Science of the Total Environment 903 (2023) 166168 Contents lists available at ScienceDirect Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv Beyond here and now: Evaluating pollution estimation across space and time from street view images with deep learning Ricky Nathvani a,b,*,1, Vishwanath D. a,b,1, Sierra N. Clark a,b, Abosede S. Alli c, Emily Muller a,b, Henri Coste a,b, James E. Bennett a,b, James Nimo d, Josephine Bedford Moses d, Solomon Baah d, Allison Hughes d, Esra Suel a,b,e, Antje Barbara Metzler a,b, Theo Rashid a,b, Michael Brauer f, Jill Baumgartner g,h, George Owusu i, Samuel Agyei-Mensah j, Raphael E. Arku c,2, Majid Ezzati a,b,k,2 a Department of Epidemiology and Biostatistics, School of Public Health, Imperial College London, London, UK b MRC Centre for Environment and Health, School of Public Health, Imperial College London, London, UK c Department of Environmental Health Sciences, School of Public Health and Health Sciences, University of Massachusetts, Amherst, USA d Department of Physics, University of Ghana, Accra, Ghana e Centre for Advanced Spatial Analysis, University College London, London, UK f School of Population and Public Health, University of British Columbia, Vancouver, Canada g Institute for Health and Social Policy, McGill University, Montreal, Canada h Department of Epidemiology, Biostatistics, and Occupational Health, McGill University, Montreal, Canada i Institute of Statistical, Social & Economic Research, University of Ghana, Accra, Ghana j Department of Geography and Resource Development, University of Ghana, Accra, Ghana k Regional Institute for Population Studies, University of Ghana, Accra, Ghana H I G H L I G H T S G R A P H I C A L A B S T R A C T • Street-view image based deep learning models can extend pollution estimation. • Image and feature-based models are complimentary in flexibility and interpretability. • Noise and air models use specific fea- tures (e.g. market umbrellas and haze). • Images and sensor networks can broaden pollution monitoring in African cities. • Data collection for model development should prioritise spatial representativeness. A R T I C L E I N F O A B S T R A C T Editor: Anastasia Paschalidou 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 * Corresponding author at: Department of Epidemiology and Biostatistics, School of Public Health, Imperial College London, London, UK. E-mail address: r.nathvani@imperial.ac.uk (R. Nathvani). 1 Joint first authors. 2 Joint senior authors. https://doi.org/10.1016/j.scitotenv.2023.166168 Received 4 May 2023; Received in revised form 7 August 2023; Accepted 7 August 2023 Available online 14 August 2023 0048-9697/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 Keywords: infrastructure for measuring the complex patterns of pollution is limited and modelling may therefore provide Deep learning the most utility. We employed convolutional neural networks (CNNs) for two complementary classification Computer vision models, in both an end-to-end approach and as an interpretable feature extractor (object detection), to estimate Air pollution spatially and temporally resolved fine particulate matter (PM2.5) and noise levels in Accra, Ghana. Data used for Noise pollution training the models were from a unique dataset of over 1.6 million images collected over 15 months at 145 Street-view images representative locations across the city, paired with air and noise measurements. Both end-to-end CNN and Environmental modelling 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. 1. Introduction interpretable image features in the form of object counts, from applying an object-detection CNN to each image. These models were applied to a The urban population in low- and middle-income countries (LMICs) bespoke dataset of over 1.6 million time-lapsed images co-located with increased from 357 million in 1950 to 3.39 billion in 2020, with the PM2.5 and noise measurements, at 145 representative locations over 15 majority of the LMIC population now living in cities (United Nations, months (Clark et al., 2020). Models were trained and evaluated on Department of Economic and Social Affairs, and Population Division, subsets of data specifically to interrogate their temporal and spatial 2019). While cities offer their inhabitants better access to infrastructure, generalisability and in order to compare and contrast strategies for data services and economic opportunity (Ezzati et al., 2018), factors such as collection with fixed resources when developing such models. We road transport and residential and commercial energy generation can further assessed model performance for both the day and night time, also increase hazardous environmental exposures, including air and different seasons, and types of urban land use. noise pollution (Kammen and Sunter, 2016; Kelly and Zhu, 2016). Although some sources of urban pollution in LMICs, such as vehicular 2. Data and methodological context and contributions traffic, are similar to those of many high income countries, there are also differences in the sources, and in their spatial and temporal patterns Some studies have predicted pollution from visual elements of the (Alli et al., 2021; Amegah and Agyei-Mensah, 2017; Clark et al., 2021; environment. Two studies, also from Accra, recorded PM2.5 and PM10 in Deng et al., 2020; Ebare et al., 2011; Weagle et al., 2018; Zhou et al., selected neighbourhoods, in a multi-week measurement campaign 2013) such as seasonal Saharan Desert dust storms (Zhou et al., 2013), (Dionisio et al., 2010; Rooney et al., 2012), together with researcher burning biomass fuels for cooking and heating, and the use of diesel observations and census data on environmental factors, such as biomass generators where there are intermittent electricity outages (Dionisio fuels and unpaved roads, to predict pollution levels. Some studies have et al., 2010). also predicted pollution using remote sensing data, which differs from Data on the patterns of air and noise pollution and their sources our study, not only in the view of the city, but also in spatial and tem- across space and time are needed to identify and evaluate mitigation poral scales and the observable features in images (Sorek-Hamer et al., measures and policies. However, collecting such data is challenging in 2022; Wei et al., 2020; Weigand et al., 2019). resource-constrained settings (Brauer and Guttikunda, 2019; Clark et al., Other studies used terrestrial images for predicting air pollution 2020; Khan et al., 2018). Recent methodological advances in image (Chakma et al., 2017; Feng et al., 2021; Ganji et al., 2020; Gu et al., processing and analysis, particularly in the form of deep convolutional 2019; Hong et al., 2020; Liu et al., 2016; Liu et al., 2015; Qi and Hankey, neural networks, have demonstrated that street-level images can help 2021; Wang et al., 2022; Won et al., 2022; Zhang et al., 2018), and one with predicting air and noise pollution levels (Ganji et al., 2020; Hong for noise (Hong et al., 2020) based on images and pollution measure- et al., 2020; Qi and Hankey, 2021; Weichenthal et al., 2019), contingent ment data though none had spatiotemporally linked image and pollution on initial data measurements needed to develop the image-based data during the night time, as we do. Previously adopted approaches pollution estimation models. So far, image-based pollution models span a variety of experimental configurations, making a unifying, have largely been developed for East Asia (Chakma et al., 2017; Feng quantitative comparison among studies infeasible. The specific metric of et al., 2021; Gu et al., 2019; Liu et al., 2016; Liu et al., 2015; Wang et al., pollution (e.g., black carbon vs PM), timescales on which pollution is 2022; Won et al., 2022; Zhang et al., 2018) and North America (Ganji predicted (single measurement in time vs variation across day), spatial et al., 2020; Hong et al., 2020; Qi and Hankey, 2021), typically based on resolution (city-wide vs local), images used (static vs time-varying), data a few weeks’ observation at selected locations, asynchronous or spatially