key: cord-0770594-po8891so authors: Talbot, Nick; Takada, Akika; Bingham, Andrew H.; Elder, Dan; Lay Yee, Samantha; Golubiewski, Nancy E. title: An investigation of the impacts of a successful COVID-19 response and meteorology on air quality in New Zealand date: 2021-06-01 journal: Atmos Environ (1994) DOI: 10.1016/j.atmosenv.2021.118322 sha: 746c0d97315a0808528463d679ce6068f032b7de doc_id: 770594 cord_uid: po8891so The COVID-19 pandemic brought about national restrictions on people's movements, in effect commencing a socially engineered transport emission reduction experiment. In New Zealand during the most restrictive alert level (Level 4), roadside concentrations of nitrogen dioxide (NO(2)) were reduced 48–54% compared to Business-as-usual (BAU) values. NO(2) concentrations rapidly returned to near mean levels as the alert levels decreased and restrictions eased. PM(10) and PM(2.5) responded differently to NO(2) during the different alert levels. This is due to particulates having multiple sources, many of natural origin and therefore less influenced by human activity. PM(10) and PM(2.5) concentrations were reduced during alert level 4 but to a lesser extent than NO(2) and with more variability across regions. Particulate concentrations increased notably during alert level 2 when many airsheds reported concentrations above the BAU means. To provide robust BAU reference concentrations, simple 5-year means were calculated along with predictions from machine learning modelling that, in effect, removed the influence of meteorology on observed concentrations. The results of this study show that latter method was found to be more closely aligned to observed values for NO(2) as well as PM(2.5) and PM(10) away from coastal regions. • COVID-19 response produced notable impacts on air quality. • Decrease in traffic emissions reduced NO 2 notably across New Zealand. • As the restrictive levels eased, pollution levels returned to near long-term means. • Particulate concentrations increased during alert level 2, attributed to more home heating emissions from wood burning. • Machine learning found good R values for NO 2 ; however, modelling results were weaker for coastal particulate data. On-road vehicles Machine learning Atmospheric pollutants The COVID-19 pandemic brought about national restrictions on people's movements, in effect commencing a socially engineered transport emission reduction experiment. In New Zealand during the most restrictive alert level (Level 4), roadside concentrations of nitrogen dioxide (NO 2 ) were reduced 48-54% compared to Businessas-usual (BAU) values. NO 2 concentrations rapidly returned to near mean levels as the alert levels decreased and restrictions eased. PM 10 and PM 2.5 responded differently to NO 2 during the different alert levels. This is due to particulates having multiple sources, many of natural origin and therefore less influenced by human activity. PM 10 and PM 2.5 concentrations were reduced during alert level 4 but to a lesser extent than NO 2 and with more variability across regions. Particulate concentrations increased notably during alert level 2 when many airsheds reported concentrations above the BAU means. To provide robust BAU reference concentrations, simple 5-year means were calculated along with predictions from machine learning modelling that, in effect, removed the influence of meteorology on observed concentrations. The results of this study show that latter method was found to be more closely aligned to observed values for NO 2 as well as PM 2.5 and PM 10 away from coastal regions. At the end of 2019, a new disease emerged from Wuhan, China initially named SARS-COVID 2 (now . Highly contagious between humans, the virus rapidly spread around the globe with over 200 countries ultimately affected. With no known cure or effective treatment at the initial stages of the outbreak, strict social restrictions were activated in the form of alert levels to help protect vulnerable citizens and prevent health systems becoming inundated (Wilder-Smith and Freedman, 2020; World Health Organisation, 2020) . Globally, research has shown that social intervention restrictions led to notable changes in air pollution levels. For example, significant reductions in nitrogen dioxide (NO 2 ) concentrations were reported in Brazil (Nakada and Urban, 2020) , Spain (Tobías et al., 2020) , Ecuador (Zambrano-Monserrate and Ruano, 2020), the USA (Berman and Ebisu, 2020) , and China (Griffith et al., 2020) . Concentrations of fine particles (PM 2.5 ) also decreased during intervention strategies in China (Wang and Su, 2020) , India (Sharma et al., 2020) , Kazakhstan (Kerimray et al., 2020) and the USA (Hudda et al., 2020) . Conversely, tropospheric ozone (O 3 ) concentrations increased in Europe, China, and South America (Sicard et al., 2020; Siciliano et al., 2020) . Particulate results are more variable according to the geographic position (e.g., next to major sources such as oceans or deserts) or the proportion of secondary particulates (Tobías et al., 2020) . The influence of natural sources along with human behaviour on particulate concentrations would therefore result in locations experiencing less reduction due to change in activities. New Zealand is uniquely placed to investigate the impact of behaviour change on air quality given it is so geographically isolated and inherits little transported pollutants from other countries. New Zealand also held one of the strictest and best observed lockdowns of any country, worldwide (Robert, 2020) . A four-level response system (https://covid19.govt.nz/alert-system/) was introduced on March 21, 2020, at which point New Zealand entered Level 2. On 23 March, New Zealand moved to Level 3, with limits put on movement, advice to work from home, and schools closed in preparation for a 4-week national strict Level 