Performance of road-traffic-based exposure proxies against personal PM2.5 measurements in three Sub-Saharan African countries Authors: Handsome Bongani Nyoni, Terence Darlington Mushore, Laura Munthali, Sibusisiwe Audrey Makhanya, Laurine Chikoko, Stanley Luchters, Matthew F Chersich, Fortunate Machingura, Liberty Makacha, Benjamin Barratt, Hiten D Mistry, Marie Laure Volvert, Peter von Dadelszen, Anna Roca, Umberto D’alessandro, Marleen Temmeran, Esperança Sevene, Tamara Rosemary Govindasamy, Prestige Tatenda Makanga, The PRECISE Network, The HEAT Center Published: 2026 ABSTRACT Particulate Matter (PM2.5) exposure contributes to the global disease burden, yet its monitoring remains sparse and uneven, with limited ground sensor infrastructure. Road-traffic proxy indicators can provide indirect estimates of PM2.5 where measurements are limited but require context-specific validation. We evaluated three PM2.5 road-traffic-related proxies: (i) population-Weighted Road Network Density (wRND), (ii) Euclidean (straight line) distance from highways (EH), and (iii) Euclidean distance from main roads (EM). We validated these proxies using high-resolution outdoor filtered PM2.5 personal exposure measurements collected over 1 year from 343 postpartum participants in The Gambia, Kenya, and Mozambique. Proxy-PM2.5 associations were assessed using Spearman correlation, and predictive utility was tested using country-specific and global Random Forest (RF) models (3-fold cross-validation), reporting R2, RMSE, and feature importance. Spatial mapping showed heterogeneous proxy–PM2.5 relationships across and within sites, with elevated PM2.5 occurring in both low- and high-proxy contexts. wRND–PM2.5 correlations were weak overall and statistically significant only in Mozambique (r = 0.351; p = 0.005), with non-significant associations in Kenya (r = − 0.041; p = 0.673) and The Gambia (r = − 0.020; p = 0.909). EH–PM2.5 correlations were positive in The Gambia (r = 0.335; p = 0.053) and Mozambique (r = 0.292; p = 0.020) but negative and significant in Kenya (r = − 0.224; p = 0.018). Single-variable RF models performed poorly across all countries (R2 < 0.45) and the Global model (R2 = 0.42). Combining proxies improved performance in Kenya (R2 = 0.52; RMSE = 31.7 µg/m3) and Mozambique (R2 = 0.60; RMSE = 8.9 µg/m3), Global R2 = 0.46; RMSE = 29.1 µg/m3), although in The Gambia, the combined model (R2 = 0.53; RMSE = 37.6 µg/m3) did not exceed the best single-proxy model. Road-network proxies provided limited but context-dependent signals of personal PM₂.₅ exposure. Their performance varied substantially across countries, indicating that road-based indicators should not be used as stand-alone exposure measures in heterogeneous sub-Saharan African settings. Instead, they are most defensible as locally validated components of hybrid exposure models that also incorporate meteorology, land use, biomass burning, household energy, and other non-traffic sources.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s11869-026-02060-y. Introduction Exposure to ambient and household air pollution, including fine Particulate Matter with an aerodynamic diameter ≤ 2.5 micrometers (PM2.5), remains a significant global public health concern in the 21st century. PM2.5 exposure contributes to an estimated 6.7 million premature deaths annually (Bachwenkizi et al. 2022; WHO, 2024). These deaths are driven by cardiopulmonary pathways, including asthma, hypertensive disorders of pregnancy such as pre-eclampsia, cardiovascular disease, and a range of acute and chronic respiratory outcomes. (Adeoye et al. 2024; Jafta et al. 2025; Lai et al. 2018). The substantial health burden is partly attributable to PM2.5 size and chemical composition, which facilitate deep penetration into the respiratory tract, translocation across biological barriers, and systemic inflammation that affects multiple organ systems. (Adeoye et al. 2024; B et al., 2014; Jafta et al. 2025). Emerging evidence also links prenatal PM2.5 exposure to preterm birth, low birth weight, and impaired neurodevelopment among children. (Issah et al. 2024; Lavigne et al. 2016). A growing body of recent works also suggests associations with poor mental well-being and general health. (Adza et al. 2022; Curto et al. 2019; Valencia et al. 2023). These impacts are amplified in sub-Saharan Africa (SSA) by rapid urbanization, increased motorization, unpaved-road dust, climate change, and landscape fire smoke, which collectively intensify both primary emissions and secondary particle formation (Greenwald et al. 2024; Velizarova et al. 2024). Effective management and mitigation of PM2.5 exposure depend on reliable air quality monitoring stations capable of capturing spatiotemporal variability. (Falzone et al. 2022). However, monitoring coverage is highly unequal across the globe. High-income countries (HIC) typically have dense ground-based monitoring stations that provide broader spatio-temporal coverage. Low- and middle-income countries (LIMCs), particularly