inputs (solely images vs inclusion of meteorological variables), temporal distant from pollution measurements. Few studies have sought to predict range (<~1 week vs multiple months of observation), synchrony be- spatially and temporally resolved pollution from images, and none in tween data sources (pollution and images <~5 min apart vs >~1 year Africa, the world’s fastest urbanising region (United Nations, Depart- apart), modelling approach (regression of continuous pollution data vs ment of Economic and Social Affairs, and Population Division, 2019). classification into different classes) and model inputs (specific features We developed and evaluated machine learning models to predict vs entire images), vary from study to study. Furthermore, within studies temporally and spatially varying noise and fine particulate matter that used images as model inputs, a variety of features and feature (PM2.5; particles <2.5 μm in diameter, with known human health im- extraction methods (object detection vs segmentation) were used, pacts (Pope and Dockery, 2006)) levels from street-level images in including in relation to stationarity of features in time (e.g., buildings Accra, Ghana. We used deep convolutional neural networks (CNNs), and trees vs vehicles and pedestrians). The majority of studies used a which learn robust and hierarchical feature representations that give single configuration from such choices depending on the available data, them superior performance for many image-processing tasks (Schmid- generating prediction tasks that are easier or more difficult relative to huber, 2015; Gu et al., 2018), in two complementary strategies. The first others. We outline the different experimental setups for previous studies used a CNN, without a priori assumptions on the image features relevant in Appendix Table A. In the specific case of cities in Africa, one study for prediction, and another used gradient boosted machines, applied to used street-view images to predict PM2.5 and NO2 across several cities, 2 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 including Accra (Suel et al., 2022). Data used for model training were 2c) Comparison of model types from 2a) and 2b): Do models perform derived from modelled estimates of annual average pollution level with better on multiple, unseen locations (remaining ~10 % of rotating sites) a model only evaluated, not trained, on data from Accra. when given an abundance of images from a few locations, or fewer Our work advances the state of knowledge in a number of ways. Our images across a variety of locations? dataset is much larger and was collected over a longer duration than For each question, we divided our data into subsets for training and most previous image-based studies, comprising 145 locations and a total testing, as illustrated in Fig. 1. The number of images belonging to each (prior to merging with our pollution data) of 2.1 million images over 15 of the datasets is given in Appendix Table B. months. We co-captured both air pollution and noise data with images in Each panel shows how data from fixed and rotating sites were allo- both day and night time. We predicted air pollution concentrations and cated to training and testing sets, for each question posed in Section 3.2. noise levels at finer classification intervals, i.e. with classes that each For the training sets, indicated in blue, each block was further divided encompass a smaller and more precise range as described in Section 3.3, into a 75–25 split with the latter being used as a validation set during than comparable previous classification-based studies. We systemati- training configuration and hyperparameter determination. Final models cally evaluated both the spatial and temporal generalisability of models were trained on the entire training set (including the validation set) and which is relevant for designing an optimal digital surveillance strategy evaluated on the testing set, indicated in red. and guiding data collection. Our study is unique in the use of both end- to-end CNN (outcome-driven) and object-based (feature-driven) models, 3.3. Modelling which both inform model selection and enhance model interpretability. Finally, to our knowledge, this is the first use of images for predicting For all research questions, we trained models to predict levels of both air and noise pollution in the context of an African city. noise (dBA) or PM2.5 (μg/m3) at a given time (1-min averaged interval) and location from a single image taken within 30 s of pollution mea- 3. Materials and methods surement. We framed the prediction task as a classification problem, i.e. the models predict specific ranges (classes) in which noise and PM2.5 fall 3.1. Data rather than as a continuous value, for two reasons. First, policy targets and guidelines, such as those of the World Health Organization (Basner We collected co-located time-lapsed images at 5-min intervals and and McGuire, 2018; World Health Organization, 2021), tend to be PM2.5 and noise measurements averaged and recorded at 1-min intervals formulated based on discrete levels. Second, a preliminary analysis in a field campaign from April 2019 to June 2020, details of which are indicated that models trained explicitly for classification outperformed described in Appendix A and the study protocol paper (Clark et al., regression models trained for continuous value prediction, as detailed in 2020). We had ten fixed sites where data were collected over 15 months, Appendix C. The classes for noise were: ≤39, 40 to <45, 45 to <50, 50 to and 135 rotating sites where data was collected for one week. The fixed <55, 55 to <60, 60 to <65, 65 to <70, 70 to <75, 75 to <80, ≥ 80 dBA. sites provided information for assessing temporal generalisability of The classes for PM2.5 were: 0 to <5, 5 to <10, 10 to <15, 15 to <20, 20 to models, and both fixed and rotating sites for assessing spatial general- <25, 25 to <30, 30 to <40, 40 to <50, 50 to <100, 100 to <150, ≥ 150 isability. Sites were grouped into four land-use classes: commercial, μg/m3. business, industrial (CBI); informal, mostly high-density, settlements For both forms of pollution, we produced two classification models and slums; formal, mostly low- and medium-density, residential areas; (Fig. 2). The first, referred to as end-to-end classification, used an entire and “other” areas that are often peri-urban or rural, and can have dense unprocessed image, with red, green and blue pixel channels, as input to a vegetation (i.e., forest, grassland) or barren land (i.e., sand, soil, dirt). CNN to predict pollution class. No assumptions were made on relevant The classes for each fixed site are detailed in Appendix Table B. image features, which were learned from the data. The second group of models used counts of objects detected from images as input for classi- 3.2. Research questions fication via gradient boosted machines (Friedman, 2001) (GBM). Other approaches to feature extraction from images, such as semantic seg- We developed two types of models that used images to predict noise mentation, could also have been employed to provide model inputs for and air pollution. We analysed how well our models’ prediction gener- pollution estimation, as used in a North American study (Qi and Hankey, alise across time and space, through the following research questions: 2021). We used objects in our second approach since the data needed to 1a) Temporal generalisability: How well do models trained on im- train a model, namely objects, were less resource intensive to generate ages taken from a single location predict noise and PM2.5 at different, within our bespoke dataset with bounding boxes (Nathvani et al., 2022) random times at the same location? as compared with pixel-level annotation, which may also be explored in 1b) Spatial generalisability: How well do models trained in 1a), future work. The object counts were obtained from training an object which are based on a single location, generalise to another unseen detection CNN, described in detail in previous work (Nathvani et al., location? 