4 'lockdown' commencing 26 March. Most New Zealanders stayed at home with only essential services operating through 28 April. Activity slowly resumed as the country eased restrictions: Level 3 (28 April -13 May) followed by Level 2 (14 May -8 June) (Table 1) . A by-product of New Zealand's COVID-19 alert levels (Table 1) was nationally regulated reductions of major air pollution emission sources. Both the intensity of the lockdown and the structured easement of restrictions through the alert levels allow for indepth analysis of observed air quality and the relative importance of key pollution sources. The relationship between changes in the emission sources and the resultant concentrations offers rare insight into each pollutant's relative importance in the ambient atmosphere. In a recent study, Patel et al. (2020) reported NO 2 reductions due to COVID-19 related reductions in traffic volume across Auckland, New Zealand's largest city. Whereas emission inventories and receptor modelling approaches confirm the dominance of traffic sources for nitrogen oxides (NOx) (86%) and black carbon (BC) (72%) across the city, observations showed consequent reductions in NO 2 of only 34-57% and in BC, 55-75%. PM 2.5 (also likely to be dominated by traffic emissions) and particulate matter (PM 10 ) (seasonal, dominated by sea salt and, to a lesser extent, traffic emissions) were reduced during Level 4 (8-17% for PM 2.5 and 7-20% for PM 10 ) (Davy et al., 2017; Patel et al., 2020; Xie et al., 2019) . Patel et al. (2020) investigated three Auckland sites representative of different airsheds-urban peak, urban background, and regional background only during the Level 4 lockdown, not the subsequent levels under which restrictions eased (Table 1) . Further investigation is required to understand the trajectory of air quality post lockdown and whether the results in Auckland are representative of a national picture. The level of reduction in air pollutant concentrations during COVID- 19 restrictions is dependent on a reasonable approximation of a business-as-usual (BAU) value for concentrations during that period. To achieve this, this paper incorporates two methods. Firstly, a simple longterm mean (LTM) for each period was calculated. Secondly, machine learning (ML) was used to account for the impacts of meteorology during each of the alert periods, given that meteorology is a key factor affecting air pollution concentrations (Ebenezer, 2019; Lolli et al., 2020) . The application of machine learning as a prediction tool in the area of atmospheric science is a growing and innovative application of the technology (Feng et al., 2019; Alimissis et al., 2018; Lautenschlager et al., 2020) . ML can offer a more robust assessment than the simple LTM to normalise the impacts of meteorological variables (Grange et al., 2018) . Meteorological normalisation is a technique that accounts for changes in meteorology in an air quality time series. Controlling for such changes helps support trend analysis because there is more certainty that the observed trends are due to changes in emissions or secondary processes rather than changes in meteorology. This method was successfully used to analyse COVID-19 related national NO 2 changes across Spain (Petetin et al., 2020) . However, the study did not compare using a 5-year mean against ML algorithm results nor did it consider how meteorological conditions influenced particulate pollution levels. To the authors' knowledge, ML has not been used for particulate measurements and therefore the response from biogenic sources such as sea salt is unknown. It is due to this uncertainty that LTM is used in conjunction with ML in this paper. This study addresses these knowledge gaps via a national assessment of air quality, via NO 2 , PM 10 and PM 2.5 pollutants, during New Zealand's first round of COVID-19 restrictions (during alert Levels 4, 3 and 2). Observed changes to air quality under COVID-19 restrictions were compared to two reference values, LTM and ML predictions. The observed changes in pollution concentrations are discussed as well as the model performance considering both location and pollutant. We collected data for daily PM 10 , PM 2.5 and NO 2 measurements from 36 stations across New Zealand and classified them into five geographical clusters ( Table 2 ). The sites were chosen based on data availability and population, where restrictions were likely to have been more evident and have had greater impacts on air quality. The period considered was from January 1, 2015 to June 8, 2020, although the amount of historical data (2015-2019) varies from station to station. The air quality sites chosen are run according to national standards (Resource Management (National Environmental Standards for Air Quality) Regulations 2004) by local government using instruments that meet regulatory standards for air quality data collection (Ministry for the Environment, 2009). Where available, meteorological data were retrieved from the same stations recording air pollutant concentrations, including air temperature, relative humidity, wind speed and wind direction. For air quality stations that did not record meteorological parameters, the closest available station was used. All data were aggregated into daily values to align with the air pollutant concentrations. Both air pollutant concentration and meteorological data were mainly extracted through the Ministry for the Environment's data pipeline, which directly accesses councils' databases. Some data were received directly from councils or downloaded through their platform when it could not be accessed directly through the pipeline. Some meteorological data were obtained from the National Institute of Water and Atmospheric science (NIWA) CLIFLO database (https://cliflo.niwa. co.nz/) when those data were not collected by councils or