Unusual response of O₃ and CH₄ to NO₂ emissions reduction in Japan during the COVID-19 pandemic

ABSTRACT

We investigate the impact of NO₂ emissions reduction on O₃ and CH₄ levels in 14 metropolitan areas of Japan in 2020. To account for meteorological variations, we employ business-as-usual air quality time series generated by machine learning models. Additionally, we use satellite observations and biogeochemical model simulations to analyse air quality changes. During the lockdown period from April 7 to May 25 in 2020, we observed a NO₂ reduction that equated to a decrease equivalent to 3.4 and 5 years of the corresponding trends in roadside and ambient air quality recorded from 2010 to 2019. After meteorological normalization, NO₂ decreased by 14.5% at ambient air stations and 19.1% at roadside stations. Surprisingly, the NO₂ reduction did not immediately lead to increased O₃. Instead, O₃ levels rose after the lockdown, specifically in August due to favorable sunny conditions. This finding is important for Japan and has not been reported in previous studies. We found that changes in NO₂ and CO marginally contributed to variations in CH₄ levels across the study areas. It is recommended to simultaneously reduce NO_x as well as non-methane volatile organic compounds to mitigate their adverse effects on future policies.

KEYWORDS:

• NO₂
• O₃
• CH₄
• Japan
• COVID-19
• weather-normalization

1. Introduction

Nitrogen dioxide (NO₂) is an important air pollutant that raises significant concerns due to its negative effects on human health (Hamra et al. 2015). Additionally, it serves as a crucial precursor to tropospheric ozone (O₃), along with volatile organic compounds (VOCs) (Akimoto and Tanimoto 2022). Nitrogen oxides (NO_x = NO + NO₂), carbon monoxide (CO) and non-methane volatile organic compounds (NMVOCs) have an influence on the methane (CH₄) lifetime by affecting the atmospheric mixing ratio of hydroxyl radicals (OH) (Akimoto and Tanimoto 2022), which act as a primary sink for CH₄ (Turner, Frankenberg, and Kort 2019). Both O₃ and CH₄ are short-lived climate pollutants (SLCPs) that contribute to positive radiative forcing, thereby intensifying global warming (Akimoto and Tanimoto 2022). Moreover, owing to its short lifetime in the atmosphere and significant signal compared to carbon dioxide (CO₂), NO₂ possesses the potential to serve as an indicator for monitoring localized fossil fuel CO₂ emissions (Miyazaki and Bowman 2023).

In 2020, the implementation of COVID-19 social distancing policies in multiple countries led to a significant decrease in human activities worldwide (de Palma, Vosough, and Liao 2022). While the general anticipation was for a reduction in NO₂ emissions in many cities due to the decline in anthropogenic activities (Bauwens et al. 2020; Barré et al. 2021; Cooper et al. 2022), the response of O₃ and CH₄ concentrations has been unexpected.

Increased levels of O₃ have been observed in northern Europe, China, and South Africa as a consequence of the COVID-19 lockdown, according to sensitivity simulations conducted using the MIROC-CHASER global chemical transport model (Miyazaki et al. 2021). This rise in O₃ can be attributed to the general reduction in NO_x, which enhances O₃ production by reducing NO titration in areas with high levels of NO_x pollution or VOC-limited areas (Akimoto and Tanimoto 2022). Furthermore, meteorological effects have played a significant role in the changes observed in O₃ levels between 2020 and the reference year (Ordóñez, Garrido-Perez, and García-Herrera 2020; Liu et al. 2021). Despite accounting for the influence of weather conditions, significant variations in O₃ level estimates have been reported across studies, particularly in European countries (Ordóñez, Garrido-Perez, and García-Herrera 2020; Grange et al. 2021), and China (Liu et al. 2021; Shi et al. 2021). The presence of sunlight is essential for the O₃ generation in response to the decrease in NO_x during the lockdown period. As a result, the lack of sunny conditions in specific urban areas at the time of the atmospheric response to NO₂ reduction may have led to differing time delays before observable changes in O₃ levels occurred (Grange et al. 2021).

In 2020, during the COVID-19 pandemic, global CH₄ emissions experienced a significant growth rate, which was contrary to the expected decrease in anthropogenic CH₄ emissions due to the implementation of lockdown measures (Peng et al. 2022). In 2020, anthropogenic CH₄ emissions only slightly decreased compared to 2019, while wetland emissions rose sharply. This increase in wetland emissions was likely influenced by unusually warm and wet weather in the Northern Hemisphere (Peng et al. 2022), which could be connected to the impact of climate change (Zhang et al. 2023). Apart from the variation in CH₄ emissions, it was found that the decrease in hydroxyl radical (OH) concentrations due to changes in air pollutants like NO_x, CO, and NMVOCs during the COVID-19 pandemic mainly accounted for approximately half (53 ± 10%) of the observed global CH₄ level growth in 2020 (Peng et al. 2022). A similar finding regarding the effects of NO_x, CO, and NMVOC emission changes on 2020 methane levels is reported by Stevenson et al. (2022). However, other studies using Greenhouse gases Observing SATellite (GOSAT) observations indicated that most of the observed increases in atmospheric CH₄ during 2020 and 2021 can be attributed to increased CH₄ emissions (Qu et al. 2022; Feng et al. 2023). Although CH₄ has a long estimated lifetime of 8–10 years and has mostly been discussed at the global level, it is important to note that policies and approaches to address CH₄ emissions may vary locally.

