32] has further demonstrated that ensemble methods, which integrate multiple learning algorithms, often outperform conventional dispersion models, especially in complex urban settings characterized by heterogeneous emission sources and microclimate conditions.

A critical factor influencing the performance of these complex approaches is the availability of high-quality, sufficiently large training datasets [33,34]. Recent studies have addressed this challenge through innovative strategies, such as transfer learning techniques [35] and the fusion of diverse data sources, including satellite imagery and traffic flow data [32,36]. These integrated approaches have enhanced the ability of models to capture the multifaceted drivers of air pollution at fine spatial and temporal resolutions.

Nevertheless, significant challenges persist, particularly in the validation of street-level predictions. The scarcity of high-resolution ground-truth measurements in urban areas complicates the evaluation of model accuracy and reliability. Addressing this limitation requires the development of robust validation frameworks, potentially leveraging emerging technologies such as low-cost sensor networks and mobile monitoring platforms. Overcoming these challenges is essential to ensure the practical application of machine learning-based air quality models for informing policy decisions and designing targeted mitigation strategies.

2.4. Environmental Justice in Urban Air Quality

Environmental justice research has consistently highlighted disparities in air pollution exposure across socioeconomic and racial groups, with recent studies uncovering increasingly complex and nuanced patterns of inequality [37]. Studies have indicated that improvements in overall air quality often fail to benefit all communities equally, with marginalized groups frequently experiencing slower or less pronounced reductions in pollution levels [26,38]. Jbaily et al. [1] demonstrated that targeted interventions are essential for ensuring the equitable protection of all people from environmental hazards.

The conventional approach to assessing environmental justice, which relies on aggregated data at the census-tract (CT) or zip-code level, has faced growing criticism for its inability to capture fine-scale variations in exposure. Recent research has underscored the limitations of these administrative boundaries, revealing significant intra-area variability in pollution exposure, particularly in urban environments characterized by complex street networks and heterogeneous building morphologies [39,40]. Additionally, studies incorporating dynamic accessibility metrics have suggested that traditional static residential exposure assessments may not fully capture environmental justice concerns. For instance, in walkable neighborhoods or areas with high mobility, daily activity patterns can lead to substantial variations in individual exposure levels, which are not captured by residential-based metrics alone [41,42].

These findings highlight the need for more granular and dynamic approaches to environmental justice analysis. Integrating high-resolution spatial data, mobility patterns, and activity-based exposure assessments can provide a more comprehensive understanding of pollution disparities and their drivers. Such advancements are critical for designing equitable air quality policies and ensuring that marginalized communities are not disproportionately burdened by urban air pollution.

2.5. Research Gaps and Challenges

A comprehensive review of the existing literature reveals several critical research gaps and methodological challenges that require further investigation:

• Existing air quality assessment methods often lack the spatial resolution required for detailed street-level analysis. This limitation impedes the accurate characterization of exposure patterns, particularly in walkable neighborhoods and urban environments characterized by complex micro-scale dynamics. The absence of high-resolution data is especially critical when evaluating environmental justice implications within emerging urban frameworks, such as the 15-minute city model, where localized pollution gradients can significantly influence equitable access to clean air.
• While machine learning techniques have shown considerable promise in air quality prediction, there is a critical need for robust, validated frameworks that integrate high-resolution pollution estimates with accessibility-based analyses. Current approaches often operate in isolation, failing to fully capture the complex interplay between pollution distribution, mobility patterns, and urban infrastructure. Developing integrated methodologies that synthesize these dimensions is essential for advancing both theoretical understanding and practical applications in urban environmental studies.
• Traditional environmental justice studies have been limited by the coarse resolution of air quality data, which may mask significant intra-neighborhood variations in exposure. This limitation is particularly pronounced when examining disparities across socioeconomic groups within accessible neighborhood ranges (15-minute city). There is a pressing need for fine-grained, dynamic assessments to better understand how pollution exposure intersects with factors such as income, race, and daily activity patterns, particularly in the context of equitable urban planning and policy-making.

These gaps underscore the need for innovative methodological approaches that integrate high-resolution pollution modeling, dynamic exposure assessments, and accessibility-based frameworks to advance urban air quality research and environmental justice analysis. By addressing these challenges, researchers can generate actionable insights to mitigate pollution disparities and foster healthier, more sustainable urban development.

3. Methodology

This study introduces a novel framework for evaluating environmental disparities in air pollution exposure through the lens of the 15-minute city concept (Figure 1). The methodology is structured around three key components: (1) data collection and preprocessing; (2) street-level air pollution estimation and 15-minute aggregation; and (3) disparity analysis across different demographic groups.

The framework initiates with the collection and preprocessing of diverse datasets, such as road network topology, air quality measurements from mobile sensing, and demographic information. These datasets are spatially and temporally harmonized to establish a consistent foundation for subsequent analysis.

The core of the framework is dedicated to generating high-resolution estimates of air pollution concentrations, a process that unfolds in three critical steps. First, the road network is represented as a graph structure, capturing both spatial connectivity and topological dynamics in urban environments. Second, a machine learning model is employed to leverage this graph-based representation to estimate street-level PM2.5 concentrations. Third, these high-resolution estimates are aggregated within 15-minute walking networks centered on each Census Block (CB), producing accessibility-based exposure metrics consistent with the principles of the 15-minute city concept.

The final component of the framework investigates environmental disparities by analyzing the relationship between estimated exposure levels and demographic characteristics. Multiple inequality metrics, including the Gini index and Coefficient of Variation (CoV), are used to quantify disparities across income levels and racial groups. The findings provide a new perspective on environmental inequality patterns, uncovering how disparities in air pollution exposure align with demographic factors in 15-minute neighborhoods and offering actionable insights for equitable urban planning and policy development.

