Evaluating TROPOMI and MODIS performance to capture the dynamic of air pollution in São Paulo state: A case study during the COVID-19 outbreak

Abstract

Atmospheric pollutant data retrieved through satellite sensors are continually used to assess changes in air quality in the lower atmosphere. During the COVID-19 pandemic, several studies started to use satellite measurements to evaluate changes in air quality in many different regions worldwide. However, although satellite data is continuously validated, it is known that its accuracy may vary between monitored areas, requiring regionalized quality assessments. Thus, this study aimed to evaluate whether satellites could measure changes in the air quality of the state of São Paulo, Brazil, during the COVID-19 outbreak; and to verify the relationship between satellite-based data [Tropospheric NO₂ column density and Aerosol Optical Depth (AOD)] and ground-based concentrations [NO₂ and particulate material (PM; coarse: PM₁₀ and fine: PM_2.5)]. For this purpose, tropospheric NO₂ obtained from the TROPOMI sensor and AOD retrieved from MODIS sensor data by using the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm were compared with concentrations obtained from 50 automatic ground monitoring stations. The results showed low correlations between PM and AOD. For PM₁₀, most stations showed correlations lower than 0.2, which were not significant. The results for PM_2.5 were similar, but some stations showed good correlations for specific periods (before or during the COVID-19 outbreak). Satellite-based Tropospheric NO₂ proved to be a good predictor for NO₂ concentrations at ground level. Considering all stations with NO₂ measurements, correlations >0.6 were observed, reaching 0.8 for specific stations and periods. In general, it was observed that regions with a more industrialized profile had the best correlations, in contrast with rural areas. In addition, it was observed about 57% reductions in tropospheric NO₂ throughout the state of São Paulo during the COVID-19 outbreak. Variations in air pollutants were linked to the region economic vocation, since there were reductions in industrialized areas (at least 50% of the industrialized areas showed >20% decrease in NO₂) and increases in areas with farming and livestock characteristics (about 70% of those areas showed increase in NO₂). Our results demonstrate that Tropospheric NO₂ column densities can serve as good predictors of NO₂ concentrations at ground level. For MAIAC-AOD, a weak relationship was observed, requiring the evaluation of other possible predictors to describe the relationship with PM. Thus, it is concluded that regionalized assessment of satellite data accuracy is essential for assertive estimates on a regional/local level. Good quality information retrieved at specific polluted areas does not assure a worldwide use of remote sensor data.

1. Introduction

The global pandemic of coronavirus disease 2019 (COVID-19) has been causing an overload on health care systems worldwide due to the high rate of hospitalized people (Driggin et al., 2020; Ortiz-Prado et al., 2020). Thus, measures to prevent the transmission of the virus have been adopted around the world. Among these measures, mobility restriction was proposed by several governments, including the imposition of a hard lockdown in regions most affected by the COVID-19 outbreak (Engle et al., 2020; Tang, 2020; The Lancet, 2020; Zhang et al., 2021; Zhou et al., 2020). These restrictions suggested by the World Health Organization (WHO, 2020) have proven helpful and were replicated in different parts of the world. In addition, they were effective not only in reducing the transmission of the virus, but also resulted in the reduction in the concentration of various air pollutants associated with vehicular and industrial emissions (Bao and Zhang, 2020; Bauwens et al., 2020; Dantas et al., 2020; Fan et al., 2020b; Krecl et al., 2020; Venter et al., 2020; Wang et al., 2021).

Mobility restrictions were mainly imposed in large urban centers, where transmission rates were higher (Li and Dai, 2020). Studies carried out during lockdown periods have shown significant improvements in air quality in several parts of the globe, especially in Europe, Asia, and North America (e.g., Bao and Zhang, 2020; Berman and Ebisu, 2020; Cicala et al., 2020; Collivignarelli et al., 2020; Dutheil et al., 2020; Gama et al., 2021; Gkatzelis et al., 2021; Lokhandwala and Gautam, 2020; Sharma et al., 2020; Tian et al., 2021). The main conclusions of the studies based on surface measurements are that, in the short term, reductions occurred in the concentrations of particulate matter (PM_2.5 and PM₁₀), carbon monoxide (CO), nitrogen oxides (NO_X), and sulfur dioxide (SO₂). On the other hand, increases in tropospheric ozone concentrations (O₃) were observed, especially in large urban centers where Volatile Organic Compounds (VOC) are commonly limiting in the atmospheric chemical regime of ozone (Collivignarelli et al., 2020; Sharma et al., 2020; Sicard et al., 2020; Siciliano et al., 2020b; Wang et al., 2020).

Globally, most urban areas do not have robust air quality monitoring networks, making it difficult to assess the continuous spatio-temporal profile of air pollutants. However, most of the pandemic-related discussions about the drop in air pollution were developed based on remote sensing information (Bauwens et al., 2020; Devara et al., 2020; Diamond and Wood, 2020; Fan et al., 2020a; Nakada and Urban, 2020). In this case, the main evaluated pollutants have been tropospheric nitrogen dioxide (NO₂) and columnar aerosol/PM loading. NO₂ data have been largely observed through the sensors Tropospheric Monitoring Instrument (TROPOMI) onboard Sentinel-5 Precursor and Ozone Monitoring Instrument (OMI) onboard Aura (e.g., Bauwens et al., 2020; Diamond and Wood, 2020; Fan et al., 2020a; Nakada and Urban, 2020; Tobías et al., 2020; Zambrano-Monserrate et al., 2020). Aerosol/PM information has been estimated through the Aerosol Optical Depth (AOD), retrieved predominantly from MODIS sensor data (Diamond and Wood, 2020; Fan et al., 2020a; Zambrano-Monserrate et al., 2020).

