Air Quality in Southeast Brazil during COVID-19 Lockdown: A Combined Satellite and Ground-Based Data Analysis

: With the current COVID-19 pandemic being spread all over the world, lockdown measures are being implemented, making air pollution levels go down in several countries. In this context, the air quality changes in the highly populated and trafﬁcked Brazilian states of São Paulo (SP) and Rio de Janeiro (RJ) were addressed using a combination of satellite and ground-based daily data analysis. We explored nitrogen dioxide (NO 2 ) and ﬁne particulate matter (PM 2.5 ) daily levels for the month of May from 2015–2020. Daily measurements of NO 2 column concentrations from the Ozone Monitoring Instrument (OMI) aboard NASA’s Aura satellite were analyzed and decreases of 42% and 49.6% were found for SP and RJ, respectively, during the year 2020 compared to the 2015–2019 average. Besides NO 2 column retrievals, ground-based data measured by the Brazilian States Environmental Institutions were analyzed and correlated with satellite retrievals. Correlation coefﬁcients between year-to-year changes in satellite column and ground-based concentrations were 77% and 53% in SP and RJ, respectively. Ground-based data showed 13.3% and 18.8% decrease in NO 2 levels for SP and RJ, respectively, in 2020 compared to 2019. In SP, no signiﬁcant change in PM 2.5 was observed in 2020 compared to 2019. To further isolate the effect of emissions reduction due to the lockdown, meteorological data and number of wildﬁre hotspots were analyzed. NO 2 concentrations showed negative and positive correlations with wind speed and temperature, respectively. PM 2.5 concentration distributions suggested an inﬂuence by the wildﬁres in the southeast region of the country. Synergistic analyses of satellite retrievals, surface level concentrations, and weather data provide a more complete picture of changes to pollutant levels.


Introduction
The new coronavirus (COVID- 19) disease was declared as a global pandemic by the World Health Organization on 11 March 2020 [1]. Since then, several countries that had not yet adopted social isolation measures have been promulgating lockdown decrees to avoid the spread of the virus. In Brazil, small towns to bigger cities and entire states have implemented lockdown regulations in varying periods and social-economic activities, depending on local virus propagation, infrastructure and regional characteristics [2]. The southeastern states of Sao Paulo (SP) and Rio de Janeiro (RJ) were the first ones to implement lockdown measures in local areas due to their high numbers of COVID-19 cases. Considering that they are highly populated and industrialized regions, and hold the largest economic clusters in the nation, any imposed changes on their regular activities greatly affects the country's GDP (gross domestic product). Therefore, no lockdown rule was introduced for industries [3]. Public and private transportations, on the other hand, were considerably reduced (e.g., a limitation of travel based on the plate number).
A number of previous studies have investigated the impact of lockdown measures on the level of air pollution in several regions of the world, including Spain [4,5], India [6,7], East Asia [8][9][10], the United States [11,12], Morocco [13], Thailand [14], Iran [15], and Italy [16]. A typical observation, common to these studies, is that the COVID-19 lockdown across the countries of the world resulted in reductions of PM 2.5 , PM 10 , NO 2 , CO, and volatile organic compounds (VOCs), a slight decrease or no change in SO 2 , and an increase in O 3 (see, e.g., [17] for a review). In Brazil, Nakada and Urban [18] highlighted the air quality impacts of a partial lockdown in SP during 24 March to 20 April 2020. They found more than 50% lower concentrations of traffic-related pollutants (CO, NO x ) during this timeframe compared to a five-year mean value based on data from four near-road ground-based stations. A series of studies [3,19,20] was conducted in RJ during partial lockdown (2 March to 16 April 2020) that showed significant reductions in CO and NO 2 , a slight reduction of PM 10 , and an increased level of Ozone.
Although there is no shortage of studies relating the COVID-19 lockdown and level of air pollution, those that have based their analysis on collective information from satellite retrievals and ground-based pollutant measurements are rather rare, especially applied to Brazil. For example, the analysis of NO 2 levels in all the above-mentioned studies in Brazil were based on a limited number of ground-based measurements. The short chemical life-time of NO 2 , however, provides a potential for surface emission observations to be well correlated with tropospheric NO 2 vertical column densities (VCDs), which are concentrations measured by satellite and converted from a slant plane to a perpendicular position in relation to the Earth's surface [21]. In fact, previous studies have shown the temporal correlation potential between NO 2 ground and satellite column concentrations [22][23][24][25]. Therefore, NO 2 VCDs may be used as a proxy of surface data especially when groundbased stations are sparse, as in some regions of Brazil. Although used in studies focusing on other regions of the world (see Table 1 for a summary), no previous studies have used NO 2 VCDs to quantify air pollution change due to COVID-19 in Brazil. Table 1. Summary of recent studies using satellite data to investigate air pollution changes during the COVID-19 lockdown.