3.1. Graph-Based Road Network Representation

Accurately assessing street-level air pollution exposure requires a framework that captures the nuanced relationship between spatial proximity and network connectivity in urban environments. Traditional buffer-based methods, which use straight-line (Euclidean) distances, oversimplify these dynamics [36,43]. Such approaches overlook two critical factors: (1) the complex dispersion of air pollution along street networks, influenced by traffic patterns, urban geometry, and microclimate conditions; and (2) the realistic movement of pedestrians navigating road networks in their daily activities. These limitations are particularly significant in the context of 15-minute cities, where defining exposure zones based on actual walking distances—rather than straight-line approximations—is essential to accurately reflect real-world accessibility. To address these challenges, this study adopts a graph-based road network representation that provides a more sophisticated and realistic model of urban spatial relationships. This approach not only enhances the precision of air pollution exposure assessment but also aligns with the principles of accessible, pedestrian-centered urban design.

As shown in Figure 2, this study constructs a mathematical graph $G=(V,E)$ to represent the urban road network, where each vertex $v\in V$ corresponds to the midpoint of a road segment and edges E represent the physical connections between adjacent segments. The transformation process is designed to ensure both topological precision and practical relevance for air pollution exposure analysis within the 15-minute city concept. It begins by dissolving the raw road network into a unified geometry and splitting it at intersections to create distinct, topologically consistent road segments. Each segment is represented as a vertex in the graph, preserving the network’s connectivity and spatial relationships. For any two connected road segments a and b, the edge distance $d_{ab}$ between their midpoints is calculated as

$$d_{ab}={\displaystyle \frac{s_a+s_b}{2}}$$

(1)

where $s_a$ and $s_b$ represent the lengths of the respective road segments and $d_{ab}$ serves as the edge weight in the graph. This distance metric is integral to the methodology, fulfilling two primary roles:

• It facilitates the identification of neighboring road segments, enabling the incorporation of broad spatial context into air pollution estimates.
• It provides a basis for computing realistic walking distances, which are essential for defining 15-minute exposure zones in alignment with the principles of accessible urban design.

By leveraging this graph-based representation, this study establishes a robust and spatially refined framework for analyzing air pollution exposure, reflecting both the physical structure of the road network and the movement patterns of pedestrians.

Figure 2. Example of graph representation.

Figure 2. Example of graph representation.

3.2. Graph-Based Machine Learning

Building upon the graph structure constructed in Section 3.1, this study develops a machine learning approach that leverages the network topology to estimate street-level air pollution concentrations. By integrating topological relationships with spatial dynamics, the method effectively captures the complex patterns of pollution dispersion while explicitly modeling spatial dependencies, which achieves high prediction accuracy and scalability across diverse urban environments [44].

The methodology begins with network-based feature engineering. For each road segment, a topological distance approach is used to identify influential neighboring segments based on network connectivity and spatial proximity. As illustrated in Algorithm 1, the process captures the structural and functional relationships within the street network, assigning greater influence to segments that are closer to the target road segment. This hierarchical design aligns with the physical principles of pollutant dispersion in urban environments, where the influence of emission sources typically weakens as distance increases through the street network.

Algorithm 1 Top R Road Segment Extraction

1: Input: 2: G: Road graph including V nodes and edge weights W 3: R: Number of road segments to extract 4: Procedure: 5: for $n=1$ to V do 6:      Extract nodes A connected to node n in G                               ▹ First level nodes 7:      Sort nodes A by edge weights W (ascending) 8:      Select $S=A[0:R]$                                                    ▹ Select up to R nodes from A 9:      while $\mathrm{len}\left(S\right)<R$ do                                                                ▹ Higher level nodes 10:          Extract nodes B connected to nodes in S          ▹ B excludes nodes in $\left\{n\right\}\cup S$ 11:          if B is empty then 12:               Break the loop                                         ▹ Exit if no new nodes are found 13:          end if 14:          Calculate the shortest path weights from n to each node in B 15:          Sort nodes B by their shortest path weights (ascending) 16:          Add top nodes from B to S 17:      end while 18: end for 19: Output: Extracted road segments S for each node n in G.

Following the identification of influential neighboring segments, the process constructs feature vectors that integrate both local characteristics and broader spatial dependencies. For each target segment a, a feature vector is constructed by combining the local features and aggregated features of its R neighboring segments:

$$F_a=[F_a^{\mathrm{local}},F_a^{\mathrm{R}-\mathrm{neighbor}}]$$

(2)

where $F_a$ represents the feature vector, $F_a^{\mathrm{local}}$ represents local features, and $F_a^{\mathrm{R}-\mathrm{neighbor}}$ represents features from R neighboring segments. Local features incorporate road-specific attributes such as road type and traffic volume, along with built environment characteristics such as building density and land use. R-neighbor features are constructed by aggregating information from influential neighboring segments. This network-based feature engineering process provides a comprehensive representation of the target segment’s characteristics and its relationship to the surrounding street network, enabling the model to effectively capture the complex patterns of pollutant dispersion in urban environments.

After graph-based feature construction, this study applies machine learning methods to estimate air pollution concentrations. LightGBM is selected as the primary machine learning algorithm based on extensive cross-validation testing, owing to its efficiency and effectiveness in handling high-dimensional and heterogeneous features in air pollution estimation. The modeling process begins with feature selection using Recursive Feature Elimination (RFE), which identifies the most significant predictors while reducing dimensionality to mitigate overfitting. Hyperparameter optimization is performed using a grid search approach, systematically balancing model complexity with predictive performance to achieve an optimal configuration. Model performance is rigorously evaluated using multiple metrics across all validation folds, including the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and R-squared ($R^2$). This comprehensive evaluation process ensures consistent performance across diverse urban contexts, enhancing the model’s generalizability and reliability [45,46].

