Personalised route guidance to reduce air pollution exposure during active commuting: a proof-of-concept study using graph-based network analysis in Leicester, UK

1 Introduction

Physical inactivity is a leading modifiable risk factor for non-communicable disease globally, contributing to an estimated 3.2 million deaths annually (World Health Organization, 2022). Active commuting, which is defined as the use of non-motorised, human-powered transport modes such as walking and cycling to work or a place of education, represents an effective strategy for integrating habitual physical activity into daily routines without requiring dedicated leisure time (Strain et al., 2020). Epidemiological studies consistently demonstrate that regular active commuting is associated with reduced risks of cardiovascular disease, type 2 diabetes, and all-cause mortality (Dinu et al., 2019; Hamer and Chida, 2008).

However, the health calculus of active commuting is complicated by the concurrent exposure to traffic-related air pollutants (TRAPs). Pedestrians and cyclists travelling along busy arterial roads may inhale substantially higher concentrations of PM₂.₅ and NO₂ than those using low-traffic alternatives, owing to proximity to vehicle exhaust emissions and reduced wind dispersion at street level (Cole-Hunter et al., 2017). Moreover, the elevated ventilation rates associated with physical exertion during active commuting significantly increase the inhaled dose per unit time compared with passive transport modes (Int Panis et al., 2010). This creates a tension between the cardiovascular and metabolic benefits of active travel and the potential respiratory and systemic harms associated with elevated pollutant inhalation.

Modelling studies suggest that for individuals commuting through highly polluted urban corridors, the net health benefit of cycling may be partially attenuated under high ambient PM₂.₅ concentrations, though net benefit is typically preserved for the majority of commuters in typical European city conditions (Tainio et al., 2021; Mueller et al., 2015). Whilst in-car concentrations of NO₂ typically exceed those experienced by pedestrians and cyclists, as shown by Panchal et al. (2022) who measured in-car NO₂ concentrations 25–30 μg m⁻³ higher than those recorded on paired cycling commutes across four routes in Leicester, the elevated ventilation rates during physical exertion mean that active commuters can nonetheless inhale substantially higher total pollutant doses than car occupants (Cepeda et al., 2017; Int Panis et al., 2010). Notably, Panchal et al. (2022) also observed that PM₂.₅ concentrations were slightly higher during active commutes than in-car across all routes studied, providing direct empirical corroboration for the spatial uniformity of PM₂.₅ relative to NO₂ that underpins the pollutant-specific findings of the present study. Strategies that preserve the physical activity benefits of active commuting while minimising pollutant exposure are therefore of considerable public health interest. Route selection is one such modifiable behavioural lever: by diverting walkers and cyclists away from major traffic corridors onto lower-trafficked streets, parks, and shared-use paths, cumulative pollutant exposure may be meaningfully reduced without prohibitive increases in journey time or distance. Empirical evidence supports the feasibility of this approach: Mitsakou et al. (2021) reviewed evidence from across Europe and found that cyclists on routes separated from motorised traffic experienced over three times lower pollutant concentrations than those cycling on-carriageway, and that exposure could be more than four times higher on high-traffic routes than on dedicated separated lanes. The importance of street-level NO₂ gradients for active commuters is further illustrated by natural experiment evidence from London, where temporary traffic restrictions on Oxford Street were associated with marked reductions in roadside NO₂ concentrations (Walton et al., 2015), underscoring the potential for routing interventions to meaningfully reduce exposure in contexts with strong spatial pollution gradients.

A growing body of computational work has sought to operationalise this route-selection lever. The dominant approach represents the street network as a weighted graph in which conventional distance or time costs are augmented or replaced by cumulative pollutant exposure, with least-cost routes then identified using shortest-path algorithms such as Dijkstra’s (Mahajan et al., 2019; Sinnott and Zhong, 2023; Wang et al., 2022). Reported reductions have been consistent but generally modest: graph-based optimisation has achieved mean PM₂.₅ exposure reductions of approximately 11–17% for small distance penalties in Taipei and Melbourne (Mahajan et al., 2019; Sinnott and Zhong, 2023), with larger reductions (up to approximately 32%) attainable only at substantial detour where cleaner routes are markedly separated from motorised traffic (Wu et al., 2021; Qiu et al., 2019). Two features of this literature are salient for the present study. First, it is overwhelmingly focused on PM₂.₅, which is comparatively straightforward to map from low-cost and mobile sensors; NO₂-specific route optimisation remains sparse (Gao et al., 2022), despite NO₂ being more tightly bound to traffic corridors and therefore, in principle, more modifiable through route choice than the spatially more uniform, background-dominated PM₂.₅ (Ramacher and Karl, 2020). Second, many analyses rely on annual-average or otherwise temporally static concentration surfaces that flatten the pronounced diurnal variation in traffic-related pollution (Willberg et al., 2023), and few pollution-aware routing tools have progressed to deployment or been evaluated for real-world exposure reduction (Helle et al., 2023; Rafiepourgatabi et al., 2021).