Study Area Reference Description
Global [26] TROPOMI NO 2 and CO, and MODIS AOD reductions were assessed during Feb/Mar 2020. Findings include a substantial reduction of NO 2 , low reduction in CO, and a low-to-moderate reduction in AOD in major hotspots of COVID-19 outbreaks.
Global [27] TROPOMI and OMI NO 2 are evaluated during Jan-April 2020 compared to the same lockdown timeframe in 2019. The most significant drop was found in Chinese cities, with a cumulative −40% in 2020 compared to 2019. Decreases in western Europe and United States are also significant (−20% to −38%).
Multiple (34 countries) [28] TROPOMI NO 2 and O 3 , as well as MAIAC AOD measurements, were evaluated during Feb/Mar 2020 compared to the same months in 2019. NO 2 decreases of 10.7% were found in remote areas, while the highest reductions (20%) were found in Europe and China. AOD increased slightly (+13.2%) overall, although local declines were evident in some parts of China.
Southeast Asia [29] AOD from the Himawari satellite and OMI NO 2 were observed during the local lockdown period, in addition to ground NO 2 , SO 2 , PM 10 , PM 2.5 , CO measurements. Larger tropospheric NO 2 reductions of 27-34 % were found over urban areas.
Europe [30] Tropospheric NO 2 columns from TROPOMI are compared over the European region between similar periods of 2019 and 2020, according to lockdown timeframes. A decrease of up to 85% in 2020 in some of the big cities was observed. Cross-correlation were performed between the NO 2 column and ground values, and a R 2 ranging between 0.5 and 0.75 was found in different locations. The Industrial Production Index and air traffic volumes are included, confirming the reason behind the study findings.

Study Area Reference Description
Thailand [14] Assessment of lockdown on the air quality of a medium-sized urban area. TROPOMI NO 2 and ground NO 2 , SO 2 , PM 10 , PM 2.5 , CO and O 3 were studied. Tropospheric NO 2 is in agreement with ground observations. California, USA [12] Ground and tropospheric OMI NO 2 is evaluated during the 2020 lockdown weeks compared to previous 2015-2019 historical data. Spatial patterns of OMI NO 2 showed a decreasing trend over powerplant locations and an increasing trend over residential areas near national highways.
In addition, conclusions of several studies investigating the relations between COVID-19 lockdown and regional air quality have been limited due to not investigating meteorological fields and/or natural emissions that might influence pollutant concentration variations [12,17,19,29].
In this paper, we investigated pollutant concentration changes before and during COVID-19 lockdown in two Brazilian states of SP and RJ using a combined analysis of satellite and surface data. In addition, the effect of meteorological variables and natural wildfire emissions on the findings was evaluated.
We focused on nitrogen dioxide (NO 2 ) and fine particulate matter (PM 2.5 ) as the two major pollutants emitted by motor vehicles [31]. It is estimated that road traffic accounts for 82% of the NO x in the metropolitan region of SP [32]. Additionally, de Fatima Andrade et al. [33] found that vehicle emissions contribute to at least 40% of the PM 2.5 mass in six metropolitan cities including Sao Paulo and Rio de Janeiro. Short-term NO 2 exposure can increase the risk of total, cardiovascular, and respiratory death, while there is evidence of a long-term effect of NO 2 on mortality [34,35]. A range of health issues can happen if particulate matters are inhaled [36]. Fine particulate matters (PM 2.5 ) are especially of interest as they may penetrate deep into the respiratory tract. Islam et al. [37] investigated diesel-exhaust particulate transport and deposition in the upper airways and found differences in deposition associated with particle size.
By evaluating NO 2 data from satellite and ground-based monitoring systems, we aim to highlight the importance of a synergistic analysis of both datasets, especially in a location that lacks ground measurements to provide accurate NO 2 trends. Since the COVID-19 pandemic has begun in Brazil, no previous scientific literature has adopted satellite data for air quality evaluations, and in fact quantitative analyses of satellite retrievals (as opposed to the common usage of composite satellite imagery) are rather rare in studies focusing on other regions of the world. Furthermore, meteorological measurements are not always included in air pollution analysis. For instance, in Brazilian studies mentioned above [3,19,20], no actual weather data were provided besides dispersion information on air quality results during the pandemic. By studying the cross-correlation between meteorological and air quality variables, this study aims to provide a broader and more quantitative understanding of the regional air quality effects due to COVID-19 lockdown. Finally, it is important to explore relationships between NO 2 and PM 2.5 concentrations with natural emissions trends.