This graph-based machine learning approach enables high-resolution estimation of air pollution concentrations across the street network, capturing the intricate spatial dependencies and topological dynamics inherent in urban environments. The robust framework not only provides accurate air pollution predictions but also establishes a solid foundation for subsequent accessibility-based exposure analysis, which is detailed in the following section.

3.3. Accessibility-Based Exposure Assessment

Building on the street-level air pollution estimation, this study develops a comprehensive methodology to assess air pollution exposure within the 15-minute walking range of each Census Block (CB). This approach provides a nuanced, neighborhood-level perspective on exposure patterns, capturing the air quality conditions residents are likely to encounter during their daily activities.

3.3.1. Walking Range Delineation

As shown in Figure 3, for each CB, this study implements a systematic process to identify accessible street segments within a 15-minute walking range. The process begins by identifying the CB centroid and locating its nearest road segment within the network. Using the graph structure developed in Section 3.1, this study then identifies all reachable street segments within a 15-minute walk, assuming a standard walking speed of 5 km/h. This network-based approach provides a more realistic representation of accessible areas compared to simple circular buffers, as it accounts for the actual connectivity and reachability of residents’ daily activities within street networks.

3.3.2. Exposure Value Calculation

For each CB’s walking range, this study calculates a length-weighted pollution concentration to represent the neighborhood exposure level. This weighted approach ensures that longer street segments contribute proportionally more to the overall exposure metric, which provides a more accurate reflection of real-world exposure. The accessibility-based exposure value (AEV) for each CB i is calculated as follows:

$$AE{V}_i={\displaystyle \frac{{\sum }_{j\in R_i}(P{M}_j\times L_j)}{{\sum }_{j\in R_i}{L}_j}}$$

(3)

where $P{M}_j$ represents the PM2.5 concentration of street segment j, $L_j$ is the length of segment j, and $R_i$ represents the set of all segments within the 15-minute walking range of CB i.

The AEV serves as a standardized metric to quantify average pollution exposure within a resident’s daily walking range, facilitating robust environmental justice analyses. By integrating high-resolution street-level pollution data with the 15-minute city concept, this approach reveals how urban design and mobility patterns contribute to air quality disparities across diverse demographic and geographic contexts. These findings emphasize the critical relationship between localized pollution exposure, public health, and social equity, providing a valuable foundation for targeted interventions and equitable policy development.

3.4. Environmental Inequality Assessment

To align with the spatial resolution of demographic data at the census-tract level, AEV values at the CB level are aggregated to the CT scale. This enables the evaluation of air pollution exposure disparities across demographic groups using two established statistical metrics: the Gini coefficient, which quantifies overall distributional disparity, and the Coefficient of Variation (CoV), which measures relative disparity. Both metrics are widely recognized in environmental inequality research and provide complementary insights for understanding exposure patterns [1,47,48]. By employing these measures, this study establishes a robust framework for quantifying disparities across socioeconomic and racial dimensions.

For a given exposure threshold $\tau$, the exposure rate $q_r$ for each demographic group r is calculated as the percentage of the population exposed to PM2.5 levels exceeding $\tau$:

$$q_r={\displaystyle \frac{{\sum }_{i\in r}{p}_i·I(P{M}_i>\tau )}{{\sum }_{i\in r}{p}_i}}\times 100$$

(4)

where $p_i$ represents the population in CT i belonging to group r, $P{M}_i$ denotes the corresponding PM2.5 concentration, and $I(·)$ is an indicator function that equals 1 when the condition is true and 0 otherwise.

To quantify environmental disparities, two distinct metrics are employed. The Gini coefficient, a widely used measure in environmental justice studies [47], captures distributional disparity among demographic groups by measuring the cumulative deviation from perfect equality in exposure rates. It is defined as

$$Gini=1-2{\int }_0^{1}L\left(F\right)dF$$

(5)

where $L\left(F\right)$ denotes the Lorenz curve [48], which plots the cumulative proportion of exposure (population-weighted q-values) against the cumulative proportion of the population, with demographic groups ordered by their exposure rates. The Gini coefficient ranges from 0, indicating perfect equality where all groups experience identical exposure rates, to 1, denoting maximum inequality.

To complement the Gini coefficient, the CoV is calculated to quantify relative disparity. The CoV, defined as the standard deviation of exposure rates divided by the mean, provides a standardized measure of “between-group” variance [1]. A higher CoV indicates greater relative disparity across demographic groups. It is defined as

$$CoV={\displaystyle \frac{\sqrt{\frac{1}{R}{\sum }_{r=1}^R{(q_r-\overline{q})}^2}}{\overline{q}}}$$

(6)

where R denotes the number of demographic groups and $\overline{q}$ represents the mean exposure rate across groups. By focusing on standardized variability, the CoV facilitates meaningful comparisons between groups, independent of absolute exposure levels.

This dual-metric approach is applied to analyze disparities across three primary demographic dimensions: income groups (per capita income categorized by percentile-derived intervals), poverty groups (poverty rate categorized by percentile-derived intervals), and racial groups (based on CT demographic statistics). This comprehensive assessment captures both the overall inequality patterns through the Gini coefficient and the relative variability of exposure levels through the CoV, providing a robust quantification of environmental disparities across demographic groups.