High-resolution urban air quality modelling platforms such as EarthSense MappAir (EarthSense, Leicester, UK), which generates street-level NO₂ and PM₂.₅ concentration maps at 10 m spatial resolution, combined with open-source graph-based road network analysis tools (e.g., OSMnx; Boeing, 2017) now makes it computationally feasible to overlay pollution data directly onto pedestrian and cycling network graphs, enabling pollution-weighted route optimisation at the city scale. Despite this technical capability, few studies have systematically quantified the magnitude of exposure reduction achievable across a representative set of urban commuting corridors using this approach or explored how these reductions vary across times of day, travel modes, and route types.

This proof-of-concept study addresses this gap by applying a graph-based pollution-weighted routing algorithm to a set of common active commuting origin–destination pairs in a medium-sized English city (Leicester, UK) with heterogeneous air quality, a mix of arterial and residential street networks, and an established active travel policy agenda. We quantify potential reductions in PM₂.₅ and NO₂ exposure for both pedestrians and cyclists across multiple times of day and characterise the trade-off between exposure reduction and route detour, providing an empirical basis for future implementation of personalised pollution-aware navigation tools. In doing so, the present study advances prior pollution-aware routing work in four respects: it evaluates NO₂ alongside PM₂.₅ within a single formulation, isolating the pollutant most modifiable through route choice; it uses time-resolved (diurnal) rather than temporally static concentration surfaces; it treats pedestrian and cycling networks separately; and it characterises the exposure–detour trade-off across a representative multi-corridor sample rather than a single route.

2 Methods

2.1 Study area

The study area is the City of Leicester, United Kingdom, a medium-sized English city with a population of approximately 370,000. Leicester presents a useful case study for active commuting research owing to its compact urban form, dense residential suburbs, and concentration of major employment and education sites within a relatively small geographic area. The study domain was defined by a bounding box in WGS84 coordinates spanning longitudes −1.217° to −1.040° and latitudes 52.58° to 52.69° (approximately 12 km east–west by 12 km north–south), capturing the urban core and the principal surrounding residential areas.

2.2 Road network acquisition

Pedestrian and cycling road networks for Leicester were obtained from OpenStreetMap using the OSMnx Python library (Boeing, 2017), using the walk and bike network types, respectively. In graph-theoretic terms, a road network is represented as a directed graph comprising nodes (street intersections and dead-ends) and edges (the street segments connecting adjacent nodes), each carrying attributes including length, road classification, and, in the present study, modelled pollutant concentrations. All spatial operations were carried out in UTM EPSG:32630. Only the largest strongly connected component of each directed graph was retained to ensure full origin–destination reachability. Edge travel times were computed assuming pedestrian and cycling speeds of 5 km h⁻¹ and 15 km h^{−1, respectively.} The resulting networks comprised 22,535 nodes and 59,330 directed edges for walking with a mean length of 50.4 m (median 34.2 m, IQR 16.6–64.1 m), and 17,835 nodes and 41,896 directed edges for cycling (mean 58.8 m, median 41.1 m, IQR 19.3–74.1 m). In both networks, the great majority of edges were short: 88% (walk) and 85% (cycle) were under 100 m, and fewer than 0.3% exceeded 500 m.