The Study Area and Time Frame
The areas of interest adopted for the present study were the Brazilian states of SP and RJ, with their hydrographic regions shown in Figure 1. We focused on two metropolitan areas, namely Alto Tiete in SP and Baia de Guanabara in RJ ( Figure 1C,E), where air pollution due to transportation is expected to be high. Monthly-averaged NO 2 and PM 2.5 concentrations during May 2020 were compared to those of May 2015-2029. This timespan was selected after carefully reviewing the timeframe of local lockdown decrees. Earlier studies in southeast Brazil [3,[18][19][20] have focused on early months of partial lockdown (March and April). However, recently, Noda et al. [38] used a social isolation index and noticed two distinct time-periods of lockdown in SP: before and after 1 May 2020. The first period was associated with greater reductions in the emissions, while the second period was characterized by greater fluctuations in pollutant concentrations. Emphasizing that the pollutant levels were not similar throughout the lockdown, Noda et al. [38] called for more studies focusing on the second interval.

Satellite Retrievals of Tropospheric NO 2
The Ozone Monitoring Instrument (OMI)-a nadir-viewing near-UV/Visible spectrometer (264-504 nm spectrum) aboard NASA's Aura satellite-was adopted for NO 2 VCD retrievals in this study. OMI was launched on 15 July 2004 and provides daily measurements of key air quality components including NO 2 over the whole globe. It orbits the Earth in accordance with Aura's polar Sun-synchronous pattern. Its field of view (FOV) is approximately 13 km × 24 km near nadir and its local equator crossing time (LECT) is 13:45 ± 0:15. For our study area in Brazil, the overpassing time was approximated to be 30 min after LECT, or 14:15 ± 0:15, in accordance with the OMNO 2 product description [39]. We used cloud-screened tropospheric column Level-3 daily global gridded (spatial reso- [40]. By choosing tropospheric instead of total column, stratospheric values are removed, which is expected to result in a better correlation between column and surface concentrations. The OMNO 2 d daily values were then averaged for the May months of 2015-2019 as well as for May 2020. Additionally, NO 2 mean column values were calculated for grid cells covering SP and RJ to be used for quantitative comparisons and correlation analyses with ground-based data.

