Assessing the transferability of landuse regression models for ultrafine particles across two Canadian cities

Combining Google traffic map with deep learning model to predict street-level traffic-related air pollutants in a complex urban environment

BACKGROUND

Traffic-related air pollution (TRAP) is a major contributor to urban pollution and varies sharply at the street level, posing a challenge for air quality modeling. Traditional land use regression models combined with data from fixed monitoring stations may be unable to predict and characterize fine-scale TRAP, especially in complex urban environments influenced by various features. This study aims to estimate fine-scale (50 m) concentrations of nitrogen oxides (NO and NO₂) in Hong Kong using a deep learning (DL) structured model.

METHODS

We collected data from mobile air quality sensors on buses and crowd-sourced Google real-time traffic status as a proxy for real-time traffic emissions. Our DL model was compared with existing machine learning models to assess performance improvements. Using an interpretable machine learning method, we hierarchically evaluated the global, local, and interaction effects for different features.

RESULTS

Our DL model outperformed existing machine learning models, achieving R2 values of 0.72 for NO and 0.69 for NO₂. The incorporation of traffic status as a key predictor improved model performance by 9% to 17%. The interpretable machine learning method revealed the importance of traffic-related features and their pairwise interactions.

CONCLUSION

The results indicate that traffic-related features significantly contribute to TRAP and provide insights and guidance for urban planning. By incorporating crowd-sourced Google traffic information, we assessed traffic abatement scenarios that could inform targeted strategies for improving urban air quality.

Hyperlocal Air Pollution Mapping: A Scalable Transfer Learning LUR Approach for Mobile Monitoring

Addressing the challenge of mapping hyperlocal air pollution in areas without local monitoring, we evaluated unsupervised transfer learning-based land-use regression (LUR) models developed using mobile monitoring data from other cities: CORrelation ALignment (Coral) and its inverse distance-weighted modification (IDW_Coral). These models mitigated domain shifts and transferred patterns learned from mobile air quality monitoring campaigns in Copenhagen and Rotterdam to estimate annual average air pollution levels in Amsterdam (50m road segments) without involving any Amsterdam measurements in model development. For nitrogen dioxide (NO2), IDW_Coral outperformed Copenhagen and Rotterdam LUR models directly applied to Amsterdam, achieving MAE (4.47 μg/m3) and RMSE (5.36 μg/m3) comparable to a locally fitted LUR model (AMS_SLR) developed using Amsterdam mobile measurements collected for 160 days. IDW_Coral yielded an R2 of 0.35, similar to that of the AMS_SLR based on 20 collection days, suggesting a minimum requirement of 20-day mobile monitoring to capture city-specific insights. For ultrafine particles (UFP), IDW_Coral's citywide predictions strongly correlated with previously published mixed-effect models fitted with 160-day Amsterdam measurements (Pearson correlation of 0.71 for UFP and 0.72 for NO2). IDW_Coral demands no direct measurements in the target area, showcasing its potential for large-scale applications and offering significant economic efficiencies in executing mobile monitoring campaigns.

Capturing Exposure Disparities with Chemical Transport Models: Evaluating the Suitability of Downscaling Using Land Use Regression

High resolution exposure surfaces are essential to capture disparities in exposure to traffic-related air pollution in urban areas. In this study, we develop an approach to downscale Chemical Transport Model (CTM) simulations to a hyperlocal level (∼100m) in the Greater Toronto Area (GTA) under three scenarios where emissions from cars, trucks and buses are zeroed out, thus capturing the burden of each transportation mode. This proposed approach statistically fuses CTMs with Land-Use Regression using machine learning techniques. With this proposed downscaling approach, changes in air pollutant concentrations under different scenarios are appropriately captured by downscaling factors that are trained to reflect the spatial distribution of emission reductions. Our validation analysis shows that high-resolution models resulted in better performance than coarse models when compared with observations at reference stations. We used this downscaling approach to assess disparities in exposure to nitrogen dioxide (NO2) for populations composed of renters, low-income households, recent immigrants, and visible minorities. Individuals in all four categories were disproportionately exposed to the burden of cars, trucks, and buses. We conducted this analysis at spatial resolutions of 12, 4, 1 km, and 100 m and observed that disparities were significantly underestimated when using coarse spatial resolutions. This reinforces the need for high-spatial resolution exposure surfaces for environmental justice analyses.