data were not available due to instrument failures. More details on the location, available air pollutant concentrations, and meteorological data from monitoring stations can be found in Appendix 1. Daily traffic count for light vehicles and heavy goods vehicles (HGV) were collected by New Zealand Transport Agency (NZTA) at key traffic sites in the main urban centres of Auckland (region NU), Wellington (region NL), and Christchurch (region SL) from December 3, 2018 until June 08, 2020. This information was used as proxy of the amount of reduction on vehicle movements that occurred during each alert level. To test the relationship between traffic volume and air quality changes, Black Carbon (BC) data have been used. BC is a component of PM mass and is a stable particulate that exists in the fine/accumulation modes in the atmosphere; it therefore has a residence time longer than NO 2 and is on a par with fine particulates. The key emission source of BC is diesel emissions typical of those emitted from HGV. The wavelength of 880 nm is used singularly as the known indicator for diesel-emitted BC. BC data were collected from online aethalometers at previously reported sites in Auckland (Henderson), Wellington (Willis Street) and Christchurch (St Albans). This wavelength allows for separating the contribution from home heating (330 nm), another common BC source across New Zealand. To assess changes in air quality during the national COVID-19 restrictions, each alert level, which poses different limitations, has been investigated separately. In this study, we examine data from the full periods for Level 4 (26 March -27 April 2020), Level 3 (28 April -13 May), and Level 2 (14 May -8 June) (Table 1) . We compared two methods to estimate the concentrations of each air pollutant that would have been observed without restriction measures: 1) the long-term mean of historical data (2015-2019), and 2) predicted values, incorporating meteorological variability, using the random forest machine learning algorithm. We grouped data into historical (LTM) (2015-2019) and observed (OB) (2020) datasets. Historical data were further grouped into average of Julian day (day of a year) to calculate the availability across historical period of each alert level. For example, air quality data on 26th of March in 2015-2019, inclusive, from each site was averaged as a historical 26th of March value and counted as one data point in the historical Level 4 period. This approach was also applied to meteorology data. Given that each alert level lasted for a short period of time (16-33 days), we only considered sites with at least 50% of air quality and meteorological data available during each alert level period and its associated historical period. Analyses were done using R 3.6.1 (R Core Team, 2020) and the caret (v6.0.86), randomForest (v4.6.14; Liaw and Wiener, 2001), openair (v2.7.4; Carslaw and Ropkins, 2012) and ranger (v0.12.1) packages. It is understood that mobility was most affected by COVID-19 restrictions, and, as such, transport volume, and therefore emissions, altered. To help understand these changes, traffic data and fleet type data were collected by New Zealand Transport Agency during the COVID-19 alert level periods (Waka Kotahi NZ Transport Agency, 2020). Major road junctions in three New Zealand cities were considered. The data from these sites are extrapolated as proxies for changes in on-road vehicle volume across New Zealand. The similarity of volume changes among the three cities in accordance with alert levels supports this assumption. To understand how 'typical' meteorological conditions were during the COVID-19 alert levels for each region, the percentage difference between the two periods has been calculated. The results provided information on the potential changes in air pollutant concentrations under BAU. We applied two methods to estimate a reference concentration value for each air pollutant that could have been expected in the absence of COVID-19 restrictions. The first was the average of historical data (2015-2019), named long-term mean (LTM). The second was derived using a random forest machine learning algorithm (ML). We then examined the differences between observed values and BAU estimates using the LTM and ML. Air quality observations during each alert level were assessed for differences compared to their respective LTM based on 5 years of daily averages (2015-2019), where available. This allows for reasonable comparisons between air quality observations during each COVID-19 alert level and what could be expected under BAU, had pandemic restrictions not been imposed. 2.2.3.2.1. Machine learning using random forest algorithm. Random forest (RF) is a machine learning algorithm used for classification and regression to unfold patterns and relationships and is considered to be one of the most powerful tools in many fields (Chen and Ishwaran, 2012; Herrera et al., 2019) . The algorithm builds many regression trees where the target variable takes continuous values to exercise binary recursive partitioning. It is an iterative process that splits the data into partitions or branches into smaller groups. These regression trees are then aggregated (combines the results of multiple predictions) with two layers of randomness to ensure it goes through a robust learning process: 1) different samples for training, and 2) randomly selected predictors that build many regression trees as a result and will collectively decide the final output values. These two modifications make fair use of all potentially predictive variables and prevent overfitting. In this study, we used daily concentrations of air pollutants (PM 10 , PM 2.5 , and NO 2 ) from the historical dataset (2015-2019) as dependent variables to grow RF models. The explanatory variables were air temperature, relative humidity, wind