in Sub-Saharan Africa (SSA), remain severely under-monitored. SSA exhibits heterogeneous PM2.5 sources that cannot be quantified or assessed using existing, sparse ground-based monitoring systems. According to the WHO ambient air pollution database of 2024, SSA accounts for less than 1% of global monitoring stations (203/40,222) (WHO Ambient Air Quality Database, 2024). This reflects constraints, including limited policy enforcement, high setup and operational costs, and shortages of the expertise required for sustained operation of a ground monitoring network (Khan et al. 2024). This results in limited spatiotemporal coverage of air pollution data, limiting exposure characterization for health assessment, and weakening the design of effective interventions and mitigation strategies (Baharane and Shatalov 2024). This persistent data gap has accelerated interest in alternative air pollution estimation methods, including proxy indicators for air pollution (Wu et al. 2016). Proxy indicators are indirect measures used to infer air pollution/quality, including PM2.5 concentration, in regions without direct air quality measurements. (Yun et al. 2023). Such indicators leverage open geospatial datasets, such as road network density, visibility, meteorological variables, and land-use/land-cover patterns, to characterize air pollution and infer pollutant concentrations (Singh et al. 2025). However, while these proxies are widely applied in HIC, their validity in SSA remains insufficiently evaluated. Robust validation of road and traffic-related PM2.5 proxy indicators is therefore necessary to determine whether they can reliably be blended with in-situ measurements to reduce data gaps and support pollution exposure characterization in SSA. In this study, we evaluated the validity of three road-network-based proxies for PM2.5 exposure: (i) population-weighted Road Network Density (wRND), (ii) Euclidean distance to Highway (EH), and (iii) Euclidean distance to Main roads (EM). The main aim of this study was to evaluate how these proxies performed against measured outdoor-filtered PM2.5 personal exposure data from The Gambia, Mozambique, and Kenya. Methods Overall framework We developed a framework to validate road-traffic proxy indicators using high-resolution personal PM2.5 exposure data (Fig. 1). The framework integrates personal PM2.5 measurements from the PREgnancy Care Integrating Translational Science Everywhere (PRECISE-DYAD) project (Craik et al. 2025), road network data from OpenStreetMap, and gridded population data from WorldPop (WorldPop and CIESIN, 2018). All analyses in the study followed the Characterizing Effects of Air Quality In Maternal, Newborn and Child Health (CHEAQI-MNCH) protocol (Munthali et al. 2026), and the use of PRECISE DYAD and PRECISE-HOME data complied with project ethical clearance and guidance. While the PRECISE-DYAD personal monitoring protocol captures integrated exposure across indoor and outdoor micro-environments, this analysis restricted observations to outdoor movement periods to better isolate the traffic-related component relevant to proxy evaluation. By evaluating these proxies against outdoor PM2.5 exposure measurements across multiple SSA settings, we provide a conceptual framework for PM2.5 estimation in regions with sparse conventional air quality networks. Fig. 1Flow chart of the validation framework developed in the study to validate the proxy indicators in Kenya, The Gambia, and Mozambique Flow chart of the validation framework developed in the study to validate the proxy indicators in Kenya, The Gambia, and Mozambique Study Sites The study was conducted in selected villages in three African countries: southern Mozambique, coastal Kenya, and the North Bank region of The Gambia (Craik et al. 2025; von Dadelszen et al. 2020) (Fig. 2). These sites represent contrasting SSA geographies, road connectivity, settlement form, and land use contexts that influence spatial patterns of PM2.5 exposure (Evangelopoulos et al. 2020; Tchanche et al. 2022). In Mozambique, the study villages are located within Manhica District and Xinavane locality, both in Maputo province, an area characterized by intensive large-scale sugarcane agricultural activity (Pinto et al. 2024). In Kenya, the study sites fall within the coastal corridor around Mariakani and the rural Rabai Sub-County located along the northern corridor, where heavy traffic volumes are a major potential source of PM2.5 (Lazaro and Baba 2023). In The Gambia, the study sites are situated in and around Farafenni and within surrounding rural districts, including Illiasa and Sabach Sanjal. These districts are characterized by dense population and a high level of informal transport (Egger et al. 2023). In general, ground air quality monitoring is limited in these countries and concentrated in a small number of locations, primarily in capital cities (Mozambique: 1 in Boane, Kenya: 1 in Nairobi and 1 in Mombasa, The Gambia: 1 in Banjul). Fig. 2PRECISE study villages