2022), for object categories relevant to the local environmental context: 2a) Spatial generalisability: How well do models trained using an persons, market vendor (a person carrying a container over their heads abundance (~1,000,000 total across sites) of images from a set of nine which is a common scene in African markets), car, taxi, pick-up truck, (long-term) fixed sites, predict noise and PM2.5 at the remaining (10th) bus, lorry, van, tro-tro (mini buses used for public transportation), unseen location? motorcycle, bicycle, market stall, loudspeaker, umbrella (commonly 2b) Spatial generalisability: How well can models trained using used to protect market and roadside vendors from the sun and rain), fewer images (~100,000 total across sites) from ~90 % of our 135 cookstove, cooking pot/bowl (which frequently contain wares for sale in rotating sites, predict noise and PM2.5 at the remaining ~10 % of sites? the marketplace), food, trash, (piece of) debris, and animal. All object The fixed sites, due to their extended data collection period, categories are those which may vary over time at a given place, since comprised seven times as much data as rotating sites in total. Since in- although other static features, such as buildings or trees, may also affect situ pollution measurements are resource intensive, especially in quan- noise and air pollution, their unchanging presence over daily timescales tities needed to train a CNN (Sun et al., 2017), there is a need to opti- is less informative for predicting temporal variation in pollution at a mally allocate the use of cameras and pollution measurement hardware, single location (such as those models developed in 1a and 1b). The ac- as well as personnel time. Therefore we also investigated whether curacy with which these objects could be detected in our images is given models trained using more data from a smaller number of (fixed) sites, or in Appendix Table C and described in previous work (Nathvani et al., fewer data from a greater number of (rotating) sites led to more spatially 2022). In this analysis, we did not use counts of cookstove, loudspeakers, generalisable CNN models: market vendors or buses, due to their sparse presence in our data (<10 3 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 Fig. 1. Data use for training and testing of models. Fig. 2. End-to-end (CNN) and object-based (GBM) modelling approaches. counts of each in 2.1 million images). The end-to-end and feature-driven pollution value was recorded within +/− 30 s of image capture. Where approaches are complementary with respect to flexibility and feature two cameras were placed at a site, both images are assigned pollution agnosticism versus prior assumption and interpretability (Zhang and data based on this procedure. Some images did not have corresponding Zhu, 2018). pollution values due to a lack of measurements when monitors failed or were unstable. For noise, 83–98 % of fixed site images and 99 % of rotating site images were assigned pollution data. For PM 3.4. Data preparation 2.5 , 68–89 % of fixed site images and 79 % of rotating site images were assigned pollution data. Full details are provided in Appendix Fig. C. As We prepared our data in the following manner for the purpose of mentioned in section 3.3 we applied a previously developed object training and evaluating both CNN and object-GBM models. First, due to detection CNN to all 2.1 million images in our unmerged data set to the Covid-19 pandemic and associated lockdown in Accra from March obtain information on the counts of different object categories within 30th to April 20th 2020, we excluded images and pollution data from each image, which are used as numerical inputs for our GBM models. March 23rd to May 11th 2020, when we were unable to attend to the Examples of the detected objects within our images may be seen in regular maintenance of monitoring hardware, and therefore data Appendix Fig. D. collection was incomplete and uneven across sites. 1.6 million images were assigned corresponding PM values and A small number of cameras experienced internal failure of their 2.5 1.9 million images with noise values. Each dataset was divided into clocks, resetting to a factory default of January 2017 at the start of their training, validation and test sets, as shown in Fig. 1. The test set was 10 deployment, which led to images recorded with incorrect timestamps. % of all data. Training and validation sets (a 25 % holdout of the We corrected the timestamps for these images by re-assigning the initial remainder of data) were used to train the model and set hyper- timestamp based on the start of the monitoring period, which was parameters. After determining hyperparameters, the final models were recorded on a log-sheet when visiting every site. Since each image trained on the combined training and validation sets and evaluated on thereafter was captured at regular five minute intervals, subsequent the test set. Both CNN and GBM models used the exact same sets of images were assigned time stamps at five minute intervals. Finally, a images, with the former using the entire image and the latter the counts small fraction (<1 %) of images and pollution data were corrupted and of objects in each image. hence unreadable by code. These data were excluded. Images and pollution data were combined by assigning each image the pollution observation nearest in time, with a requirement that the 4 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 3.5. Model training 3.6. Model evaluation 3.5.1. End-to-end CNN We compared the performance of both end-to-end and feature-driven We used a Residual Network with 101 layers (ResNeXt 101) (Xie approaches to infer plausible contributors to how they predict pollution. et al., 2017) as the convolutional neural network (CNN) architecture for We calculated classification accuracy for exact class prediction as well as classifying noise and PM2.5 levels from an entire image. The algorithm for when the model classified into the ground truth or adjacent class was implemented and trained in PyTorch (Paszke et al., 2019) and was (shown as “same and ±1 class accuracy”). We evaluated all our models pretrained on ImageNet data to enable the CNN to recognise low level against a null model which measures whether the models do better than features, e.g., edges, which improved model performance, as seen in simply taking the average from a distribution of training data. We also other computer vision tasks (Huh et al., 2016). calculated the models’ accuracy for specific subsets of data under During training, CNNs were given images resized to 224 ⨉ 224 different environmental conditions, including day and night time, and pixels with layer depths of 3 to accommodate the Red, Green and Blue the dry and dusty Harmattan season (November–February). During the channels, with Z-score normalisation applied across all images to assist Harmattan season dust from the Saharan Desert is carried by trade winds with the gradient descent process of learning (LeCun et al., 2012). A (Adetunji et al., 1979), and there is haze and “redness”(Adetunji et al., modified cross-entropy loss function with a log-barrier constraint (Bel- 1979; Anuforom, 2007; Ette and Olorode, 1988; McTainsh, 1980; Ochei harbi et al., 2020) was used to account for the ordinal nature of pollution and Adenola, 2018; Pinker et al., 1994), caused by absorption and classes, as described in Appendix B. Training was performed on two scattering of light (Groblicki et al., 1981; Waggoner and Weiss, 1980) NVIDIA Quadro RTX 6000 GPUs (48GB memory), with models taking and the dust itself, which has a red-brown colour (Breuning-Madsen and approximately 1 h per epoch and lasted for 30 epochs. The batch size Awadzi, 2005; Lafon et al., 2004). These changes in visibility can inform was 32 images, using stochastic gradient descent with an initial learning air pollution estimation (Hyslop, 2009; Ozkaynak et al., 1985). rate of 0.001, a momentum of 0.9 and a step size of 40. Final models For GBM models in 1a, we quantified the importance of each object were those which performed best on the validation