2.3 Air quality data

Pollution concentration data were obtained from the EarthSense MappAir near-real-time modelled product, a machine-learning-derived system fusing fixed-sensor networks, mobile sensing, satellite retrievals, and land-use regression to produce spatially continuous concentration estimates at 10 m resolution across UK cities. PM₂.₅ and NO₂ concentrations (μg m⁻³) were retrieved via the EarthSense Web Coverage Service API as GeoTIFF rasters for four diurnal time windows: Morning Peak (07:00–09:00), Midday (11:00–13:00), Evening Peak (16:00–18:00), and Off-Peak (20:00–22:00). These windows were selected to capture systematic diurnal variation in traffic-related emissions (Table 1). Concentrations were assigned to road network edges by sampling each raster at the geometric midpoint of every edge; the small number of edges falling outside the raster extent were filled by averaging valid concentrations from immediately adjacent edges. This procedure yielded eight pollution attributes per edge (PM₂.₅ and NO₂ for each time window) with 100% edge coverage across both networks.

Table 1. Algorithm performance across blending parameter (α, where α ∈ [0, 1]; α = 1 recovers pure shortest-distance routing, α = 0 minimises pollution irrespective of distance) sensitivity analyses.

2.3.1 Air quality data validation

To assess the plausibility of the MappAir concentration fields, modelled NO2 and PM2.5 values were extracted at the 10 m grid cells containing the two Leicester AURN reference monitors (Leicester University, urban background; Leicester A594 Roadside, St Matthews Way) and compared with hourly measurements matched to the four diurnal windows. Over a continuous 14-day window (2026-06-19 to 2026-07-02), modelled and measured concentrations agreed well for NO2 (n = 206, r = 0.69, RMSE = 7.8 μg m⁻³, mean bias = +2.7 μg m⁻³, NMB = 0.18) and PM2.5 (n = 189, r = 0.79, RMSE = 2.3 μg m⁻³, mean bias = +1.0 μg m⁻³, NMB = 0.21). For the eight snapshot hours used in the routing analysis, agreement was r = 0.71 (RMSE = 10.4 μg m⁻³) for NO2 and r = 0.88 (RMSE = 6.3 μg m⁻³) for PM2.5 (n = 16 site-hours each) (see Supplementary Figure S2). Critically for the routing application, MappAir reproduced the roadside-minus-background NO2 gradient between the two sites: over the extended window the measured gradient was 11.8 μg m⁻³ against a modelled 9.5 μg m⁻³ (81% of the measured value), with the modelled gradient positive in 92% of paired hours. Across the eight snapshot hours the modelled gradient remained positive but was attenuated (1.3 vs. 15.1 μg m⁻³ measured), indicating that the spatial contrast exploited by the routing is directionally correct and, if anything, conservative in magnitude. AURN data for 2026 are provisional (not yet ratified); all comparisons were made in UTC at matched hours.

2.4 Origin–destination pairs

Twenty representative commuting corridors were defined to reflect a range of home-to-work and home-to-education journeys across Leicester’s principal residential suburbs. Origins were located in the following suburbs: Oadby, Evington, Knighton, Stoneygate, Clarendon Park, Aylestone, Eyres Monsell, Braunstone Town, Westcotes, New Parks, Belgrave, Beaumont Leys, Rushey Mead, and Humberstone. Destinations comprised the main employment and education clusters: Leicester City Centre, University of Leicester, De Montfort University, Leicester Royal Infirmary, and Leicester General Hospital. The 20 pairs were selected to provide spatial coverage across all cardinal directions relative to the city centre and to include corridors of varying length, road type composition, and expected pollution exposure.

Each origin and destination was snapped to its nearest graph node using ox.nearest_nodes. Any OD pair for which the origin and destination snapped to the same graph node was automatically excluded from the analysis (none were excluded in practice). The resulting node indices were stored alongside each pair definition and reused across all routing experiments, ensuring consistency between walk and bike analyses.

2.5 Pollution-weighted routing

For each routing computation, a composite edge cost was computed that trades off physical travel distance against cumulative pollution exposure. In brief, the algorithm proceeded in the following steps outlined in Box 1. Min–max normalisation was chosen because it directly amplifies relative spatial gradients in concentration regardless of the absolute pollutant level, making the routing sensitive to intra-urban variation rather than to the magnitude of concentrations relative to guideline values. The shortest-path route used the OSM length attribute as the edge weight. The pollution-minimising route used the pollution_cost attribute computed by the procedure described above. Normalisation and cost assignment were recomputed freshly for each combination of time window and α-value, since the global min and max concentrations may differ across time windows.

