Predicting Traffic-Related Air Pollution Using Feature Extraction from Built Environment Images

Portraying on-road CO2 concentrations using street view panoramas and ensemble learning

A significant reduction in carbon dioxide (CO2) emissions caused by transportation is essential for attaining sustainable urban development. Carbon concentrations from road traffic in urban areas exhibit complex spatial patterns due to the impact of street configurations, mobile sources, and human activities. However, a comprehensive understanding of these patterns, which involve complex interactions, is still lacking due to the human perspective of road interface characteristics has not been taken into account. In this study, a mobile travel platform was constructed to collect both on-road navigation Street View Panoramas (OSVPs) and the corresponding CO2 concentrations. >100 thousand sample pairs that matched "street view-CO2 concentration" were obtained, covering 675.8 km of roads in Shenzhen, China. In addition, four ensemble learning (EL) models were utilized to establish nonlinear connections between the semantic and object features of streetscapes and CO2 concentrations. After performing EL fusion modeling, the predictive R2 in the test set exceeded 90 %, and the mean absolute error (MAE) was <3.2 ppm. The model was applied to Baidu Street View Panoramas (BSVPs) in Shenzhen to generate a map of average on-road CO2 with a 100 m resolution, and the Local Indicator of Spatial Association (LISA) was then used to identify high CO2 intensity spatial clusters. Additionally, the Light Gradient Boost-SHapley Additive exPlanation (LGB-SHAP) analysis revealed that vertically planted trees can reduce CO2 emissions from on-road sources. Moreover, the factors that affect on-road CO2 exhibit interaction and threshold effects. Street View Panoramas (SVPs) and Artificial Intelligence (AI) were adopted here to enhance the spatial measurement of on-road CO2 concentrations and the understanding of driving factors. Our approach facilitates the assessment and design of low-emission transportation in urban areas, which is critical for promoting sustainable traffic development.

Space-specific flux estimation of atmospheric chemicals from point sources to soil

Predicting chemical flux to soil from industrial point sources accurately at a regional scale has been a significant challenge due to high uncertainty in spatial heterogeneity and quantification. To address this challenge, we developed an innovative approach by combining California Air Resources Board Puff (CALPUFF) and mass balance models, leveraging their complementary strengths in quantitative accuracy and spatial precision. Specifically, CALPUFF was used to predict the polycyclic aromatic hydrocarbons (PAHs) flux to soil due to industrial sources. Additionally, the spatial distribution coefficient of PAHs flux (e.g., si for spatial unit i) was calculated by neural network and combined with the mass balance model to obtain the results of total PAHs fluxes, which were then combined with the results predicted by CALPUFF to effectively estimate the contribution of industrial sources to soil PAHs flux. Taking a petrochemical industry region located in Zhejiang province, China as a case study, results showed the input Phenanthrene (Phe) and Benzo(a)pyrene (BaP) fluxes predicted by CALPUFF were generally lower than those by the mass balance model, with slightly different distribution patterns. CALPUFF results, based on 36 industrial sources, partially represent those of the mass balance model, which includes all sources and pathways. It was suggested that industrial sources contributed 49%-89% and 65%-100% of soil Phe and BaP, respectively across the study area. The average Phe flux from point sources by deposition averaged 2.68 mg m-2∙a-1 in 2021, accounting for approximately 60% of the total Phe flux to soil. The average BaP flux from point sources by deposition averaged 0.0755 mg m-2∙a-1, accounting for only 0.1%-3.65% of the total BaP flux to soil. Thereby, our approach fills up a gap between the relevance to point sources and the accuracy of deposition quantification in estimating chemical flux from specific point sources to soil at a regional scale.