Remote sensing-based measurements are essential in areas not covered by air quality monitoring networks, especially in developing countries of Africa, Asia, and Latin America. For instance, until 2014, air quality measures were carried out in a few major Brazilian cities (Réquia et al., 2015), corresponding to only 0.03 air quality monitoring stations per 1000 km², a number much smaller than in the USA (0.5), Japan (4.9), and Germany (5.18) (Réquia et al., 2015). Satellite remote sensing is a tool with enormous potential to fill this gap. However, it is known that, despite the retrieved information being useful for a given region, it could present several limitations when applied in a different region, restricting its use for decision-making purposes. Retrieval uncertainties are regionally dependent on several environmental factors, such as those related to particle emission from different geo-climatic and/or degrees of urbanization, particle composition and size range, prevailing synoptic conditions, and local meteorological conditions (e.g., humidity, wind, temperature), affecting both satellite and ground-level measurements (Falah et al., 2021). In the case of Brazilian urban areas, for example, the transport of biomass burning emissions from the Amazon and Cerrado produces very high concentrations of pollutants, frequently exceeding air quality standards (Martins et al., 2018; Pereira et al., 2011; Rudke et al., 2021; Targino et al., 2019). Therefore, although many studies have already conducted quality assessments for multiple satellite data in China, the USA, and some European countries (He et al., 2017; Ialongo et al., 2020; Liu et al., 2019; Nichol and Bilal, 2016; Tack et al., 2021; Verhoelst et al., 2021), there is still a need for regionalized evaluations in areas with distinctive local characteristics. In this context, the sharp drops in air pollutant levels observed during the COVID-19 outbreak (Collivignarelli et al., 2020; Gkatzelis et al., 2021; Nakada and Urban, 2020; Rudke et al., 2021, Rudke et al., 2022) can configure an ideal scenario to evaluate the performance of satellite remote sensing in capturing surface air quality variability.

The COVID-19 pandemic is a rare event, which has shown significant variations in the concentrations of air pollutants over several locations worldwide. Therefore, it serves as an open-air laboratory for assessing the applicability of remote sensing products to map the regional dynamic of air pollutants, especially sudden changes in pollution levels. In this direction, this study aims to evaluate the performance of two satellite sensors, Moderate Resolution Imaging Spectroradiometer (MODIS) (Justice et al., 1998) and TROPOspheric Monitoring Instrument (TROPOMI) (Veefkind et al., 2012), designed to monitor the variability of AOD and tropospheric NO₂, respectively. For this purpose, the variability of AOD and tropospheric NO₂ column density data were evaluated against ground-based data during the partial mobility restriction imposed by the COVID-19 outbreak over the state of São Paulo, Brazil. This area was selected, as it is the Brazilian state with the most robust air quality monitoring network.

2. Methodology

2.1. Description of the study area

The study was conducted in the state of São Paulo (Fig. 1 ), the most populous Brazilian state - more than 46 million inhabitants living in 645 cities spread over approximately 250,000 km² (CETESB, 2020; IBGE, 2020). The state is located in an essentially humid climate zone, with tropical climate areas in the northwest and on the coast (Alvares et al., 2013). São Paulo presents different economic vocations and is considered the most economically developed state in Brazil, primarily due to its industrial, farming, and livestock activities (CETESB, 2020). Together with industries and biomass burning, the large vehicle fleet makes the region the primary source of air pollutants in Brazil (de Andrade et al., 2017; CETESB, 2020; Pereira et al., 2017).

Fig. 1.

Location of the study area. The dots over the state of São Paulo (SP) represent the locations of the CETESB monitoring stations used in this study, and the area emphasized with the largest number of stations represents the Metropolitan Area of São Paulo (MASP). The urban areas were based on Visible Infrared Imaging Radiometer Suite (VIIRS) Nighttime Light data from 2020 (https://eogdata.mines.edu/products/vnl/).

Although it is considered the state that produces the most pollutants in the country, São Paulo has experienced a significant reduction in the concentrations of various pollutants over the years (de Andrade et al., 2017; CETESB, 2020; Rudke et al., 2021). It is currently known that a large amount of this reduction is related to Brazilian air pollution control policies, namely motor vehicle air pollution control programs (PROCONVE) and air pollution control programs for motorcycles and similar vehicles (PROMOT), in addition to local mobile and fixed emission control programs (de Andrade et al., 2017; Carvalho et al., 2015; Gómez Peláez et al., 2020). Besides local sources, long-range transport of air pollutants from biomass burning (both forest and agricultural) plays a significant role in the study area, as suggested by recent research (e.g., Martins et al., 2018; Pereira et al., 2011; Squizzato et al., 2021; Targino et al., 2019). Therefore, during the biomass burning season (August to November), the real pollution levels in the region are highly dependent on the national policy to control fires in the Amazon region, Cerrado and, more recently, the Pantanal.

The Metropolitan Area of São Paulo (MASP, Fig. 1) is the largest urban area in Brazil. Approximately half of the state of São Paulo's population lives in this region (21.6 million inhabitants). The main city of MASP is São Paulo (capital of the State), which is among the most populated cities in the world (≈ 12.3 million inhabitants) and is the foremost industrial center in Latin America. Similarly to other major urban centers in the State, this region has the most significant air quality degradation related to mobile sources (CETESB, 2020). Mobile sources in the MASP are responsible for around 97% of the emissions of CO, 84% of hydrocarbons (HC), 64 % of NO_X, 26% of PM, and 17% of sulfur oxides (SO_X) (CETESB 2019). Other critical air pollution hotspots in the State of São Paulo are the industrial centers of Cubatão and Santa Gertrudes (CETESB, 2020; Rudke et al., 2021). Cubatão is known for its essential petrochemical, steel, chemical, and fertilizer sectors, which account for most of the national industry primary products (Vieira-Filho et al., 2015). The industrial sector is the main source of pollutants in Cubatão, a city known worldwide in the past as one of the most polluted places in the world. The rugged topography and unfavorable weather conditions have also hindered the dispersion of pollutants (CETESB, 2020; Vieira-Filho et al., 2015). The industrial hub of Santa Gertrudes is also a significant source of air pollutants for the region, mainly particulate matter, since this is the largest producer of ceramics in the country (CETESB, 2020).

2.2. Data sets

The data used in this study are ground-based concentrations of PM₁₀, PM_2.5, and NO₂, besides satellite-based Tropospheric NO₂ column density and AOD. The ground-based concentrations were obtained from the Environmental Company of São Paulo State (CETESB) air quality monitoring network (https://qualar.cetesb.sp.gov.br). NO₂ satellite data are from the TROPOspheric Monitoring Instrument (TROPOMI), accessed through the Copernicus Open Access Hub. AOD data from the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm were obtained from the National Aeronautics and Space Administration (NASA) webpage (https://ladsweb.modaps.eosdis.nasa.gov). Among the algorithms used to retrieve AOD from MODIS sensor (Deep Blue, Dark Target, and MAIAC), MAIAC was used because it has the best spatial resolution (1 km) and is widely used to estimate PM in epidemiological studies (e.g., Cohen et al., 2017; Just et al., 2015; Kloog et al., 2014; Li et al., 2021; Maheshwarkar and Sunder Raman, 2021).