speed, wind direction, Unix time (number of seconds since January 1, 1970), week of the year, and day of the week (Fig. 1) . The last three temporal variables are used to account for the seasonal variability, and potential changes in long-term emissions. We used k-nearest neighbours' imputation (Batista et al., 2002) for pre-processing the explanatory variables. For an arbitrary sample, the five closest neighbours (values from five observations that are like the current observation) was found in the training set and the value for the predictor is imputed using these values (using the mean). This approach also centred and scaled the data to ensure that all explanatory variables give equal contribution to the analysis. For each site, we developed a model for each air pollutant. We used 80% of data as a training set for identifying the best model and 20% of data as a test set for testing the final model. Three hyper parameters were tuned to find the best learners: the number of variables (called mtry), node size (1, 5 and 10) and the number of trees (300, 500 and 800). Ten-fold cross-validation was used to evaluate the RF models, which randomly splits the training data to ten sets of approximately equal size. The model was run ten times for validation, during which each unique group was used as a test set and the remaining as training set. This means each sample has an equal chance of being used as training and test samples, minimising the bias that may occur in a simple, one-off training and test set split. Root Mean Square Error (RMSE), R squared, and Mean Absolute Error (MAE) were calculated in each run to evaluate the model performance. The results from the ten runs were then averaged as the results for the model. After retrieving the results of model performance, the RF algorithm used the whole dataset as a training set to build the final model. Finally, we used the test set to evaluate the model and evaluate the expected accuracy of the models with Normalised Root Mean Square Error (NRMSE), R squared, and MAE. The potential deviation from the predicted value was quantified using the mean residuals (predicted minus observed concentration of air pollutants) for each site. We computed both average and 7-day running average of the daily residuals, and the associated 5th and 95th percentiles were derived as the uncertainty interval. 2.2.3.2.2. Evaluation of machine learning models. We used MAE, NRMSE and R squared to examine whether the model predictions are valid reference values. We compared the estimated values from the models for the historical period (2015-2019) to its historical observed value (testing dataset) to calculate these metrics. MAE and RMSE both measure the accuracy of the predictions versus observed values. The main differences between two metrics are: 1) RMSE penalises large errors and increases with the variance of the frequency distribution of error magnitudes and 2) RMSE tends to be much larger as test sample size increases (Chai & Draxler, 2014; Willmott & Matsuura, 2005) . Here, we applied NRMSE to remove the scale-dependent nature from RMSE which allows it to compare between accuracies of different datasets. In general, the lower the values, the higher the accuracy. However, zero error is not what machine learning models are aiming for as the model should be flexible and be able to predict in a set of new values instead of memorising the whole training dataset. On the other hand, R squared represents the proportion of the variance for an air pollutant concentration that is explained by meteorological and time variables that we used. These three metrics together explain whether the models have considered relevant variables that affect air pollutant concentrations from the site and reasonably predicted air pollutant concentrations. Average variable importance was also examined for the models for each air pollutant to justify the inclusion of the variables and to understand their impacts. During the Level 4 lockdown, vehicle flow reduced by approximately 79% for light vehicles and 72% for heavy goods vehicles (HGV) (Waka Kotahi NZ Transport Agency, 2020) across the country (Fig. 2) . During Level 3, HGV levels returned to 19% lower than the previous year. However, light vehicle volumes were about 46% lower than the previous year. During Level 2, traffic levels remained lower than 2019 for light vehicles (− 20%) and for heavy good vehicles (− 7%). Linear regression (Fig. 3) showed strong relationships between BC concentrations at monitoring stations and traffic counts for both light and heavy vehicles at all three key traffic sites, whereas the relationships in South Lower appeared to be weaker. The South Lower site showed relatively weak relationships between vehicle number and BC concentrations. The lower correlations indicate the likelihood of confounding factors influencing the data, most plausibly local meteorology, with the South Lower collection site close to the Pacific Ocean. Slight deviations of the airflow would bring clean air from the Pacific rather than more polluted city air over the measurement site. Another possibility is that the South Lower on-ramp vehicle count location was less representative of traffic flow across that region during different alert levels. However, given the similarity in traffic volume changes across regions, this appears unlikely. Overall, the results here show that extrapolation of traffic volume from one road can be used as a proxy for emitted pollution elsewhere in the same city, albeit with an understanding of meteorology and the pollutant's characteristics. Meteorological conditions during COVID-19 alert levels in 2020 differed from past years (2015-2019) to varying degrees, with the most meteorological anomalies occurring during Level 2 ( On the other hand, meteorological conditions during Level 2 were markedly