4. Case Study and Results

4.1. Data Collection

This study investigates environmental inequities within the 15-minute city framework, with a focus on NYC. The city’s diverse built environment, characterized by variations in development density, street configurations, and socioeconomic profiles, provides an ideal setting for assessing disparities in air pollution exposure across communities within daily activity ranges. This approach offers critical insights into the interplay between urban design, accessibility, and environmental justice, enabling informed decision-making in urban planning and policy development.

4.1.1. Street-Level Air Quality

The analysis was grounded in high-resolution PM2.5 data obtained from the MIT Senseable City Lab’s ‘City Scanner’ mobile sensing campaign [19]. Conducted between September and December 2021, the campaign employed five municipal vehicles equipped with calibrated air quality sensors that traversed NYC’s street network. These sensors collected geo-located PM2.5 measurements at five-second intervals, resulting in a comprehensive dataset of 515,917 unique observations across the city’s diverse urban landscape.

Data quality was ensured through a rigorous calibration protocol, which included two key phases: a four-week initial co-location with a reference-grade station and statistical adjustment using reference measurements. The calibrated measurements exhibited strong agreement with the regulatory monitoring station, achieving a Pearson correlation coefficient of 0.97 [19]. This high-quality dataset enabled a detailed examination of air pollution patterns at the street level, facilitating the investigation of exposure disparities within 15-minute walking networks.

4.1.2. Supplementary Datasets

To characterize urban morphological features that influence street-level air pollution patterns, multiple urban datasets were integrated across four key dimensions (

Table 1. Supplementary urban datasets.

Aspect	Dataset	Source
Traffic and Road Networks	Annual Average Daily Traffic (AADT)	New York State DOT (nysdottrafficdata.drakewell.com)
	Truck Routes	NYC Open Data (opendata.cityofnewyork.us)
	Bus Routes	NYC Open Data (opendata.cityofnewyork.us)
	Street Centerline	NYC Open Data (opendata.cityofnewyork.us)
	Road Networks	OpenStreetMap (openstreetmap.org)
Census and Built Form	Census of Population and Housing	U.S. Census Bureau (census.gov)
	PLUTO (Primary Land Use Tax Lot Output)	NYC Department of City Planning (nyc.gov)
POIs and	Energy and Water Data	NYC Open Data (opendata.cityofnewyork.us)
Permitted Emissions	CATS (Clean Air Tracking System) Permits	NYC Open Data (opendata.cityofnewyork.us)
	NYC Restaurant	NYC Open Data (opendata.cityofnewyork.us)
	Facilities	NYC Open Data (opendata.cityofnewyork.us)
	POIs	OpenStreetMap (openstreetmap.org)
Street View	Google Street View	Google Maps (developers.google.com/maps)

): traffic and road networks, census and built form, points of interest (POIs) and permitted emissions, and street-view imagery.

Traffic and road network data offer essential insights into emission sources and pollution dispersion pathways, while built-form datasets capture urban canyon effects that shape local air quality dynamics. POI and permitted emission data facilitate the identification of specific pollution sources and activity hubs within neighborhoods. Additionally, street-view imagery enhances these datasets by providing granular details on street-level features, such as vegetation density and building configurations, which influence localized air pollution patterns. The integration of these multifaceted datasets formed a robust foundation for the machine learning-based estimation of street-level air pollution and enabled a comprehensive analysis of exposure disparities within 15-minute walking networks.

4.2. Data Preprocessing

4.2.1. Road Graph Construction and Air Pollution Processing

Accurately assessing street-level air pollution exposure requires a framework that captures the nuanced relationship between spatial proximity and network connectivity in urban environments. To achieve this, a road graph was constructed from NYC’s street network through a systematic process, as introduced in Section 3.1. First, overlapping segments in the raw road network were dissolved, and roads were split at intersections to create distinct, topologically consistent segments as fundamental analysis units. These segments were then transformed into a graph structure, with each segment represented as a node and the edges denoting physical connectivity between adjacent segments. This graph served as the spatial framework for analyzing air pollution exposure, reflecting both the physical structure of the road network and the movement patterns of pedestrians.

The integration of air quality measurements into the road graph structure involved careful spatial averaging. Multiple sensor readings within each road segment were aggregated to create segment-level representations. Log transformation was applied to these measurements to address the non-normal distribution patterns typically observed in urban air pollution data, with transformations later reversed for result interpretation and visualization [19].

4.2.2. Feature Processing

The feature engineering process characterized 282 features across five major aspects, as summarized in

Table 2. Feature categories.

Aspect	Representative Features	Number
Traffic and Road Networks	AADT, Truck AADT, bus route	81
Census and Built Form	Building floor area, commercial floor area, population	35
POIs and Permitted Emissions	POIs (e.g., amenities, education), energy use by fuel type, facilities (e.g., transportation facilities, waste facilities)	136
Street View	Streetscape (road, sidewalk, building, vegetation, sky)	25
Meteorology	Temperature, dew point, humidity, wind speed, air pressure	5
Total		282

. The process began with missing data imputation, where continuous variables were filled using mean values, and categorical variables were imputed using mode values. All geospatial datasets were standardized to a unified coordinate system and integrated through spatial joins or unique identifiers. For each road segment, two types of features were created: buffer-based density metrics, which quantified both short- and long-range environmental influences, and distance-based metrics, which measured proximity to key urban features such as highways and industrial zones. This multi-dimensional approach effectively captured the diverse influences of urban form and function on air pollution patterns, ensuring a comprehensive and robust representation of the urban environment.