located in Kenya, Mozambique, and The Gambia, representing East, West, and Southern Africa regions. Data used to validate proxy indicators were collected in these villages PRECISE study villages located in Kenya, Mozambique, and The Gambia, representing East, West, and Southern Africa regions. Data used to validate proxy indicators were collected in these villages Data: Personal exposure and proxies This study used a secondary high-resolution personal PM2.5 exposure dataset from the PRECISE-DYAD project comprising 1,048,576 one-minute observations collected over 12 months (2022–2023). In total, 343 women were recruited in PRECISE-DYAD to carry a personal exposure monitor bag that recorded PM2.5 concentrations (µg/m3), alongside relative humidity and air temperature, capturing both indoor and outdoor microenvironments. However, this study focused on outdoor exposure as it better reflects road traffic-related pollution. The de-identified PRECISE DYAD/PRECISE-HOME dataset was accessed for research purposes on 15 May 2024. The authors did not have access to information that could identify individual participants during or after data access. To identify outdoor mobility periods and reduce contamination from stationary and indoor time, we applied a spatial displacement threshold of 100 m/min (1.55 m/s). This threshold exceeds the commonly used static-cluster threshold (~ 0.83 m/s) used in other studies (Hine 2010), thereby reducing misclassification of non-outdoor periods as outdoor. The selected value was consistent with typical adult walking speed (1.0–1.8 m/s) (Gault et al. 2017; Yokote and Kishimoto 2021), supporting the use of 1.5 m/s to identify sustained outdoor movement. To improve the robustness of this classification, a spatial validation step was implemented using footprint polygons from OpenStreetMap. Building geometries were retrieved through the Overpass API and used to perform point-in-polygon analysis to determine whether monitoring points occurred inside or outside building footprints. Speed-based classifications were compared with building-based labels using confusion matrix analysis, and discrepancies were corrected using the building-derived classification as ground truth to minimize misclassification of indoor and outdoor environments. After applying these filtering and validation procedures, the recorded observations were retained for subsequent exposure analysis. We derived a population-weighted road network density (wRND) indicator by integrating road density with population distribution. The inputs included OpenStreetMap road network data (May 2024), village boundary polygons, village centroid point locations, and WorldPop (2018) gridded population data at 100 m spatial resolution. We projected the road-network dataset to the WGS84 UTM coordinate system to support spatial distance and area calculations. We used ArcMap 10.8 to compute wRND at the neighborhood level by first clipping the road network to community boundaries and generating a regular fishnet grid aligned to the 100 m population raster. The total road length within each grid cell was summarized using the cell ID as the aggregation unit, and road network density (RND) was calculated per unit area (km/km2). Population counts from WorldPop were extracted and assigned to the corresponding grid cell, producing a weighted Road Network Density score. The wRND score was computed using Eq. 1. wRND_village = Σ (RND_cell × Population_cell) / Σ(Population_cell) [Eq. 1]. We also derived distance proxies in ArcMap 10.8 by calculating Euclidean (Suwanda et al. 2020) (straight-line) Distances from village centroids to the nearest main roads (EM) and highways (EH) based on the OpenStreetMap road network. We calculated Euclidean distance because prior LUR studies in LMICs validate Euclidean distance as an effective proxy when vehicle count data are absent. Additionally, the rationale of this study is to test the hypothesis that exposure to PM2.5 from motor vehicles decreases as you move away from the road, hence the need to calculate the Euclidean distance. The road network was reprojected to WGS84 UTM Zones to ensure accurate metric distance estimation, and roads were classified by hierarchy to isolate main roads and highways for separate analysis. The Near tool was used to calculate the shortest straight-line distance from each village centroid to the closest road segment, identifying the nearest location along the road geometry. Data Analysis: Spatial Association and Predictive Modelling of Personal PM2.5 Exposure Using Proxy Indicators We mapped the spatial distribution of the proxy-PM2.5 relationship using proportional symbols in ArcMap 10.8. For each study region, a 5 km hexagonal tessellation grid (Gold 2016) was generated, and village locations were spatially joined to the tessellation to aggregate village-level attributes within each hexagon. Within each hexagon, PM2.5 and proxy indicators (wRND, EM, and EH) were summarized using the mean, producing an aggregated surface of village