different than the reference period: RH and wind speed. Concentrations of PM 10 were reduced during Level 4 across all regions, ranging from 11.5% (South Lower) to 34.1% (South Upper), all lower than predicted by LTM (Table 4) . During Level 3, PM 10 concentrations were closer to the LTM with a regional spread from 6.6% above predicted for the North Central region and 8.1% lower for the North Lower. During Level 2, PM 10 concentrations were more varied according to region with North Upper reporting observed concentrations 13.8% lower than LTM, whereas the adjoining North Central region showed concentrations almost 30.9% higher than LTM (Table 4) . During Level 4 all regions had lower PM 2.5 concentrations than the LTM; the largest reduction was in South Upper with a 22.6% reduction whilst North Upper had just a 2.0% decrease on LTM (Table 5 ). PM 2.5 NO 2 concentrations markedly reduced during Level 4: all regions reported similar proportional reductions compared to LTM, ranging 48.0%-54.5% below the LTM (Table 6 ). All regions remained below the LTM during Level 3, although to a lesser degree, with a range from 3.7% (North Lower) to 44.7% (North Upper). The upwards trend in NO 2 concentrations continued during Level 2, NO 2 concentrations were 11.3% higher than LTM for the North Lower region during Level 2: however, concentrations for North Upper (19.9%) and South Lower (11.9%) were still reduced (Table 6 ). For PM 10 , MAE ranges from 2.7 to 6.3 (μg/m 3 ) and NRMSE ranges from 0.3 to 0.9 (μg/m 3 ) (Fig. 4) . MAE was generally higher in South Lower and lower in North Lower. Large variation of NRMSE is seen in most regional groups, except for South Upper. R squared varied widely from 0.07 to 0.84 with a median of 0.47. R squared values were relatively high at all stations in the South Island, whereas lower R squared were mostly found in the North Island. ML modelling error results for PM 2.5 showed MAE ranges from 1.3 to 4.1 (μg/m 3 ) and NRMSE ranges from 0.2 to 0.5 (μg/m 3 ) (Fig. 5) . A wide range of R squared values were found, from 0.2 to 0.8 with the median at 0.6. All measures showed large variations in the North Island but less variation from the South Island. Error measures in NO 2 ranged similarly to PM 10 and PM 2.5 , whereas R squared values were generally higher, with all values over 0.5 (Fig. 6) . For NO 2 , MAE had a range of 1.6-6.7 (μg/m 3 ) across regions and NRMSE were between 0.2 and 0.5 (μg/m 3 ), where variations in each region were found to be the highest in North Upper. NO 2 models yielded overall high R-squared from 0.5 to 0.8. The values were higher at stations in the South and lower at stations in the North. For all air pollutants targeted in this study, the week of the year had been the most important variable for modelling, followed by wind speed (Fig. 7) . Wind direction was an important factor as well. Mixed results were shown in air temperature, Unix time and RH, which are all more important for particulates than NO 2 . Conversely, the day of the week had only been important for NO 2 . Using the RF machine learning algorithm, a predicted air quality concentration for each alert level was calculated based on observed meteorological conditions (see Section 2.2.3.2). These were then summarised for each region and compared to 2020 measurements and LTM. This comparison helps to support the estimated scale of changes in observed concentrations for different pollutants during the COVID-19 alert levels. For PM 10 , OB values were notably reduced during Level 4 when compared to estimates from both LTM and ML (Fig. 8) , indicative of reduced emissions during this strict lockdown period. The ML predicted much higher concentration for the three North regions than was observed or suggested by LTM. For South Upper and South Lower, ML predictions closely matched the LTM but were, again, higher than observed concentrations. During Level 3, OB concentrations of PM 10 For Level 4, PM 2.5 OB values were lower than reference values produced by LTM or ML. (Fig. 9 ). ML modelling estimated higher values than OB or LTM for all Northern regions and were like LTM for Southern regions. During Level 3, PM 2.5 concentrations increased across all areas. ML estimates were close to OB for North Upper, North Lower and South Upper. During Level 2, OB PM 2.5 concentrations were higher than either LTM or ML estimates in the South. North Upper and North Lower OB data were close to LTM, whilst OB levels were slightly lower than LTM for the North Central region. For NO 2, OB levels were notably lower than either LTM or ML estimates (Fig. 10) . Observed NO 2 concentrations increased across all regions during Level 3, however, the OB values remained below the LTM and ML values (Fig. 10) . For the Level 2 period, ML and OB results were similar for the North Upper and North Lower regions with LTM and ML estimates like each other but above the observed concentrations for the South Lower region. During the most severe social restrictions of Level 4, NO 2 concentrations were most reduced when compared to LTM (between 48.0 and 54.5%). This compares to 34.0-57.0% reduction during Level 4 restrictions across the Auckland region (Patel et al., 2020) . The results are also comparable to research findings reported globally (Tobías et al., 2020) . NO 2 reduction can be associated directly with the removal of on-road vehicle emissions. Patel et al. (2020) found that the NO 2 reduction during Level 4 in Auckland was not linear with reductions in traffic volumes. The fact that there was proportionally more HGVs on the road than light vehicles during Level 4 (Fig. 2) could explain this. Diesel-fuelled bus services were also operating throughout, at a slightly lower frequency, for the use of essential workers. Buses have been found to be major contributors to both black carbon and NO 2 concentrations, especially