Ground Measurements
The NO 2 surface-level concentrations were obtained from the State's Environmental Agencies Air Quality Information Systems, CETESB (Companhia Ambiental do Estado de São Paulo) in SP and INEA (Instituto Estadual do Ambiente) in RJ [41,42]. The NO 2 measurement technique is consistent between the two datasets both using a chemiluminescence method. The PM 2.5 ground measurements were only available from the CETESB Agency in SP. The temporal resolution of NO 2 and PM 2.5 data was hourly and daily, respectively.
The ground stations of interest were located within the metropolitan areas in each state (see Figure 1C,E). All SP stations are within the Water Resources Management Unit nº 6 (WRMU 6) which encompasses the Alto Tiete hydrographic region, and all RJ stations are within the West Side of the Baia de Guanabara Hydrographic Region. Stations of interest were chosen objectively and according to a defined set of criteria. Six stations in SP and four stations in RJ were selected. All of these stations recorded NO 2 mass concentrations higher than 60 µg⁄m 3 at least once for each year. The stations in SP had PM 2.5 measurements available for at least 15 days of the month. All stations in SP and RJ were situated strategically close to busy highways. Therefore, it can be assumed that NO 2 emissions were due to vehicles. The only exception was the RJ station Campos Eliseos, which is located close to a petrochemical pole that is a potential NO 2 emitter. Table 2 summarizes the selected stations in each state.
Additionally, daily ground weather data (temperature, humidity, wind speed) was collected for the main airports in the cities of SP and RJ.  Figure 2 shows a schematic of the workflow to process and analyze data in this study. To obtain local NO 2 mean columns and averaged ground NO 2 and PM 2.5 concentrations, different approaches were adopted. For NO 2 satellite data, the overall daily values within the geographic coordinates were averaged for the May months of each year (2015-2020). Then, the averaged data frame was retrieved for each state with its limits defined using shapefiles for each local hydrographic region. In addition, the arithmetic mean of groundlevel concentrations were averaged using the hourly maximum value of NO 2 for each day, and PM 2.5 daily values, respectively, for each chosen ground station. Finally, the averaged values for all SP and RJ stations were calculated.

Data Processing and Analysis
The Pearson correlation coefficient (r s ) was considered for this study. This choice was made based on its wide use and efficacy in measuring linear relations between satellite and ground-based data, as well as previous studies investigating air quality during the COVID-19 lockdown [30,[43][44][45]. Accounting for sample covariance and standard deviation, r s values give information about magnitudes and directions of correlation. Its range goes from −1 to 1, resulting from the following equation: where, x i = yearly May OMNO2 mean satellite columns;x = average of yearly May OMNO2 mean satellite columns; y i = yearly May ground NO 2 mean for all local stations from 14:00-15:00; andȳ = Average of yearly May ground NO 2 means from 14:00-15:00. Note that OMI's overpassing time in Brazil is around 14:30 local time.
To add on the weather data analysis and check on relationships amongst weather variables and ground NO 2 and PM 2.5 concentrations, pairwise correlation using the same Pearson coefficient was performed for the SP region.

Data Limitations
Retrieval errors are present when the OMNO 2 d product is conceived [39]. These include a fitting error in the slant column, estimated to be 0.3 − 1 × 10 15 no. molecules/cm 2 for all data product levels. In addition, there are missing values in the daily dataset, but since the data was averaged for the whole month of each year, the gaps do not represent a significant error within the product datasets. Ground data limitations include gaps in daily and hourly concentrations. If gaps were present in 15 days or more in a month, the corresponding station would be discarded from the analysis. In addition, some hours did not have data available, but if there were data available for at least half of a day, the station would be included in the daily averaging. This range for RJ was 5.28 × 10 15 molecules/cm 2 (87.7 mol/km 2 ) during 2015-2019, and 1.75 × 10 15 molecules/cm 2 (29 mol/km 2 ) for 2020. These observations suggest that the partial lockdown had a more pronounced impact on the reduction of NO 2 in RJ when compared to SP. The May-month-average ground-based observations for each year were transcribed into a boxplot for the six SP and four RJ ground stations and are shown in Figures 4 and 5 for NO 2 and PM 2.5 , respectively. The cross in each box represents the average value for the month of May based on daily data from all stations, and the data within the boxes are within the 25% and 75% percentiles.  In SP, there was a slight variation in the NO 2 mean concentration throughout the years, with the highest relative difference observed between 2016 and 2017 (5.76 ppb or 15%) and the second highest between 2019 and 2020 (5.44 ppb or 13.3%). The former disparity was attributed to an unusual favorability in air pollutant dispersion during the winter months (May-September) of 2016 over the state of SP, as described in the CETEB's 2016 Annual Air Quality Report [46]. The latter difference might be attributed to reduction in traffic due to the local COVID-related lockdown decree. A decrease of 18.8% (6.01 ppb) was observed in NO 2 mean concentrations between years 2019 and 2020 in RJ as shown in Figure 4B. A student's t-Test was performed between the data in years 2020 and 2019 and a difference within monthly means was found to be statistically significant for both states assuming a 90% confidence interval with p-values of 0.099 and 0.077 for SP and RJ, respectively. Data normality was tested for each yearly dataset before the statistical test was conducted.

