key: cord-318437-tzp33iw7
authors: Lovrić, Mario; Pavlović, Kristina; Vuković, Matej; Grange, Stuart K.; Haberl, Michael; Kern, Roman
title: Understanding the true effects of the COVID-19 lockdown on air pollution by means of machine learning()
date: 2020-11-06
journal: Environ Pollut
DOI: 10.1016/j.envpol.2020.115900
sha: 
doc_id: 318437
cord_uid: tzp33iw7

During March 2020, most European countries implemented lockdowns to restrict the transmission of SARS-CoV-2, the virus which causes COVID-19 through their populations. These restrictions had positive impacts for air quality due to a dramatic reduction of economic activity and atmospheric emissions. In this work, a machine learning approach was designed and implemented to analyze local air quality improvements during the COVID-19 lockdown in Graz, Austria. The machine learning approach was used as a robust alternative to simple, historical measurement comparisons for various individual pollutants. Concentrations of NO(2) (nitrogen dioxide), PM(10) (particulate matter), O(3) (ozone) and O(x) (total oxidant) were selected from five measurement sites in Graz and were set as target variables for random forest regression models to predict their expected values during the city's lockdown period. The true vs. expected difference is presented here as an indicator of true pollution during the lockdown. The machine learning models showed a high level of generalization for predicting the concentrations. Therefore, the approach was suitable for analyzing reductions in pollution concentrations. The analysis indicated that the city’s average concentration reductions for the lockdown period were: -36.9 to -41.6%, and -6.6 to -14.2% for NO(2) and PM(10,) respectively. However, an increase of 11.6 to 33.8% for O(3) was estimated. The reduction in pollutant concentration, especially NO(2) can be explained by significant drops in traffic-flows during the lockdown period (-51.6 to -43.9%). The results presented give a real-world example of what pollutant concentration reductions can be achieved by reducing traffic-flows and other economic activities.

The COVID-19 pandemic has caused disastrous health and socio-economic crises across the 50 globe (Alabdulmonem et al., 2020;  McKee and Stuckler, 2020). Questions have been raised 51 whether atmospheric pollution is a co-factor in disease development causing a higher lethality 52 rate, especially in highly populated and polluted areas such as those in Italy (Conticini et al., 53 2020; Fattorini and Regoli, 2020) . A study from China suggests there is a statistically confirmed 54 relationship between air pollution by means of elevated concentrations of PM 2.5 , PM 10 , CO, NO 2 55

and O 3 and the COVID-19 infection rate (Zhu et al., 2020) . Another study from Italy supports the 56 insight by providing causal relationships between the COVID-19 spread and air quality (Delnevo 57 et al., 2020 ). An interplay of air quality and the pandemic seems obvious. 58

On the other side, lockdowns have caused significant changes in air quality (Dutheil et al., 59 2020). A study on 44 Chinese cities (Bao and Zhang, 2020) showed a decrease in main air 60 pollutants from 5.93-24.67% during the lockdown while megacities such as Sao Paulo showed 61 even higher concentration drops (40-70%) for some pollutants (Krecl et al., 2020) . A study on 62 PM 2.5 in capital cities showed concentration drops of 20-60% during the COVID-19 crisis 63

(Rodríguez-Urrego and Rodríguez-Urrego, 2020). It is suggested that the pollution drop was 64 mainly driven by a reduction in traffic (Kerimray et al., 2020) and industrial activities (Li et al., 65 2020). Even if lockdowns hinder economic growth and might cause various negative effects in 66 the long term, drops in pollution concentrations may act as another factor which slows disease 67 transmission in tandem with limiting human contact. Lockdowns in Europe were instituted 68 gradually by means of governmental interventions (Desvars-Larrive et al., 2020). This massive 69

intervention also poses a unique opportunity to study the change in various aspects of air quality, 70 thus motivating our study. 71

We discuss and explore that for complete understanding of the true factors influencing 72 pollutant concentrations, pure statistical tests or observational comparisons might be inadequate 73 since weather conditions, particle persistence and seasonality affect concentrations by linear and 74 non-linear processes (Šimić et al., 2020) . Furthermore, transport pathways and source distribution 75

can also play a role in analyzing the effects of the lockdown on pollution by means of trajectory 76 models (Zhao et al., 2020) .Therefore, a comparison of air quality in 2020 vs other years may be 77 biased since other independent factors such as shifts in heating seasons or weather conditions can 78 affect air quality (Schiermeier, 2020) . To be able to solve independent factor some authors 79

proposed fixed effect models (Liu et al., 2020; Venter et al., 2020) . Our proposal is that the 80 pollution level can be solved as a multivariate problem predicted by independent variables 81 elevated from environmental variables and seasonal trends, i.e. there are many effects and they 82 might not have fixed effects because the atmosphere is a very dynamic system. Moreover, if one 83 wants to return a full time series, observational and fixed effect methods might fail if not 84 accounted for environmental dependencies. For predicting the pollutant concentrations, we 85 employed the Random Forest algorithm, a non-linear regression method which has the power to 86

solve multivariate problems independent of the variable type. A complementary approach found 87 J o u r n a l P r e -p r o o f in literature is a forecasting method which accounted for atmospheric and other effects but using 88 a mechanistic instead of a data-driven machine learning approach (Menut et al., 2020) . 89