along busy urban canyons (Talbot and Lehn, 2018) . It should also be considered that complex physico-chemical processes within the atmosphere would mean a simple linear reduction in overall emissions and real-time observed concentrations was unlikely (Gulia et al., 2015) . The rapid increase in measured concentrations once the alert levels eased showed how quickly NO 2 concentrations respond to changed emissions, with the link to societal behaviour apparent. Still, this finding offers interesting insight into transport policy given the current push to reduce private vehicle transport in populated areas and to increase electric vehicles. Across New Zealand, PM 10 were reduced 11.5-34.1% across the country and PM 2.5 reduced 2.0-22.6%. Patel et al. (2020) reported a 7-20% reduction for PM 10 and 8-17% for PM 2.5 across Auckland during Level 4. Here we report a reduction of 12.1% for PM 10 and 2.1% reduction for PM 2.5 across the north upper region which contains the Auckland airshed. The variability of concentrations is in accordance with the key source(s) of particulate. PM 10 concentrations are dominated by sea salt near coasts, dust and resuspended road dust, and wood smoke from wood burning for fuel. There are also notable emissions from outdoor burning and a smaller fraction from tailpipe emissions (Ministry for the Environment, 2018). The mixture of natural and anthropogenic particulate sources should reduce PM 2.5 , and to a lesser degree PM 10 , concentrations in response to human behaviour pattern changes. PM 10 concentrations were reduced compared to the LTM across New Zealand during Level 4, even though only a small proportion of PM 10 derives from on-road vehicle combustion. The decreases in observed concentrations may also be explained by less resuspended road dust loading from lower vehicle numbers using the roads. PM 10 concentrations are also very heavily influenced by meteorology across New Zealand: higher winds can bring high sea salt concentrations elevating PM 10 mass across coastal areas (Davy et al., 2017) . Conversely, lower wind speeds can help elevate PM 10 and PM 2.5 concentrations due to a build-up of wood smoke from home heating sources, and this appears to be the case for the South Upper and South Lower during Level 2, when temperatures lowered, and wind speeds were found to be anomalously lower than LTM. It is likely that during Level 2, household emissions increased due to higher home occupancy during this period when compared to pre-COVID times. Wind direction was anomalous during Level 2 across most of New Zealand (Table 3) , which likely influenced PM 2.5 concentrations. A North Upper regional site located in Auckland in the Henderson area provides a good example. The location experienced mostly normal wind conditions during Levels 4 and 3 (Fig. 11 ). During these alert levels, NO 2 concentrations were reduced in line with most roadside measurement stations across New Zealand (Fig. 10) . However, PM 10 concentrations were close to LTM (Fig. 8) . During Level 2, an unusual wind direction prevailed emanating from the southeast quadrant instead of the usual south westerlies (Fig. 11) . The result of anomalous wind direction was a reduction in PM 10 concentrations (Fig. 11C ) due to the air flow no longer coming from the Tasman Sea and the associated sea salt loading that dominates Auckland's air during this season (Davy et al., 2017) . Conversely, NO 2 concentrations increased during Level 2 compared to the LTM ( Fig. 11A and B). The increase in NO 2 can be attributed not just to the increase in traffic flow during this period compared to Levels 4 and 3 ( Fig. 2) but could also be a product of air being drawn from the south easterly sector, a more urbanised neighbourhood of Auckland with light industry and many busy roads as compared to the prevailing south westerly, which draws air from the Waitakere Forest and, ultimately, the Tasman Sea (Fig. S3 ). In PM 10 and PM 2.5 , MAE reflected the reasonable amount of prediction errors in the daily average, which showed a good level of prediction from the models (Figs. 4 and 5) . The larger errors found in the Southern regions were likely to results from the cooler autumn and early winter period, when particulate concentrations were higher due to residents lighting more fires for home heating purposes, a common practice in New Zealand (Ministry for Environment, 2018) . NRMSE has taken the scale of the air pollutant concentration at each site into account, which is a more comparable model accuracy measure. It appeared that accuracy of PM 10 models is similar across regional clusters, where outliers were found at most of the regions. With fewer PM 2.5 sites than PM 10 across New Zealand, PM 2.5 results were much more scattered within each region dependent on the key local emission soirce. High variation in R squared indicated that meteorology and time variables may not be sufficient in explaining much of the concentrations, especially for Northern stations. Furthermore, many coastal stations, especially in the North Upper region, showed poor (<0.3) R squared values (Fig. 11 ). This suggested that there might be other variables that are significantly affecting PM 10 and PM 2.5 concentrations. An explanation could be due to the dominant natural source of particulates, sea salt, which is hard to predict by meteorology or trend variables (Davy et al., 2017) . The temporal aggregation (daily average) used in this study may have increased the difficulties in differentiating between the impacts of sea salt from the changeable meteorology at coastal stations, where the diurnal pattern of wind speed and direction is often caused by the changes in overland air temperature and sea surface temperature (Tian et al., 2018) . The diurnal pattern of sea breezes may have reduced the interpretability of the model