The dataset adopted for this assessment considers the first semester of 2019 and 2020. To represent the first half of 2020 the period from December 2019 to June 2020 was considered and to represent the first half of 2019 the period from December 2018 to June 2019. This practice was followed due to lower data availability before the mobility restrictions (March 16), since this is the rainy season in the study region, when clouds significantly reduce data retrieval. The characteristics of the data used in this study are briefly described below.

2.2.1. Ground-based data

According to CETESB, in 2019, São Paulo had 62 fixed automatic stations, two automatic mobile stations, and 23 manual monitoring points (CETESB, 2020). Thus, São Paulo has a density of 0.35 stations per 1000 km², while Brazil has 0.03. The State of São Paulo is, therefore, the best choice of location for assessing the quality of remote sensing data in Brazil.

In this study, we considered only 50 automatic air quality monitoring stations providing measurements of at least one pollutant of interest (PM₁₀, PM_2.5, and/or NO₂), and with data available for the analysis period (first semesters of 2019 and 2020). Of these, 39 stations were used to measure PM₁₀, 20 for PM_2.5, and 32 for NO₂ (Fig. 1). The information of the selected stations is available in Table S1 (Supplementary Material). The data provided by the CETESB automatic stations correspond to data constantly taken from the atmosphere at intervals of five seconds, automatically accumulated in the form of an hourly average and made available for access.

2.2.2. Satellite data

2.2.2.1. TROPOMI

TROPOMI was launched in October 2017 and is a passive hyperspectral nadir-viewing imager aboard the ESA's Copernicus Sentinel-5 Precursor (S—5P) satellite. The swath width is approximately 2600 km, resulting in almost daily global coverage (equator crossing time at 13:30 LT) (Veefkind et al., 2012). This Instrument has received particular attention during the pandemic, which ended up impacting the concentration of pollutants (Faridi et al., 2021; Gkatzelis et al., 2021).

For the obtention of the NO₂ column density, TROPOMI measures the Earth radiance at 0.2-0.4 nm resolution in the visible absorption band (van Geffen et al., 2019). The procedure used to retrieve NO₂ vertical column density (VCD) is based on three steps: (i) the amount of NO₂ along the slant path (Slant Column Density - SCD) is derived using the Differential Optical Absorption Spectroscopy (DOAS) technique (Platt and Perner, 1983) (Verhoelst et al., 2021); (ii) the SCDs are incorporated by the TM5-MP Chemical Transport Model (CTM) that allocates the SCDs to a vertical profile of the NO₂ concentration to separate the stratospheric and tropospheric amounts (Verhoelst et al., 2021); and finally (iii) the SCDs are transformed to vertical column densities using Air Mass Factors (AMFs) (van Geffen et al., 2019). Depending on the CTM mode used, forecast mode (1-day forecast meteorological) or delayed processing mode (12-h forecast meteorological data), the near-real-time (NRTI) or offline (OFFL) TROPOMI measurements are generated (Verhoelst et al., 2021).

Although the S—5P satellite was launched in October 2017, NO₂ data are only available from the end of April 2018, after the end of the commissioning phase (the period used to verify the data chain, i.e., the instruments and ground-processing algorithms, before releasing the data to the users) (ESA, 2020). This study used tropospheric NO₂ column density data from Level 2 OFFL (processor version 1.3.2), which had a spatial resolution at the nadir of 7.0 × 3.5 km² until August 6, 2019 and a resolution of 5.5 × 3.5 km² afterward. The data are made available together with detailed quality flags. We followed the guidelines recommended by the producers for tropospheric applications that indicate to use only NO₂ information with a quality assurance value >0.75, which removes very cloudy scenes, snow- or ice-covered scenes, and problematic retrievals (Eskes et al., 2019).

2.2.2.2. MODIS

Onboard the Terra and Aqua satellites, MODIS acquires data at the top of the atmosphere (TOA) since 1999 and 2002, respectively. The MODIS sensor has 36 spectral bands, which vary in wavelength from 0.4 μm to 14.4 μm in both satellites. The temporal resolution of the sensor is 1 to 2 days and moderate spatial resolutions of 250 m, 500 m, and 1000 m. Of the 36 spectral bands available on the sensor, 7 have primary use aimed at aerosols, with wavelengths ranging from 0.47 to 2.13 μm (referring to the visible spectrum region, near-infrared, and short-wave infrared) (Nichol and Bilal, 2016). Due to the almost daily data and the long data series, the sensor has been used in numerous aerosol-related studies, mainly by using AOD data at 550 nm (Hsu et al., 2004, Hsu et al., 2013; Levy et al., 2007, Levy et al., 2013; Lyapustin et al., 2011, Lyapustin et al., 2018; Remer et al., 2005; Tanré et al., 1997).

AOD represents how much the aerosol layer can prevent the sunlight direct transmission through the atmosphere. Therefore, it is not possible to directly determine the concentration of the particles from satellite (or remote sensing), such as the traditional PM_2.5 and PM₁₀ monitored by air quality monitoring stations. However, the relationship between AOD and surface level particulate matter has been evaluated for PM_2.5 and PM₁₀, and the results indicate a good correlation (e.g., Just et al., 2015; Kloog et al., 2011, Kloog et al., 2015; Li et al., 2021; Wang and Christopher, 2003). Therefore, although the relationship between the concentration of particles and AOD may depend on several environmental factors, there is sufficient evidence for the use of AOD as an indicator of the greater or lesser presence of particles near the surface. Thus, in this study, we evaluated AOD estimates at a wavelength of 550 nm, as a predictor for particulate matter.

In this research, the MODIS MCD19A2 product was used for obtaining AOD data. MCD19A2 is a MODIS-AOD data retrieved using the MAIAC algorithm that combines information from the Aqua and Terra satellites (Lyapustin et al., 2011, Lyapustin et al., 2018). The MAIAC algorithm uses a multi-day time series analysis (4 days at the poles and 16 days at the equator) and different spatial scales for processing [e.g., 1 km grid cells (pixels), 25-km blocks, and 150 mesoscale areas km] (Lyapustin et al., 2018). In the analysis, MAIAC uses 12 spectral bands that are nested in a 1-km grid. The MAIAC AOD retrieval has been evaluated over South America, demonstrating robust performance over heterogeneous surfaces (Martins et al., 2017).