We investigate the effects of lockdown on air quality in an urbanized area in Graz, Styria, 90

Austria. Due to the high degree of traffic influence, we have included traffic data into our 91

analysis. Furthermore, we have investigated in detail which of the pollutants' concentrations were 92 influenced by the lockdown. As such, the outcome of our study serves as a guide for future 93 interventions and their expected associated change in the pollutants' concentration changes. 94

Our study contains traditional exploratory statistical analysis, including the utilization of principal 96 component analysis (PCA) to explore key attributes. However, the primary analysis is based on 97 machine learning (ML) models which were used to capture historical relationships between the 98 attributes and compare the predictions to true pollution values after the COVID-19 lockdowns 99

were imposed. We utilize historical data which matches the time frame of the lockdown for the 100 preceding years, but also include traffic flow data to represent the drop in mobility. 101 Data description 102 We collected environmental, pollution and weather data from publicly available sources provided 103 by the Austrian government 1 . In order to obtain a realistic picture of air quality during the 104 lockdown, we analyzed the long term measurement data from January 2014 to May 2020 from 105 five measurement sites in the Austria city of Graz (Süd (eng. South) -S, Nord (eng. North) -N, 106

West (eng. West) -W, Don Bosco -D, Ost (eng. East) -O); Figure 1 ). Graz is a medium-sized 107

European city which has much in common in respect to size and layout to many other European 108 urban areas. The latter two measurement sites are situated on arterial roads with high traffic 109 volumes, especially during morning and evening rush hours. The most polluted measurement site 110

of Graz is Don Bosco that struggles to meet the annual NO 2 and PM 10 regulatory limits of the 111 EU-Council directive 96/62/EC. This is primarily because of the traffic related emissions, but 112 also because of the emissions from a nearby steel-and iron-mill (Hinterhofer, 2014) . Although 113 Graz East is located at a heavily frequented commuter-arterial, mean pollutant concentrations are 114 lower than at Don Bosco. Graz South is situated at a secondary road segment but also records 115 higher pollutant concentrations due to an industrial complex nearby. Graz North and West are 116 classified as urban background sites and are located near minor roads with no specific emission 117 contributors in immediate vicinity. A more detailed site description, photos of the sites and 118 historical overview of the sites is given in Moser et al., 2019. 119 With the intention of understanding the potential effects of traffic, the traffic flow for the 120 city of Graz was accessed. The traffic flow data were mainly measured with inductive loop 121 detectors where the detectors measure the change in field when objects pass over them. Once a 122 vehicle drives over a loop sensor, the loop field changes which allows the detection of the 123 presence of an object (a vehicle). The "Traffic control and street lighting unit of the city of Graz" 124 monitors and records the data at one-minute time frequency and provided data from January 2017 125

to May 2020 for two sites, namely Don Bosco and Ost. 126

To determine the start, end, and duration of the Austrian lockdown, we extracted these 127 data from a dataset which contains a collection 

The air quality data covers PM 10 and NO 2 , from five sites (D, N, O, S, W) described in Figure 1 values were imputed by backfilling (see missing value counts in Table 1 ). The processed data 160

consists of 2324 days and 60 variables in total and is provided in table format within a persistent 161 data repository (Lovrić et al., 2020b) . The traffic data were aggregated to a daily frequency and 162 stored as a time series for the two sites (O, D). The processed traffic data ranges from January 163 2017 to May 2020. values (values from the previous two days). These predictive variables allowed the machine 190 learning model to capture seasonal behavior from activities such industrial production and traffic 191 flows and therefore, can be thought of as surrogate variables. 192

The machine learning algorithm used was Random Forest regression (RF) (Breiman, 193 2001) which has been utilized in a number of previous air pollution models and air quality data shown in Figure 4 . 273 one can see that for PM 10 the ML models show a larger concentration drop whereas for NO 2 we 345 see a smaller concentration drop with the ML models (Table 3, which are located close to roads (Figure 1 ). The traffic measured at the detector loops at these 356 measurement sites showed a reduction of 45.6% at O and 51.6% at D respectively (Table 3, industrial processes were also significantly curtailed in and around Graz during the lockdown 362 period, it could be expected that PM 10 would decrease more drastically. The lack of reduction 363 outlines the more numerous and complicated processes which drive PM 10 concentrations, and/or 364 a high lag in the relationships involved in the PM 10 related variables, i.e., the duration of the 365 intervention was too short to lead to a significant drop in this particular pollutant. 366

Our results agree regarding a traffic-related drop in NO 2 and an increase in O 3 with a 367 variety of studies conducted on air quality issues regarding the COVID-19 lockdown. A study 368 from northern China shows that a reduction in NO 2 and PM was most likely caused by a 369 reduction in traffic and industrial activities . Furthermore, they also show an 370 increase in O 3 consistent with our analysis as well. Another study from China where lockdown 371 measures were introduced earlier than in Austria (1 st January -29 th February 2020) (Shi and 372