The feature categories are described as follows:

• Traffic and Road Network Features: A total of 81 features characterizing the transportation network were generated, including AADT, truck AADT, bus routes, truck routes, and road network characteristics. These features were processed using a multi-buffer approach (50 m, 100 m, 250 m, 500 m, and 1000 m) to capture both immediate and broader-scale traffic influences on air quality. Distance-based metrics from major transportation infrastructure, such as bus routes and truck corridors, were computed to quantify the impact of different emission sources.
• Census and Built-Form Features: The built environment is represented through 35 features incorporating multiple dimensions of urban development, such as building floor area, residential density, factory area, and open space. Population density measures were also incorporated to capture the spatial distribution of human activity and its influence on urban air quality.
• POI and Permitted Emission Features: The most extensive feature set (136 features) was developed to represent POIs and emission sources, reflecting the complex nature of urban activities and their impact on air quality. These features encompass educational facilities, permitted combustion sources, energy consumption patterns, restaurant density, transportation facilities, and waste management infrastructure. This comprehensive coverage ensured that both direct emission sources and proxy indicators of urban activity intensity were accounted for.
• Street-View Features: Through systematic processing of Google Street View images, 25 features characterizing the street-level environment were developed across five distance buffers. Each downloaded panoramic image was transformed into four 90-degree perspective views using equirectangular projection at a resolution of 1024 × 1024 pixels. The SegFormer model was used for semantic segmentation [49] to quantify five key environmental elements (road, sidewalk, building, vegetation, and sky), providing crucial information about street canyon geometry and urban greenery.
• Meteorological Features: Five meteorological features, including temperature, dew point, humidity, wind speed, and air pressure, were incorporated to account for atmospheric conditions that affect pollution dispersion. These features were temporally aligned with air quality measurements to ensure an accurate representation of weather-related effects.

This comprehensive feature set provided a robust foundation for capturing local characteristics that influence air pollution concentrations. As introduced in Section 3.2, to account for the spatial dynamics of pollutant dispersion, this study identified 10 neighboring segments for each road segment based on network connectivity and spatial proximity. By aggregating information from these influential neighbors, network-based features were constructed to model the hierarchical influence of emission sources and the complex spatial patterns of pollutant dispersion in urban environments.

4.3. Modeling Performance

The graph-based machine learning framework for street-level PM2.5 prediction was rigorously evaluated through a comprehensive model selection and validation pipeline. As outlined in Section 3.2, RFE was used to identify the optimal feature subset, ensuring the inclusion of the most relevant predictors while reducing dimensionality. Hyperparameter tuning was systematically conducted using grid search to optimize model configurations and enhance predictive performance. Additionally, 10-fold cross-validation was implemented to ensure robust and reliable performance assessment, mitigating the risk of overfitting and providing a more generalized evaluation of model accuracy.

The analysis revealed significant variations in model performance across different machine learning approaches, as detailed in

Table 3. Performance metrics of machine learning models.

Model Category	Algorithm	${\mathit{R}}^2$	MAE	RMSE
Basic Models	Decision Tree	0.243	0.161	0.220
	k-Nearest Neighbors	0.384	0.145	0.198
	Ridge Regression	0.415	0.145	0.194
Advanced Models	Deep Neural Network	0.434	0.143	0.190
	Support Vector Regression	0.543	0.129	0.171
	Gradient Boosting	0.583	0.123	0.163
Ensemble Methods	XGBoost	0.592	0.121	0.161
	Random Forest	0.625	0.117	0.155
	LightGBM †	0.647	0.113	0.150

^† Graph-based LightGBM, which was selected as the final model for air pollution estimation.

. The basic models exhibited limited capability in capturing the complex spatial dynamics of urban air pollution patterns. For instance, Decision Tree achieved the lowest performance ($R^2$ = 0.243), while k-Nearest Neighbors and Ridge Regression showed moderate enhancements, with $R^2$ values of 0.384 and 0.415, respectively.

The advanced algorithms demonstrated significantly improved performance, with the Deep Neural Network achieving an $R^2$ of 0.434, Support Vector Regression attaining an $R^2$ of 0.543, and Gradient Boosting reaching 0.583. The ensemble methods consistently outperformed other approaches, with XGBoost and Random Forest delivering robust results ($R^2$ = 0.592 and 0.625, respectively). The superior performance of the ensemble methods underscores their ability to effectively process heterogeneous inputs and capture the complex, multidimensional relationships inherent in air pollution modeling.

As shown in Table 3, the graph-based LightGBM model emerged as the top-performing approach, achieving the highest scores across all evaluation metrics ($R^2$ = 0.647, MAE = 0.113, and RMSE = 0.150). This represents an improvement over previous work by MIT’s Senseable City Lab [19], which achieved an $R^2$ of 0.62 on the same dataset. The model’s success stems from its ability to integrate local characteristics with network-based spatial dependencies, effectively capturing the complex patterns of pollutant dispersion in urban environments. Furthermore, LightGBM provides enhanced interpretability and transparency, offering urban planners and policymakers actionable insights for evidence-based pollution mitigation strategies. Given its superior accuracy and practical utility, LightGBM was selected as the final model for pollution estimation and exposure disparity analysis.

4.4. Distribution Analysis

4.4.1. Street-Level PM2.5 Patterns

Following the validation of the model’s performance, the optimized LightGBM framework was applied to estimate street-level PM2.5 concentrations across NYC from September to November. This analysis provided an unprecedented resolution of air pollution exposure, revealing detailed spatial patterns, as illustrated in Figure 4.