2.3. Comparison between satellite- and ground-based air quality measures

Satellite images have extensive spatial coverage at a short time interval, while ground-based measurements provide a high sampling rate at a local point. Thus, a direct comparison using the location of the air quality monitoring station or using only the satellite overpass time could restrict the probability of matchup data due to cloud cover or the time delay between datasets. For comparison, only the nearest pixels from each air quality monitoring station were used, considering extraction radii of 3, 5, and 10 km. Extraction radii smaller than 10 km were chosen not to excessively degrade the satellite data, allowing a more significant number of collocated data that appropriately represent air quality stations. Besides, the closest hourly ground-based measurement from the satellite overpass time was selected, considering a maximum temporal window of ±1 h. Data from December of the previous year of 2019 and 2020 were used to increase the amount of data in the period before the COVID-19 outbreak in Brazil (the first case was confirmed on February 26, 2020, and the beginning of mobility restrictions was on March 16), which is a rainy period in the region, reducing data collection probability by satellite. Therefore, to represent the year 2019, data from December 2018 to June 2019 were used, and the same strategy was used for 2020. Besides, March 16 was used as the framework for mobility restrictions, so the period before this date was called “before” the COVID-19 outbreak (December 1 to March 15), and the following period (March 16 to June 30) was considered as “during” COVID-19. Based on the extracted data, comparisons were made using Spearman's correlation coefficient, since the data do not have the same units, making it impossible to use other statistical approaches.

2.4. Change in air quality assessed by satellite observations

Changes in air quality over the State of São Paulo during the mobility restrictions were evaluated for Tropospheric NO₂ column density data. This analysis considered a short data series, 30 days before and after the beginning of the restrictions imposed by the COVID-19 (day of year 76), seeking to emphasize only the period with the most restrictive control practices (Rudke et al., 2021). Thus, satellite data from 2020 (February 15th to April 14th) was used in comparison with data from 2019 (February 15th to April 15th).

In addition to the comparison made for periods before and after the restrictions, the effect of the measures adopted to contain the COVID-19 outbreak was quantitatively estimated. In this sense, the equation, previously applied by Fan et al. (2020a), was used to estimate the effect of the mobility restrictions. Thus, based on data for pollutants before and after the restrictions equivalent period, but in 2019, together with data for the period before restrictions in 2020, the expected value for the period after the beginning of mobility restrictions in 2020 was estimated (Eq. 1).

$$X_{2020\mathit{after},\mathit{est}}=X_{2020\mathit{before}}\frac{X_{2019\mathit{after}}}{X_{2019\mathit{before}}}$$ (1)

Where X represents the evaluated pollutant. This equation assumes that the effect of emissions and the impact of meteorological factors are similar in the two periods compared.

Finally, based on the observed and expected values for the period after the beginning of the restrictions, the relative difference was calculated.

2.5. Atmospheric conditions

The role of meteorology on air quality variability is well known (Elminir, 2005; Pearce et al., 2011) and must be considered to avoid misinterpretations regarding air quality changes. In some cases, the expected effects of the mobility restrictions on air pollutants may be reduced or intensified due to unconsidered variations in meteorological variables. In their review work, (Gkatzelis et al., 2021) present several methodologies that apply different techniques to estimate the variability induced by atmospheric conditions to better quantify the impact of mobility restrictions. The authors mention that the most recurrent analyses are performed by normalizing pollutants using long-lived species such as CO, normalizing by dilution corrections using meteorological variables, benchmarking periods of similar meteorology, analyzing synoptic meteorological conditions, and performing back trajectory analysis using models as the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT).

Considering the importance of atmospheric circulation for the concentration of pollutants, we analyzed some meteorological fields and parameters associated with them [Convective Available Potential Energy (CAPE) and Precipitable Water (PWater), temperature, wind speed, total precipitation, Planetary Boundary Layer Height (PBLH), and Relative Humidity (RH)]. Weather condition is particularly important to consider possible factors related to the degradation of satellite data (AOD and Tropospheric NO₂ data retrievals), and/or affect the variation in the concentrations of atmospheric pollutants (weather variables have a great impact on the intra- and interannual variability of atmospheric pollutants). The analysis of the synoptic conditions for the State of São Paulo during the study period was based on the ECMWF Reanalysis v5 (ERA5), the fifth generation of the ECMWF atmospheric reanalysis (Hersbach et al., 2020). The global reanalysis is provided by Copernicus Climate Change Service (C3S) and covers the period from January 1950 to the present. The ECMWF model provides hourly estimates of atmospheric variables at a 30 km grid resolution, 137 levels (from surface up to 80 km). The data can be freely accessed on the following website: https://www.ecmwf.int/en/forecasts/dataset/ecmwf-reanalysis-v5.

3. Results and discussion

The State of São Paulo was divided into Hydrographic Units for Water Resources Management (HUWRMs) for better presentation and discussion of results. The convenience of HUWRMs considers the topography and climatic types/subtypes, which directly or indirectly affect the concentration of pollutants, and should be considered in air quality studies. Furthermore, each of the HUWRMs was previously identified according to their economic vocation by CETESB, which contributes to the discussion of the main factors related to local air pollution. The division by HUWRM and its assigned identification number are shown in Fig. 2 , which will be used to discuss the results. More information about the characteristics of HUWRMs can be found in CETESB (2020) and Rudke et al. (2021).

Fig. 2.

Map of the 22 Hydrographic Units for Water Resources Management (HUWRMs) in the State of São Paulo. The dot colors indicate the economic vocation of each HUWRM.