In 2020, Japan also experienced the impact of the COVID-19 pandemic, and in response to prevent the virus’s spread, a state of emergency was declared from April 7 to May 25. This measure resulted in the suspension of various economic activities and imposed restrictions on people’s mobility. As a consequence, there was a significant decline in mobility trends (Figure 1a). The reduction was more pronounced during the weekend compared to the weekday (Damiani et al. 2022), as illustrated in Figure 1b.

Figure 1. Mobility changes for six prefectures in Japan (Aichi, Fukuoka, Tokyo, Osaka, Kyoto, and Hyogo) in 2020 based on Google’s mobility indices for time series (a) and for days of the week (b).

Although the primary aim of the lockdown was not specifically to address air pollution and greenhouse gas emissions, the implementation of these measures offers valuable insights for atmospheric modeling. It can provide practical knowledge and first-hand experience to develop more efficient strategies for mitigating air pollution and reducing greenhouse gas emissions in the future (Grange et al. 2021). It is important to note that the changes in air pollutant concentrations during this period varied across regions and were strongly influenced by meteorological conditions. A regional analysis of these changes could provide evidence to support the formulation of appropriate regional policies in the future. In this study, our objective is to evaluate the impacts of changes in anthropogenic activities during the COVID-19 pandemic (from April 7 to December 31) on NO₂, O₃, CO and CH₄ concentrations in metropolitan areas (MAs) of Japan in 2020, a topic which has not yet been thoroughly investigated.

In the first phase (Section 2), we gather data from ground observations, satellite sources, and biogeochemical model simulations. Subsequently, we construct a weather normalization model under business-as-usual (BAU) conditions utilizing machine learning techniques, incorporating meteorological, spatial, and temporal predictors (Section 3). We investigate variations in air pollution levels by analysing the BAU predictions alongside additional data in Section 4. Lastly, we provide discussions in Section 5, while in Section 6, we present our findings, conclusions, and recommendations for future policy considerations.

2. Data

2.1. Study area

Prior research has primarily focused on assessing the impact of pandemic lockdown measures on air quality within the Greater Tokyo Area, being the most densely populated metropolitan area globally (Damiani et al. 2022; Zoran et al. 2023). Nevertheless, there is a notable absence of similar analyses for other MAs. Our study covers 14 MAs in Japan, from Sapporo in the north to Kagoshima in the south (Figure 2). We focus on these metropolitan areas as they house Japan’s most highly populated and vibrant cities, where we can observe the intricate connections between human activities and air pollution in Japan.

Figure 2. The locations of 14 metropolitan areas and the distribution of ground observations for air quality monitoring in Japan.

2.2. Ground observation

To acquire air quality data, we gathered ground observations for NO₂, O₃, CO, and CH₄ concentrations from the air quality monitoring data archive published by the National Institute for Environmental Studies (NIES). These observations spanned a ten-year period from 2010 to 2020 and were collected from 1,180 stations for NO₂, 835 stations for O₃, 383 stations for CH₄, and 237 stations for CO. The study utilized two types of stations: roadside air monitoring stations (RsAMS), which are placed in areas prone to air pollution from vehicle exhaust caused by traffic congestion, like intersections, roads, and near road edges, and ambient air monitoring stations (AAMS), which are established to assess air pollution in general living spaces such as residential areas. These station types have been categorized by NIES, and the data can be readily acquired from the original downloadable dataset.

Apart from air quality data, we incorporated ground observations of meteorological data from the Japan Meteorological Agency (JMA) as input features for the BAU models used in the study. Specifically, we obtained daily records from 52 weather stations located within the same 14 MAs. At each weather station, we gathered temperature, wind direction and speed, local atmospheric pressure, and relative humidity, as suggested by Grange et al. (2021). The corresponding meteorological parameters were extracted from the nearest weather observation site for each air quality station.

2.3. ERA5 reanalysis dataset

Alongside the weather data collected from the ground stations in the NIES database, for the features of the BAU models, we incorporated additional daily data pertaining to boundary layer height, total cloud cover, downward solar radiation (SR), and total precipitation, as recommended by Shi et al. (2021). This supplementary information was sourced from the ERA5 reanalysis dataset (ERA5 hourly data on single levels from 1940 to the present) obtained from the Climate Data Store of the Copernicus Climate Change Service. Additionlly, the ERA5 2 m temperature variable (T2M) and SR will be utilized to assess the variation of sunny conditions during both the lockdown and post-lockdown periods within the study area. The original ERA5 data possesses a spatial resolution of 0.25°×0.25°.