Harnessing AI to unmask Copenhagen's invisible air pollutants: A study on three ultrafine particle metrics

Ultrafine particles (UFPs) are airborne particles with a diameter of less than 100 nm. They are emitted from various sources, such as traffic, combustion, and industrial processes, and can have adverse effects on human health. Long-term mean ambient average particle size (APS) in the UFP range varies over space within cities, with locations near UFP sources having typically smaller APS. Spatial models for lung deposited surface area (LDSA) within urban areas are limited and currently there is no model for APS in any European city. We collected particle number concentration (PNC), LDSA, and APS data over one-year monitoring campaign from May 2021 to May 2022 across 27 locations and estimated annual mean in Copenhagen, Denmark, and obtained additionally annual mean PNC data from 6 state-owned continuous monitors. We developed 94 predictor variables, and machine learning models (random forest and bagged tree) were developed for PNC, LDSA, and APS. The annual mean PNC, LDSA, and APS were, respectively, 5523 pt/cm3, 12.0 μm2/cm3, and 46.1 nm. The final R2 values by random forest (RF) model were 0.93 for PNC, 0.88 for LDSA, and 0.85 for APS. The 10-fold, repeated 10-times cross-validation R2 values were 0.65, 0.67, and 0.60 for PNC, LDSA, and APS, respectively. The root mean square error for final RF models were 296 pt/cm3, 0.48 μm2/cm3, and 1.60 nm for PNC, LDSA, and APS, respectively. Traffic-related variables, such as length of major roads within buffers 100-150 m and distance to streets with various speed limits were amongst the highly-ranked predictors for our models. Overall, our ML models achieved high R2 values and low errors, providing insights into UFP exposure in a European city where average PNC is quite low. These hyperlocal predictions can be used to study health effects of UFPs in the Danish Capital.

A comprehensive review of the development of land use regression approaches for modeling spatiotemporal variations of ambient air pollution: A perspective from 2011 to 2023

Land use regression (LUR) models are widely used in epidemiological and environmental studies to estimate humans' exposure to air pollution within urban areas. However, the early models, developed using linear regressions and data from fixed monitoring stations and passive sampling, were primarily designed to model traditional and criteria air pollutants and had limitations in capturing high-resolution spatiotemporal variations of air pollution. Over the past decade, there has been a notable development of multi-source observations from low-cost monitors, mobile monitoring, and satellites, in conjunction with the integration of advanced statistical methods and spatially and temporally dynamic predictors, which have facilitated significant expansion and advancement of LUR approaches. This paper reviews and synthesizes the recent advances in LUR approaches from the perspectives of the changes in air quality data acquisition, novel predictor variables, advances in model-developing approaches, improvements in validation methods, model transferability, and modeling software as reported in 155 LUR studies published between 2011 and 2023. We demonstrate that these developments have enabled LUR models to be developed for larger study areas and encompass a wider range of criteria and unregulated air pollutants. LUR models in the conventional spatial structure have been complemented by more complex spatiotemporal structures. Compared with linear models, advanced statistical methods yield better predictions when handling data with complex relationships and interactions. Finally, this study explores new developments, identifies potential pathways for further breakthroughs in LUR methodologies, and proposes future research directions. In this context, LUR approaches have the potential to make a significant contribution to future efforts to model the patterns of long- and short-term exposure of urban populations to air pollution.