clusters rather than individual village points. Proxies were then visualized as proportional symbol sizes, while PM2.5 was represented using a sequential color gradient (µg/m³), enabling comparison of PM2.5 intensity across proxy ranges within and between study regions. Proxy-PM2.5 associations were evaluated using Spearman’s rank correlation (Beckerman et al. 2008). We modeled PM2.5 as a function of road-traffic proxies wRND, EH, and EM, fitting them individually and jointly within a Land Use Regression (LUR) model Framework. LUR is widely used to estimate spatial variations in pollutant concentrations using proxy geospatial predictors (Ji et al. 2010; Mölter and Lindley 2021). LUR was implemented using a random forest (RF) regressor and an ensemble method that combines multiple decision-tree models on bootstrapped samples and averages predictions to reduce variance and overfitting (Sharma et al. 2023). At each node split within a tree, RF considers a random subset of predictors (mtry), which decorrelates between trees and improves robustness. Country-specific RF models were developed for the study countries (Fig. 3). We set the number of trees (ntree) to 100, informed by a parameter sensitivity analysis that demonstrated R2 stabilizes beyond ntree = 100 across all country-pollutant combinations (Supplementary Figures S4 and S5; Supplementary Table S1). The variables considered at each split (mtry) were set to the square root of the number of predictors. The model performance was evaluated using 3 cross-validation folds and summarized by the coefficient of determination (R2), root mean squared error (RMSE), mean absolute error (MAE), and mean bias error (Belitz et al., 2021; Hua et al. 2019; Velizarova et al. 2024). Feature importance was computed for the combined models, and the results are presented in Supplementary Figure S3. Fig. 3Diagrammatic representation of the random forest framework used to estimate PM2.5 across study sites. For each country, four models were fitted (wRND, EH, EM, and Combined), each with 100 trees. Model outputs for PM2.5 were averaged across trees within each model and evaluated separately Diagrammatic representation of the random forest framework used to estimate PM2.5 across study sites. For each country, four models were fitted (wRND, EH, EM, and Combined), each with 100 trees. Model outputs for PM2.5 were averaged across trees within each model and evaluated separately Results Proxies and Personal PM2.5 exposure – Summary statistics and spatial distribution Kenya recorded PM2.5 mean = 31.3 µg/m³, exceeding the WHO AQG 2021 annual mean (5 ug/m3) more than 6-fold. Mozambique (PM2.5 mean = 19.6 µg/m³) exceeded the WHO AQG annual guideline by 3.9-fold. The Gambia had the highest PM2.5 burden overall (PM2.5 mean = 49.4 µg/m³), exceeding the WHO AQG annual guideline by 9.9-fold. A WHO guideline comparison Table S2 shows a comparison of daily and annual guidelines. Table 1: Highlights the summary statistics for the proxies in the 3 countries. Figure S1 shows the Violin-box plots for the distribution and heterogeneity of personal PM2.5 concentrations across the three study sites. Table 1Summary statistics PM2.5 (µg/m³), wRND, EM (km), and EH (km) across study sites over the full analysis periodVariableCountryMeanStdMin25%Median75%MaxPM2.5Gambia49.442.09.024.535.566.5220.7Kenya31.335.74.89.616.334.1173.1Mozambique 19.6 11.8 2.3 10.4 18.2 24.1 60.7wRNDGambia26.125.15.917.020.725.7152.5Kenya74.846.8 0.3 41.575.098.2198.0Mozambique 16.1 7.62.5 8.8 17.421.2 36.5 EM(km)Gambia1.51.60.20.40.72.16.7Kenya0.7 0.9 0.0 0.1 0.50.97.0Mozambique 1.1 1.0 0.0 0.5 0.7 1.5 5.4 EH (km)Gambia1.81.60.30.51.72.5 6.7 Kenya1.34.2 0.0 0.2 0.6 1.242.5Mozambique4.04.20.00.81.97.017.86 Table 1Summary statistics PM2.5 (µg/m³), wRND, EM (km), and EH (km) across study sites over the full analysis periodVariableCountryMeanStdMin25%Median75%MaxPM2.5Gambia49.442.09.024.535.566.5220.7Kenya31.335.74.89.616.334.1173.1Mozambique 19.6 11.8 2.3 10.4 18.2 24.1 60.7wRNDGambia26.125.15.917.020.725.7152.5Kenya74.846.8 0.3 41.575.098.2198.0Mozambique 16.1 7.62.5 8.8 17.421.2 36.5 EM(km)Gambia1.51.60.20.40.72.16.7Kenya0.7 0.9 0.0 0.1 0.50.97.0Mozambique 1.1 1.0 0.0 0.5 0.7 1.5 5.4 EH (km)Gambia1.81.60.30.51.72.5 6.7 Kenya1.34.2 0.0 0.2 0.6 1.242.5Mozambique4.04.20.00.81.97.017.86 Summary statistics PM2.5 (µg/m³), wRND, EM (km), and EH (km) across study sites over the full analysis period The bivariate spatial distribution of PM2.5 and wRND across the village study villages was aggregated into tessellations as shown in Fig. 4. In The Gambia (Fig. 4a), villages along Njaba Kunda/No Kunda and Farafenni have average to higher PM2.5 concentrations, ranging from low to moderate (0–20). In Kenya (Fig. 4b), villages in Mariakani and Rabai show mixed spatial patterns in which elevated PM2.5 occurs across multiple wRND classes. In Mozambique (Fig. 4c), Xinavane exhibits co-located higher values of PM2.5 and wRND; these clusters extend through some villages in Palmeria but