2.4. Sentinel 5P TROPOMI

In this study, we utilized the Sentinel 5P (S5P) Tropospheric Monitoring Instrument (TROPOMI) data to evaluate the tropospheric formaldehyde-to-NO₂ ratio (FNR) specifically for the year 2020. This ratio serves as a key indicator for the sensitivity of tropospheric ozone production. The tropospheric NO₂ and formaldehyde (HCHO – as a proxy for NMVOCs) data was obtained from the S5P L3 product ‘OFFL/L3_NO2’ (based on processor version 1.2.x and 1.3.x) and ‘OFFL/L3_HCHO’ (based on processor version 1.1.x) collections from Google Earth Engine, respectively. To generate the comprehensive L3 S5P product, each operational level (L2) product underwent preprocessing and mosaicking using the harpconvert tool. Low-quality pixels were filtered out in the L3 NO₂ product by excluding those with QA values below 75% for the band ‘tropospheric_NO2_column_number_density.’ The resulting data, ready for download, is available with a spatial resolution of about 1×1 km²_.

2.5. Biogeochemical modeled CH₄ budget

In our assessment of CH₄ emission variations, with a specific focus on emissions from natural sources such as wetlands, we utilized CH₄ budget data obtained from the Vegetation Integrative SImulator for Trace gases (VISIT) (Ito et al. 2019). VISIT is a biogeochemical model that takes into account historical land use and climatic conditions to estimate CH₄ emissions (Ito et al. 2019). The CH₄ budgets generated by the VISIT model are now available and accessible through the Global Environmental Database provided by NIES, Japan (Ito et al. 2019). We utilized the global data versions ‘ver.2021.1_CH4Wetl_Cao’ (Ito 2021a), and ‘ver.2021.1_CH4Wetl_WH’ (Ito 2021b), which incorporate the Cao scheme (Cao, Marshall, and Gregson 1996), and the Walter and Heimann scheme (WH scheme) (Walter and Heimann 2000), to estimate CH­₄ emission for each MA, which offers CH₄ emission information at a spatial resolution of 0.5°×0.5°.

3. Method

3.1. Weather normalization model under BAU conditions

To accurately quantify the actual change in the levels of the four pollutants, we developed a weather normalization model under BAU conditions using machine learning. This model was specifically designed to simulate pollutant levels without the influence of COVID-19 restriction measures, using meteorological, spatial, and temporal features as inputs. The meteorological predictors utilized in our model include ground observation data such as temperature, wind direction and speed, local atmospheric pressure, and relative humidity. Additionally, we incorporated data from the ERA5 reanalysis dataset, which comprises boundary layer height, total cloud cover, downward solar radiation, and total precipitation. Temporal predictors included the Julian date (the number of days since January 1) and the day of the week. Furthermore, latitude and longitude coordinates of each station were utilized as spatial predictors. To develop the weather normalization models for each pollutant at both AAMS and RsAMS, we utilized data from the years 2016–2019, which offers a comprehensive timeframe to account for the diverse air pollution concentration fluctuations experienced across various meteorological conditions. Extending the period, such as from 2010 to 2019, would not accurately represent recent air quality trends due to the impact of past air pollution reduction policies. Conversely, a shorter timeframe, such as the pre-lockdown period months would not adequately capture the full range of meteorological variations. Overall, four separate weather normalization models were developed for each pollutant (NO₂, O₃, CO, and CH₄), taking into account the specific station type (RsAMS and AAMS).

We employed the LightGBM machine learning model (Ke et al. 2017), a gradient boosting decision tree algorithm which is suggested by Phan and Fukui (2023), to construct the BAU model using the aforementioned predictors. To fine-tune the model’s hyperparameters, we utilized Fast and Lightweight AutoML Library (FLAML) (Wang et al. 2021), a lightweight library specifically designed for accurately identifying optimal hyperparameters for models. During the training process, we utilized 70% of the station data within each metropolitan area (MA), while the remaining 30% was reserved for validating the model’s performance. Both the training and test data sets were randomly selected for each MA, ensuring unbiased representation across the dataset.

In order to evaluate the performance of the BAU model we utilized the following metrics mean bias error (MBE), normalized mean bias error (NormMBE), root mean square error (RMSE), normalized root mean square error (NormRMSE) and Pearson correlation coefficient (R) as suggested by Grange et al. (2021). The detailed results are presented in Figure 3 for each pollutant and station, average scores are shown in Table 1. In general, the model demonstrated strong performance with high R values (mostly R > 0.8) and low MBE and RMSE scores when applied to the test set for NO₂, O₃, and CH₄. Regarding CO, the model achieved a satisfactory R value (R > 0.73).

Figure 3. Detailed score of each station in the test set. For each station in the test set, we calculated the following scores and display them in the figure: Pearson correlation coefficient (R), root mean square error (RMSE), normalized root mean square error (NormRMSE), mean bias error (MBE), and normalized mean bias error (NormMBE).

Table 1. The performance of BAU model on the test set (30% station data) with the following metrics: Pearson correlation coefficient (R), root mean square error (RMSE), normalized root mean square error (NormRMSE) and mean bias error (MBE), normalized mean bias error (NormMBE). For the normalized MBE and RMSE, we normalize values for each station and then compute the mean.