Brasseur, 2020) reveals a reduction of NO 2 up to 60% and a O 3 increase of 40-100%. A two-stage 373 lockdown in India, which was introduced concurrently with the Austrian lockdown (Mahato et  374 al., 2020), and data from this region shows a PM 10 concentration drop around 60% compared to 375 the same period in 2019. These results are contrary to ours, since Graz did not experience such a 376 PM 10 drop, revealed by both historical comparison and machine learning prediction. Reasons for 377 that may be domestic heating which is difficult to evaluate since data on heating and stays in the 378 city are difficult data to obtain. The same study observed concordant results in the drop of NO 2 (-379 52.68%). A study from the UK, which employed also a historical comparison shows a reduction 380 of 48% in NO 2 concentration and an average increase in O 3 concentration of 11% across 126 381 urban sites, which is also consistent with our analysis. A maybe more relevant comparison to our 382 study is given by (Menut et al., 2020) shows a decline in PM 10 during the lockdown we believe the fixed-effect and observational 397 methods may not be enough to deliver a definite conclusion on the concentration reductions. It is 398 a limitation of this study that long-range transport data and chemical speciation of the particulate 399 matter is not available. More efforts must be put into chemical speciation of PM at the individual 400 sites, especially measurement techniques which deliver "online" data. 401

In this work, we have explored the changes in air pollutant concentrations during the COVID-19 403 lockdown for the city of Graz, Austria. The exploration illuminated the relative influences of 404 observed meteorological variables on a wide range of pollutants for an unpresented historic event 405

of human society. Besides using explorative methods, we employed random forest regression to 406 analyze the differences between predicted (expected) and observed (true) pollution levels based 407 on environmental data. 408

Our prediction models showed good generalization and performance for the analyzed 409 pollutants indicating that the selection of independent variables (predictors) was sufficient to 410 explain changes in pollutant concentrations. For PM 10 

Principal component analysis

COVID-19: A global public health disaster

Does lockdown reduce air pollution? Evidence from 44 cities in 503 northern China

Random Forests

Can atmospheric pollution be considered a co-factor in 508 extremely high level of SARS-CoV-2 lethality in Northern Italy?

Particulate Matter and COVID-19 Disease Diffusion 511 in Emilia-Romagna (Italy). Already a Cold Case? Computation

A structured open dataset of 518 government interventions in response to COVID-19, medRxiv. Cold Spring Harbor 519 Laboratory Press

COVID-19 as a factor influencing air pollution? 521

Role of the chronic air pollution levels in the Covid-19 outbreak 524 risk in Italy

Using meteorological normalisation to detect interventions in 529 air quality time series

Random forest 532 meteorological normalisation models for Swiss PM10 trend analysis

Random forest 535 meteorological normalisation models for Swiss PM10 trend analysis

More Dust Blows Out from North Africa [WWW Document

Anteil der verkehrsbedingten PM10 und PM2,5 Emissionen aus Abrieb 541 und Wiederaufwirbelung an der Feinstaubbelastung in Österreich

Assessing air quality changes in large cities during COVID-19 lockdowns: 545 The impacts of traffic-free urban conditions in Almaty

Drop in urban air pollution 548 from COVID-19 pandemic: Policy implications for the megacity of São Paulo

Air quality changes during the COVID-19 lockdown over the Yangtze 553 River Delta Region: An insight into the impact of human activity pattern changes on air 554 pollution variation

Abrupt decline in tropospheric nitrogen dioxide over China after the outbreak of 558 COVID-19

Prediction of 560 anode lifetime in electro galvanizing lines by big data analysis

(including COVID-19 lockdown) data from Graz

566 Machine learning in prediction of intrinsic aqueous solubility of drug-like compounds: 567 generalization, complexity or predictive ability? chemrxiv

Effect of lockdown amid COVID-19 pandemic on air 570 quality of the megacity Delhi

If the world fails to protect the economy, COVID-19 will damage 573 health not just now but also in the future

Impact of 576 lockdown measures to combat Covid-19 on air quality over western

Statistische Analyse der Luftqualitätin Graz anhand von 579 Feinstaub und Stickstoffdioxid

Air quality during the COVID-19: PM2.5 581 analysis in the 50 most polluted capital cities in the world

Why pollution is plummeting in some cities -but not others

The Response in Air Quality to the Reduction of Chinese 586 Economic Activities During the COVID-19 Outbreak

Applying machine learning methods 589 to better understand, model and estimate mass concentrations of traffic-related pollutants at 590 a typical street canyon

COVID-19 lockdowns cause global 593 air pollution declines

Source apportionment of particulate matter in Europe: A review of methods and results

Changes in air quality related to 601 the control of coronavirus in China: Implications for traffic and industrial emissions

Air pollution episodes during the COVID-604 19 outbreak in the Beijing-Tianjin-Hebei region of China: An insight into the transport 605 pathways and source distribution

Association between short-term exposure to air 608 pollution and COVID-19 infection: Evidence from China

☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.☐The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:J o u r n a l P r e -p r o o f