3.1. Mobility restrictions

In mid-February 2020, confirmed cases of COVID-19 began to emerge in Brazil. The first death caused to the detriment of COVID-19 occurred on March 12 in São Paulo State and, due to its vast population and the significant movement of people, São Paulo quickly became one of the main epicenters of COVID-19 in the country. To curb the transmission of the virus, the first mobility restrictions were applied on March 16. From March 16 to April 3, the first partial lockdown took place, where there was the closure of commerce in general, suspension of non-essential services, and the establishment of home office work (Rudke et al., 2021, Rudke et al., 2022). In this period, there was a significant reduction in public transport displacements (80%) in the State and less vehicle circulation (about 60% reduction). Consequently, a decrease in air pollutant concentrations was observed in many cities across São Paulo for the first 30 days of restrictions (Rudke et al., 2021, Rudke et al., 2022). After this period with the more significant restrictions, governments began to “relax” their sanctions, seeking to partially resume economic activities in the region. Thus, in mid-April, commerce and other non-essential activities started to work on shorter hours than those adopted before the pandemic. This fact caused the number of vehicles on the road to increase and, consequently, the concentrations of air pollutants began to rise (Rudke et al., 2022). Fig. S1 shows how mobility restrictions reduced the number of displacements in the State of São Paulo. More information on how the sanctions adopted by the São Paulo government to reduce the transmission of COVID-19 impacted mobility and, consequently, the concentrations of pollutants can be obtained in the studies of (Rudke et al., 2021, Rudke et al., 2022).

3.2. Atmospheric conditions

This section summarizes atmospheric conditions for the first semesters of 2019 and 2020 and their correlations with the studied air pollutants (PM₁₀, PM_2.5, and NO₂); Complete details are given in Section 1 of the Supplementary Material. The results of the t-Student test using CAPE and PWater indicated that, generally, the energy available for deep convection was statistically different between January and June of 2019 in comparison to the same period of 2020. This is valid both for both São Paulo city and state. However, São Paulo city and state show statistically similar conditions regarding moisture for the first six months of the two years. These two variables indicate that different atmospheric conditions may have predominated in the region during the two years analyzed in this study. In addition, a visual estimate of the number of synoptic cyclones indicates 24 synoptic cyclonic vortices and associated cold fronts passing over São Paulo in 2019, against 31 such systems passing in the same region in 2020. This significant difference shows the predominance of different atmospheric conditions for the two years.

Although the first half of 2020 presents a greater activity of atmospheric systems crossing over the study region, such events were concentrated in the first months of the year, i.e., before the restrictions due to the pandemic. This differential temporal distribution between the two years significantly impacted the concentration of pollutants. As we can easily see in Fig. 3 , while in 2019, rainfall events remained reasonably distributed throughout the first half of the year, in 2020, there was a greater concentration in the first weeks of the year, cleaning the atmosphere in a period that precedes the mobility restrictions due to the pandemic. As a result, the atmosphere was relatively clean when the lockdown took effect. In the weeks following the lockdown, there was a significant reduction in convective activity, thus reducing the chance of pollutant dispersion/removal. Consequently, the expected decreases in the concentration of pollutants due to decreased human activities did not manifest as intensely as observed in other urban regions worldwide. There was even an increase in some of the pollutants in some stations (Rudke et al., 2021). This change in the behavior of the atmosphere, which occasionally coincided with the lockdown period, occurred throughout the State of São Paulo, but was much more pronounced for MASP.

Fig. 3.

Observed weekly rainfall and air pollutant levels for (a) MASP and (b) State of São Paulo during the study period. Rainfall data were obtained from the National Institute of Meteorology (INMET) automatic stations (https://portal.inmet.gov.br/). The vertical dashed black line indicates the beginning of mobility restrictions that started in 2020.

The correlations between NO₂, PM₁₀, and PM_2.5 measured at ground level and temperature, wind speed, precipitation, PBLH, and RH show that precipitation and relative humidity significantly impact air pollutant levels, especially for particulate matter. Comparisons between the period studied for 2019 and 2020 do not show significant differences in the correlations of meteorological variables and pollutant concentrations measured at ground level. However, it is possible to observe that, in general, the 2020 correlations were more significant than those of 2019. This fact may be related to reducing very localized sources due to mobility restrictions, especially on high-traffic roads.

Biomass burning is an important source of air pollutant emissions in Brazil. Analyzing the first semester of 2019 and 2020, it is observed that fire outbreaks have the same behavior for both years, but with a more expressive density mapped for 2020. The evaluation of the primary sources of air pollutants demonstrates that the places with an intense density of fires tend to impact the air quality of São Paulo. However, it is observed that the impact of long-range transport of aerosols and gases is lower when the entire period is evaluated than the more localized sources (which include local fire emissions). It should be noted that the period analyzed here (January to June) is not the typical period of biomass burning in Brazil, which extends from August to November.

3.3. Quantitative spatiotemporal evaluation of satellite estimates

Due to uncertainties related to satellite data retrieval, in this section, we will evaluate tropospheric NO₂ and AOD data by comparing them to ground-based observations.

3.3.1. Satellite estimates - Aerosols (AOD)

The correlations between MAIAC-AOD and concentrations of PM₁₀ were calculated considering all available CETESB stations (Table S1) and are presented in Fig. 4 and Tables S2 and S3 of the Supplementary Material. The AOD data were extracted considering the nearest satellite data (pixel centroid) into three extractions radii (10 km, 5 km, and 3 km from the CETESB station). In general, the correlations were largely non-significant (α > 0.05) and, when significant, only weak or moderate correlations were found. The highest correlation values are associated with estimates based on the smallest area around the station (3 km). This relationship with the extraction radius reinforces the fact that, although distance increases the chance of obtaining samples for a given day, it also increases the inclusion of pixels that do not represent the station measurements. It is also observed that the correlations for 2020 are slightly higher than for 2019. In addition, the correlation results are consistent with previous research carried out in the region (e.g., Damascena et al., 2021; Natali, 2008).

Fig. 4.

Scatterplot of MODIS-AOD at 550 nm (Aqua and Terra satellites) and ground-based measurements of PM₁₀ for the periods: (A) before March 16, 2019; (B) after March 16, 2019; (C) before March 16, 2020; and (D) after March 16, 2020 (mobility restrictions). The blue, orange, and green lines indicate the regression line (Theil–Sen estimator) for AOD data extracted at 3, 5, and 10 km. The same colors indicate the correlation coefficient (R), p-value (p), intercept, and slope for each extraction radius. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Concerning PM_2.5, it is observed that the correlations with AOD were higher and more significant (Fig. 5 ), when compared to PM₁₀ data, reaching 0.61 at Santa Gertrudes station (3 km radius) (Tables S4 and S5). Although the correlations are relatively higher, it should be considered that the number of samples is lower; of the 213 daily data for the year 2020 (from 12‐01-2019 to 06-30-2020), less than a quarter was used to calculate the correlations. One factor that should be mentioned is that, in general, the correlations were more significant for 2020 than for 2019. Furthermore, as observed for PM₁₀, the correlations considering data from the MODIS sensor coupled to the Terra satellite are smaller (and less significant) when compared to the sensor coupled to the Aqua satellite. This fact was also observed by Damascena et al., 2019, Damascena et al., 2021 and Natali (2008) and may be related to satellite overpass times. The Terra satellite crosses the study region in mid-morning when the Planetary Boundary Layer (PBL) starts to evolve, and the aerosols are still not well mixed in the atmospheric column. In contrast, the Aqua crosses in the early afternoon, when the aerosols are well integrated into the atmospheric column (Damascena et al., 2021).