Results and Discussion
The ground-level PM 2.5 mass concentrations were investigated for SP only because, for RJ, there were no consistent data for all stations throughout the years passing the criteria described in Section 2. As shown in Figure 5, there was not a remarkable drop in PM 2.5 concentrations between 2019 and 2020 (only 0.2 µg⁄m 3 or 0.01%). Similar to the NO 2 trend, however, the PM 2.5 mean concentrations showed a clear decrease in 2016-in fact, 2016 represents the lowest May-month-average across the six years.
To further investigate multi-year trends in satellite data, the mean values of NO 2 VCDs were retrieved for the grid cells covering the selected ground stations, and the results are illustrated in Figure 6. By averaging the 2015-2019 NO 2 values above and comparing the result to the 2020 mean, we found a relative decrease of ∼42% (or 151 mol/km 2 ). A Pearson correlation between year-to-year changes in satellite column NO 2 and surface data from six SP stations collected between 14:00 and 15:00 (associated with OMI's local overpassing time) was calculated and a value of 0.77 was obtained. For RJ, the 2020 mean value was 2.21 × 10 15 no. molecules/cm 2 (∼36.71 mol/km 2 ), and constitutes a decrease of 48.1% from the 2015-2019 mean. A Pearson correlation coefficient of 0.53 was found between year-to-year trends of RJ stations and the corresponding satellite column values. Wind speed and temperature data measured at the Congonhas and Galeao airports for SP and RJ, respectively, were also studied. To assess the role of atmospheric dispersion in pollutant concentrations, we adopted the criteria suggested by the CETESB in objectively defining favorable/unfavorable conditions for dispersion. Unfavorable conditions for air pollutant dispersion were considered [41] as (1) no rainfall, and (2) wind speed less than 1 m/s. All daily rainfall data for SP and RJ averaged for the May months of 2015-2020 showed less than one inch of rainfall, so the effect of rain was considered to be negligible. Daily wind speeds averaged during the month of May were higher than 1 m/s for all years in both SP and RJ with no significant difference seen between years (see Figure 7). Therefore, no considerable unfavorable conditions for pollutant dispersion were detected for May 2015-2020 with a potential to impact concentration levels observed in this study. Temperature variations, however, showed a noticeable decrease in May 2020 in both SP and RJ compared to May 2015-2019. For both states, this decrease was ∼3 • C when compared to 2019, and it was 1.3 • C and 1.4 • C in SP and RJ, respectively, when compared to the average of 2015-2019. It should be noted that the standard deviation of temperature variations over 2015-2020 was 1.16 • C and 1.07 • C for SP and RJ, respectively. Further analysis to understand implications of this lower temperature in 2020 on pollutant levels should be conducted in future studies.
We further analyzed cross-correlations amongst various variables in SP throughout May 2015-2020, with results summarized in Table 3. We observed a high negative correlation (−0.76) between wind speed and NO 2 . On the contrary, the temperature showed a positive correlation with NO 2 and a negative relationship with wind speed. This result reflects the influence of wind speed and temperature on the dispersion, and the retention of downwind NO 2 concentrations. Several studies have shown this relationship by using a noise barrier as an inducer and high-resolution modeling [47][48][49]. Atmospheric stability is also mentioned on these studies, as it is directly related to environmental air temperature, humidity and pressure, and can have an influence on atmospheric pollutant concentrations. As shown in Table 3, PM 2.5 showed a small positive correlation with NO 2 which could be associated with secondary PM, originated in NO x chemical reactions. There is also a relatively high negative correlation of PM 2.5 with ambient humidity, and small positive correlations with temperature and wind speed. The latter could be increasing PM 2.5 by long-range transport from fire hotspots. This is evident by the correlation between PM 2.5 and number of fire hotspots (0.59), in addition to PM 2.5 and wind speed.  In addition to meteorological parameters, the number of fire hotspots reported by the National Institute for Space Research (Instituto Nacional de Pesquisas Espaciais) [50] were also investigated for southeast Brazil, including both states of SP and RJ. The emissions from wildfires are known to contribute to particulate matter and NO x levels [51], hence it is important to assess whether there is a potential interference of fire burnings in our yearly data of PM 2.5 and NO 2 .
As shown in Figure 8, a clear increase (46%) in number of fires during May 2020 was observed in comparison to the same month in 2019. Compared to the average counts for 2015-2019, this increase in 2020 was 35%. These results indicate that ground PM 2.5 and NO 2 levels in May 2020 would potentially be higher than in previous years if only wildfires were considered to impact pollutants' concentrations. This seems to be more pronounced for PM 2.5 , as we did not see any significant decrease in 2020 values during the COVID-19 lockdown period. Furthermore, the distribution of PM 2.5 values shown in Figure 5 seems to follow the same yearly pattern of the fire counts in the southeast region, suggesting that the long-range transport of pollutants emitted from fires within southeastern states could have impacted PM 2.5 values in SP. The relatively high positive correlation between PM 2.5 and number of hotspots (Table 3), as well as the dominant northwesterly wind direction (i.e., from the fire hotspots towards the Sao Paulo metropolitan area) measured at the Congonhas Airport station during May 2020 further supports this argument.

