key: cord-0709340-73lolfls authors: Wijnands, Jasper S.; Nice, Kerry A.; Seneviratne, Sachith; Thompson, Jason; Stevenson, Mark title: The impact of the COVID-19 pandemic on air pollution: A global assessment using machine learning techniques date: 2022-04-28 journal: Atmos Pollut Res DOI: 10.1016/j.apr.2022.101438 sha: b6be7cfd96256cb3077563ab00ed2da2cb8ad85d doc_id: 709340 cord_uid: 73lolfls In response to the COVID-19 pandemic, most countries implemented public health ordinances that resulted in restricted mobility and a resultant change in air quality. This has provided an opportunity to quantify the extent to which carbon-based transport and industrial activity affect air quality. However, quantification of these complex effects has proven to be difficult, depending on the stringency of restrictions, country-specific emission source profiles, long-term trends and meteorological effects on atmospheric chemistry, emission levels and in-flow from nearby countries. In this study, confounding factors were disentangled for a direct comparison of pandemic-related reductions in absolute pollutions levels, globally. The non-linear relationships between atmospheric processes and daily ground-level NO [Formula: see text] , PM(10), PM(2.5) and O [Formula: see text] measurements were captured in city- and pollutant-specific XGBoost models for over 700 cities, adjusting for weather, seasonality and trends. City-level modelling allowed adaptation to the distinct topography, urban morphology, climate and atmospheric conditions for each city, individually, as the weather variables that were most predictive varied across cities. Pollution forecasts for 2020 in absence of a pandemic were generated based on weather and formed an ensemble for country-level pollution reductions. Findings were robust to modelling assumptions and consistent with various published case studies. NO [Formula: see text] reduced most in China, Europe and India, following severe government restrictions as part of the initial lockdowns. Reductions were highly correlated with changes in mobility levels, especially trips to transit stations, workplaces, retail and recreation venues. Further, NO [Formula: see text] did not fully revert to pre-pandemic levels in 2020. Ambient PM(2.5) pollution, which has severe adverse health consequences, reduced most in China and India. Since positive health effects could be offset to some extent by prolonged exposure to indoor pollution, alternative transport initiatives could prove to be an important pathway towards better health outcomes in these countries. Increased O [Formula: see text] levels during initial lockdowns have been documented widely. However, our analyses also found a subsequent reduction in O [Formula: see text] for many countries below what was expected based on meteorological conditions during summer months (e.g., China, United Kingdom, France, Germany, Poland, Turkey). The effects in periods with high O [Formula: see text] levels are especially important for the development of effective mitigation strategies to improve health outcomes. (2020) reported a 53% reduction compared to the pre-lockdown period in Delhi. 51 In Wuhan, reported reductions were 53.3% compared to the pre-lockdown pe-52 riod (Lian et al., 2020) and 57% compared to 2017-2019 (Sicard et al., 2020) . In contrast, the average PM 2.5 reduction in Europe was found to be only using a machine learning approach to incorporate meteorological variability, as 150 opposed to directly comparing measurements in previous years to 2020 levels. 151 Importantly, machine learning is especially beneficial for accurately predicting 152 pollution at finer spatial and temporal scales (Petetin et al., 2020) . Therefore, 153 for a robust comparison to daily city-level mobility data (i.e., fine spatial and should be seen in light of these developments. In this paper, we present a 164 city-level analysis at a global scale, including adjustment for weather variables 165 such as temperature, wind speed and precipitation. City-and pollution-specific 166 modelling based on historical atmospheric data and annual trends, provides 167 air pollution estimates for 2020 in absence of a pandemic (i.e., a counterfac-168 tual, 'business as usual' scenario). Various methods were explored to optimise 169 forecast accuracy. The identified pollution anomalies were then compared to 170 country-specific government policies intended to limit the spread of COVID-19. Hence, our research investigates air pollution both at the micro-and the macro- The following global