Development of Spatio-Temporal Land Use Regression Models for Fine Particulate Matter and Wood-Burning Tracers in Temuco, Chile

Biomass burning is common in much of the world, and in some areas, residential wood-burning has increased. However, air pollution resulting from biomass burning is an important public health problem. A sampling campaign was carried out between May 2017 and July 2018 in over 64 sites in four sessions, to develop a spatio-temporal land use regression (LUR) model for fine particulate matter (PM) and wood-burning tracers levoglucosan and soluble potassium (Ksol) in a city heavily impacted by wood-burning. The mean (sd) was 46.5 (37.4) μg m-3 for PM2.5, 0.607 (0.538) μg m-3 for levoglucosan, and 0.635 (0.489) μg m-3 for Ksol. LUR models for PM2.5, levoglucosan, and Ksol had a satisfactory performance (LOSOCV R2), explaining 88.8%, 87.4%, and 87.3% of the total variance, respectively. All models included sociodemographic predictors consistent with the pattern of use of wood-burning in homes. The models were applied to predict concentrations surfaces and to estimate exposures for an epidemiological study.

Predicting spatial variations in annual average outdoor ultrafine particle concentrations in Montreal and Toronto, Canada: Integrating land use regression and deep learning models

BACKGROUND

Concentrations of outdoor ultrafine particles (UFP; <0.1 µm) and black carbon (BC) can vary greatly within cities and long-term exposures to these pollutants have been associated with a variety of adverse health outcomes.

OBJECTIVE

This study integrated multiple approaches to develop new models to estimate within-city spatial variations in annual median (i.e. average) outdoor UFP and BC concentrations as well as mean UFP size in Canada's two largest cities, Montreal and Toronto.

METHODS

We conducted year-long mobile monitoring campaigns in each city that included evenings and weekends. We developed generalized additive models trained on land use parameters and deep Convolutional Neural Network (CNN) models trained on satellite-view images. Using predictions from these models, we developed final combined models.

RESULTS

In Toronto, the median observed UFP concentration, UFP size, and BC concentration values were 16,172pt/cm3, 33.7 nm, and 1225 ng/m3, respectively. In Montreal, the median observed UFP concentration, UFP size, and BC concentration values were 14,702pt/cm3, 29.7 nm, and 1060 ng/m3, respectively. For all pollutants in both cities, the proportion of spatial variation explained (i.e., R2) was slightly greater (1-2 percentage points) for the combined models than the generalized additive models and a greater (approximately 10 percentage points) than the deep CNN models. The Toronto combined model R2 values in the test set were 0.73, 0.55, and 0.61 for UFP concentrations, UFP size, and BC concentration, respectively. The Montreal combined model R2 values were 0.60, 0.49, and 0.60 for UFP concentration, UFP size, and BC concentration models respectively. For each pollutant, predictions from the combined, deep CNN, and generalized additive models were highly correlated with each other and differences between models were explored in sensitivity analyses.

CONCLUSION

Predictions from these models are available to support future epidemiological research examining long-term health impacts of outdoor UFPs and BC.

Development of transferable neighborhood land use regression models for predicting intra-urban ambient nitrogen dioxide (NO2) spatial variations

Land use regression (LUR) models have been extensively used to predict air pollution exposure in epidemiological and environmental studies. The lack of dense routine monitoring networks in big cities places increased emphasis on the need for LUR models to be developed using purpose-designed neighborhood-scale monitoring data. However, the unsatisfactory model transferability limits these neighborhood LUR models to be then applied to other intra-urban areas in predicting air pollution exposure. In this study, we tackled this issue by proposing a method to develop transferable neighborhood NO₂ LUR models with comparable predictive power based on only micro-scale predictor variables for modeling intra-urban ambient air pollution exposure. Taking Auckland metropolis, New Zealand, as a case study, the proposed method was applied to three neighborhoods (urban, central business district, and dominion road) and compared with the corresponding counterpart models developed using pools of (a) only macro-scale predictor variables and (b) a mixture of both micro- and macro-scale predictor variables (traditional method). The results showed that the models using only macro-scale variables achieved the lowest accuracy (R²: 0.388-0.484) and had the worst direct (R²: 0.0001-0.349) and indirect transferability (R²: 0.07-0.352). Those models using the traditional method had the highest model fitting R² (0.629-0.966) with lower cross-validation R² (0.495-0.941) and slightly better direct transferability (R²: 0.0003-0.386) but suffered poor model interpretability when indirectly transferred to new locations. Our proposed models had comparable model fitting R² (0.601-0.966) and the best cross-validation R² (0.514-0.941). They also had the strongest direct transferability (R²: 0.006-0.590) and moderate-to-good indirect transferability (R²: 0.072-0.850) with much better model interpretability. This study advances our knowledge of developing transferable LUR models for the very first time from the perspective of the scale of the predictor variables used in the model development and will significantly benefit the wider application of LUR approaches in epidemiological and environmental studies.