Air pollution prediction and backcasting through a combination of mobile monitoring and historical on-road traffic emission inventories

An important challenge for studies of air pollution and health effects is the derivation of historical exposures. These generally entail some form of backcasting, which refers to a range of approaches that aim to project a current surface into the past. Accurate backcasting is conditional upon the availability of historical data for predictor variables and the ability to capture spatial and temporal trends in these variables. This study proposes a method to backcast traffic-related air pollution surfaces developed using land-use regression models by including temporal variability of traffic and emissions and trends in concentrations measured at reference stations. Nitrogen dioxide (NO2) concentrations collected in the City of Toronto using the Urban Scanner mobile platform were adjusted for historical trends captured at reference stations. The Bayesian Estimator of Abrupt change, Seasonal change, and Trend (BEAST), a powerful tool for time series decomposition, was employed to isolate seasonal variations, annual trends, and abrupt changes in NO2 at reference stations, hence decomposing the signal. Exposure surfaces were generated for a period extending from 2006 to 2020, exhibiting decreases ranging from 10 to 50 % depending on the neighborhood, with an average of 20.46 % across the city. Yearly surfaces were intersected with mobility patterns of Torontonians extracted from travel survey data for 2006 and 2016, illustrating strong spatial gradients in the evolution of NO2 over time, with larger decreases along major roads and highways and in the central core. These findings demonstrate that air pollution improvements throughout the 14 years are inhomogeneous across space.

High-Precision Microscale Particulate Matter Prediction in Diverse Environments Using a Long Short-Term Memory Neural Network and Street View Imagery

In this study, we propose a novel long short-term memory (LSTM) neural network model that leverages color features (HSV: hue, saturation, value) extracted from street images to estimate air quality with particulate matter (PM) in four typical European environments: urban, suburban, villages, and the harbor. To evaluate its performance, we utilize concentration data for eight parameters of ambient PM (PM1.0, PM2.5, and PM10, particle number concentration, lung-deposited surface area, equivalent mass concentrations of ultraviolet PM, black carbon, and brown carbon) collected from a mobile monitoring platform during the nonheating season in downtown Augsburg, Germany, along with synchronized street view images. Experimental comparisons were conducted between the LSTM model and other deep learning models (recurrent neural network and gated recurrent unit). The results clearly demonstrate a better performance of the LSTM model compared with other statistically based models. The LSTM-HSV model achieved impressive interpretability rates above 80%, for the eight PM metrics mentioned above, indicating the expected performance of the proposed model. Moreover, the successful application of the LSTM-HSV model in other seasons of Augsburg city and various environments (suburbs, villages, and harbor cities) demonstrates its satisfactory generalization capabilities in both temporal and spatial dimensions. The successful application of the LSTM-HSV model underscores its potential as a versatile tool for the estimation of air pollution after presampling of the studied area, with broad implications for urban planning and public health initiatives.

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.

Beyond here and now: Evaluating pollution estimation across space and time from street view images with deep learning

Advances in computer vision, driven by deep learning, allows for the inference of environmental pollution and its potential sources from images. The spatial and temporal generalisability of image-based pollution models is crucial in their real-world application, but is currently understudied, particularly in low-income countries where infrastructure for measuring the complex patterns of pollution is limited and modelling may therefore provide the most utility. We employed convolutional neural networks (CNNs) for two complementary classification models, in both an end-to-end approach and as an interpretable feature extractor (object detection), to estimate spatially and temporally resolved fine particulate matter (PM2.5) and noise levels in Accra, Ghana. Data used for training the models were from a unique dataset of over 1.6 million images collected over 15 months at 145 representative locations across the city, paired with air and noise measurements. Both end-to-end CNN and object-based approaches surpassed null model benchmarks for predicting PM2.5 and noise at single locations, but performance deteriorated when applied to other locations. Model accuracy diminished when tested on images from locations unseen during training, but improved by sampling a greater number of locations during model training, even if the total quantity of data was reduced. The end-to-end models used characteristics of images associated with atmospheric visibility for predicting PM2.5, and specific objects such as vehicles and people for noise. The results demonstrate the potential and challenges of image-based, spatiotemporal air pollution and noise estimation, and that robust, environmental modelling with images requires integration with traditional sensor networks.

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.