occur across both lower and higher wRND classes. Fig. 4Spatial distribution of aggregated PM2.5 and weighted road network density (wRND) across hexagonal tessellations in (a) The Gambia, (b) Kenya, and (c) Mozambique. Hexagons represent aggregation units of the study villages Spatial distribution of aggregated PM2.5 and weighted road network density (wRND) across hexagonal tessellations in (a) The Gambia, (b) Kenya, and (c) Mozambique. Hexagons represent aggregation units of the study villages The yellow-to-orange color gradient represents PM2.5 (µg/m³), and proportional symbols represent wRND classes (0–10, 10–20, and > 20) within each hexagon, the spatial distribution of the PM2.5-EM relationship in the study villages as shown in Fig. 5. In The Gambia (Fig. 5a), villages in Sabach and Farafenni show higher PM2.5 concentrations, including low EM classes, while villages more than 20 km from Sanjal show elevated PM2.5 at greater distances. In Kenya (Fig. 5b), Mariakani and Rabai exhibit mixed patterns in which higher PM2.5 occurs in both low EM (0–2 km) and intermediate EM (2–4 km) hexagons. The results highlight that in Mozambique (Fig. 5c), villages in Xinavane exhibit higher PM2.5 concentrations across multiple EM classes, including some hexagons with larger EM (> 3 km), contradicting a distance-decay effect in PM2.5 with increasing EM. Overall, across the study villages, the PM2.5–EM relationship is heterogeneous across and within sites, consistent with context-dependent proxy performance. Fig. 5Spatial distribution of aggregated PM2.5 and Euclidean distance to main roads (EM) across hexagonal tessellations in (a) The Gambia, (b) Kenya, and (c) Mozambique. Hexagons represent aggregation units of the study villages. The yellow-to-orange color gradient represents mean PM2.5 (µg/m³), and proportional symbols represent mean EM classes (0–2 km, 2–4 km, and > 3 km) within each hexagon Spatial distribution of aggregated PM2.5 and Euclidean distance to main roads (EM) across hexagonal tessellations in (a) The Gambia, (b) Kenya, and (c) Mozambique. Hexagons represent aggregation units of the study villages. The yellow-to-orange color gradient represents mean PM2.5 (µg/m³), and proportional symbols represent mean EM classes (0–2 km, 2–4 km, and > 3 km) within each hexagon The spatial distribution of the PM2.5–EH relationship in the study villages is shown in Fig. 6 as hexagon tessellations. In the Gambia corridor, villages in Njaba Kunda and Sabach show moderately elevated PM2.5 at low EH, aligning more closely with the proximity-to-highways assumption. In Kenya (Fig. 6b), Mariakani shows a proximity effect, with aggregated low-EH values corresponding to higher PM2.5 concentrations. In contrast, village Rabai shows higher PM2.5 at higher EH. In Mozambique (Fig. 6c), Xinavane shows a direct positive relationship between higher PM2.5 and greater highway distances, contradicting the expected decline in PM2.5 with increasing EH. Fig. 6Spatial distribution of aggregated PM2.5 and Euclidean distance to highways (EH) across hexagonal tessellations in (a) The Gambia, (b) Kenya, and (c) Mozambique. Hexagons represent the aggregation units. The yellow to orange color gradient represents mean PM2.5 (µg/m³), and proportional symbols represent mean EH (km) within each hexagon Spatial distribution of aggregated PM2.5 and Euclidean distance to highways (EH) across hexagonal tessellations in (a) The Gambia, (b) Kenya, and (c) Mozambique. Hexagons represent the aggregation units. The yellow to orange color gradient represents mean PM2.5 (µg/m³), and proportional symbols represent mean EH (km) within each hexagon Proxy-PM2.5 associations in Figs. 7 and 8 indicate limited and context-specific statistical relationships, showing the scatter plots of the proxies against personal PM2.5. wRND was positively and significantly associated with PM2.5 in Mozambique (r = 0.351; p = 0.005), but associations were weak and non-significant in Kenya (r = − 0.041; p = 0.673), and The Gambia (r = − 0.020; p = 0.909); the pooled (“global”) association was near zero (r = 0.032; p = 0.648). EM showed weak, non-significant associations across all settings (Kenya r = − 0.103; p = 0.280; The Gambia r = 0.146; p = 0.410; Mozambique r = − 0.038; p = 0.766), with an overall weak negative Global association (r = − 0.008; p = 0.912). EH exhibited contrasting behavior, with a significant negative association in Kenya (r = − 0.224; p = 0.018) and a positive association in Mozambique (r = 0.292; p = 0.020); whilst the association in The Gambia had positive but marginal (r = 0.335; p = 0.053). Fig. 7Spearman correlations between proxy indicators (wRND, EH EM) and pollutants (PM2.5) in Mozambique, Kenya, and the Gambia. The shades of blue indicate negative correlations, while those of red indicate positive correlations. Significance is denoted by asterisks: * p < 0.05, ** p < 0.01, *** p < 0.001 Spearman correlations between proxy indicators (wRND, EH EM) and pollutants (PM2.5) in Mozambique, Kenya, and the