Three values of α were evaluated: 0.1, 0.3, and 0.5. The primary analysis used α = 0.3 (weighting pollution avoidance at 70% and distance at 30%). We emphasise that α is a preference parameter rather than an empirically estimated quantity; in the absence of an established basis for any single value, we selected an interior value giving a moderate pollution preference and report outcomes across α ∈ {0.1, 0.3, 0.5} as a sensitivity analysis. In a deployed tool, α would be exposed as a user-tunable setting. The values α = 0.1 and α = 0.5 bracket this primary value to characterise how route choice and exposure outcomes respond to different priorities.

Pollution-weighted routing algorithm: procedural summary

Step 1. Retrieve edge-level PM₂.₅ and NO₂ concentration values for the selected time window and transport mode network.

Step 2. Min–max normalise concentrations across all graph edges, separately for each pollutant normₓ(e) = (cₓ(e) − cₓ,min)/(cₓ,max − cₓ,min), where x ∈ {PM₂.₅, NO₂}.

Step 3. Compute a length-weighted pollution score per edge pollution_score(e) = (W_pm × norm_pm(e) + W_no2 × norm_no2 (e)) × ℓ(e), where W_pm = W_no2 = 1.0 (equal weighting) and ℓ(e) is edge length (m).

Step 4. Form a composite edge cost as a convex combination of distance and pollution score cost(e) = α × ℓ(e) + (1 − α) × pollution_score(e), where α ∈ [0, 1]. Setting α = 1 recovers the shortest-distance route; α = 0 minimises pollution irrespective of distance.

Step 5. Run Dijkstra’s algorithm (nx.shortest_path) on cost(e) to generate the pollution-minimising route; run on ℓ(e) alone for the shortest-path comparator.

Step 6. Compute mean route concentration (MRC), estimated inhaled dose (ID), cumulative exposure (CE), time-weighted exposure (TWE), and detour statistics on both routes (Section 2.6).

Steps 1–6 were repeated for each combination of time window, α value, and weight scheme, since min/max concentrations differ across time windows. Primary analysis: α = 0.3 (equal weights); sensitivity analyses: α = 0.1 and α = 0.5.

2.6 Outcome metrics

The following metrics were computed for both the shortest-path and pollution-minimising routes on each OD pair.

• Mean Route Concentration (MRC, μg m⁻³): The length-weighted mean pollutant concentration along the route: MRC = Σ(c_e × ℓ_e)/Σ(ℓ_e). MRC is independent of route length, enabling direct comparison between routes of different distances. Reported separately for PM₂.₅ and NO₂.
• Estimated Inhaled Dose (ID, μg): The mass of pollutant inhaled over the journey, following Borghi et al. (2020) and Panchal et al. (2022): ID = MRC × T × VE, where T is journey time (min) and VE is the mode-specific ventilation rate. Ventilation rates were taken from Zuurbier et al. (2009): VE = 0.023 m³ min⁻¹ (walking) and 0.024 m³ min⁻¹ (cycling). Because VE is constant within a mode, ΔID (%) differs from ΔMRC (%) only when the two routes have different travel times. Because VE enters multiplicatively and is identical on both routes for a given individual, it cancels in the percentage dose reduction; the reported ΔID (%) is therefore invariant to the absolute VE value and to between-individual variation in ventilation, which rescales only the absolute inhaled mass. This holds under the constant-speed assumption; route-specific terrain effects on ventilation are not modelled (see Section 4.3).
• Supplementary metrics: Cumulative exposure (CE = Σ(c_e × ℓ_e), μg m⁻³ m) and time-weighted exposure (TWE = Σ(c_e × t_e), μg m⁻³ s) were retained for completeness but are not the primary reported outcomes.
• Route detour: Additional distance and travel time of the pollution-minimising route relative to the shortest-path route, expressed as absolute values and as percentage of the shortest-path distance. For each metric, the percentage reduction achieved by pollution-minimising routing was calculated as Δmetric (%) = (metric_shortest − metric_pollmin)/metric_shortest × 100.

Results are summarised as median and interquartile range, grouped by transport mode and time window, for the primary α = 0.3 configuration.

2.7 Experimental design

The full experimental matrix crossed 20 OD pairs, 2 transport modes (walk, bike), 4 time windows (morning, midday, evening, off-peak), and 3 α values (0.1, 0.3, 0.5). For each of these 480 configurations, both a shortest-path route and a pollution-minimising route were computed and evaluated on all metrics described in Section 2.6, yielding 960 evaluated routes in total.