Analysis of driving factors of spatial distribution of heavy metals in soil of non-ferrous metal smelting sites: Screening the geodetector calculation results combined with correlation analysis

Heavy metals (HMs) discharged from smelting production may pose a major threat to human health and soil ecosystems. In this study, the spatial distribution characteristics of HMs in the soil of a non-ferrous metal smelting site were assessed. This study employed the geodetector (GD) by optimizing the classification condition and supplementing the correlation analysis (CA). The contribution of driving factors, such as production workshop distributions, hydrogeological conditions, and soil physicochemical properties, to the distribution of HMs in soil in the horizontal and vertical dimensions was assessed. The results showed that the main factors underlying the spatial distribution of As, Cd, Hg, Pb, Sb, and Zn in the horizontal direction were the distance from the sintering workshop (the maximum q value of that factor, q=0.28), raw material yard (q=0.14), and electrolyzer (q=0.29), while those in the vertical direction were the soil moisture content (q=0.17), formation lithology (q=0.12), and soil pH (q=0.06). The findings revealed that the CA is a simple and effective method to supplement the GD analysis underlying the spatial distribution characteristics of HMs at site scale. This study provides useful suggestions for environmental management to prevent HMs pollution and control HMs in the soil of non-ferrous metal smelting sites.

Design, calibration, and testing of a mobile sensor system for air pollution and built environment data collection: The urban scanner platform

This paper describes a mobile air pollution sampling system, the Urban Scanner, which aims at gathering dense spatiotemporal air quality data to support urban air quality and exposure science. Urban Scanner comprises custom vehicle-mounted sensors for air pollution, meteorology, and built environment data collection (low-cost sensors, wind anemometer, 360 deg camera, LIDAR, GPS) as well as a server to store, process, and map all gathered geo-referenced sensory information. Two levels of sensor calibration were implemented, both in a chamber and in the field, against reference instrumentation. Chamber tests and a set of mathematical tools were developed to correct for sensor noise (wavelet denoising), misalignment (linear and nonlinear), and hysteresis removal. Models based on chamber testing were further refined based on field co-location. While field co-location captures natural changes in air pollution and meteorology, chamber tests allow for simulating fast transitions in these variables, like the transitions experienced by a mobile sensor in an urban environment. The best suite of models achieved an R² higher than 0.9 between sensor output and reference station observations and an RMSE of 2.88 ppb for nitrogen dioxide and 4.03 ppb for ozone. A mobile sampling campaign was conducted in the city of Toronto, Canada, to further test Urban Scanner. We observe that the platform adequately captures spatial and temporal variability in urban air pollution, leading to the development of land-use regression models with high explanatory power.

Characterisation of urban environment and activity across space and time using street images and deep learning in Accra

The urban environment influences human health, safety and wellbeing. Cities in Africa are growing faster than other regions but have limited data to guide urban planning and policies. Our aim was to use smart sensing and analytics to characterise the spatial patterns and temporal dynamics of features of the urban environment relevant for health, liveability, safety and sustainability. We collected a novel dataset of 2.1 million time-lapsed day and night images at 145 representative locations throughout the Metropolis of Accra, Ghana. We manually labelled a subset of 1,250 images for 20 contextually relevant objects and used transfer learning with data augmentation to retrain a convolutional neural network to detect them in the remaining images. We identified 23.5 million instances of these objects including 9.66 million instances of persons (41% of all objects), followed by cars (4.19 million, 18%), umbrellas (3.00 million, 13%), and informally operated minibuses known as tro tros (2.94 million, 13%). People, large vehicles and market-related objects were most common in the commercial core and densely populated informal neighbourhoods, while refuse and animals were most observed in the peripheries. The daily variability of objects was smallest in densely populated settlements and largest in the commercial centre. Our novel data and methodology shows that smart sensing and analytics can inform planning and policy decisions for making cities more liveable, equitable, sustainable and healthy.