set during the for prediction via its permutation importance, which calculates the training process, ranging from the 16–25th epoch. reduction in the model’s accuracy on the test set, before and after At training time, data augmentation was used to improve model randomly shuffling the values of an input feature (in our case, counts in a generalisability and mitigate overfitting to the data (Shorten and given object category) across images. Object counts in our data were Khoshgoftaar, 2019), by uniformly, randomly cropping the image bor- correlated among different object categories across images (Nathvani ders, with the central 90 % area of the image always preserved, random et al., 2022), e.g., people and cars. Therefore a given object’s importance rotations of the image between 10◦ anti-clockwise to 10◦ clockwise, and score might be lower when multiple objects are used for prediction than evenly random flipping of the image in the horizontal plane. These when single objects are used, because other correlated objects capture transformations correspond to the variance seen between camera images some of the same information. and the placement at different sites, which had different fields of view and camera orientations. 4. Results and discussion 3.5.2. Object-based gradient boosted machines For noise, classification accuracy at different times in the same We used gradient boosting machines or GBMs as the algorithm for location (i.e. Question 1a) for both the end-to-end (CNN) and object- classifying noise and PM2.5 from specific, interpretable features, which based (GBM) models ranged from 40 to 70 % across sites, consider- were counts of objects detected within each image from a separate CNN ably outperforming their null models (Fig. 3). Accuracy increased to (Nathvani et al., 2022). GBMs are ensemble tree-based models which use 80–90 % for neighbouring (±1) class classification. The performance of “boosting”, i.e. adaptively changing the weights of data points in the the two models was similar, with CNNs slightly outperforming GBM training distribution during the learning process to improve perfor- models. Predictions at sites with high road-traffic, such as Asylum Down, mance on less easily predicted data (Friedman, 2001), which were Tema Motorway and N1 West Motorway, had higher accuracy than implemented XGBoost (Chen and Guestrin, 2016) in Python and Scikit- those at other sites. Noise predictions using CNN models were often Learn (Pedregosa et al., 2011). GBMs have high efficacy across many more accurate in the daytime (59.9 % average classification accuracy problem domains with structured data inputs, due to their ability to across all sites) than night time (49.8 %) (Appendix Fig. E) which may learn non-linear relationships between features with robustness to out- result from predictive features such as people, traffic and marketplace liers in a flexible and scalable manner (Chen and Guestrin, 2016). They indicators (e.g., umbrellas) being present, and more visible in the day, as offer advantages compared to linear models, which are more biased in in Appendix Fig. D, since street lighting conditions vary across our sites. complex data domains, and computationally expensive models such as PM2.5 classification had lower accuracy than that of noise in most support vector machines, and artificial neural networks which are model and site combinations (Fig. 3). There was also a larger discrep- typically more cumbersome to optimise. Furthermore, in a preliminary ancy between the predictive performance of CNN and GBM models for analysis, GBMs had better performance, as measured by classification PM2.5, with CNN models achieving 30–55 % classification accuracy and accuracy, than comparable tree-based methods such as decision trees GBM models 15–25 %, though both outperformed null model bench- and random forests. marks. Sites with higher classification accuracy for noise performed less The input to the GBM models were vectors representing the counts of well for air, and vice versa; for example, the three poorest performing different objects in each image, e.g., (cars: 2, people: 3, umbrellas: 0…). sites for noise (University of Ghana, Ashaiman and East Legon) had the The model hyperparameters were determined with Bayesian optimisa- greatest accuracy for CNN models when predicting PM2.5. Accuracy of tion, with the validation set being used for fixed sites’ data and with 3- CNN models reached 70–90 % for neighbouring class classification, and fold cross validation when training on 9 folds of rotating site data that of GBM models 30–50 %. Unlike noise, performance of PM2.5 (Fig. 1) and a cross-entropy loss function. The search range for the pa- classification using CNN models differed little between day (40.2 % rameters is given in Appendix Table D. Training was performed during average classification accuracy across all sites) and night time (39.6 %) the Bayesian hyperparameter tuning process and was stopped when 5 and had higher accuracy during the Harmattan period (57.7 %) than in iterations of tuning yielded no further improvements in overall class other times (35.1 %) (Appendix Fig. E), whereas GBM models had no prediction accuracy on the independent validation set. consistent advantage during the Harmattan (17.5 %) than in other times (19.9 %). When trained at one fixed site and tested at a different fixed site (Question 1b), accuracy dropped compared with same-site testing and 5 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 Fig. 3. The classification accuracy achieved by CNN and GBM models trained and tested on images from the same fixed site (Question 1a) is shown for noise and PM2.5. CNN models for noise and PM2.5 performed similar to null model 4.1. Object feature importance benchmarks, demonstrating the inability of models to generalise from the measurements at a single site (Fig. 4). For noise, the variation in In GBM models developed in 1a, cars, people, taxis, umbrellas and accuracy of GBM models was greater than that of CNNs, but on average tro-tros contributed most to predictions for both noise and PM2.5 achieved greater improvement over null model benchmarks (+2.4 %) (Fig. 6). For PM2.5, debris and trucks also contributed to prediction ac- than did CNNs (+0.9 %). For PM2.5 accuracy ranged 7–20 % for both curacy. These object categories were frequently detected in fixed site GBM and CNN models, with the latter achieving greater improvement images. We calculated the Spearman correlation between object counts over the null model benchmarks (+0.9 %) than did GBMs (+0.7 %). In and noise and PM2.5 levels across images at each fixed site (Appendix all cases, accuracy and null model performance broadly increased with Table E) in order to test whether this correlation explained an object’s Bhattacharya coefficient, a measure of the similarity of pollution dis- permutation importance. The explained variance, calculated as the tributions between training and testing site (Bhattacharyya, 1943). square of Pearson correlation between object-pollution correlations and Training on nine fixed sites with abundant (~1,000,000) data the permutation importance for each object was 0.86 for noise and 0.76 (Question 2a) produced similar results for generalising to a single, un- for PM2.5. Heuristically, the greater the correlation of an objects’ counts seen fixed site as for models trained at single fixed sites (Appendix Fig. with that of pollution, the greater its contribution to model prediction F). For noise, accuracies were at best 30–40 % for both CNN and GBM accuracy, which may also indicate why noise models in 1a performed models, and in some instances similar to the null model; neither CNN nor better than those for PM2.5. Since objects visible in the images had GBM models had a distinct advantage. For PM2.5, both CNN and GBM greater impact on the accuracy of noise prediction compared to PM2.5, model accuracy remained similar to null model performance. CNN and CNN models for noise may have learned to rely on the same features as GBM models trained using fewer images (~100,000) from ~90 % of those