Gambia. The shades of blue indicate negative correlations, while those of red indicate positive correlations. Significance is denoted by asterisks: * p < 0.05, ** p < 0.01, *** p < 0.001 Fig. 8Scatter plots of the proxies against personal PM2.5 Scatter plots of the proxies against personal PM2.5 Regression analysis of Personal PM2.5 exposure and proxy indicators The observed versus predicted values for the RF models, using single and combined models across the study sites, are shown in Fig. 9. The performance metrics of the models are in Table S4, and the model sensitivity is shown in Figure S3. EM only models’ regression ranged from The Gambia (R2 = 0.37; RMSE = 36.97 µg/m3), Kenya (R2 = 0.43; RMSE = 32.3 µg/m3), and Mozambique (R2 = 0.42; RMSE = 9.5 µg/m3) and Global performing moderately as well (R2 = 0.42, RMSE = 32.3 µg/m3). EH-PM2.5 models were offered better single proxy in performance (R2= 0.58; RMSE = 39.0 µg/m3) and remained comparable in Kenya (R2 = 0.43; RMSE = 31.7 µg/m3) and Mozambique (R2 = 0.45; RMSE = 9.8 µg/m3), with Global R2 = 0.44 (RMSE = 29.1 µg/m3). wRND-only models showed weaker performance in The Gambia (R2 = 0.32; RMSE = 40.0 µg/m3) but were like EM/EH in Kenya (R2= 0.44; RMSE = 30.8 µg/m3) and moderate in Mozambique (R2= 0.35; RMSE = 9.9 µg/m3), with Global R2 = 0.36 (RMSE = 29.6 µg/m3). Combining indicators improved performance in Kenya (R² = 0.52; RMSE = 31.7 µg/m³) and Mozambique (R2 = 0.60; RMSE = 8.9 µg/m3) and increased pooled performance (Global R2= 0.46; RMSE = 29.1 µg/m3), although in The Gambia the combined model (R2 = 0.53; RMSE = 37.6 µg/m3) did not exceed the EM-only model. Fig. 9Observed versus predicted PM2.5 from Random Forest models using single and combined proxy indicators across study countries and the whole dataset. The panels show the PM2.5 model performance for each country, annotated with R2 and RMSE Observed versus predicted PM2.5 from Random Forest models using single and combined proxy indicators across study countries and the whole dataset. The panels show the PM2.5 model performance for each country, annotated with R2 and RMSE Mean bias Error (MBE) plots for PM2.5 are shown in Fig. 10. Combined proxy models showed the smallest and most stable biases across countries in Kenya; the combined biases ranged from mean ± 2–5 µg/m3 for low and mid-range (0–40 µg/m3) concentration bins. Whilst in The Gambia and Mozambique, MBE exceeded 40 µg/m³ in the combined model at higher PM2.5 concentrations; the deviations were + 20 µg/m³ at higher-range concentrations. Fig. 10Mean bias error (MBE) plots for random forest PM2.5 predictions across study countries and the global dataset. The black line shows mean bias, the red line shows median bias, and the grey shading indicates the 95% confidence interval for mean bias. The plots show that combined proxy models exhibit the lowest overall bias, though bias increases to + 10 µg/m³ at the highest concentrations Mean bias error (MBE) plots for random forest PM2.5 predictions across study countries and the global dataset. The black line shows mean bias, the red line shows median bias, and the grey shading indicates the 95% confidence interval for mean bias. The plots show that combined proxy models exhibit the lowest overall bias, though bias increases to + 10 µg/m³ at the highest concentrations Discussion This study evaluated whether commonly used road-traffic proxy indicators can represent personal PM₂.₅ exposure in three sub-Saharan African settings where routine air quality monitoring is limited. The findings show that road-based proxies capture some exposure-relevant spatial signals, but their performance is inconsistent across countries and insufficient as stand-alone indicators. This is an important validation finding: in heterogeneous SSA environments, where PM₂.₅ may arise from multiple sources, including traffic, unpaved road dust, biomass burning, informal transport, household energy use, agricultural activity, and local settlement patterns, road proximity alone cannot be assumed to reflect personal exposure. The main contribution of this study, beyond testing three accessible road-based proxies, is to demonstrate the need for local calibration and hybrid exposure models before such proxies are used in epidemiological analyses or public health decision-making. Beyond validating the study site-specific proxies, our analysis provides a structured framework for evaluating the utility of these indicators in representing PM2.5 in areas with limited monitoring infrastructure. Spatial distribution of proxy indicators and Personal PM2.5 exposure The personal PM2.5 exposures recorded in this study exceed WHO AQG 2021 guidelines by 3.9- to 6.3-fold across countries, underscoring the urgent public health burden of air pollution in these settings. These levels are associated with increased risks of cardiovascular disease, preterm birth, and respiratory conditions, particularly for pregnant women who formed the study population (Bochenski et al., 2022; WHO, 2024). This