All analyses were conducted in Python using the OSMnx, NetworkX, rasterio, pyproj, NumPy, and pandas libraries. The dataset is available to explore as an online digital tool https://jamesad90.github.io/leicester-pollution-routing/.

3 Results

3.1 Road network characteristics

The Leicester walking network comprised a strongly connected graph covering the full city extent, with route lengths across the 20 OD pairs ranging from 1,578 m to 5,342 m (median 3,539 m). The cycling network yielded comparable OD pair distances (range: 1,653–5,644 m; median 3,625 m), reflecting the similar spatial extents of accessible pedestrian and cycling infrastructure. The network encompassed a broad range of street classifications, from high-traffic primary and trunk roads concentrated in the city centre and along major radial corridors, to residential streets and off-road paths predominating in peripheral areas (Figure 1).

3.2 Pollution concentration distribution across the network

Modelled NO₂ concentrations during the morning peak exhibited marked spatial heterogeneity across the Leicester network, with the highest values concentrated along primary arterial routes converging on the city centre (median MRC 50.2 μg/m³ across OD pairs; Figure 1). Concentrations declined substantially towards residential and off-road edges in suburban areas, with values as low as approximately 30–35 μg/m³ in eastern and southern peripheral zones.

In contrast, PM₂.₅ concentrations showed substantially less spatial variability across the network, with no strong differentiation between arterial and residential edges. Time-of-day multipliers produced the expected diurnal pattern: concentrations were highest during the evening peak (16:00–18:00) and morning peak (07:00–09:00), with midday concentrations at baseline and off-peak concentrations approximately 35% lower for NO₂ and 30% lower for PM₂.₅.

3.3 Route characteristics and algorithm performance

At the primary blending parameter (α = 0.3), pollution-minimising routes diverged from the shortest-path route for 129 of 160 route combinations (81%), generating a unique alternative in the majority of OD pair × mode × time window combinations. Sensitivity analyses demonstrated a clear dose–response relationship between pollution-prioritisation and route divergence: at α = 0.1, 92% of routes (148/160) differed from shortest-path; at α = 0.5, this fell to 62% (99/160), as the cost function increasingly favoured distance minimisation (Table 1).

3.4 NO₂ exposure reductions

Pollution-minimising routing achieved substantively larger reductions in NO₂ than PM₂.₅, reflecting the stronger spatial gradients in NO₂ across the Leicester network. At α = 0.3, NO₂ MRC reductions were most pronounced during the evening peak for both modes (Table 2; Figure 2). For cycling, median NO₂ MRC reductions were 3.4% (IQR: 0.2–19.6%) at midday and 4.4% (IQR: 0.0–27.3%) during the evening peak, with maximum reductions of 48.3 and 59.9%, respectively. Walking routes achieved median MRC reductions of 3.0% (IQR: 0.2–14.0%) at midday and 5.9% (IQR: 1.7–23.1%) during the evening peak, with maxima of 48.4 and 48.5%. Estimated inhaled NO₂ dose reductions were modestly smaller: evening peak ID reductions were 3.2% (IQR: 1.1–17.8%; max 44.4%) for walking and 3.2% (IQR: 0.0–23.3%; max 54.9%) for cycling.

Table 2. Median (IQR) percentage reductions in NO₂ mean route concentration (MRC) exposure and route detour by mode and time window (α = 0.3; n = 20 OD pairs per cell).

Morning peak and off-peak reductions were more modest. During morning peak hours, median NO₂ MRC reductions were 1.8% for cycling (IQR: 0.0–13.9%; max 36.6%) and 1.6% for walking (IQR: 0.1–9.4%; max 31.4%). During off-peak hours, median MRC reductions were 0.5% for cycling and 0.7% for walking, consistent with lower absolute NO₂ concentrations and reduced spatial gradients at that time of day.

ID reductions followed the same directional pattern as MRC reductions across all conditions but were uniformly smaller, as the longer travel times incurred on pollution-minimising routes increase the inhaled dose independently of concentration (Table 2). The largest individual NO₂ reductions were observed for routes connecting Clarendon Park and Belgrave to the city centre during midday and evening peak hours (Figure 3).