National Land Use Regression Model for NO2 Using Street View Imagery and Satellite Observations

Land use regression (LUR) models are widely applied to estimate intra-urban air pollution concentrations. National-scale LURs typically employ predictors from multiple curated geodatabases at neighborhood scales. In this study, we instead developed national NO₂ models relying on innovative street-level predictors extracted from Google Street View [GSV] imagery. Using machine learning (random forest), we developed two types of models: (1) GSV-only models, which use only GSV features, and (2) GSV + OMI models, which also include satellite observations of NO₂. Our results suggest that street view imagery alone may provide sufficient information to explain NO₂ variation. Satellite observations can improve model performance, but the contribution decreases as more images are available. Random 10-fold cross-validation R² of our best models were 0.88 (GSV-only) and 0.91 (GSV + OMI)─a performance that is comparable to traditional LUR approaches. Importantly, our models show that street-level features might have the potential to better capture intra-urban variation of NO₂ pollution than traditional LUR. Collectively, our findings indicate that street view image-based modeling has great potential for building large-scale air quality models under a unified framework. Toward that goal, we describe a cost-effective image sampling strategy for future studies based on a systematic evaluation of image availability and model performance.

Prediction of Short-Term Ultrafine Particle Exposures Using Real-Time Street-Level Images Paired with Air Quality Measurements

Within-city ultrafine particle (UFP) concentrations vary sharply since they are influenced by various factors. We developed prediction models for short-term UFP exposures using street-level images collected by a camera installed on a vehicle rooftop, paired with air quality measurements conducted during a large-scale mobile monitoring campaign in Toronto, Canada. Convolutional neural network models were trained to extract traffic and built environment features from images. These features, along with regional air quality and meteorology data were used to predict short-term UFP concentration as a continuous and categorical variable. A gradient boost model for UFP as a continuous variable achieved R² = 0.66 and RMSE = 9391.8#/cm³ (mean values for 10-fold cross-validation). The model predicting categorical UFP achieved accuracies for "Low" and "High" UFP of 77 and 70%, respectively. The presence of trucks and other traffic parameters were associated with higher UFPs, and the spatial distribution of elevated short-term UFP followed the distribution of single-unit trucks. This study demonstrates that pictures captured on urban streets, associated with regional air quality and meteorology, can adequately predict short-term UFP exposure. Capturing the spatial distribution of high-frequency short-term UFP spikes in urban areas provides crucial information for the management of near-road air pollution hot spots.

What you see is what you breathe? Estimating air pollution spatial variation using street level imagery

High spatial resolution information on urban air pollution levels is unavailable in many areas globally, partially due to high input data needs of existing estimation approaches. Here we introduce a computer vision method to estimate annual means for air pollution levels from street level images. We used annual mean estimates of NO2 and PM2.5 concentrations from locally calibrated models as labels from London, New York, and Vancouver to allow for compilation of a sufficiently large dataset (~250k images for each city). Our experimental setup is designed to quantify intra and intercity transferability of image-based model estimates. Performances were high and comparable to traditional land-use regression (LUR) and dispersion models when training and testing on images from the same city (R2 values between 0.51 and 0.95 when validated on data from ground monitoring stations). Like LUR models, transferability of models between cities in different geographies is more difficult. Specifically, transferability between the three cities i.e., London, New York, and Vancouver, which have similar pollution source profiles were moderately successful (R2 values between zero and 0.67). Comparatively, performances when transferring models trained on these cities with very different source profiles i.e., Accra in Ghana and Hong Kong were lower (R2 between zero and 0.21) suggesting the need for local calibration with local calibration using additional measurement data from cities that share similar source profiles.