The model predictions showed PM2.5 concentrations ranging from 3.23 to 6.46 μg/m³, with a mean value of 4.55 μg/m³ across the street network. SHAP (SHapley Additive exPlanations) analysis revealed that both meteorological factors and urban morphology significantly influence street-level PM2.5 concentrations. Among urban features, transportation-related elements, such as bus lanes, truck routes, bus depots, railyards, and facilities like waste processing sites, were identified as key predictors of local air quality. Notably, truck routes were found to significantly influence PM2.5 levels, with elevated concentrations consistently observed along major transportation corridors, highlighting the impact of heavy vehicle traffic on local air pollution (Figure 4). These findings highlight the critical role of urban planning and transportation systems in shaping air quality patterns, underscoring the need for targeted strategies to mitigate pollution in high-traffic areas.

4.4.2. Network Analysis Within the 15-Minute City Framework

To contextualize street-level PM2.5 predictions within the 15-minute city framework, pollution estimates were aggregated using the length-weighted averaging method described in Section 3.3. This approach generated CB-level exposure metrics, capturing realistic pollution exposure patterns within residents’ daily activity spaces. Walking speed, a critical factor in the 15-minute city concept, determines the spatial activity zones and related exposure patterns. A sensitivity analysis, conducted by testing walking speeds ranging from 4.5 km/h to 5.5 km/h, revealed minimal variability in CB-level exposure metrics (with a variation of 0.0021), demonstrating that reasonable changes in speed do not significantly bias citywide exposure patterns. Based on these findings, a typical walking speed of 5 km/h was selected for this study, with the resulting exposure distribution illustrated in Figure 5.

The 15-minute exposure revealed pronounced spatial heterogeneity in PM2.5 concentrations across NYC’s boroughs. High-pollution hotspots, defined as areas with PM2.5 concentrations in the upper quartile, are predominantly clustered in specific urban zones, such as eastern Brooklyn, northwestern Queens, and northern Staten Island. Notably, the central-south Bronx exhibits exceptionally elevated pollution levels, likely attributable to the confluence of major transportation corridors and limited green space.

In contrast, several regions demonstrated consistently lower PM2.5 levels. Central Manhattan, despite its high urban intensity, maintains moderate pollution concentrations, potentially due to stringent traffic regulations and efficient urban planning. The northern Bronx, characterized by greater green space coverage, shows improved air quality, while a distinct north–south corridor in Brooklyn and southern Staten Island also exhibits lower pollution levels.

The observed pollution patterns exhibit significant spatial clustering, indicating that systematic, neighborhood-scale factors—such as proximity to transportation hubs and densely developed areas—play a critical role in shaping local air quality dynamics. This non-random distribution underscores the influence of built-environment characteristics on PM2.5 concentrations, with high-exposure zones disproportionately affecting specific communities within their 15-minute neighborhoods. Such spatial heterogeneity raises important environmental justice concerns, emphasizing the need for more inclusive decision-making in urban planning and policy development to foster sustainable environmental equity across communities.

5. Discussion

5.1. Environmental Justice Within the 15-Minute City Concept

This study examines exposure disparities across socioeconomic groups, with a focus on income and racial differences, using 15-minute exposure values aggregated from the CB level to the CT level. Income and race are analyzed because they are widely recognized in environmental justice research as key factors linked to unequal environmental outcomes and as critical for informing policies and interventions aimed at promoting equity and supporting vulnerable communities [1,19]. In alignment with the methodology detailed in Section 3.4, exposure patterns are evaluated using both distributional (Gini coefficient) and variability (CoV) metrics across multiple PM2.5 thresholds. This approach provides a robust framework for identifying and quantifying disparities in environmental exposure, offering insights into the intersection of socioeconomic status and environmental justice.

5.1.1. Income and Poverty Disparities

The analysis of income-based disparities reveals distinct patterns of environmental inequality. As illustrated in Figure 6, census tracts with per capita income exceeding USD 47,930 (highest income quintile) consistently exhibit the lowest exposure rates across all pollution thresholds. As shown in

Table 4. Income- and poverty-based inequality metrics across PM2.5 exposure thresholds.

Percentile	Threshold	${\mathrm{Gini}}_{\mathit{Income}}$	${\mathrm{CoV}}_{\mathit{Income}}$	${\mathrm{Gini}}_{\mathit{Poverty}}$	${\mathrm{CoV}}_{\mathit{Poverty}}$
50%	4.473	0.198	0.386	0.098	0.169
60%	4.559	0.218	0.417	0.103	0.171
70%	4.665	0.260	0.489	0.097	0.164
80%	4.776	0.267	0.512	0.116	0.232
90%	4.891	0.326	0.639	0.203	0.397

, the income-based Gini coefficients demonstrate a progressive increase from 0.198 at the 50th percentile threshold (PM2.5 > 4.473 μg/m³) to 0.326 at the 90th percentile threshold (PM2.5 > 4.893 μg/m³). This trend becomes particularly pronounced at the 90th percentile threshold, where a notable jump in inequality metrics is observed, with the Gini coefficient increasing to 0.326 and the CoV reaching 0.639.

The spatial analysis reveals distinct geographic patterns in these income-based disparities, as illustrated in Figure 7. High-income areas, marked by green circles, form distinct clusters of environmental privilege, particularly evident in certain neighborhoods of Manhattan and outer borough residential districts. In contrast, areas with the lowest income quintile bear disproportionate pollution burdens, creating identifiable hotspots of environmental vulnerability (marked by red circles) in parts of the southern Bronx and eastern Brooklyn.