Pollutants	Station type	R	RMSE	NormRMSE	MBE	NormMBE
NO2	AAMS	0.89	3.13	0.15	−0.12	−0.07
	RsAMS	0.88	4.84	0.10	0.30	−0.03
O3	AAMS	0.96	3.75	0.02	−0.37	−0.02
	RsAMS	0.96	4.92	0.06	−3.18	−0.16
CO	AAMS	0.73	0.84	0.17	0.00	−0.07
	RsAMS	0.77	1.23	0.13	0.39	0.04
CH4	AAMS	0.82	3.75	0.00	−0.29	0.00
	RsAMS	0.80	3.82	0.00	−0.26	0.00

3.2. Design of experiments

Our aim is to assess the alterations in NO₂ levels within 14 MAs during both the lockdown and post-lockdown periods in 2020. We also intend to explore how changes in NO₂ may influence the shifts in O₃ and CH₄ levels in each of these timeframes. Notably, we were motivated to undertake this investigation based on an observation of an unusual O₃ response to NO₂ reduction in the Greater Tokyo Area (Damiani et al. 2022), prompting us to study the response of O₃ and CH₄ in all 14 MAs across Japan.

We conducted three experiments to assess the impacts of NO₂ changes on O₃ and CH₄ levels. In the first experiment, we focused solely on quantifying changes in NO₂ levels using the time series observations and OBS-BAU estimate, which involved subtracting the BAU prediction from the observed data (OBS). In the second experiment, we expanded the analysis to include O₃, incorporating additional variables from the ERA5 (temperature – T2M and SR) and S5P datasets (FNR and HCHO). The last experiment included CH₄, incorporating the OBS-BAU estimate for CH₄ and NO₂, as well as the OBS-BAU estimate for CO, and simulated CH₄ emissions from wetlands using the VISIT model.

For the experiments, we selected April 7 to May 25 as the lockdown period, August 1 − 31 as the post-lockdown period for O₃ analysis, and June 1 to December 31 for CH₄ analysis. We selected these timeframes to better understand how the four air pollutants changed in response to the COVID-19 lockdown measures and the period after the lockdown.

4. Results

4.1. NO₂ level changes

We initially examined the monthly trend of observed NO₂ concentration levels across 1,180 stations in the 14 MAs from 2010 to 2019, and we compared these trends with the NO₂ levels observed during the lockdown in 2020 as depicted in Figure 4a. The results indicate that actual reductions in NO₂ levels during the lockdown in 2020 were lower than the trend observed during 2010–2019, specifically 2.7 parts per billion (ppb) for RsAMS and 2.2 ppb for AAMS. This implies that the NO₂ levels observed during the lockdown were equivalent to those in 2023 for RsAMS and 2025 for AAMS, based on trends observed during the period 2010–2019.

Figure 4. (a) Mean ground observation trend with the reduction in NO₂ due to the lockdown in 2020 for AAMS and RsAMS, (b) map visualization of the OBS-BAU estimate for NO₂ during the lockdown period, (c) seven-day rolling mean of 2020 observation (OBS), BAU prediction (BAU), and mean level of NO₂ from 2016 to 2019 for four MAs.

Prior studies have indicated the importance of considering meteorological factors when evaluating the effects of intervention measures (Ordóñez, Garrido-Perez, and García-Herrera 2020; Grange et al. 2021; Shi et al. 2021). In order to accurately assess the impact of the lockdown while isolating the effects of weather conditions, we computed the OBS-BAU estimates for all MAs as depicted in Figure 4b. Additionally, Figure 4c presents the complete time series of NO₂ levels in 2020 (OBS), the expected levels without the lockdown (BAU), and the average data from 2016–2019 for four MAs (Kanto, Kinki, Chukyo, Fukuoka). We only show the figures for four MAs to avoid overwhelming complexity and to provide a more manageable representation of the figures.

Overall, NO₂ levels exhibited a decline across most MAs. The decline in emissions was particularly significant in RsAMS compared to AAMS in most MAs, with an average reduction of 19.1% and 14.5% respectively. However, these reductions were smaller compared to those observed in European cities (Barré et al. 2021; Grange et al. 2021). Additionally, we observed that the reduction in NO₂ levels during weekends was more significant than on weekdays, primarily due to a substantial decrease in mobility during weekends compared to weekdays (Figure 1b). During the lockdown the average reduction in NO₂ levels for AAMS was 12.9% on weekdays and 18.4% on weekends. As for RsAMS, the average reduction stood at 18% on weekdays and 21.9% on weekends. For most MAs, even though the lockdown was lifted at the end of May 2020, NO₂ levels continued to decline until the end of December 2020. This continued decrease may be attributable to the sustained reduction in mobility from the start of the lockdown through the end of 2020 (Figure 1a). These findings are summarized in Table 2 and Table 3.

Table 2. OBS-BAU estimates for NO₂ and O₃ during the lockdown (April 7 – May 25) and post-lockdown (August 1 − 31). For time series estimate, we considered all days of the week. However, when considering weekdays, we only included Monday to Friday, while for weekends, we only accounted for Sunday and Saturday. The values are represented as mean (standard deviation).