data sources were selected to provide information on air 177 pollution levels, weather, the severity of government restrictions, and mobility 178 patterns of city residents. Data was checked for completeness, cleaned and 179 processed using Java, Python and R. were converted back to their original unit for comparative analysis in this study. Governments adopted a broad variety of (initial) approaches to deal with the 213 COVID-19 pandemic, ranging from trying to maintain business as usual (e.g., XGBoost has several hyperparameters that influence the learning process. that is used for model calibration. In this study, hyperparameters were tuned using grid search. Initial ex-300 periments used a large grid for a small sample of models (n = 45; 15 cities 301 and 3 pollutants). The initial grid, presented in Table 1 Table 2 . T Air temperature at 2 meters altitude, day t T Air temperature at 2 meters altitude, mean(t-3, t-2, t-1) S Net solar radiation at the surface, day t S Net solar radiation at the surface, mean(t-3, t-2, t-1) Table 3 , the 410 corresponding gains of all 2888 models were averaged to obtain the mean gain. Note that the gain is an indicator of relative feature importance (i.e., how im- The top features per pollutant were not always the best-performing features, 432 as considerable variation in feature importance was observed across cities (see 433 Table 4 ). For example, wind direction was the most important feature for sev-434 eral coastal cities (e.g., Liverpool, Nantong, Jakarta). In these cities, clean Wind speed was a slightly more informative feature in NO 2 models than in 448 PM 10 , PM 2.5 and O 3 models. In contrast, the total precipitation over the past 449 three days (P ) was more useful for predicting PM 10 and PM 2.5 than NO 2 and Fig. 4) . In other countries, the peaks were observed around early 479 April. In larger countries such as China and India, regional lockdowns led to Table 6 shows the variation of model fit across cities. Fur-548 ther, a time series modelling approach without weather variables was explored. As the time series forecasting approach differed substantially from our preferred 550 XGBoost method described in Section 2.2, it is described here in more detail. Machine learning methods such as XGBoost tend to overfit on training data. To provide an indication of out-of-sample performance, Fig. 7 residents ofÜrümqi were locked inside their homes from 18 July to 31 August Importantly, when aggregating results to country level, the alternative ap-593 proaches led to similar results, but with an increased level of noise (see Fig. 8 ). Table 9 provides a comparison to studies that investigated a specific country, spent over 30% more time at home during the initial lockdown (see Fig. 6 ). Although targeting mobility reductions in these countries can be effective in Air pollution in Ontario, Canada during the COVID-19 Apple Maps Mobility Trends Reports The world air quality index project Changes in air quality and human mobility in the USA during the COVID-767 19 pandemic Impact of coronavirus outbreak on NO2 pollution assessed 772 using TROPOMI and OMI observations Changes in criteria air pollution levels in the 776 US before, during, and after COVID-19 stay-at-home orders: evidence from 777 regulatory monitors Changes in U.S. air pollution during the COVID-780 19 pandemic openair -an R package for air quality data 783 analysis. Environmental Modelling & Software 27-28 XGBoost: a scalable tree boosting system STL: a 790 seasonal-trend decomposition procedure based on Loess Estimates and 25-year trends of the global burden of disease 798 attributable to ambient air pollution: an analysis of data from the Global 799 Burden of Diseases Study Air quality during COVID-19 in four 803 megacities: lessons and challenges for public health. International Jour-804 nal of Environmental Research and Public Health Contribution of air conditioning adoption to 807 future energy use under global warming Lower air pollu-811 tion during COVID-19 lock-down: improving models and methods estimat-812 ing ozone impacts on crops Changes in outdoor air pollution due to COVID-816 19 lockdowns differ by pollutant: evidence from Scotland Current and future 821 global climate impacts resulting from COVID-19 Greedy function approximation: a gradient boosting 824 machine Short-term and long-term health impacts of air pollution reductions from 828 COVID-19 lockdowns in China and Europe: a modelling study Reference and equivalent methods used to mea-831 sure National Ambient Air Quality Standards (NAAQS) criteria air pollu-832 tants -Volume I The global impacts of COVID-19 lockdowns on urban air 837 pollution: a critical review and recommendations. Elementa: Science of the 838 Disentangling the impact of the COVID-19 lockdowns on 841 urban NO2 from natural variability Google COVID-19 Community Mobility Reports Impact of weather types on 847 UK ambient particulate matter concentrations Using meteorological normalisation to detect 850 interventions in air quality time series A global panel database of pandemic policies (Oxford COVID-855 19 Government Response Tracker) The elements of statistical learn-858 ing: Data mining, inference, and prediction The short-term impacts of COVID-19 861 lockdown on urban air pollution in China The ERA5 global reanalysis Insignificant impact of the "stay-874 at-home" order on ambient air quality in the Memphis metropolitan area Atmosphere 11 Spartan performance and 877 flexibility: an HPC-cloud chimera An-880 drew Temporary reduction in daily the COVID-19 forced confinement The 885 contribution of outdoor air pollution sources to premature mortality on a 886 global scale Impact 888 of city lockdown on the air quality of COVID-19-hit of Wuhan city. Science 889 of The Total Environment Identification 892 of the most influential areas for air pollution control using XGBoost and Importance Rank Application of the XGBoost 896 machine learning method in PM2.5 prediction: a case study of Shanghai Effect of lockdown amid COVID-19 900 pandemic on air quality of the megacity Delhi, India. Science of The Total 901 Environment 730 Examining effects of the COVID-903 19 national lockdown on ambient air quality across urban India Air pollution removal by urban 906 trees and shrubs in the United States. Urban Forestry & Urban Greening 4 Early spring near-909 surface ozone in Europe during the COVID-19 shutdown: meteorological efof The Total Environment Meteorology-normalized impact of the COVID-19 914 lockdown upon NO 2 pollution in Spain Comparison of machine learning 917 and land use regression for fine scale spatiotemporal estimation of ambient 918 air pollution: modeling ozone concentrations across the contiguous United 919 Pattern of early human-to-human transmission 922 of Wuhan Early observations on the impact of the COVID-926 19 lockdown on air quality trends across the UK. Science of The Total Envi-927 ronment 754 Chemistry and the linkages between air quality 931 and climate change Atmospheric chemistry and physics: from air 934 pollution to climate change Effect of 936 restricted emissions during COVID-19 on air quality in India Amplified ozone pollution in cities 940 during the COVID-19 lockdown The relation between ozone, NO x and hydrocarbons in urban 943 and polluted rural environments Impacts of modifiable factors on ambient air pollution: a case study 947 of COVID-19 shutdowns Mongolia -Air pollution in Ulaanbaatar: initial assess-950 ment of current situation and effects of abatement measures. Discussion paper 951 52970 PM2.5 source allocation in Eu-955 ropean cities: a SHERPA modelling study World urbanization prospects: the 2014 revision A/366. Department of Economic and Social Affairs Population Division lockdowns cause global air pollution declines Severe air pollution 964 events not avoided by reduced anthropogenic activities during COVID-19 965 outbreak. Resources, Conservation and Recycling 158 Air quality guidelines: global update 2005 968 -particulate matter, ozone, nitrogen dioxide, and sulfur dioxide. WHO 969 Regional Office for Europe Household air pollution and 971 health Coronavirus (COVID-19) dashboard Study of the meteorological 976 influence on ozone in urban areas and their use in assessing ozone trends in all 977 seasons from 2009 to 2015 in Tianjin New indices for wet scavenging of air 981 pollutants (O3, CO, NO2, SO2, and PM10) by summertime rain prediction based on random forest, XGBoost, and deep learning 985 using multisource remote sensing data Substantial changes in nitrogen dioxide and ozone after excluding 989 meteorological impacts during the COVID-19 outbreak in mainland China PM 2.5 pollution should be carefully considered in any intervention. Note that 743 in many countries, NO 2 reductions were highly correlated with changes in mo-744 bility levels, especially trips to transit stations, workplaces, retail and recreation 745 venues. Therefore, alternative transport initiatives for these type of trips, such 746 as electric vehicles, bicycles and ride sharing, could be an important pathway 747 towards improved health outcomes.