Fig. 5.

Same as in Fig. 4, but for PM_2.5.

PBLH is considered a key factor with a direct influence on the relationship between satellite-derived AOD and ground-level PM since AOD is a measure of the total vertical column atmosphere, while PM is a measure performed at the surface. Thus, assuming that aerosols are homogeneously distributed under PBL, some researchers suggested that AOD values normalized by PBLH can be considered extinction at the surface level (Gong et al., 2017). In China, using MAIAC-AOD and PM_2.5, low correlations were observed for several urban agglomerations (R < 0.4) (He et al., 2021). The authors corrected the AOD data based on RH and PBLH, leading the correlations to >0.6 (He et al., 2021). Damascena et al. (2021) corrected the AOD data considering these variables (RH and PBLH) for MASP, but the correlations remained weak (<0.4). To assess the influence of PBL in the region and period studied, the AOD data were corrected considering the PBLH (AOD/PBLH). Table 1 shows the correlations between PM and AOD (observed and corrected). Overall, little change in correlations was observed for PBLH-corrected AOD. Slight improvements in correlations were kept mainly for the period before March 16 (beginning of mobility restrictions) and correlations with PM₁₀. When the correlations are observed for stations and clustering classes (Table S2 to S6), it is verified that the correction by PBLH has different effects throughout the state of São Paulo, with improvements in correlations in some regions and reductions in others. The low improvement or even decrease in correlations when AOD was corrected may demonstrate that the PBL is not well-mixed, as suggested by Damascena et al. (2021) for MASP. Furthermore, it is important to mention that the PBLH data used to correct AOD come from reanalysis (ERA5) and have a resolution of about 30 km, which may impact the accuracy of the data used for the correction.

Table 1.

Correlation between particulate matter (PM₁₀ and PM_2.5) measured at ground level by CETESB stations and AOD data (observed and corrected) retrieved using MAIAC algorithm.

			2019-2020		2019 - Bef.		2019 - Aft.		2020 - Bef.		2020 - Aft.
		ER	AOD	AOD/PBLH	AOD	AOD/PBLH	AOD	AOD/PBLH	AOD	AOD/PBLH	AOD	AOD/PBLH
PM2.5	AQUA	3 km	0.21***	0.19***	0.13	0.07	0.08	0.06	0.24*	0.29*	0.28***	0.27***
		5 km	0.2***	0.17***	0.12	0.08	0.07	0.06	0.35***	0.36***	0.27***	0.24***
		10 km	0.17***	0.15***	0.01	0.01	0.07	0.06	0.24***	0.28***	0.27***	0.23***
	TERRA	3 km	0.02	0.02	-0.07	-0.11*	-0.02	-0.03	-0.04	-0.03	0.13***	0.15***
		5 km	0.01	0.01	-0.04	-0.09*	-0.03	-0.03	-0.04	-0.04	0.12***	0.13***
		10 km	0	0.01	-0.03	-0.07*	-0.01	-0.01	-0.03	-0.01	0.08**	0.1**
PM10	AQUA	3 km	0.08***	0.09***	0.2***	0.22***	0.13***	0.05	0.19**	0.21**	0.14***	0.16***
		5 km	0.05**	0.07***	0.12*	0.17***	0.14***	0.07*	0.18**	0.18**	0.13***	0.14***
		10 km	0.01	0.05**	0.06	0.14***	0.12***	0.06*	0.17***	0.2***	0.11***	0.12***
	TERRA	3 km	0.03*	0.03*	0	-0.01	0.04	0.01	0	0.02	0.09***	0.1***
		5 km	0.03*	0.03**	0.01	-0.01	0.04	0.02	0.03	0.05	0.08***	0.09***
		10 km	0.01	0.02*	0	0	0.04	0.03	0	0.03	0.06**	0.07***

Note: ER – Extraction radius; AOD – correlation between PM and observed AOD; AOD/PBLH - correlation between PM and corrected AOD (AOD/PBLH). Before (Bef.) and after (Aft.) indicate the correlation for the period before and after March 16 - 2019 and 2020.

By disaggregating the concentrations of PM₁₀ and PM_2.5 into class intervals (Figs. S2 and S3), it is observed that there is no great variation in the AOD values, even with high variations in the concentrations of particulate matter (PM₁₀ and PM_2.5). Temporally, there is a low relationship between AOD at 550 nm and PM₁₀/PM_2.5 data. Considering the average of AOD and PM₁₀/PM_2.5 data for 2019 and the difference related to this period, it is observed that there is no great similarity between data behavior (Fig. 6 ). Furthermore, it is difficult to assertively assess whether the AOD data were influenced by the mobility restrictions implemented during the COVID-19 outbreak or whether the behavior is related to some other behavior inherent in the data. Spatially, there was no significant pattern regarding the variation in correlations, both for PM₁₀ and PM_2.5 (Table S6). However, regions with industrial economic vocation were the only ones that presented significant correlations (α > 0.05), specifically HUWRMs 2, 5, and 7.

Fig. 6.