Using an Exposome-Wide Approach to Explore the Impact of Urban Environments on Blood Pressure among Adults in Beijing-Tianjin-Hebei and Surrounding Areas of China

Existing studies mostly explored the association between urban environmental exposures and blood pressure (BP) in isolation, ignoring correlations across exposures. This study aimed to systematically evaluate the impact of a wide range of urban exposures on BP using an exposome-wide approach. A multicenter cross-sectional study was conducted in ten cities of China. For each enrolled participant, we estimated their urban exposures, including air pollution, built environment, surrounding natural space, and road traffic indicator. On the whole, this study comprised three statistical analysis steps, that is, single exposure analysis, multiple exposure analysis and a cluster analysis. We also used deletion-substitution-addition algorithm to conduct variable selection. After considering multiple exposures, for hypertension risk, most significant associations in single exposure model disappeared, with only neighborhood walkability remaining negatively statistically significant. Besides, it was observed that SBP (systolic BP) raised gradually with the increase in PM_2.5, but such rising pattern slowed down when PM_2.5 concentration reached a relatively high level. For surrounding natural spaces, significant protective associations between green and blue spaces with BP were found. This study also found that high population density and public transport accessibility have beneficially significant association with BP. Additionally, with the increase in the distance to the nearest major road, DBP (diastolic BP) decreased rapidly. When the distance was beyond around 200 m, however, there was no obvious change to DBP anymore. By cluster analysis, six clusters of urban exposures were identified. These findings reinforce the importance of improving urban design, which help promote healthy urban environments to optimize human BP health.

Development of season-dependent land use regression models to estimate BC and PM1 exposure

Reliable estimation of exposure to black carbon (BC) and sub-micrometer particles (PM₁) within a city is challenging because of limited monitoring data as well as the lack of models suitable for assessing the intra-urban environment. In this study, to estimate exposure levels in the inner-city area, we developed land use regression (LUR) models for BC and PM₁ based on specially designed mobile monitoring surveys conducted in 2019 and 2020 for three seasons. The daytime and nighttime LUR models were developed separately to capture additional details on the variation in pollutants. The results of mobile monitoring indicated similar temporal variation characteristics of BC and PM₁. The mean concentrations of pollutants were higher in winter (BC: 4.72 μg/m³; PM₁: 56.97 μg/m³) than in fall (BC: 3.74 μg/m³; PM₁: 33.29 μg/m³) and summer (BC: 2.77 μg/m³; PM₁: 27.04 μg/m³). For both BC and PM₁, higher nighttime concentrations were found in winter and fall, whereas higher daytime concentrations were observed in the summer. A supervised forward stepwise regression method was used to select the predictors for the LUR models. The adjusted R² of the LUR models for BC and PM₁ ranged from 0.39 to 0.66 and 0.45 to 0.80, respectively. Traffic-related predictors were incorporated into all the models for BC. In contrast, more meteorology-related predictors were incorporated into the PM₁ models. The concentration surface based on the LUR models was mapped at a spatial resolution of 100 m, and significant seasonal and diurnal trends were observed. PM₁ was dominated by seasonal variations, whereas BC showed more spatial variation. In conclusion, the development of season-dependent diurnal LUR models based on mobile monitoring could provide a methodology for the estimation of exposure and screening of influencing factors of BC and PM₁ in typical inner-city environments, and support pollution management.

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.

Using Street View Imagery to Predict Street-Level Particulate Air Pollution

Land-use regression (LUR) models are frequently applied to estimate spatial patterns of air pollution. Traditional LUR often relies on fixed-site measurements and GIS-derived variables with limited spatial resolution. We present an approach that leverages Google Street View (GSV) imagery to predict street-level particulate air pollution (i.e., black carbon [BC] and particle number [PN] concentrations). We developed empirical models based on mobile monitoring data and features extracted from ∼52 500 GSV images using a deep learning model. We tested theory- and data-driven feature selection methods as well as models using images within varying buffer sizes (50-2000 m). Compared to LUR models with traditional variables, our models achieved similar model performance using the street-level predictors while also identifying additional potential hotspots. Adjusted R² (10-fold CV R²) with integrated feature selection was 0.57-0.64 (0.50-0.57) and 0.65-0.73 (0.61-0.66) for BC and PN models, respectively. Models using only features near the measurement locations (i.e., GSV images within 250 m) explained ∼50% of air pollution variability, indicating PN and BC are strongly affected by the street-level built environment. Our results suggest that GSV imagery, processed with computer vision techniques, is a promising data source to develop LUR models with high spatial resolution and consistent predictor variables across administrative boundaries.