The analysis of poverty-based patterns offers further insights into environmental disparities. As shown in Figure 8, census tracts with poverty rates below 3% exhibit significantly lower exposure rates compared to those with higher poverty concentrations. While income-based disparities increase progressively across pollution thresholds, the Gini coefficient and CoV for poverty-based disparity remain relatively stable between the 50th and 80th percentiles (Gini: 0.098 to 0.116; CoV: 0.169 to 0.232). A notable increase occurs at the 90th percentile (Gini: 0.203; CoV: 0.397), where CTs with poverty rates above 21% experience substantially higher exposure rates. This contrast is further supported by the spatial analysis in Figure 7, which reveals that low-poverty CTs consistently experience lower PM2.5 concentrations, whereas high-poverty CTs face elevated levels. These findings highlight the disproportionate impact of extreme pollution on high-poverty communities, emphasizing their increased vulnerability under severe pollution conditions.

The geographic clustering of exposure disparities highlights the uneven 15-minute activity environments experienced by different socioeconomic groups, with low-income communities facing disproportionately higher pollution levels in their daily living and activity spaces. These patterns reveal systemic inequities, as environmental burdens align closely with socioeconomic vulnerability, undermining the health and equity goals of the 15-minute city concept. Addressing these disparities requires urban planning strategies that integrate air quality improvements and equitable resource distribution, ensuring that the 15-minute city model promotes both accessibility and environmental justice for all residents.

5.1.2. Racial Disparities

The analysis identifies significant racial disparities in PM2.5 exposure that parallel the socioeconomic inequalities discussed above. As shown in Figure 9, Black communities experience the highest exposure rates across most pollution thresholds, while White communities exhibit substantially lower exposure rates. These disparities are particularly stark at higher pollution levels, where the exposure rates for Black communities are more than double those of White communities at the 80th and 90th percentile thresholds. Other racial groups, including Asian, Hawaiian, and Indian communities, generally exhibit intermediate exposure levels.

Quantitative measures of inequality, presented in

Table 5. Racial inequality metrics across PM2.5 exposure thresholds.

Percentile	Threshold	Gini	CoV
50%	4.473	0.243	0.128
60%	4.559	0.253	0.150
70%	4.665	0.285	0.197
80%	4.776	0.306	0.259
90%	4.891	0.328	0.331

, further underscore these racial disparities. Both the Gini coefficient and CoV exhibit progressive increases from the 50th to 90th percentile thresholds (Gini: 0.243 to 0.328; CoV: 0.128 to 0.331), indicating that racial inequalities in PM2.5 exposure become more pronounced as pollution levels rise.

The spatial analysis, as illustrated in Figure 10, reveals distinct patterns of racial segregation in pollution exposure. Census tracts with White population concentrations exceeding the citywide average (43%) consistently demonstrate better air quality. In contrast, census tracts where Black populations surpass the citywide average (26%) systematically experience higher pollution levels. To explore the potential drivers of these disparities, this study employs Cohen’s d to examine the relationship between urban features (averaged within a 15-minute walking distance) and racial composition at the CT level. The results indicate that Black-dominant CTs are more likely to be located near pollution sources, such as bus depots (Cohen’s d = −0.54) and waste processing facilities (Cohen’s d = −0.35), while having limited access to green spaces and recreational areas (Cohen’s d = −0.24). In contrast, White-dominant CTs tend to be further from pollution sources and have better access to green spaces, as indicated by positive associations (Cohen’s d = 0.54, 0.29, and 0.30, respectively). These findings suggest that Black-dominant communities face disproportionate exposure to environmental hazards and limited access to mitigating resources, while White-dominant communities benefit from lower exposure and greater access to amenities that enhance quality of life. While these relationships provide valuable insights, they are exploratory and do not establish causality. Future research could build on this foundation to investigate causal mechanisms, such as historical zoning laws and urban planning policies, and inform targeted interventions to address systemic inequities.

5.2. Disparities Across Urban Environments

To examine 15-minute exposure disparities across urban environments, this study evaluates average exposure levels and inequality metrics across NYC’s five boroughs. As shown in

Table 6. Population-weighted average PM2.5 concentrations by borough.

Borough	PM2.5 (μg/m3)
Manhattan	4.31
Bronx	4.48
Brooklyn	4.50
Queens	4.54
Staten Island	4.55
NYC	4.47

, population-weighted average PM2.5 concentrations demonstrate significant spatial variability, ranging from 4.31 μg/m³ in Manhattan to 4.55 μg/m³ in Staten Island, with a citywide average of 4.47 μg/m³. These results underscore the spatial heterogeneity in air quality exposure within residents’ daily activity spaces, offering critical insights into how environmental inequalities manifest under the 15-minute city framework across diverse urban contexts.

The distribution of environmental disparities across boroughs, as depicted in Figure 11, illustrates a complex interplay of socioeconomic and racial inequalities within distinct urban contexts. Manhattan, despite having the lowest average PM2.5 concentration, exhibits the highest income-based disparity (Gini: 0.49) and significant poverty-based inequality (Gini: 0.22). This apparent paradox suggests that while Manhattan benefits from superior overall air quality, these benefits are not equitably distributed among its residents.

In contrast, the Bronx demonstrates a distinct pattern, characterized by pronounced income-based disparities (CoV: 0.58) but relatively lower Gini coefficients across all three aspects. Brooklyn and Queens display more moderate disparities across all measures, although their average PM2.5 concentrations (4.50 and 4.54 μg/m³, respectively) exceed the citywide average. Staten Island presents a unique profile, with high racial disparity (Gini: 0.58) despite relatively consistent income- and poverty-based metrics. This suggests that racial segregation may play a particularly significant role in shaping environmental exposure patterns in this borough.

These borough-level variations highlight the critical influence of localized urban development patterns and demographic distributions on environmental justice outcomes within the 15-minute city framework. The findings reveal the necessity for policy interventions to be specifically designed to address the unique environmental inequality patterns characteristic of each borough, accounting for the distinct socioeconomic and racial dynamics that drive exposure disparities. Implementing such context-sensitive strategies is vital for advancing equitable environmental outcomes and ensuring that the benefits of urban living are more fairly distributed across diverse communities.