Station type	Lockdown (April 7 – May 25)			Post-lockdown (August 1 − 31)
	Time series (%)	Weekday (%)	Weekend (%)	Time series (%)	Weekday (%)	Weekend (%)
NO2
AAMS	−14.5 (12.1)	−12.9 (14.3)	−18.4 (8.6)	−10.2 (7.3)	−6.8 (7.8)	−17.2 (8.3)
RsAMS	−19.1 (13.5)	−18.0 (14.2)	−21.9 (13.9)	−18.1 (11.2)	−13.6 (12.3)	−27.4 (10.0)
O3
AAMS	−2.3 (2.7)	−2.7 (3.2)	−1.2 (2.7)	2.2 (15.6)	3.2 (15.3)	0.0 (18.8)
RsAMS	−0.6 (2.7)	−1.4 (2.7)	1.4 (3.7)	8.9 (10.7)	8.9 (12.3)	8.6 (12.7)

Table 3. OBS-BAU estimates for NO₂ and CO and CH₄ during the lockdown (April 7 – May 25) and the post-lockdown (June 1 – December 31). For CH₄ analysis we only consider time series estimates which include all days of the week. The values are represented as mean (standard deviation).

Pollutant	Station type	Lockdown (April 7 – May 25)	Post-lockdown (June 1 – December 31)
		Time series (%)	Time series (%)
NO2	AAMS	−14.5 (12.1)	−12.8 (4.3)
	RsAMS	−19.1 (13.5)	−18.3 (6.4)
CO	AAMS	−10.9 (11.0)	−5.7 (9.4)
	RsAMS	−8.8 (24.6)	−5.5 (25.2)
CH4	AAMS	0.6 (0.3)	1.3 (0.2)
	RsAMS	0.8 (0.6)	1.1 (0.9)

4.2. O₃ level changes

In this experiment, we investigated various parameters to gain a better understanding of the changes in O₃ in response to the reduction of NO₂ caused by COVID-19 social distancing policies. Alongside the OBS-BAU estimates, we examined standardized anomalies of T2M and SR between 2020 and the 2016–2019 period, S5P FNR in 2020, and changes in S5P HCHO between 2020 and 2019. These parameters were analyzed for two distinct periods: the lockdown period and the post-lockdown (August 1–31, 2020).

4.2.1. Changes during the lockdown period

During the lockdown period (April 7 to May 25), we observed a slight change in O₃ levels across most MAs (Figure 5 second row, and Figure 6). On average, there was a reduction of 2.3% in AAMS and 0.6% in RsAMS, as indicated in Table 2. Although the overall trend showed a decrease, we did find instances of increased O₃ levels in certain MAs, particularly in RsAMS such as Kanto (1.6%), Kinki (2.2%), and Fukuoka (3.5%), as depicted in Figure 5 (second row). Moreover, we observed the existence of an ‘ozone weekend effect’ in the changes of O₃ levels, indicating a higher increase in O₃ mixing ratios during weekends in comparison to weekdays (Akimoto and Tanimoto 2022). This effect was observed in the OBS-BAU estimates for RsAMS in Fukuoka (increased 8.8% on weekends, 1.3% on weekdays) and Kinki (increased 4.9% on weekends, 1.2% on weekdays).

Figure 5. The first and third columns show the plots for the lockdown (April 7 to May 25). The second and last columns show the plots for August 1 − 31. First row: OBS-BAU estimates of NO₂ for AAMS and RsAMS. Second row: OBS-BAU estimates of O₃ for AAMS and RsAMS. Third row: Standardized anomalies of downward solar radiation (SR) and temperature (T2M) from ERA5 dataset. Last row: Formandehyle-to-NO₂ (FNR) ratio in 2020 and HCHO change between 2020 and 2019 from Sentinel 5P data.

Figure 6. Seven-day rolling mean of 2020 observation (OBS), BAU prediction (BAU), and mean level of O₃ from 2016 to 2019 for 4 MAs (Kanto, Kinki, Chukyo, and Fukuoka).

The observed slight decrease in O₃ levels across most MAs in Japan contrasts with trends observed in many other major cities worldwide (Shi et al. 2021; Grange et al. 2021), where significant increases in O₃ levels have been observed. For instance, after accounting for weather effects, notable increases have been reported in Beijing (28.9%), Wuhan (44.5%), Milan (66.8%) Rome (55.8%), New York (17.4%), Los Angeles (14.8%), and Delhi (26.2%) (Shi et al. 2021).

To explore these variations further, we analyzed the disparity in T2M and SR between the corresponding period of 2020 and the reference period 2016–2019 (Figure 5, third row). We observed small positive SR anomalies in the southeast region of Japan and negative SR anomalies in the northeast region. Additionally, across the entire country, negative T2M anomalies were observed. The presence of negative T2M anomalies and fluctuating SR levels suggests that the prevailing weather conditions during this period impeded the production of O_3.