used by the GBM models prediction. This may in turn explain the rotating sites (121–122 sites) for classifying noise and PM2.5 at the similar performance between the two models in 1a, which we further remaining rotating sites (Questions 2b) outperformed their respective examine in the section below. null models (Fig. 5). As in temporal transferability (Question 1a), models As shown in Fig. 6, the objects with the highest permutation performed better in classifying noise than PM2.5, but the advantage of importance were various types of vehicles, consistent with previous CNN over GBM models disappeared. Noise models had 25 % accuracy research on the significant contributions from road traffic to both air and for exact class and 65 % when allowing for neighbouring class classifi- noise pollution (Dionisio et al., 2010; Onuu, 2000; Rooney et al., 2012). cation, with little variance (± 2.5 %) between folds, whilst PM2.5 models In addition, umbrellas were frequently present in images due to their had 17.5 % accuracy, and 47.5 % allowing for neighbouring class clas- extended use in the daytime to protect market vendors and their sification. For noise, but not for PM2.5, CNN accuracy for daytime images merchandise from the sun and rain. Markets also attract high levels of was significantly higher than on night time images (Appendix Fig. G). vehicular traffic (Agyapong and Ojo, 2018), people (Asante, 2020; CNN models had similar performance across all land use categories for Asante and Mills, 2020), and roadside cooking and food vending, which both forms of pollution (Appendix Fig. G). collectively increase noise and air pollution (Alli et al., 2021; Clark et al., When comparing the approaches in Questions 2a (fewer sites with 2021). Furthermore, for PM2.5 models, debris had higher than average more data per sites) and 2b (more sites with fewer data per sites) with feature importance. Although not a source, debris is more visible and consistent test data (Question 2c), models trained on a smaller amount readily detected by our object detection algorithm during daylight of data from many (rotating) sites performed better than those trained hours, serving as a proxy for time of day, and for sites with diurnal on many times more data from a smaller number of (fixed) sites, for both patterns of PM2.5. In addition, debris may be more visible when unob- noise and air pollution models (Appendix Fig. H). This suggests that with scured by other objects, acting as an implicit indicator for a lack of finite monitoring capacity, a diversity of locations for data gathering is crowds or traffic, and instances where the road surface, from which dust more likely to produce spatial generalisability for pollution prediction particles may be resuspended, is exposed. using CNN models than long-term capture of data at fewer locations, highlighting the importance of optimising the spatial as well as temporal representativeness of data within cities for pollution modelling, as seen 4.2. Harmattan influence in other domains of computer vision (Schat et al., 2020). To probe why CNN models in 1a performed better during the Harmattan season, when PM2.5 levels were much higher, we derived characteristics of images related to changes in hue and haze between 6 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 Fig. 4. The classification accuracy achieved by the CNN (left) and GBM (right) models trained and tested from images at one fixed site and tested at a different fixed site (Question 1b) for both noise (top) and PM2.5 (bottom) prediction. Points are coloured by the Bhattacharya coefficient between the pollution distributions be- tween the training and testing sites, which is a measure of the overlap between the distributions. Data points with a star indicate testing and training performed at the same site, as in 1a. All null models are from the fixed site used for training. Below each scatterplot the relative improvements in classification accuracy over the null model accuracy is given for each data point (i.e. the vertical distance between the round points and the dashed diagonal line in the scatterplot); the purple dashed line shows the average across all data points, illustrating whether models achieved improvement over their benchmarks overall. Harmattan and non-Harmattan periods, since previous work has have also used these features (Feng et al., 2021; Wang et al., 2022; Liu demonstrated that changes due to Harmattan dust, such as an increase in et al., 2015; Ganji et al., 2020) and we therefore created feature metrics “redness” and haze are indicators and predictive factors for pollution which relate to qualities of hue and haze in our images in order to study (Adetunji et al., 1979; Anuforom, 2007; Ette and Olorode, 1988; our models. For daytime images, we compared mean pixel intensity in McTainsh, 1980; Ochei and Adenola, 2018; Pinker et al., 1994), as well each colour channel between these periods: red, green and blue. For as light scattering from dust (Groblicki et al., 1981; Waggoner and night time images in single-channel grayscale, we used mean pixel in- Weiss, 1980). Other approaches for predicting pollution from images tensity and pixel intensity standard deviation (SD). Red pixel intensity 7 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 Noise (dBA) PM2.5 (μg/m 3) Fig. 5. Classification accuracy achieved by CNN and GBM models trained and tested on images from rotating sites (Question 2b) for both noise and PM2.5 prediction. Accuracies are shown as the average over the folds of training data, as shown in Fig. 1. The bars show the standard deviation of the accuracy across different folds. Fig. 6. Permutation importance for each object used as inputs to the noise and PM2.5 GBM models in Question 1a, for each fixed site. Permutation importance is calculated on the test set for each model, as shown in Fig. 1. was greater during the Harmattan period, whilst the opposite was seen correspondence between redness of image and air pollution, the better for blue (Appendix Fig. I). Furthermore, night time pixel intensity was the model performed on average. higher at eight of ten sites during Harmattan, while pixel SD was lower at seven sites. To infer to what extent this information was used by our 4.3. Model interpretability across time and space CNN models for PM2.5 classification, we calculated Spearman correla- tions between mean red and blue pixel intensity and each image’s Our results suggest that specific features, such as the objects selected associated PM2.5 value (Appendix Table F). The average across sites was for the GBM model, are better attuned at predicting noise, whilst pre- 0.11 for red pixel intensity and − 0.21 for blue pixel intensity. Similarly, dictive performance for PM2.5 is somewhat improved by leveraging grayscale pixel intensity tended to be positively correlated with air more complex visual features from images such as red pixel channel pollution, 0.10, whilst grayscale pixel SD was negatively correlated, intensity and haziness. This is supported by the discrepancy between day − 0.18, consistent with light sources appearing more diffuse in hazy and night image accuracy for CNN models for noise in 1a and 2b, which conditions due to light scattering. The magnitude of these correlations may result from objects being more visible and present in the day than was also greater during the Harmattan period, during which CNN night, allowing CNN models to make more accurate predictions of noise, models in 1a had greater accuracy (Appendix Table F). The Pearson but not PM2.5, using daytime images. correlation across all 10 sites between the aforementioned Spearman These observations also suggest that features learned by the CNN for correlations for red pixel intensity and the accuracy of a CNN model at noise classification are likely similar to the objects used by the GBM that site in 1a, was 0.73. In heuristic terms, the greater the models, whilst being of less importance for CNN models for PM2.5. To 8 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 investigate this, we generated a sample of gradient