Conclusions
Qualitative and quantitative differences in NO 2 and PM 2.5 concentrations were investigated for timeframes before and during the COVID-19 lockdown determined by the Brazilian government. To analyze if there were any significant differences in NO 2 and PM 2.5 concentrations in 2020 during lockdown compared to previous years, satellite and ground-based measurements were synergistically analyzed. After finding a clear visual difference in gridded NO 2 column densities for the month of May in 2015-2019 and 2020 within the hydrographic regions of interest (where the selected ground stations were located), a quantitative analysis was executed. From OMI satellite retrievals, we found a decrease of 42% in 2020 compared to the 2015-2019 period in SP's Alto Tiete. An even higher decline of 49.6% was found in RJ's Baia de Guanabara for the same time interval.
Aside from satellite NO 2 columns, ground concentrations also presented differences within years. However, the difference seemed to be smaller than that for satellite retrievals. In May 2020, NO 2 levels were 13.3% lower (p-value = 0.099 with a 90% confidence interval) than for the same month in 2019 for SP, while in RJ this drop was equal to 18.8% (p-value = 0.077 with a 90% confidence interval) in 2020. PM 2.5 mass concentrations, however, did not show a significant contrast between 2020 and the previous years. The correlation coefficients between year-to-year trends of satellite column NO 2 and ground-based measurements for multiple selected stations were ∼0.77 for SP and ∼0.53 for RJ.
To examine if there was any weather interference on the statistical comparison performed on this study, the dispersion conditions and mean air temperature for selected stations were studied for SP and RJ. Both SP and RJ stations showed mostly favorable dispersion days for all days during the month of May in all years, so this factor was not considered to be affecting the NO 2 decrease found for 2020. In addition, a high negative correlation between wind speed and NO 2 (−0.76) further supports this statement. Nevertheless, the average temperature during May 2020 was lower than that for the same period during 2015-2019, while we observed a positive correlation between temperature and NO 2 /PM 2.5 . Together, these suggest that temperature might have a role in reducing concentrations observed during May 2020 in southeast Brazil, warranting additional studies. Finally, natural emissions by wildfire were examined using the number of hotspots within the southeast region, and it was concluded that PM 2.5 concentrations were potentially affected by those fires.
The results of this work can serve as a reference for local air quality compliance, in particular at locations that lack ground monitoring stations. Moreover, the combined results with meteorology observations can give a more precise overlook of NO 2 and PM 2.5 scenarios during the COVID-19 lockdown, also serving as a base of analysis for future extreme air pollution events. Suggestions for future work include extending these analyses to other pollutants, as well as other Brazilian regions. This is important because groundbased stations are temporally and spatially sparse in Brazil. Data Availability Statement: All the data used in this study can be obtained from references given in the text.

Conflicts of Interest:
The authors declare no conflict of interest.