reinforces the need for scalable, low-cost proxy-based monitoring systems as proposed in this study. In The Gambia, areas characterized by higher wRND and low EM exhibited higher PM2.5 concentrations, consistent with evidence linking dense road networks with elevated PM2.5 levels in urban and peri-urban environments (Yu et al. 2023). Similar spatial relationships between proximity to major roads and PM2.5 have been demonstrated in previous proximity-based analyses in Ireland. (Richmond-Bryant et al. 2017). Additionally, traffic dispersion modelling in road-dense corridors yielded the same results, supporting the use of road proximity as a proxy of PM2.5 exposure (Huang et al. 2018). In contrast, Mozambique’s PM2.5 concentrations were less dependent on EM. They remained elevated even at greater distances from main roads, highlighting that local sources were dominant and emanated from non-traffic processes. This finding mirrors findings from the same Xinavane region, where sugarcane burning is a major contributor to PM2.5. (Wang et al. 2017). In Kenya, higher PM2.5 levels from villages, i.e., Rabai, further from main roads, also suggest on-road traffic sources, as identified in other studies, including biomass burning from sugar cultivation in the adjacent Sabaka River valley, a well-documented PM2.5 source (Omondi et al. 2024). Statistical Relationship of proxy indicators and Personal PM2.5 exposure Overall, wRND showed moderate alignment with PM2.5, in line with studies reporting positive road-PM2.5 associations in urban areas. (Jin et al. 2019; Shi et al. 2024). However, weaker wRND-PM2.5 correlations in Kenya and The Gambia indicate that wRND alone is insufficient in some settings, suggesting stronger roles for land use, meteorology, and non-traffic emissions in shaping PM2.5 variability (Ahmad et al. 2023). Notably, the positive correlation between PM2.5 and EH in The Gambia and Mozambique contradicts the typical distance-decay pattern observed in many LMICs (Shams et al. 2021). The negative country-level correlation between EH and PM2.5 in Kenya should be interpreted alongside local spatial departures; higher PM2.5 in the Rabai/Sabaki River valley at greater highway distances likely reflects the influence of biomass burning from sugarcane cultivation, which can elevate PM2.5 in areas farther from highways. However, this reflects confounding non-traffic sources (biomass and agricultural burning) that elevate PM2.5 away from major highways. We therefore expand this interpretation to account for setting-specific mechanisms that explain the variations. Earlier research highlights that biomass and agricultural burning can generate episodic PM2.5 peaks that do not co-vary with road proximity; in southern Mozambique. (Curto et al. 2019; Lazaro and Baba 2023). These studies identify biomass burning as a dominant contributor to PM2.5, which provides a plausible explanation for elevated PM2.5 at larger EH distances and the unexpected positive EH–PM2.5 relationship. Secondly, these study sites are predominantly heavy on informal transport systems such as matatus and boda-bodas. (Kerzhner 2023), including two-stroke engines, which can concentrate along minor or locally important roads that are poorly represented by simplified proximity metrics (EH/EM). These may not be captured by road-class hierarchies, thereby diluting the expected distance-decay relationship. Lastly, we establish that in rural and peri-rural villages, resuspended dust from unpaved roads may contribute materially to PM2.5, particularly during dry seasons, and this source may be weakly correlated with mapped road type, road density, or distance-to-road indicators. Overall, road proxies can inform PM2.5, but validity is context-dependent and further gains are likely if models integrate environmental covariates such as satellite-derived fire activity (hotspots/burned area) and land-use/land-cover indicators to explicitly represent biomass-burning and dust-related contributions. The regression analysis of Proxy indicators in predicting Personal PM2.5 exposure Single proxy RF models showed limited predictive power for PM2.5 across all study sites, reflecting exposure variability unlikely to be captured by a single road-based variable. (Jin et al. 2022; Mathew et al. 2023). Further, this finding aligns with studies in Italy, which also found that single proxy models had limited predictive capacity for PM2.5. (Shi et al. 2024). In particular, reliance on proximity-only measures (e.g., EM) demonstrated limited robustness, reflecting sensitivity to local spatial heterogeneity and shifting emission sources that complicate exposure modelling in urban and peri-urban environments. (Kamana et al. 2024). In contrast, combining proxy indicators significantly improved PM2.5 prediction, highlighting the value of integrating spatial indicators, such as road network density and proximity to major roads, to represent exposure variability. This finding aligns with work from Cape Town, South Africa, and other