class activation maps and PM2.5 with either CNN or GBM models had moderate performance. (Selvaraju et al., 2017) (Grad-CAMs) for our CNN models in 1a. Grad- No model surpassed 80 % classification accuracy, and performance was CAMs use gradient descent to work backwards from a network’s class considerably lower when testing on previously unseen locations. This prediction from a given image to the regions in the image itself which may be due to a combination of factors including the static nature of most contributed to that prediction. Fig. 7 shows that noise models focus images, which may fail to capture transitory sources, e.g., emergency either on visible objects or the location of main thoroughfares, whilst vehicle sirens, compared with exposures which are averaged over one PM2.5 models tend to focus on either fixed features of the built envi- minute’s observation. Furthermore some pollution sources (and pre- ronment, or the sky, supporting the possibility of the object-driven na- dictors) are non-local or out of the field of view of the corresponding ture of noise models and the reliance of PM2.5 CNN models on complex image’s camera. Since the models developed in questions 1b to 2c rely visual features. This may be due to the fact that noise is transitory and on a single images from consumer-grade digital cameras to predict spatially linked to sources or their proxies such as market stalls and their pollution in an unseen location, our dataset and methodology also in- associated umbrellas. By contrast, PM2.5 persists at a location longer forms on the viability of pollution modelling from comparable camera than its sources, and is composed of both local (e.g., traffic) and non- technology, including CCTV networks which are increasingly deployed local (dust, neighbouring regions’ emissions) sources. in African cities and mobile phone camera capture, as has been used elsewhere in the literature. In addition, we intentionally tested models 5. Conclusions under the challenging condition of estimating pollution from a single image in time and space, in order to independently and conservatively We used a unique dataset of co-located images and pollution mea- assess the additional benefit images may confer above data extrapola- surements, and two different modelling approaches, to investigate how tion. Future work could improve model accuracy by making use of CNN images predict spatially and temporally resolved noise and air pollution architectures with multiple image inputs across time, or in the case of measurements in a major city in Africa. Most models developed in this feature-driven prediction, with object counts across a series of time- work surpassed null model accuracy baselines, indicating that models points prior to the moment in time whose pollution levels are esti- can learn information latent to images, beyond extrapolation from mated. Similarly, image and/or object counts from neighbouring sites outcome data. Models had similar performance across different land-use might also help predict pollution at a given location. These may prove categories within the city and across both day and night time, with a especially beneficial for the prediction of PM2.5, whose presence is more slight advantage to model accuracy in the daytime for noise. However, persistent than many of its transitory sources, such as travelling cars. even when training and testing models at single sites, classifying noise Where pollution sources are complex and vary widely across small Fig. 7. Grad-CAM visualisation heat maps for four fixed sites of different land-use categories, obtained from CNN models. The highlighted regions in each image (red) indicate the features most salient to the model’s prediction, for that image. The same image is shown as used by the corresponding noise and PM2.5 models, developed in Question 1a. 9 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 spatial scales, the inclusion of data from many sites improved the spatial George Thurston for their comments, references and insight. generalisability of CNN models in comparison to an abundance of data from a small number of locations. Our results highlight the importance Appendix A. Supplementary data of optimising the spatial as well as temporal representativeness of data gathered within cities for pollution modelling, as seen in other domains Supplementary data to this article can be found online at https://doi. of computer vision. Although spatially representative data improved org/10.1016/j.scitotenv.2023.166168. model performance, for CNNs model accuracy also varied under particular times of day and seasons, related to time-varying environ- References mental factors. We also find that models are capable of making comparably accurate estimates for night time air and noise pollution, Adetunji, J., McGregor, J., Ong, C.K., 1979. Harmattan Haze. Weather 34, 430–436. particularly for PM estimation, when such data is gathered and used https://doi.org/10.1002/j.1477-8696.1979.tb03389.x. 2.5 Agyapong, F., Ojo, T.K., 2018. Managing traffic congestion in the Accra central market, for modelling, alongside corresponding images. This shows the need for Ghana. J. Urban Manag. 7, 85–96. https://doi.org/10.1016/j.jum.2018.04.002. temporally diverse, paired pollution and image data that capture urban Alli, A.S., Clark, S.N., Hughes, A., Nimo, J., Bedford-Moses, J., Baah, S., Wang, J., environmental change on both short ( 1h) and long (~1 year) Vallarino, J., Agyemang, E., Barratt, B., Beddows, A., Kelly, F., Owusu, G., < Baumgartner, J., Brauer, M., Ezzati, M., Agyei-Mensah, S., Arku, R.E., 2021. Spatial- timescales. temporal patterns of ambient fine particulate matter (PM2.5) and black carbon (BC) Overall, our results show that inference of noise and PM2.5 from pollution in Accra. Environ. Res. Lett. 16, 074013 https://doi.org/10.1088/1748- imagery is a feasible but challenging task, especially as the spatial and 9326/ac074a. Amegah, A.K., Agyei-Mensah, S., 2017. Urban air pollution in sub-Saharan Africa: time temporal scale of prediction becomes smaller, which is relevant for for action. Environ. Pollut. Barking Essex 1987 (220), 738–743. https://doi.org/ detailed policy formulation, such as dynamic smart congestion pricing, 10.1016/j.envpol.2016.09.042. and evaluation of impacts on human exposure. Therefore, accurate and Anuforom, A.C., 2007. Spatial distribution and temporal variability of Harmattan dust haze in sub-Sahel West Africa. Atmos. Environ. 41, 9079–9090. https://doi.org/ generalisable estimates of short timescale pollution in cities continues to 10.1016/j.atmosenv.2007.08.003. require primary data collection at representative and diverse scales to Asante, L.A., 2020. Urban governance in Ghana: the participation of traders in the support these efforts. redevelopment of Kotokuraba market in Cape Coast. Afr. Geogr. Rev. 39, 361–378. https://doi.org/10.1080/19376812.2020.1726193. Asante, L.A., Mills, R.O., 2020. Exploring the socio-economic impact of COVID-19 Author contributions pandemic in marketplaces in urban Ghana. Afr. Spectr. 55, 170–181. https://doi. org/10.1177/0002039720943612. RN, VDG, SNC, EM, MB, ES and ME conceptualised the study. SNC, Basner, M., McGuire, S., 2018. WHO environmental noise guidelines for the European region: a systematic review on environmental noise and effects on sleep. Int. J. EM, ASA, JN, JBM, SB, AH, REA and ME designed and implemented the Environ. Res. Public Health 15, E519. https://doi.org/10.3390/ijerph15030519. field campaign to collect data. RN, VDG, JEB, HC, and ME developed Belharbi, S., Ayed, I.B., McCaffrey, L., Granger, E., 2020. Non-parametric Uni-modality analytical methods. RN and VDG implemented methods and conducted constraints for deep ordinal classification. ArXiv191110720 Cs stat. Bhattacharyya, A., 1943. On a measure of divergence between two statistical populations analyses. RN, VDG, SNC, and ME developed the presentation of results. defined by their probability distributions. Bull Calcutta Math Soc 35, 99–109. RN, VDG and ME wrote the original draft. RN, VDG, EM, ES, JB, ABM, Brauer, M., Guttikunda, S.K.K.A.N., Dey, S., Tripathi, S.N., Weagle, C., Martin, R.V., MB, AH, TR, EG, SM, JEB, EA, GO, SA-M and REA provided input and 2019. Examination of monitoring approaches for ambient air pollution: a case study for India. Atmos. Environ. 