to better predict PM 10 and PM 2.5 at coastal stations. This would be especially true where competing PM emissions, such as wood smoke from home heating, would possibly peak when wind speeds (and therefore sea salt loadings) are low. On the other hand, overall high R squared values from NO 2 models have suggested that R squared values were especially low at coastal stations for both PM 10 and PM 2.5 , whereas they were high in general for all NO 2 stations. meteorology and time variables account for most of the variation in NO 2 concentrations in the historical dataset (Fig. 12) . The R squared values described here are commensurate to those of international studies (Petetin et al., 2020) . Both metrics saw a larger variation of results in North Upper which may be driven in some part by the higher traffic volume in heavy vehicles. While North Central yielded the lowest MAE, models in South Lower seemed to be more accurate due to lower NRMSE. Results have indicated that the developed RF algorithm performed better at predicting NO 2 . While error metrics implied that predictions from all models were acceptable, R squared have hinted that there are more local factors affecting concentration of particulates. More research is needed to reveal other factors that might be helpful in predicting particulates concentrations. The Henderson site is situated in a location with oceanic influences, such as sea salt on its PM 10 concentrations. In fact, during summer months sea salt dominates the PM 10 fraction at this site (Davy et al., 2018) . The stability of the sea salt component and the response of the model might indicate that the model is over-predicting the sea salt component. It has been shown that ML methods had lower R-squared values for coastal influenced sites. To test this, Henderson PM 10 data can be compared to an air quality site located inland -Alexandra, a South Island township in the South Lower region that has high wintertime PM 10 concentrations dominated by wood smoke from home heating. Observed PM 10 concentrations in Henderson were lower than BAU, as indicated by LTM or ML ( Fig. 13A and B) . The case of ML predictions overestimating PM 10 concentrations occurred for several other coastal stations. By contrast, for the Alexandra South Lower site, ML estimates and LTM are more closely aligned to OB for the alert level periods (Fig. 13C and D) . This exemplifies the better model performance on inland stations that have a stronger single PM 10 emission source, which in Alexandra's case is wood smoke from domestic fires. With significantly lighter winds across many regions during Level 2, coastal regions experienced lower sea salt concentrations, reducing mass loading in the PM 10 fraction ( Fig. 13A and B) , with a marked drop in the observed data when compared to the LTM. By contrast, the same reduction of wind speed for inland areas when cooler conditions prevail would reduce dispersion of wood smoke from home heating, allowing higher concentrations. For air quality monitoring sites that are influenced by both wood smoke and sea salt, such as Henderson, the confounding factors most likely explain, in part, low R squared values of ML models. To test the ML performance results for NO 2 , the Henderson site was compared to a North Lower air quality station, Wellington Central (Willis Street). The Wellington Central permanent air quality station is next to a busy road in the city. During Level 4, a sharp drop in NO 2 concentrations occurred (Fig. 14A-D) , a result of the removal of the direct emission source for the location, namely on-road vehicles. The attribution of NO 2 reductions to human behaviour changes in emissions is simpler than that of PM 10 given that almost all NO 2 derives directly from diesel-and petrol-powered vehicle emissions, with low concentrations from natural sources or imported into New Zealand (Xie et al., 2019) . During Level 3, Wellington Central concentrations rapidly returned to values close to the LTM, whilst Henderson remained lower than LTM. In Level 2, observed concentrations were close to LTM and ML estimates. ML is more closely aligned to observed NO 2 concentrations than LTM, even at the same Henderson site under the same meteorological conditions (Fig. 14) . ML modelling also appears to reflect the nuanced increase in concentrations with the country moving into colder winter conditions as alert levels decreased. Wintertime in New Zealand generally has the highest NO 2 concentrations with increased traffic volume and reduced vertical mixing due to less soil heating of the surface. This study considers the COVID-19 alert level restrictions that, in effect, created a "natural experiment" on New Zealand's air quality. As noted elsewhere, the sudden disruption of public movement brought about by lockdowns reduced air pollution significantly. However, the amount of initial reduction varied considerably among different sites and pollutants. Transport-related NO 2 concentrations reduced 48.0% to 54.5% during the most restrictive alert level period, reflecting the immediate reduction in combustion emissions and replicating findings reported in Auckland (Patel et al., 2020) and like reductions measured in other locations with COVID-19 restrictions (e.g., Tobías et al., 2020) . Nationally, reductions in NO 2 concentrations were similar across the country during Level 4, but some regional differences emerged in NO 2 levels as restrictions eased through Levels 3 and 2. These differences are most likely explained by changes in traffic volumes during each alert level along with the proportion of vehicle types (diesel and petrol) using the roads during these periods. PM 10 and PM 2.5 was found to decrease less than NO 2 compared to both LTM and ML during Level 4. The scale and regionality of their proportional changes in particulate loading are a