Weekly dynamics of MODIS-AOD at 550 nm (Aqua) and ground-based PM₁₀ for 2019 (A) and 2020 (B); and MODIS-AOD and ground-based PM_2.5 for 2019 (C) and 2020 (D). The vertical dashed black line indicates the 76th day of 2019 and 2020, when the first restrictions started in 2020. The averages represent data obtained through the highest extraction radius (10 km). The relative difference was calculated using variable averages for 2019 as a comparison. The light blue lines indicate weekly concentrations of particulate matter (PM₁₀ and PM_2.5) that have matching satellite data. The dark blue lines indicate the weekly concentrations of particulate matter for all the data available in the satellite overpass, even without data for the satellite. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

In general, the direct relationship between PM and AOD showed low correlations for the state of Sao Paulo (in many cases, not statistically significant), diverging from studies that demonstrated good correlations for these variables (Gong et al., 2017; He et al., 2021; Kong et al., 2016; Zheng et al., 2017). Damascena et al. (2021) indicate that the low sensitivity of MAIAC-AOD for MASP may be related to the background aerosol model used to characterize South America. The background aerosol model used to characterize South America is also used in AOD data retrieval for Northern North America and the Eastern United States (Lyapustin et al., 2018). In these regions, several studies have shown good results in converting MAIAC-AOD to PM_2.5 (e.g., Chudnovsky et al., 2014; Di et al., 2016; Goldberg et al., 2019; Hu et al., 2014; Pu and Yoo, 2021). These differences between the state of São Paulo and the North American region may be related to the different chemical composition of aerosols for the areas, as shown for MASP by Damascena et al. (2021). The model used for these regions is based on the climatic characteristics of the Goddard Space Flight Center (GSFC), an Aerosol Robotic Network (AERONET) site in eastern USA (Damascena et al., 2021; Lyapustin et al., 2018).

Evaluating the aerosol classification based on the AERONET data proposed by Logothetis et al. (2020) (Fig. S4), there are very different characteristics between the AERONET sites in GSFC and São Paulo (HUWRM 6) and Cachoeira Paulista (HUWRM 2). For GSFC, there is a homogeneous characteristic of aerosols, which are largely classified as “fine non-absorbing”. However, for São Paulo (HUWRM 6) and Cachoeira Paulista (HUWRM 2), there is great heterogeneity in the types of aerosols, ranging from “fine highly absorbing”, “fine moderately absorbing”, and “fine slightly absorbing”, which demonstrates a large discrepancy between the types of aerosols defined for GSFC. Furthermore, through aerosol volume size distribution (Fig. S4), it is observed that fine particles are dominant at GSFC, mainly in summer, while in São Paulo (HUWRM 6) and Cachoeira Paulista (HUWRM 2), the fine and coarse fractions are similar in almost all seasons. For the state of São Paulo, the peak volume concentrations are observed in winter and spring, a period with a large presence of fire outbreaks that have a wide range of particle size distribution (Martin et al., 2010; Rudke et al., 2021; Umo et al., 2015).

The fine particulates present in the atmosphere of eastern US have a large contribution of Sulfate and Nitrate [(SO₄ ⁻² + NO₃ ⁻)/PM_2.5 (%): Boston ≈ 40%, New York (NH₄ ⁺, SO₄ ⁻², and NO₃ ⁻) ≈ 50%], non-absorbing fine-mode aerosols, and lower participation of carbonaceous such as black carbon (BC), absorbing fine-mode aerosols [BC/PM_2.5 (%): Boston ≈ 7%, New York ≈ 8-11%] (Ito et al., 2004; Lee et al., 2014; Venkatachari et al., 2006). The fine particulate from the state of São Paulo has variable concentrations, with different sources and chemical compositions but, in general, it also has a large contribution of Sulfate and Nitrate [São Paulo city (HUWRM 6) (SO₄ ⁻² + NO₃ ⁻ + NH₄ ⁺) ≈ 50%; Araraquara (HUWRM 13): water-soluble ions accounted for 12% of PM_2.5; Cubatão (HUWRM 7): SO₄ ⁻² + NO₃ ⁻ accounted for ≈ 60% of the ions evaluated by Valarini, 2011], but with a greater contribution of carbonaceous [BC/PM_2.5 (%): São Paulo city (HUWRM 6) = 38%, Cubatão (HUWRM 7) = 15-20%; Araraquara (HUWRM 13) = organic carbon (OC) accounted for 18% of PM_2.5] (Gonçalves et al., 2017; Miranda et al., 2012; Pereira et al., 2017; Valarini, 2011). Thus, part of the divergences between the types of aerosols found for GSFC and in São Paulo (Fig. S4) may be related to the greater presence of carbonaceous (elemental and organic carbon) in the atmosphere of São Paulo, since BC is the most absorbing component of the atmospheric aerosol and a primary pollutant (Fierce et al., 2016). This fact may explain the low correlation between AOD data and particulate matter for the state of São Paulo. In addition, this fact may also explain the higher correlations for 2020, compared to 2019, since the main source of pollutants in São Paulo is vehicular, which has a large presence of carbonaceous in its combustion (Brito et al., 2013). Thus, with the reduction in mobility due to the COVID-19 outbreak, there was a decrease in the amount of carbonaceous present in the atmosphere.

It is essential to mention that the low correlation between AOD and PM does not necessarily indicate the low accuracy of MAIAC but rather the low ability to represent this relationship directly. Comparisons made with AOD data from MAIAC and AERONET demonstrate good agreement globally (Qin et al., 2021) and over South America (Martins et al., 2017). However, in urban areas (bright surfaces), the MAIAC algorithm presents poorer results than those in vegetated areas (Martins et al., 2017; Qin et al., 2021). For MASP, the comparisons between MAIAC and AERONET showed correlations of 0.74 and 0.64 and root mean square error (RMSE) of 0.087 and 0.079 for Aqua and Terra (Fig. S5), respectively, like those found by Martins et al. (2017) for urban areas. As observed by Martins et al. (2017), the retrievals presented dispersion points scattered in low AOD values and with a slight tendency to underestimate the values (Bias Terra: − 0.024 and Bias Aqua: −0.044). The retrieval fraction within the expected error (EE) envelope of Aqua (72.28%) was better than that of Terra (69.49%), suggesting that Aqua product is more suitable for urban retrievals. According to Martins et al. (2017), the multiple pollutants sources contribute to aerosol microphysics, which represents a difficulty for aerosol models used in satellite retrievals.