216, 116940 https://doi.org/10.1016/j. revisions. AH, SA-M, REA and ME supervised data collection and atmosenv.2019.116940. analysis. Breuning-Madsen, H., Awadzi, T.W., 2005. Harmattan dust deposition and particle size in Ghana. CATENA 63, 23–38. https://doi.org/10.1016/j.catena.2005.04.001. Chakma, A., Vizena, B., Cao, T., Lin, J., Zhang, J., 2017. Image-based air quality analysis Funding sources using deep convolutional neural network, in: 2017 IEEE international conference on image processing (ICIP). In: Presented at the 2017 IEEE International Conference on This work was supported by the Pathways to Equitable Healthy Cities Image Processing (ICIP), pp. 3949–3952. https://doi.org/10.1109/ ICIP.2017.8297023. grant from the Wellcome Trust [209376/Z/17/Z]. This work was also Chen, T., Guestrin, C., 2016. XGBoost: A Scalable Tree Boosting System, in: Proceedings supported by a GCRF Digital Innovation for Development in Africa of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and network grant from UKRI [EP/T029145/1]. SC, ABM and TR were Data Mining, KDD ‘16. Association for Computing Machinery, New York, NY, USA, supported by the Imperial College President’s PhD scholarship. SC was pp. 785–794. https://doi.org/10.1145/2939672.2939785. Clark, S.N., Alli, A.S., Brauer, M., Ezzati, M., Baumgartner, J., Toledano, M.B., Hughes, A. supported by a Canadian Institutes for Health Research (CIHR) Foreign F., Nimo, J., Moses, J.B., Terkpertey, S., Vallarino, J., Agyei-Mensah, S., Study Doctoral Scholarship. For the purpose of Open Access, the author Agyemang, E., Nathvani, R., Muller, E., Bennett, J., Wang, J., Beddows, A., Kelly, F., has applied a CC BY public copyright license to any Author Accepted Barratt, B., Beevers, S., Arku, R.E., 2020. High-resolution spatiotemporal measurement of air and environmental noise pollution in sub-Saharan African cities: Manuscript version arising from this submission. pathways to equitable health cities study protocol for Accra, Ghana. BMJ Open 10, e035798. https://doi.org/10.1136/bmjopen-2019-035798. Declaration of competing interest Clark, S.N., Alli, A.S., Nathvani, R., Hughes, A., Ezzati, M., Brauer, M., Toledano, M.B., Baumgartner, J., Bennett, J.E., Nimo, J., Bedford Moses, J., Baah, S., Agyei- Mensah, S., Owusu, G., Croft, B., Arku, R.E., 2021. Space-time characterization of The authors declare that they have no known competing financial community noise and sound sources in Accra, Ghana. Sci. Rep. 11, 11113. https:// interests or personal relationships that could have appeared to influence doi.org/10.1038/s41598-021-90454-6. Deng, L., Kang, J., Zhao, W., Jambrošić, K., 2020. Cross-National Comparison of the work reported in this paper. soundscape in urban public open spaces between China and Croatia. Appl. Sci. 10, 960. https://doi.org/10.3390/app10030960. Data availability Dionisio, K.L., Rooney, M.S., Arku, R.E., Friedman, A.B., Hughes, A.F., Vallarino, J., Agyei-Mensah, S., Spengler, J.D., Ezzati, M., 2010. Within-neighborhood patterns and sources of particle pollution: mobile monitoring and geographic information Our analysis code, trained models, object count data and site meta- system analysis in four communities in Accra, Ghana. Environ. Health Perspect. 118, data can be downloaded from http://globalenvhealth. 607–613. https://doi.org/10.1289/ehp.0901365. Ebare, M.N., Omuemu, V.O., Isah, E.C., 2011. Assessment of noise levels generated by org/code-data-download/ and http://equitablehealthycities. music shops in an urban city in Nigeria. Public Health 125, 660–664. https://doi. org/data-download/ upon publication of the paper. Requests for re- org/10.1016/j.puhe.2011.06.009. analysis of images should be sent to the corresponding authors. Ette, A.I.I., Olorode, D.O., 1988. Technical note the effects of the harmattan dust on air conductivity and visibility at Ibadan, Nigeria. Atmospheric Environ. 1967 (22), 2625–2627. https://doi.org/10.1016/0004-6981(88)90499-4. Acknowledgement Ezzati, M., Webster, C.J., Doyle, Y.G., Rashid, S., Owusu, G., Leung, G.M., 2018. Cities for global health. BMJ 363, k3794. https://doi.org/10.1136/bmj.k3794. We thank Giulia Mangiameli and Abeer Arif for project management and coordination of activities and Professor John Spengler and Professor 10 R. Nathvani et al. S c i e n c e o f t h e T o t a l E n v i r o n m e n t 903 (2023) 166168 Feng, L., Yang, T., Wang, Z., 2021. Performance evaluation of photographic Rooney, M.S., Arku, R.E., Dionisio, K.L., Paciorek, C., Friedman, A.B., Carmichael, H., measurement in the machine-learning prediction of ground PM2.5 concentration. Zhou, Z., Hughes, A.F., Vallarino, J., Agyei-Mensah, S., Spengler, J.D., Ezzati, M., Atmos. Environ. 262, 118623 https://doi.org/10.1016/j.atmosenv.2021.118623. 2012. Spatial and temporal patterns of particulate matter sources and pollution in Friedman, J.H., 2001. Greedy function approximation: a gradient boosting machine. four communities in Accra, Ghana. Sci. Total Environ. 435–436, 107–114. https:// Ann. Stat. 29, 1189–1232. https://doi.org/10.1214/aos/1013203451. doi.org/10.1016/j.scitotenv.2012.06.077. Ganji, A., Minet, L., Weichenthal, S., Hatzopoulou, M., 2020. Predicting traffic-related air Schat, E., van de Schoot, R., Kouw, W.M., Veen, D., Mendrik, A.M., 2020. The data pollution using feature extraction from built environment images. Environ. Sci. representativeness criterion: predicting the performance of supervised classification Technol. 54, 10688–10699. https://doi.org/10.1021/acs.est.0c00412. based on data set similarity. PLoS One 15, e0237009. https://doi.org/10.1371/ Groblicki, P.J., Wolff, G.T., Countess, R.J., 1981. Visibility-reducing species in the journal.pone.0237009. Denver “brown cloud”—I. relationships between extinction and chemical Schmidhuber, J., 2015. Deep learning in neural networks: an overview. Neural Netw. 61, composition. Atmospheric environ. 1967. Plumes and Visibility Measurements and 85–117. https://doi.org/10.1016/j.neunet.2014.09.003. Model Components-Supplement 15, 2473–2484. https://doi.org/10.1016/0004- Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D., 2017. Grad- 6981(81)90063-9. CAM: Visual explanations from deep networks via gradient-based localization, in: Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai, B., Liu, T., Wang, X., Wang, G., 2017 IEEE international conference on computer vision (ICCV). In: Presented at the Cai, J., Chen, T., 2018. Recent advances in convolutional neural networks. Pattern 2017 IEEE International Conference on Computer Vision (ICCV), pp. 618–626. Recogn. 77, 354–377. https://doi.org/10.1016/j.patcog.2017.10.013. https://doi.org/10.1109/ICCV.2017.74. Gu, K., Qiao, J., Li, X., 2019. Highly efficient picture-based prediction of PM2.5 Shorten, C., Khoshgoftaar, T.M., 2019. A survey on image data augmentation for deep concentration. IEEE Trans. Ind. Electron. 