3.5 PM₂.₅ exposure

In contrast to NO₂, pollution-minimising routing produced modest and directionally mixed changes in PM₂.₅ MRC. The pooled median PM₂.₅ MRC reduction was 1.0% at α = 0.3, with wide IQRs spanning negative values in several conditions (Figure 4). The largest positive median MRC reductions were observed for walking during the evening peak (11.5%) and off-peak (3.5%). Estimated inhaled PM₂.₅ dose reductions were near-zero across all conditions (pooled median 0.07%), consistent with the reduced spatial variability of PM₂.₅ relative to NO₂.

3.6 Detour analysis

Route detours associated with pollution-minimising routing were modest across the primary analysis. At α = 0.3, the overall median detour was 0.94% of shortest-path distance (IQR: 0.0–4.3%). Walking routes had a median detour of 0.8% (IQR: 0.0–3.2%); cycling routes had a median detour of 1.0% (IQR: 0.0–4.5%). Several OD pairs achieved NO₂ MRC reductions exceeding 20% at detours of under 5%, demonstrating that substantial exposure benefits are available in specific corridors with modest additional distance.

The trade-off between detour and NO₂ reduction exhibited a positive and approximately linear relationship for NO₂ (Figure 5), confirming that larger reductions were generally available when the algorithm was permitted to take longer alternative routes. In contrast, the detour–PM₂.₅ scatter showed no systematic positive relationship and clustered near zero, consistent with the reduced spatial variability of PM₂.₅ relative to NO₂ (Figure 6). Notably, several OD pairs achieved NO₂ reductions of 5–10% with detours of under 2%, suggesting that substantial pollution gains are available in certain corridors with minimal time or distance penalty (Figure 3).

As α decreased from 0.5 to 0.1, route divergence increased from 62 to 92% of combinations, while maximum NO₂ reductions remained stable at 54.9% across all α values, confirming that the most beneficial routes are identified consistently and that pollution prioritisation primarily increases divergence frequency rather than maximum achievable reduction (Figure 5). A sensitivity analysis comparing equal weighting of PM₂.₅ and NO₂ against a NO₂-only cost function showed that whilst 45% of individual routes differed between schemes at α = 0.3, aggregate outcomes were virtually identical: median NO₂ MRC reductions of 2.89 and 2.30% respectively, and median inhaled NO₂ dose reductions of 1.92% vs. 1.91% (Supplementary Figure S1).

3.7 Air quality data validation

Over a continuous 14-day window (2026-06-19 to 2026-07-02), modelled and measured concentrations agreed well for NO2 (n = 206, r = 0.69, RMSE = 7.8 μg m⁻³, mean bias = +2.7 μg m⁻³, NMB = 0.18) and PM2.5 (n = 189, r = 0.79, RMSE = 2.3 μg m⁻³, mean bias = +1.0 μg m⁻³, NMB = 0.21). For the eight snapshot hours used in the routing analysis, agreement was r = 0.71 (RMSE = 10.4 μg m⁻³) for NO2 and r = 0.88 (RMSE = 6.3 μg m⁻³) for PM2.5 (n = 16 site-hours each). Critically for the routing application, MappAir reproduced the roadside-minus-background NO2 gradient between the two sites: over the extended window the measured gradient was 11.8 μg m⁻³ against a modelled 9.5 μg m⁻³ (81% of the measured value), with the modelled gradient positive in 92% of paired hours. Across the eight snapshot hours the modelled gradient remained positive but was attenuated (1.3 vs. 15.1 μg m⁻³ measured), indicating that the spatial contrast exploited by the routing is directionally correct and, if anything, conservative in magnitude. AURN data for 2026 are provisional (not yet ratified); all comparisons were made in UTC at matched hours.

4 Discussion

4.1 Principal findings

This proof-of-concept study demonstrates that graph-based pollution-weighted routing, applied to a real urban road network with high-resolution modelled air quality data, is technically feasible and yields meaningful reductions in NO₂ exposure for active commuters across a range of Leicester city corridors. At the primary blending parameter (α = 0.3), pollution-minimising routes diverged from shortest-path routes for 81% of OD pair × mode × time window combinations. NO₂ MRC reductions were most pronounced during the evening peak, reaching a median of 5.9% for walking and 4.4% for cycling, with maximum MRC reductions of 48.5 and 59.9% respectively, and individual OD pairs achieving MRC reductions of 10–25% at route detours of under 5%. A clear and consistent finding was the asymmetry between NO₂ and PM₂.₅: while NO₂ exposure was substantially reducible through route selection, PM₂.₅ exposure showed negligible and directionally inconsistent changes, reflecting the reduced spatial variability of PM₂.₅ relative to NO₂ in Leicester’s pollution landscape. Route detours were modest across the primary analysis (median 0.94%), suggesting that meaningful NO₂ MRC reductions are achievable without prohibitive increases in journey time or distance for the majority of commuters.