4.2.2. Changes during August, 2020

In August 2020, the NO₂ levels continued to decline in all MAs, albeit at a slower rate compared to the lockdown period, as shown in Table 2. However, during this period, we observed a more noticeable increase in O₃ levels across most MAs compared to the lockdown. On average, there was an 8.9% increase for RsAMS and a 2.2% increase for AAMS. Notably, the increase in O₃ levels during weekends was more significant than on weekdays in Niigata, Okayama, Kinki and Sendai. Specifically, for AAMS of Niigata, O₃ levels experienced a 9.4% increase on weekends and a 5.8% increase on weekdays. In RsAMS of Okayama, O₃ levels saw a 13% increase on weekends, exceeding the 10.6% increase observed on weekdays. Similarly, in AAMS in the Kinki region, O₃ levels exhibited a weekend increase of 19.8%, surpassing the 17.4% increase observed on weekdays. In Sendai, the increase during weekends was even more pronounced, with a 15.6% increase for AAMS and a 22% increase for RsAMS, whereas on weekdays the increase was 5.1% for AAMS and 9.8% for RsAMS. This observation could be attributed to the greater reduction in movement during weekends compared to weekdays in these MAs (Figure 1b).

In order to investigate the differences in O₃ levels between August and the lockdown period, we examined the standard anomalies of SR and T2M in August 2020, comparing them to the 2016–2019 period. Our analysis revealed positive anomalies in both SR and T2M across all MAs (Figure 5, third row). These favorable weather conditions, combined with the reduced levels of NO₂, likely facilitated increased O₃ production.

Although there was an overall trend of increasing O₃ levels during this period, we did observe a reduction in O₃ levels in five MAs in the southern region of Japan: Hiroshima (AAMS 13.7%), Matsuyama (AAMS 1%, RsAMS 3%), Fukuoka (AAMS 12.5%, RsAMS 12.3%), Kumamoto (AAMS 20.7%), and Kagoshima (AAMS 29.9%). To understand the decrease in O₃ levels observed in these five MAs, we utilized the S5P FNR for 2020, as well as the changes in HCHO as a proxy for NMVOCs between 2020 and 2019. The FNR is commonly used to assess the sensitivity of near-surface O₃ levels (Martin, Fiore, and Van Donkelaar 2004). As suggested by Duncan et al. (2010), when the FNR is below 1, the O₃ production regime is considered VOC-limited, and when it exceeds 2, it is considered NO_x-limited. When the FNR values fall within the range of 1 − 2, O₃ is expected to be in the transition regime (Duncan et al. 2010). However, it has been observed that the FNR can vary by region (Jin et al. 2020; Irie et al. 2021; Souri et al. 2023; Ren, Guo, and Xie 2022), and the assumption that it lies within the 1 − 2 range may not hold true at the global level (Schroeder et al. 2017). Hence, it would seem reasonable to calculate this ratio on a regional scale (Damiani et al. 2022; Schroeder et al. 2017). Despite the FNR showing high variability in the region, it still provides information about the trend of O₃ production regimes in our study.

Figure 5 (last row) presents the FNR across all MAs indicating a shift in the O₃ production regime from VOC-limited during the initial lockdown to NO_x-limited in August. This transition is evident as the FNR changes from 0 < FNR < 2 during the lockdown to FNR > 4 in August. During the VOC-limited regime, a decrease in NO_x typically leads to an increase in O₃ levels (Duncan et al. 2010). However, in the NO_x-limited regime, a reduction in NO_x can also result in a decrease in O₃ levels (Duncan et al. 2010). In Figure 5 (last row), we can observe that the NO_x-limited regime dominates the five MAs of Hiroshima, Matsuyama, Fukuoka, Kumamoto, and Kagoshima. Despite NO₂ levels continuing to decline during this period, the HCHO levels exhibited a more significant increase in these MAs compared to the lockdown period. This could explain the reduction in O₃ levels observed in these five southern MAs.

We elucidated the difference in O₃ levels between major MAs in Japan and other large urban areas worldwide by examining meteorological changes (T2M, SR), and variations in O₃ precusor levels by utilizing S5P FNR derived from S5P NO₂ and HCHO measurements. The difference can be attributed to the absence of sunny conditions during the lockdown period. However, in August, when sunny conditions became more prevalent, we observed an increase in O₃ levels in response to the sustained reduction in NO₂ levels across most MAs, which are likely VOC-limited areas. Based on the analysis of S5P data, it appears that the southern MAs exhibited a predominant NO_x-limited trend during August 2020, potentially due to the increased presence of biogenic VOCs (BVOCs). However, the monitoring of BVOC emissions remains challenging due to limited observations (Tani and Mochizuki 2021; Ito and Ichii 2021). Therefore, it is also important to pay attention to those NO_x-limited areas, as future reductions in anthropogenic NMVOCs may have minimal effectiveness in reducing O₃ levels (Akimoto and Tanimoto 2022).