Predicting Within-City Spatial Variations in Outdoor Ultrafine Particle and Black Carbon Concentrations in Bucaramanga, Colombia: A Hybrid Approach Using Open-Source Geographic Data and Digital Images

Outdoor ultrafine particles (UFP, <0.1 μm) and black carbon (BC) vary greatly within cities and may have adverse impacts on human health. In this study, we used a hybrid approach to develop new models to estimate within-city spatial variations in outdoor UFP and BC concentrations across Bucaramanga, Colombia. We conducted a mobile monitoring campaign over 20 days in 2019. Regression models were trained on land use data and combined with predictions from convolutional neural networks (CNN) trained to predict UFP and BC concentrations using satellite and street-level images. The combined UFP model (R² = 0.54) outperformed the CNN (R² = 0.47) and land use regression (LUR) models (R² = 0.47) on their own. Similarly, the combined BC model also outperformed the CNN and LUR BC models (R² = 0.51 vs 0.43 and 0.45, respectively). Spatial variations in model performance were more stable for the CNN and combined models compared to the LUR models, suggesting that the combined approach may be less likely to contribute to differential exposure measurement error in epidemiological studies. In general, our findings demonstrated that satellite and street-level images can be combined with a traditional LUR modeling approach to improve predictions of within-city spatial variations in outdoor UFP and BC concentrations.

Potential for developing independent daytime/nighttime LUR models based on short-term mobile monitoring to improve model performance

Land use regression model (LUR) is a widespread method for predicting air pollution exposure. Few studies have explored the performance of independently developed daytime/nighttime LUR models. In this study, fine particulate matter (PM_2.5), inhalable particulate matter (PM₁₀), and nitrogen dioxide (NO₂) concentrations were measured by mobile monitoring during non-heating and heating seasons in Taiyuan. Pollutant concentrations were higher in the nighttime than the daytime, and higher in the heating season than the non-heating season. Daytime/nighttime and full-day LUR models were developed and validated for each pollutant to examine variations in model performance. Adjusted coefficients of determination (adjusted R²) for the LUR models ranged from 0.53-0.87 (PM_2.5), 0.53-0.85 (PM₁₀), and 0.33-0.67 (NO₂). The performance of the daytime/nighttime LUR models for PM_2.5 and PM₁₀ was better than that of the full-day models according to the results of model adjusted R² and validation R². Consistent results were confirmed in the non-heating and heating seasons. Effectiveness of developing independent daytime/nighttime models for NO₂ to improve performance was limited. Surfaces based on the daytime/nighttime models revealed variations in concentrations and spatial distribution. In conclusion, the independent development of daytime/nighttime LUR models for PM_2.5/PM₁₀ has the potential to replace full-day models for better model performance. The modeling strategy is consistent with the residential activity patterns and contributes to achieving reliable exposure predictions for PM_2.5 and PM₁₀. Nighttime could be a critical exposure period, due to high pollutant concentrations.

Ambient ultrafine particle concentrations and incidence of childhood cancers

BACKGROUND

Ambient air pollution has been associated with childhood cancer. However, little is known about the possible impact of ambient ultrafine particles (<0.1 μm) (UFPs) on childhood cancer incidence.

OBJECTIVE

This study aimed to evaluate the association between prenatal and childhood exposure to UFPs and development of childhood cancer.