4.2 Comparison with existing literature

Our findings are broadly consistent with a growing body of evidence demonstrating the exposure reduction potential of route selection for active commuters. Cole-Hunter et al. (2017) demonstrated significantly lower PM₂.₅ concentrations on bicycle routes using low-traffic streets compared with major arterials in Copenhagen and Barcelona, with reductions of 20–40% reported in some corridors. De Nazelle et al. (2012) showed that cyclists travelling on designated cycle routes experienced substantially lower black carbon and PM₂.₅ exposures than those using main roads, and Mitsakou et al. (2021) confirmed in a comprehensive UK evidence review that separation from motorised traffic was the single most effective route-level determinant of active commuter exposure; a factor not directly modelled in the present study, but implicitly captured through the MappAir concentration surfaces, which are themselves a function of traffic volume and proximity. At the population scale, Hankey et al. (2017) demonstrated using land use regression modelling in Minneapolis that a substantial proportion of active commuter journeys traverse blocks with simultaneously high traffic volumes and high particulate concentrations, underscoring the public health case for pollution-sensitive routing. The most directly comparable computational work to ours is the route-planning study by Wang et al. (2022), conducted in Cangzhou, China using taxi-based mobile monitoring data, which reported that low-exposure routing could reduce cumulative pollutant exposure by up to 16.8% compared with shortest-path routes, substantially below our maximum NO₂ reductions of up to 59.9% MRC (54.9% inhaled dose), likely reflecting the more pronounced street-level NO₂ gradients characteristic of Leicester’s mixed arterial and residential network. The more modest PM₂.₅ reductions we observed compared with prior empirical work likely reflects the reduced spatial variability of PM₂.₅ relative to NO₂ in Leicester’s pollution landscape, where regional background PM₂.₅ contributes proportionally more to total concentrations than in cities with higher local primary emission densities.

The use of EarthSense MappAir data in this study represents an advance over prior work, which has typically relied on fixed-site monitoring data or coarser national modelled surfaces (e.g., LAQM grids) that may inadequately capture within-street and between-street variability in pollutant concentrations. The approximately 10 m spatial resolution of MappAir enables edge-level differentiation of pollution exposure across the road network, making it particularly well-suited to identifying low-pollution alternative routes within a small spatial footprint.

The time-of-day sensitivity analysis revealed that NO₂ exposure reductions were greatest during midday and evening peak windows, when spatial gradients in NO₂ concentrations across the Leicester network were most pronounced, with comparatively modest reductions during the morning peak, consistent with the particularly strong evening traffic build-up characteristic of mixed-use UK urban centres, and with the known diurnal profile of road transport NO₂ emissions, which peak during the evening commute period (DEFRA, 2022).

4.3 Strengths and limitations

This study has several notable strengths. First, the use of physics-informed, sensor-fused modelled air quality data from the EarthSense MappAir platform enhances the ecological validity of the exposure estimates; MappAir has been demonstrated to capture intra-urban spatial gradients at 10 m resolution across Leicester (Panchal et al., 2022), representing a finer spatial scale than conventional land-use regression or national modelled surfaces such as LAQM grids. Second, the application of separate walking and cycling network graphs acknowledges the fundamental differences in accessible infrastructure between modes, which influences both route geometry and exposure. Third, the systematic evaluation across multiple OD pairs and temporal windows provides more generalisable insights than single-corridor case studies.