4.3. CH₄ level changes

In this experiment, we analyze the OBS-BAU estimates for NO₂, CO, and CH₄, and incorporate the VISIT model’s simulated CH₄ emissions from wetlands to investigate the changes in CH₄ levels during the 2020 lockdown and post-lockdown period. Our focus is on understanding the relationship between the reduction in NO₂ and its potential impact on OH (hydroxyl radicals), as well as the contrasting effect of CO. The decrease in NO₂ levels is expected to result in a reduction in OH, while reductions in CO can increase OH levels and shorten the lifetime of CH₄ (Akimoto and Tanimoto 2022).

During the lockdown period, we observed a marginal rise in CH₄ levels across most MAs (Figure 7, third row, and Figure 8b), with an average increase of 0.6% for AAMS and 0.8% for RsAMS (Table 3). While NO₂ levels decreased in most MAs (Figure 7, first row), the trend for CO varied (Figure 7, second row, and Figure 8a). AAMS showed an average decrease of 10.9% in CO levels, while RsAMS saw a slightly smaller reduction of 8.8%. Notably, CO levels significantly increased in RsAMS of Kagoshima (60.6%), while slight increases were observed in Kanto AAMS, and in Matsuyama for both RsAMS and AAMS. It is worth noting that although the increases in CO levels in Kagoshima were significant, this region had among the lowest natural CH₄ emissions in Japan (Figure 7, last row), which explains the slight increase in CH₄ observed in this MA.

Figure 7. First and third columns show the plots for the lockdown (April 7 to May 25). Second and last columns show the plots for the post-lockdown (June to December). First row: OBS-BAU estimates of NO₂ for AAMS and RsAMS. Second row: OBS-BAU estimates of CO for AAMS and RsAMS. Third row: OBS-BAU estimate of CH₄ for AAMS and RsAMS. Last row: CH₄ emissions from wetlands based on simulation of the VISIT model with Walter and Heimann scheme and Cao scheme.

Figure 8. Seven-day rolling mean of 2020 observation (OBS), BAU prediction (BAU), and mean level of CO (a) and CH₄ (b) from 2016 to 2019 for 4 MAs (Kanto, Kinki, Chukyo, and Fukuoka).

During the post-lockdown period from June to December 2020, NO₂ levels continued to decrease, showing an average reduction of 12.8% for AAMS and 18.3% for RsAMS (Table 3) which is smaller than during the lockdown period. In contrast, CO levels started to recover as the COVID-19 lockdown was lifted, with a smaller reduction of 5.7% for AAMS and 5.5% for RsAMS. Notably, significant increases in CO levels were still evident at RsAMS in Kagoshima (62.2%). In Fukuoka we also observed a steady rise of CO levels in both RsAMS (13%) and AAMS (11.5%). In response to these changes in NO₂ and CO, we observed a greater increase in CH₄ levels during this period, with a rise of 1.3% for AAMS and 1.1% for RsAMS.

In general, we saw a slight increase in CH₄ levels both during the lockdown and the post-lockdown periods, based on the OBS-BAU estimates. However, a more pronounced increase in CH₄ was observed during the post-lockdown phase in AAMS when compared to RsAMS, which can be attributed to the more substantial recovery of CO levels in AAMS relative to the lockdown period. Although it has been reported that global CH₄ growth in 2020 is primarily attributed to the atmospheric sink resulting from lower anthropogenic NO_x emissions (Stevenson et al. 2022; Peng et al. 2022), our findings regarding the contribution of NO_x reduction to the CH₄ growth in Japan in 2020 align with a previous study (Akimoto and Tanimoto 2022; Qu et al. 2022; Feng et al. 2023), indicating that the impacts of changes in NO_x and CO levels on the increase in CH₄ growth in Japan during the lockdown and post-lockdown period are not as significant as the impacts of the direct CH₄ emissions themselves.

5. Discussion

5.1. Variations in spatial resolution of multisource data

Since we utilized multisource data for the analysis, we acknowledge that variations in spatial resolution among input data can influence the consistency and reliability of data analysis. In certain situations, the need for interpolation to achieve a uniform grid may arise, particularly when generating inputs for a Convolutional Neural Network (CNN). This interpolation process inadvertently introduces uncertainty into the results. However, in this study, we refrained from any data interpolation and used data at the original resolution provided. The multisource data was employed for two primary objectives: weather-normalization model development and visual examination.

For weather-normalization model development, we used ERA5 data and ground station data to construct the model. Certain variables, such as total cloud cover and boundary layer height, are exclusively available from ERA5. The ERA5 data we employed has a resolution of 0.25° × 0.25°, meaning that some stations might share identical ERA5 records. This can influence the model development, even though, ideally, local ERA5 values for each station should be distinct, albeit not significantly deviating from the 0.25° × 0.25° spatial resolution value. To mitigate this effect on the model development, we have integrated spatial context values (latitude and longitude) and station types as additional inputs. Since these features are distinct for each station, we anticipate that they can help minimize the impact of the coarse spatial resolution from ERA5 on the model.

To visually inspect the sensitivity of tropospheric O₃ production utilizing S5P HCHO and NO₂, as well as CH₄ emission estimates from wetlands, we rely exclusively on original data with consistent spatial resolution. It is important to note that our primary focus is to visually inspect the prevailing trends at the MA level, which has a spatial resolution coarser than that of any input data we utilized. Therefore, we believe that the dominant trends at the MA level remain unaffected by these spatial disparities in this particular MA-level context.