result of multiple sources contributing to particulate mass concentrations, including substantial contributions from oceanic sources (sea salt and secondary sulfates) along coastal fringes (Davy et al., 2017) . As the alert levels eased, concentrations in both PM 10 and PM 2.5 increased; however, the scale of the increase was more notable in South regions where the burning of wood for home heating is more prevalent during this cooler period of the year. While public movements decreased significantly during the COVID-19 alert levels, there was an increase in time spent at home, indoors. This is reflected in increases in PM 10 and PM 2.5 concentrations during Level 2, most likely due to increased wood burning for domestic heating. These domestic emissions occurred as the South regions became colder during this period and home occupancy was higher than usual due to travel restrictions. BAU reference values were developed using LTM and ML modelling. The ML modelling closely represented observed NO 2 concentrations across the country once social restrictions were eased. However, during Level 4 both LTM and ML were much higher than observations, a direct result of the reduction of on-road vehicle transport during this time. The closeness of the ML and LTM results during the return of more normal traffic volume during lower alert levels adds confidence to the scale of reduction during Level 4. The ML modelling for PM 10 and PM 2.5 results are more mixed, with closer analysis showing that the R squared values of models were lower along coastal fringes, where the predictions were often higher than the observed values. The complexity of the land-ocean interaction on particulate concentrations and, more broadly, the larger number of particulate sources most likely help explain this important finding. Contrary to this, for monitored inland regions, ML predictions more closely matched LTM. The results favour a mixed approach to creating BAU reference values, with ML modelling favoured for gas pollutants and particulates away from coastal influences, while LTM is better for those coastal environments. Using such an approach, future work could include land use factors and direct emissions in modelling computation. Writing -Review & Editing. Samantha Lay Yee: Conceptualization, Investigating, Writing -Review & Editing. Declaration of competing interest No known competing interests Spatial estimation of urban air pollution with the use of artificial neural network models A study of K-nearest neighbour as an imputation method Changes in U.S. air pollution during the COVID-19 pandemic Openair -an R package for air quality data analysis Root mean square error (RMSE) or mean absolute error (MAE)?-Arguments against avoiding RMSE in the literature Random forests for genomic data analysis Source Apportionment and Trend Analysis of Air Particulate Matter in the Auckland Region The Role of Meteorology in Atmospheric Processes and Air Pollution Studies Recurrent Neural Network and random forest for analysis and accurate forecast of atmospheric pollutants: a case study in Hangzhou Random forest meteorological normalisation models for Swiss PM10 trend analysis Long-range air pollution transport in East Asia during the first week of the COVID-19 lockdown in China Urban air quality management-A review Random forest implementation and optimization for Big Data analytics on LexisNexis's high performance computing cluster platform Reductions in traffic-related black carbon and ultrafine particle number concentrations in an urban neighborhood during the COVID-19 pandemic Assessing air quality changes in large cities during COVID-19 lockdowns: the impacts of traffic-free urban conditions in Almaty OpenLUR: off-the-shelf air pollution modeling with open features and machine learning Classification and regression by randomforest Impact of meteorological conditions and air pollution on COVID-19 pandemic transmission in Italy Std_n_ Guides/good-practice-guide-for-air-quality_NZealand_usedby_SAWS.pdf. Ministry for the Environment COVID-19 pandemic: impacts on the air quality during the partial lockdown in São Paulo state The national climate database (CliFlo). n.d, Retrieved Implications for air quality management of changes in air quality during lockdown in Auckland ( New Zealand ) in response to the 2020 SARS-CoV-2 epidemic Meteorology-normalized impact of COVID-19 lockdown upon NO2 pollution in Spain R: A language and environment for statistical computing. R Foundation for Statistical Computing Lessons from New Zealand's COVID-19 outbreak response Effect of restricted emissions during COVID-19 on air quality in India Amplified ozone pollution in cities during the COVID-19 lockdown The impact of COVID-19 partial lockdown on primary pollutant concentrations in the atmosphere of Rio de Janeiro and São Paulo Megacities (Brazil) The Impacts of Transport Emissions on Air Quality in Auckland's City Centre Spatial, seasonal and diurnal patterns in physicochemical characteristics and sources of PM2.5 in both inland and coastal regions within a megacity in China Changes in air quality during the lockdown in Barcelona (Spain) one month into the SARS-CoV-2 epidemic Weekly traffic count information. Weekly traffic count information | Waka Kotahi NZ transport agency Isolation, quarantine, social distancing and community containment: pivotal role for old-style public health measures in the novel coronavirus (2019-nCoV) outbreak Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance Coronavirus disease (COVID-19) situation reports. Coronavirus disease (COVID-2019) situation reports Auckland Air Emissions Inventory 2016 Has air quality improved in Ecuador during the COVID-19 pandemic? A parametric analysis. Air Quality, Atmos Supplementary data to this article can be found online at https://doi. org/10.1016/j.atmosenv.2021.118322.