LMIC regions, showing that multi-indicator models improve PM2.5 predictions relative to isolated proxies (Ma et al. 2024; Morapedi and Obagbuwa 2023; Ndletyana and Madonsela 2023). Importantly, the pooled (“global”) analysis indicated moderate overall skill and further improvement when proxies were combined, suggesting that aggregating data across settings can stabilize estimation and capture shared proxy–PM2.5 structure that is not consistently recoverable within individual sites. Predictive performance varied across countries, with weaker results in Kenya suggesting that road-traffic proxies alone may not adequately capture PM2.5 spatio-temporal variability. It is important to note that the maximum R2 of 0.60 and substantial RMSE values (up to 40.0 µg/m3 in The Gambia) indicate that road-traffic proxies explain at most 60% of personal PM2.5 variance in these settings. The remaining variance reflects non-traffic sources, temporal variability, and measurement error. These results should be interpreted as an exploratory validation of proxy utility, not as evidence that road-traffic proxies are sufficient for quantitative exposure assessment without local calibration and supplementary covariates. The feature of importance results further indicated that the dominant proxy contribution differs across countries. This reinforced the idea that proxy validity is context-specific and that modelling strategies should be tailored to local sources, mixtures, and spatial structures. (Houdou et al. 2024; Padhy 2024). Consistent with this are previous studies in Thailand that demonstrated that integrating satellite, environmental variables, and traffic data improves ground-level PM2.5 estimates, supporting a combined-data approach for exposure assessment (Baharane and Shatalov 2024). Implications of Validation These findings underscore the importance of validating proxy indicators of PM2.5 exposure. While some proxies demonstrated utility, inconsistent associations across countries indicate that single indicators are insufficient, and a multi-proxy approach is more robust. Road network-related proxies are of particular interest as traffic and associated resuspension are important contributors to urban PM2.5, and because road data are widely accessible. (Baharane and Shatalov 2024). Additionally, literature suggests that integrating traffic type, volume, and distance to major routes using a linear regression model strongly predicts PM2.5 concentrations. (Kumar et al. 2024). However, prior evidence shows that incorporating meteorological variables, land-use patterns, and satellite-derived aerosol optical depth (AOD) data can improve estimates of PM2.5 exposure from diverse pollution sources (Franke et al. 2021; Tang et al. 2023). Moreover, our study demonstrates the effectiveness of RF-based LUR models for estimating PM2.5 exposure in resource-constrained settings. However, between-country variability highlights the need for local calibration and expanding predictor sets where road proxies alone incompletely represent exposure dynamics. We acknowledge that network distance traffic-volume-weighted measures would offer higher precision and recommend these for future studies as data availability improves. The LUR RF model results show that combining road proxies improved PM2.5 exposure estimates, but performance is likely to improve further through the integration of fire point data, NDVI, land use, and meteorological variables in a hybrid LURF model. A key limitation is that temporal variation in PM2.5 - including diurnal cycles (cooking/traffic peaks) and seasonal variation (dry season biomass burning in The Gambia and Mozambique; wet/dry seasons in Kenya was not explicitly modelled. All observations are treated as temporally exchangeable, which may introduce noise in measurements. Future work should incorporate time-of-day and season-specific models. Conclusion and Recommendations This study integrated personal PM₂.₅ exposure measurements with open geospatial road and population data to evaluate the utility of road network-derived pollution proxy indicators in three sub-Saharan African settings with limited monitoring infrastructure. Although mean personal PM₂.₅ exposures exceeded WHO Air Quality Guideline values across all three countries, road-based proxies showed heterogeneous and context-specific relationships with measured exposure. Single-proxy models had limited predictive value, while combined proxy models improved performance in some settings but still left substantial unexplained variation. Therefore, road-network proxies should not be used as stand-alone indicators of personal PM₂.₅ exposure in heterogeneous SSA environments. Future exposure assessment and epidemiological studies in SSA should prioritize locally calibrated, hybrid models that combine road-network indicators with meteorology, land use, vegetation, satellite-derived aerosol indicators, fire activity, and other local pollution sources. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2