5.2. Limitations

In this research, we utilized the S5P FNR to examine the sensitivity of O₃ production. Although HCHO could be an alternative indicator for the presence of NMVOCs, the significant uncertainty in the FNR threshold from previous studies, along with the lack of NMVOC observations and reliable satellite HCHO and NO₂ data, poses challenges in understanding O₃ level variations during and after the lockdown period. This issue is particularly crucial and warrants in-depth exploration in future studies.

Additionally, it is important to mention that the study did not include an analysis of long-range air pollution transport from China to western MAs of Japan following the Chinese economic recovery from the pandemic (Itahashi et al. 2022). This aspect was beyond the scope of the current research but should be considered in future investigations.

6. Conclusion

This study presents an air quality analysis that examines the changes in four air pollutants, namely NO₂, O₃, CO, and CH₄, during the COVID-19 pandemic in 14 MAs of Japan from April 7 to December 31 in 2020. First, we developed a machine learning BAU model that incorporates meteorological, spatial, and temporal features to account for weather variability in air quality time series. Next, we utilized the BAU model predictions and observation data to estimate the actual reduction (OBS-BAU estimate) in NO₂ levels. We then integrated temperature and solar radiation anomalies from ERA5 reanalysis data and S5P TROPOMI data (FNR and HCHO) along with the OBS-BAU estimate to investigate the unique response of O₃ to reduced NO₂ levels during the lockdown and post-lockdown period (August 1–31, 2020). Finally, we evaluated the impact of NO₂ and CO changes on the CH₄ levels using a combination of OBS-BAU estimates and wetland CH₄ emission simulations from the VISIT model. The main findings of the study can be summarized as follows.

Based on ground observations of NO₂, the reduction of NO₂ levels during the lockdown period in 2020 corresponds to a decrease equivalent to 3.4 and 5 years of the 2010–2019 trend of NO₂ levels for roadside and ambient air monitoring stations, respectively. After normalizing the meteorological effects based on BAU predictions, the NO₂ reduction was 14.5% for AAMS and 19.1% for RsAMS. The decrease in NO₂ levels was more pronounced during the weekend than on weekdays.

By analyzing ground observations of NO₂ and O₃, along with BAU simulations and meteorological data from ERA5, as well as FNR and HCHO data from S5P TROPOMI, we found that the reduction in NO₂ levels during the lockdown did not immediately result in an increase in O₃. Instead, we observed that an increase in O₃ occurred after the lockdown, specifically in August, when sunny conditions were reinforced. This finding is significant for Japan, as it has not been previously reported in other studies.

Furthermore, when analyzing the ground observations of NO₂, CO, and CH₄ alongside BAU simulations and model-simulated CH₄ emissions from wetlands, we found that the changes in NO₂ and CO contributed marginally to the variations in CH₄ levels, ranging from 0.6% to 1.3%, across the study areas. This finding aligns with previous studies (Akimoto and Tanimoto 2022; Qu et al. 2022; Feng et al. 2023), but also differs from others where a reduction in the atmospheric sink has been reported as a major contributor to increased CH₄ levels (Stevenson et al. 2022; Peng et al. 2022).

Based on the findings of this study, we recommend simultaneous reduction of NO_x and NMVOCs (both anthropogenic and biogenic VOCs), to mitigate their adverse effects on future policies related to air pollution reduction and climate change mitigation at local level.

Acknowledgements

We express our gratitude to National Institute for Environmental Studies for providing ground observation data for NO₂, O₃, CO, and CH₄ levels. We would like to thank the Japan Meteorological Agency for providing meteorological ground observation data. We acknowledge the Copernicus Climate Change Service for archiving ERA5 reanalysis data and making it accessible to the public free of charge. We extend our sincere appreciation to the European Space Agency (ESA) and Google Earth Engine for granting us free access to satellite derived NO₂ and HCHO data from Sentinel 5P. We would like to thank NIES for the distribution of CH₄ emission simulations from the VISIT model.

Lastly, we would also like to thank the two anonymous reviewers and the editor for their valuable feedback to enhance the quality of this manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The availability of the data used in the study is summarized as follows:

• Ground observation for air pollution: https://tenbou.nies.go.jp/download/ (Last access: 17 May 2023)

• Meteorological ground observation data: https://www.data.jma.go.jp/gmd/risk/obsdl/index.php (Last access: 4 May 2023)

• Sentinel 5P NO₂ data: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S5P_OFFL_L3_NO2 (Last access: 24 April 2023)

• Sentinel 5P HCHO data: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S5P_OFFL_L3_HCHO (Last access: 24 April 2023)

• ERA5 data reanalysis: https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels (Last access: 23 March 2023)

• CH₄ emission simulations from VISIT model: 10.17595/20210521.001 (Last access: 7 June 2022)

Additional information

Funding

This work was supported by Ministry of Education, Culture, Sports, Science and Technology - Japan.