METHODS

We conducted a population-based cohort study of within-city spatiotemporal variations in ambient UFPs across the City of Toronto, Canada using 653,702 singleton live births occurring between April 1, 1998 and March 31, 2017. Incident cases of 13 subtypes of paediatric cancers among children up to age 14 were ascertained using a cancer registry. Associations between ambient air pollutant concentrations and childhood cancer incidence were estimated using random-effects Cox proportional hazards models. We investigated both single- and multi-pollutant models accounting for co-exposures to PM_2.5 and NO₂.

RESULTS

A total of 1,066 childhood cancers were identified. We found that first trimester exposure to UFPs (Hazard Ratio (HR) per 10,000/cm³ increase = 1.13, 95% CI: 1.03-1.22) was associated with overall cancer incidence diagnosed before 6 years of age after adjusting for PM_2.5, NO₂, and for personal and neighborhood-level covariates. Association between UFPs and overall cancer incidence exhibited a linear shape. No statistically significant associations were found for specific cancer subtypes.

CONCLUSION

Ambient UFPs may represent a previously unrecognized risk factor in the aetiology of cancers in children. Our findings reinforce the importance of conducting further research on the effects of UFPs given their high prevalence of exposure in urban areas.

Development and transferability of ultrafine particle land use regression models in London

Due to a lack of routine monitoring, bespoke measurements are required to develop ultrafine particle (UFP) land use regression (LUR) models, which is especially challenging in megacities due to their large area. As an alternative, for London, we developed separate models for three urban residential areas, models combining two areas, and models using all three areas. Models were developed against annual mean ultrafine particle count cm^-3 estimated from repeated 30-min fixed-site measurements, in different seasons (2016-2018), at forty sites per area, that were subsequently temporally adjusted using continuous measurements from a single reference site within or close to each area. A single model and 10 models were developed for each individual area and combination of areas. Within each area, sites were split into 10 groups using stratified random sampling. Each of the 10 models were developed using 90% of sites. Hold-out validation was performed by pooling the 10% of sites held-out each time. The transferability of models was tested by applying individual and two-area models to external area(s). In model evaluation, within-area mean squared error (MSE) R² ranged from 14% to 48%. Transferring individual- and combined-area models to external areas without calibration yielded MSE-R² ranging from -18 to 0. MSE-R² was in the range 21% to 41% when using particle number count (PNC) measurements in external areas to calibrate models. Our results suggest that the UFP models could be transferred to other areas without calibration in London to assess relative ranking in exposures but not for estimating absolute values of PNC.

Drivers of seasonal and annual air pollution exposure in a complex urban environment with multiple source contributions

Outdoor air pollution is a global health concern, but detailed exposure information is still limited for many parts of the world. In this study, high-resolution exposure surfaces were generated for annual and seasonal fine particulate matter (PM_2.5), coarse particulate matter (PM₁₀), and carbon monoxide (CO) for the Greater Beirut Area (GBA), Lebanon, an urban zone with a complex topography and multiple source contributions. Land use regression models (LUR) were calibrated and validated with monthly data collected from 58 locations between March 2017 and March 2018. The annual mean (±1 SD) concentrations of PM_2.5, PM₁₀, and CO across the monitoring locations were 68.1 (±15.7) μg/m³, 83.5 (±19.5) μg/m³, and 2.48 (±1.12) ppm, respectively. The coefficients of determination for LUR models ranged from 56 to 67% for PM_2.5, 44 to 63% for the PM₁₀ models, and 50 to 60% for the CO. LUR model structures varied significantly by season for both PM_2.5 and PM₁₀ but not for CO. Traffic emissions were consistently the main source of CO emissions throughout the year. The relative importance of industrial emissions and power generation sources towards predicted PM levels increased during the hot season while the contribution of the international airport diminished. Moreover, the complex topography of the study area along with the seasonal changes in the predominant wind directions affected the spatial predicted concentrations of all three pollutants. Overall, the predicted exposure surfaces were able to conserve the inter-pollution correlations determined from the field monitoring campaign, with the exception of the cold season. Our pollution surfaces suggest that the entire population of Beirut is regularly exposed to concentrations exceeding the World Health Organization (WHO) air quality standards for both PM_2.5 and PM₁₀.