5.3. Comparative Analysis of Environmental Disparities Across Walking Ranges

Building on the identification of environmental disparities within the 15-minute city framework, this study conducts a sensitivity analysis to examine how these disparities vary across different walking distances. Accessibility-based exposure patterns are evaluated and compared across 5-, 10-, and 15-minute walking ranges, with the traditional CT-based averaging method serving as the baseline for comparison.

The results reveal a systematic increase in disparity metrics as the walking range expands, as illustrated in Figure 12. Income-based disparities exhibit the most pronounced trend, with the Gini coefficient rising from 0.177 (traditional method) to 0.180 (5-minute range), 0.187 (10-minute range), and 0.198 (15-minute range). The corresponding CoV values show even more significant increases, climbing from 0.327 to 0.386 across the same range progression. Comparable patterns, albeit with varying magnitudes, are observed for poverty- and race-based disparities, indicating a consistent trend across different socioeconomic dimensions.

The spatial coverage of different walking distances is further illustrated in Figure 13, using a census tract in Brooklyn as an example. The figure highlights how expanding walking ranges progressively encompass larger portions of the urban environment. While the traditional CT method (panel b) is confined to road segments within administrative boundaries, the walking-based approaches (panels c–e) incorporate increasingly broader areas based on actual pedestrian accessibility. This expansion captures a more diverse set of urban environments, including residential neighborhoods, major transportation corridors, and pollution hotspots such as busy intersections and industrial zones.

The growing disparities observed as the walking range expands can be attributed to several factors. First, larger walking ranges encompass a more heterogeneous mix of urban environments, reflecting the diverse exposures residents encounter during daily activities. Second, extended ranges are more likely to include pollution hotspots and other environmental stressors that may not be captured within smaller administrative units. These findings underscore a critical limitation of traditional CT-based analyses, which may underestimate environmental inequalities by failing to account for the spatial extent of residents’ actual activity spaces. In contrast, walking-based approaches provide a more nuanced understanding of exposure patterns, aligning more closely with the lived experiences of urban populations. This analysis highlights the importance of incorporating pedestrian accessibility into environmental equity assessments, particularly in the context of the 15-minute city model.

6. Conclusions

This study advances the understanding of environmental justice within the 15-minute city framework by introducing a novel methodology for assessing air pollution exposure inequalities. By integrating graph-based machine learning with spatial network analysis, this research demonstrates that environmental disparities are more pronounced when evaluated within the context of residents’ daily walking ranges.

The findings reveal significant socioeconomic and racial disparities in PM2.5 exposure across NYC neighborhoods. Census tracts in the highest income quintile consistently exhibit lower exposure rates, with income-based inequality metrics showing progressive increases at higher pollution thresholds. Racial disparities are also evident, with Black communities experiencing disproportionately higher exposure rates, particularly at elevated pollution levels. These disparities exhibit distinct spatial patterns, as the borough-level analysis uncovers complex relationships between urban development and environmental inequality.

The methodological contribution of this study lies in demonstrating that traditional exposure analyses may underestimate environmental inequalities by not accounting for exposure patterns within residents’ daily activity spaces. The comparative analysis reveals that disparity metrics systematically increase as the walking range expands, indicating that accessibility-based exposure assessment offers a more comprehensive understanding of environmental justice implications in urban settings.

The findings carry significant implications for urban planning and policy development. First, they underscore the necessity of incorporating environmental justice metrics into 15-minute city initiatives to ensure that improvements in neighborhood accessibility do not inadvertently reinforce existing exposure disparities. Second, they highlight the value of high-resolution environmental monitoring and analysis in identifying specific areas requiring targeted intervention. Third, they suggest that effective policy responses may need to be tailored to address the distinct patterns of environmental inequality characteristic of different urban contexts.

This study has several limitations that present opportunities for future research. First, reliance on municipal vehicles for mobile sensing may introduce spatial and temporal biases, as fixed routes and schedules can result in the underrepresentation of certain areas and times, potentially limiting the comprehensive capture of pollution variations across the city. Additionally, the dataset’s limited temporal coverage (autumn, fixed hours) restricts the analysis of how exposure inequalities vary across different times (e.g., rush hours vs. nighttime) and seasons. To address these limitations, future work will integrate complementary data sources, such as longer-term monitoring campaigns, satellite-based pollution estimates, and alternative mobile platforms (e.g., drones or bikes), to enhance spatial and temporal coverage. Advanced data fusion techniques will also be explored to harmonize these heterogeneous datasets and improve the robustness of pollution modeling.

Furthermore, future studies will examine the temporal and seasonal dynamics of exposure inequalities to better understand how disparities evolve over time. Additionally, further investigation of alternative environmental inequality metrics, such as Theil’s Index and the Atkinson Index, will be conducted to provide deeper insights and a more comprehensive understanding of disparities, particularly in terms of between-group and within-group inequalities. These efforts will also focus on vulnerable populations (e.g., the elderly) and areas with poor infrastructure, aiming to inform targeted decision-making and support sustainable urban development strategies. While the specific findings of this study are tied to NYC’s unique urban context, the methodological framework is designed to be transferable. Future research will explore its application in diverse urban settings to assess generalizability and refine the approach for varied environments.

Declaration of Generative AI and AI-Assisted Technologies in the Writing Process

During the preparation of this work, the authors used Claude to improve readability and language. After using this tool, the authors reviewed and edited the content as needed. The authors take full responsibility for the content of this publication.