Several limitations warrant acknowledgement. This is a computational proof-of-concept, and exposure estimates are modelled rather than empirically measured; personal monitors worn during actual commutes would be required to validate these predictions. The routing algorithm does not currently incorporate real-time pollution data, although integration with live air quality APIs would be feasible in a deployed application. Modelled concentrations from MappAir represent temporal averages and may not fully capture episodic pollution events (e.g., high-emission vehicle passes, local wood-burning). The analysis applies fixed population-average ventilation rates (walking: 0.023 m³ min⁻¹; cycling: 0.024 m³ min⁻¹; Zuurbier et al., 2009) for dose estimation; future work should incorporate personalised ventilation rates based on individual physiology, fitness, and real-time heart rate to refine inhaled dose estimates at the individual level. Finally, route acceptability from a user perspective, which would potentially include features such as route attractiveness, subjective safety, and surface quality, was not modelled and would need to be addressed in a subsequent user-facing implementation study.

4.4 Implications for practice and future research

The findings of this study have direct implications for the development of pollution-aware active travel navigation tools. Smartphone applications integrating real-time or near-real-time air quality data with active travel routing already exist in nascent form [e.g., the P-STEP platform developed at the University of Leicester (Worboys et al., 2025)] but have rarely been subject to rigorous computational validation of achievable exposure reductions. The present analysis provides an empirical basis for the design of routing algorithms within such tools and establishes a methodological framework for their evaluation.

Future work should: (i) validate modelled exposure estimates against personal monitor measurements obtained during real commuting trips along both shortest and pollution-minimised routes; (ii) conduct user acceptance testing to evaluate willingness to adopt longer or time-delayed routes in exchange for reduced pollution exposure; (iii) extend the approach to additional UK cities with contrasting network structures and pollution profiles; and (iv) incorporate personalised ventilation rate models, for example using real-time heart rate data, to refine inhaled dose estimates beyond the population-average approximations applied here.

5 Conclusion

This proof-of-concept study demonstrates that graph-based pollution-weighted routing using high-resolution urban air quality data is a technically feasible approach to reducing NO₂ exposure for active commuters in a UK city. At α = 0.3 with equal pollutant weighting, pollution-minimising routes diverged from shortest-path routes in 81% of OD pair × mode × time window combinations, achieving NO₂ MRC reductions of up to 59.9% (inhaled dose reductions up to 54.9%) with a median route detour of 0.94%. NO₂ emerges as the principal target pollutant for route-based exposure reduction in Leicester, with PM₂.₅ largely insensitive to route selection in terms of inhaled dose (pooled median 0.07% reduction) owing to its reduced spatial variability relative to NO₂ across the road network. These findings provide computational evidence base for integrating NO₂-sensitive, time-varying pollution routing into personalised active travel navigation tools, including mHealth platforms such as P-STEP (Worboys et al., 2025), as a scalable public health strategy to optimise the health benefits of active commuting in polluted urban environments.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsc.2026.1869860/full#supplementary-material

Sensitivity of routing outcomes to the pollutant weighting scheme. Box plots compare the three weighting schemes (Section 2.5.3) pooled across all transport modes and time windows at the primary α = 0.3 (n = 480 per scheme): (a) percentage reduction in cumulative PM₂.₅ exposure; (b) percentage reduction in cumulative NO₂ exposure; (c) percentage detour (extra distance of the pollution-minimising route relative to the shortest route). “equal” denotes the primary analysis (W_PM = W_NO₂ = 1.0); “no2_only” sets W_PM = 0, W_NO₂ = 1.0; “original” retains the prior specification W_PM = 1.0, W_NO₂ = 0.5. Boxes show median (central line), IQR (box), 1.5 × IQR whiskers, and outliers (open circles); the dashed horizontal line marks zero. The “equal” and “no2_only” schemes produce near-identical NO₂ reductions despite differing on individual route choices in 45% of cases, reflecting the strong spatial correlation of the two pollutants along the Leicester road network. Supplementary material at the source

Comparison of modelled MappAir concentrations with AURN observations at Leicester monitoring sites. Scatter plots show hourly modelled versus measured concentrations for (a) NO₂ and (b) PM₂.₅ at Leicester University (urban background) and Leicester A594 Roadside (St Matthews Way). Points are extracted from 10 m MappAir grid cells corresponding to monitor locations and matched to hourly AURN observations across four diurnal windows. Grey dashed line indicates the 1:1 line. Supplementary material at the source

Footnotes

Edited by: Sreekanth Vakacherla, WRI India, India

Reviewed by: Aniebietabasi Ackley, Victoria University of Wellington - Te Aro Campus, New Zealand

Xuanpeng Zhao, Zoox Inc, United States