66, 3176–3184. https://doi.org/10.1109/ learning. J. Big Data 6, 60. https://doi.org/10.1186/s40537-019-0197-0. TIE.2018.2840515. Sorek-Hamer, M., Von Pohle, M., Sahasrabhojanee, A., Akbari Asanjan, A., Deardorff, E., Hong, K.Y., Pinheiro, P.O., Weichenthal, S., 2020. Predicting outdoor ultrafine particle Suel, E., Lingenfelter, V., Das, K., Oza, N.C., Ezzati, M., Brauer, M., 2022. A deep number concentrations, particle size, and noise using street-level images and audio learning approach for meter-scale air quality estimation in urban environments data. Environ. Int. 144, 106044 https://doi.org/10.1016/j.envint.2020.106044. using very high-spatial-resolution satellite imagery. Atmosphere 13, 696. https:// Huh, M., Agrawal, P., Efros, A.A., 2016. What Makes ImageNet Good for Transfer doi.org/10.3390/atmos13050696. Learning? ArXiv160808614 Cs. Suel, E., Sorek-Hamer, M., Moise, I., von Pohle, M., Sahasrabhojanee, A., Asanjan, A.A., Hyslop, N.P., 2009. Impaired visibility: the air pollution people see. Atmos. Environ. Arku, R.E., Alli, A.S., Barratt, B., Clark, S.N., Middel, A., Deardorff, E., Atmospheric Environment - Fifty Years of Endeavour 43, 182–195. https://doi.org/ Lingenfelter, V., Oza, N.C., Yadav, N., Ezzati, M., Brauer, M., 2022. What you see is 10.1016/j.atmosenv.2008.09.067. what you breathe? Estimating air pollution spatial variation using street-level Kammen, D.M., Sunter, D.A., 2016. City-integrated renewable energy for urban imagery. Remote Sens. 14, 3429. https://doi.org/10.3390/rs14143429. sustainability. Science 352, 922–928. https://doi.org/10.1126/science.aad9302. Sun, C., Shrivastava, A., Singh, S., Gupta, A., 2017. Revisiting unreasonable effectiveness Kelly, F.J., Zhu, T., 2016. Transport solutions for cleaner air. Science 352, 934–936. of data in deep learning era. In: In: 2017 IEEE International Conference on Computer https://doi.org/10.1126/science.aaf3420. Vision (ICCV). Presented at the 2017 IEEE International Conference on Computer Khan, J., Ketzel, M., Kakosimos, K., Sørensen, M., Jensen, S.S., 2018. Road traffic air and Vision (ICCV). IEEE, Venice, pp. 843–852. https://doi.org/10.1109/ICCV.2017.97. noise pollution exposure assessment – a review of tools and techniques. Sci. Total United Nations, Department of Economic and Social Affairs, & Population Division, Environ. 634, 661–676. https://doi.org/10.1016/j.scitotenv.2018.03.374. 2019. World urbanization prospects: the 2018 revision. Lafon, S., Rajot, J.-L., Alfaro, S.C., Gaudichet, A., 2004. Quantification of iron oxides in Waggoner, A.P., Weiss, R.E., 1980. Comparison of fine particle mass concentration and desert aerosol. Atmos. Environ. 38, 1211–1218. https://doi.org/10.1016/j. light scattering extinction in ambient aerosol. Atmospheric Environ. 1967 (14), atmosenv.2003.11.006. 623–626. https://doi.org/10.1016/0004-6981(80)90098-0. LeCun, Y.A., Bottou, L., Orr, G.B., Müller, K.-R., 2012. Efficient BackProp. In: Wang, X., Wang, M., Liu, X., Zhang, X., Li, R., 2022. A PM2.5 concentration estimation Montavon, G., Orr, Geneviève B., Müller, K.-R. (Eds.), Neural Networks: Tricks of the method based on multi-feature combination of image patches. Environ. Res. 211, Trade: Second Edition, Lecture Notes in Computer Science. Springer, Berlin, 113051 https://doi.org/10.1016/j.envres.2022.113051. Heidelberg, pp. 9–48. https://doi.org/10.1007/978-3-642-35289-8_3. Weagle, C.L., Snider, G., Li, C., van Donkelaar, A., Philip, S., Bissonnette, P., Burke, J., Liu, C., Tsow, F., Zou, Y., Tao, N., 2016. Particle pollution estimation based on image Jackson, J., Latimer, R., Stone, E., Abboud, I., Akoshile, C., Anh, N.X., Brook, J.R., analysis. PLoS One 11, e0145955. https://doi.org/10.1371/journal.pone.0145955. Cohen, A., Dong, J., Gibson, M.D., Griffith, D., He, K.B., Holben, B.N., Kahn, R., Liu, X., Song, Z., Ngai, E., Ma, J., Wang, W., 2015. PM2:5 monitoring using images from Keller, C.A., Kim, J.S., Lagrosas, N., Lestari, P., Khian, Y.L., Liu, Y., Marais, E.A., smartphones in participatory sensing. In: 2015 IEEE Conference on Computer Martins, J.V., Misra, A., Muliane, U., Pratiwi, R., Quel, E.J., Salam, A., Segev, L., Communications Workshops (INFOCOM WKSHPS), pp. 630–635. https://doi.org/ Tripathi, S.N., Wang, C., Zhang, Q., Brauer, M., Rudich, Y., Martin, R.V., 2018. 10.1109/INFCOMW.2015.7179456. Presented at the 2015 IEEE Conference on Global sources of fine particulate matter: interpretation of PM2.5 chemical Computer Communications Workshops (INFOCOM WKSHPS). composition observed by SPARTAN using a global chemical transport model. McTainsh, G., 1980. Harmattan dust deposition in northern Nigeria. Nature 286, Environ. Sci. Technol. 52, 11670–11681. https://doi.org/10.1021/acs.est.8b01658. 587–588. https://doi.org/10.1038/286587a0. Wei, X., Chang, N.-B., Bai, K., Gao, W., 2020. Satellite remote sensing of aerosol optical Nathvani, R., Clark, S.N., Muller, E., Alli, A.S., Bennett, J.E., Nimo, J., Moses, J.B., depth: advances, challenges, and perspectives. Crit. Rev. Environ. Sci. Technol. 50, Baah, S., Metzler, A.B., Brauer, M., Suel, E., Hughes, A.F., Rashid, T., Gemmell, E., 1640–1725. https://doi.org/10.1080/10643389.2019.1665944. Moulds, S., Baumgartner, J., Toledano, M., Agyemang, E., Owusu, G., Agyei- Weichenthal, S., Hatzopoulou, M., Brauer, M., 2019. A picture tells a thousand… Mensah, S., Arku, R.E., Ezzati, M., 2022. Characterisation of urban environment and exposures: opportunities and challenges of deep learning image analyses in exposure activity across space and time using street images and deep learning in Accra. Sci. science and environmental epidemiology. Environ. Int. 122, 3–10. https://doi.org/ Rep. 12, 20470. https://doi.org/10.1038/s41598-022-24474-1. 10.1016/j.envint.2018.11.042. Ochei, M.C., Adenola, E., 2018. Variability of Harmattan dust haze over northern Weigand, M., Wurm, M., Dech, S., Taubenböck, H., 2019. Remote sensing in Nigeria. J. Pollut. 1, 8. environmental justice research—a review. ISPRS Int. J. Geo-Inf. 8, 20. https://doi. Onuu, M.U., 2000. Road traffic noise in Nigeria: measurements, analysis and evaluation org/10.3390/ijgi8010020. of nuisance. J. Sound Vib. 233, 391–405. https://doi.org/10.1006/jsvi.1999.2832. Won, T., Eo, Y.D., Sung, H., Chong, K.S., Youn, J., Lee, G.W., 2022. Particulate matter Ozkaynak, H., Schatz, A.D., Thurston, G.D., Isaacs, R.G., Husar, R.B., 1985. Relationships estimation from public weather data and closed-circuit television images. KSCE J. between aerosol extinction coefficients derived from airport visual range Civ. Eng. 26, 865–873. https://doi.org/10.1007/s12205-021-0865-4. observations and alternative measures of airborne particle mass. J. Air Pollut. World Health Organization, 2021. WHO Global Air Quality Guidelines: Particulate Control Assoc. 35, 1176–1185. https://doi.org/10.1080/00022470.1985.10466020. Matter (PM2.5 and PM10), Ozone, Nitrogen Dioxide, Sulfur Dioxide and Carbon Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Monoxide. World Health Organization. Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K., 2017. Aggregated residual transformations Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S., 2019. PyTorch: for deep neural networks. In: In: 2017 IEEE Conference on Computer Vision and An Imperative Style, High-Performance Deep Learning Library, in: Advances in Pattern Recognition (CVPR). Presented at the 2017 IEEE Conference on Computer Neural Information Processing Systems. Curran Associates, Inc. Vision and Pattern Recognition (CVPR). IEEE, Honolulu, HI, pp. 5987–5995. https:// Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., doi.org/10.1109/CVPR.2017.634. Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Zhang, C., Yan, J., Li, C., Wu, H., Bie, R., 2018. End-to-end learning for image-based air Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, É., 2011. Scikit-learn: machine quality level estimation. Mach. Vis. Appl. 29, 601–615. https://doi.org/10.1007/ learning in Python. J. Mach. Learn. Res. 12, 2825–2830. s00138-018-0919-x. Pinker, R.T., Idemudia, G., Aro, T.O., 1994. Characteristic aerosol optical depths during Zhang, Q., Zhu, S., 2018. Visual interpretability for deep learning: a survey. Front. Inf. the Harmattan season on sub-Sahara Africa. Geophys. Res. Lett. 21, 685–688. Technol. Electron. Eng. 19, 27–39. https://doi.org/10.1631/FITEE.1700808. https://doi.org/10.1029/93GL03547. Zhou, Z., Dionisio, K.L., Verissimo, T.G., Kerr, A.S., Coull, B., Arku, R.E., Koutrakis, P., Pope, C.A., Dockery, D.W., 2006. Health effects of fine particulate air pollution: lines that Spengler, J.D., Hughes, A.F., Vallarino, J., Agyei-Mensah, S., Ezzati, M., 2013. connect. J. Air Waste Manag. Assoc. 56, 709–742. https://doi.org/10.1080/ Chemical composition and sources of particle pollution in affluent and poor 10473289.2006.10464485. neighborhoods of Accra, Ghana. Environ. Res. Lett. 8, 044025 https://doi.org/ Qi, M., Hankey, S., 2021. Using street view imagery to predict street-level particulate air 10.1088/1748-9326/8/4/044025. pollution. Environ. Sci. Technol. 55, 2695–2704. https://doi.org/10.1021/acs. est.0c05572. 11