Finally, it should be noted that different environmental factors can influence the AOD-PM relationship. Among these factors, it is known that the long-range transport of biomass-burning emissions has a substantial impact. Although the burning season in Brazil does not occur during the observed period, it is possible that isolated events were included in the study, generating different impacts for AOD and PM, depending on the height at which emissions are injected into the atmosphere (see Section 2, Supplementary Material). In addition, as there is variability in the various geographic (e.g., land cover, population density, altitude, traffic density) and meteorological factors (e.g., rainfall, temperature, PBLH, wind speed, RH), a variable spatial relationship between AOD and MP is expected, requiring a multivariate evaluation. As an example of such an assertive, (Li et al., 2018) found low correlations between AOD and PM_2.5 (R: 0.19) and were still able to build a model to estimate PM_2.5 with an R² of 0.87 by using mixed-effect models. In this case, the authors used not only MAIAC-AOD to estimate PM_2.5 but also meteorological parameters, proxy variables for emission sources, Normalized difference vegetation index (NDVI), and Land cover.

3.3.2. Satellite estimates – tropospheric NO₂

The correlations between the tropospheric NO₂ vertical column density (TROPOMI) and NO₂ concentration measured by CETESB stations are shown in Fig. 7 and Table S7. The results suggest that the two distinct ways of monitoring NO₂ show good correlation in most stations. As with AOD, NO₂ data showed, in general, a higher correlation when considering the smallest spatial window (3 km) (Table S7). Although there was good agreement between values calculated for 2019 and 2020, there was an exception to this behavior for some stations caused by low correlation and disagreements from the different years.

Fig. 7.

Scatterplot of satellite-based tropospheric NO₂ vertical column density (VCD) derived from TROPOMI (S—5P) and ground-based measurements of NO₂ for the periods: (A) before March 16, 2019; (B) after March 16, 2019; (C) before March 16, 2020; and (D) after March 16, 2020 (mobility restrictions). The blue, orange, and green lines indicate the regression line (Theil–Sen estimator) for NO₂ data extracted at 3, 5, and 10 km. The same colors indicate the correlation coefficient (R), p-value (p), intercept, and slope for each extraction radius. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 2 presents the averages, standard deviations (SD), and correlation results (3 km spatial window) between the ground based NO₂ and tropospheric NO₂ density. Through Table 2, it is possible to verify that, in general, stations with the highest mean concentrations and the highest standard deviations were the ones with the highest correlations [e.g., S.Bernardo-Centro, Parque D.Pedro II, Ibirapuera, and São Caetano do Sul (HUWRM 6)]. Furthermore, it is observed that, with the decrease in average and standard deviations in 2020, there were also reductions in correlations. The relationship with standard deviations can be verified when observing the 2019 correlations for Marília (HUWRM 21) and Pico do Jaraguá (HUWRM 6) stations, which have an average of around 6 to 7 μg∙m^-3, but different standard deviation values. Marília presented a smaller data variation when compared to Pico do Jaraguá, showing that the concentrations at Marília are more homogeneous, while those of Pico do Jaraguá station are more heterogeneous. Similarly to Marília, the Araraquara station (HUWRM 10) presented more homogeneous concentrations (2019 and 2020), making the variability in tropospheric NO₂ low and generating lower correlations. The Ibirapuera station had more heterogeneous concentrations, which made the correlations higher. This demonstrates that the sensor has difficulty capturing small variations at the ground level. This fact is also demonstrated through the disaggregation of NO₂ concentrations into class intervals (Fig. S6), since there is a clear distinction between values up to 20 μg∙m^-3 and higher, but a great overlap between the other categories. This fact demonstrates that good correlations will be obtained when there is a greater variation range in NO₂ concentrations.

Table 2.

Mean, standard deviation (SD), and correlation between tropospheric NO₂ density (TROPOMI) (Pmolec/cm²) and NO₂ concentration measured by CETESB stations (μg∙m^-3). The non-significant correlations (p-value >0.05) are shown in bold. Data represents only the smallest extraction radius (3 km).

		TROPOMI-2019		Air Quality Station-2019				TROPOMI-2020		Air Quality Station-2020
HUWRM	Station name	mean	SD	mean	SD	R		mean	SD	mean	SD	R
2	Guaratinguetá	1.82	0.80	5.0	3.4	0.19		1.08	0.68	4.3	2.0	0.11
	S.José Campos	3.13	1.73	11.8	6.5	0.5		1.13	0.68	4.8	2.3	0.36
	S.José Campos-Jd.Satelite	3.23	1.78	10.0	6.1	0.46		2.65	1.43	3.5	2.6	0.18
	Taubaté	2.03	1.03	8.1	6.3	0.5		1.23	0.75	4.6	2.9	0.22
5	Campinas-Taquaral	3.15	1.83	6.0	3.5	0.44		6.45	4.40	22.6	14.5	0.58
	Piracicaba	2.24	1.34	6.1	3.3	0.38		6.35	3.91	52.2	21.3	0.23
	Santa Gertrudes	1.85	0.99	13.3	6.4	0.38		6.65	5.13	18.1	11.5	0.81
6	Cerqueira César	7.76	4.23	36.0	16.0	0.52		6.15	3.63	47.1	20.1	0.56
	Congonhas	8.39	4.30	64.6	19.1	0.68		1.39	0.91	3.9	2.8	0.46
	Guarulhos-Pimentas	5.02	6.92	9.3	8.2	0.45		4.01	2.92	9.8	4.9	0.47
	Ibirapuera	9.91	7.82	15.3	10.2	0.81		7.10	3.88	15.0	11.0	0.7
	Itaim Paulista	6.08	3.85	12.1	6.2	0.63		4.56	2.33	10.5	4.9	0.71
	Marg.Tietê-Pte Remédios	6.15	3.88	69.9	24.6	0.4		0.96	0.80	5.6	2.1	-0.08
	Osasco	6.11	4.55	37.3	16.8	0.59		5.62	3.77	56.0	25.9	0.41
	Parque D.Pedro II	8.73	4.55	25.4	10.6	0.86		0.92	0.87	4.2	2.2	0.27
	Pico do Jaraguá	4.87	5.03	6.9	7.0	0.71		5.24	3.93	31.6	16.3	0.81
	Pinheiros	7.29	3.14	33.0	12.4	0.56		6.53	4.35	19.4	10.7	0.71
	S.André-Capuava	12.03	10.55	21.7	9.8	0.63		4.17	2.46	7.5	6.4	0.74
	S.Bernardo-Centro	8.57	7.47	20.4	10.1	0.86		6.18	4.14	20.3	11.1	0.73
	São Caetano do Sul	11.09	10.80	22.3	7.9	0.74		1.85	0.94	6.0	3.4	0.31
7	Cubatão-Centro	11.82	10.03	26.7	18.1	0.73		1.03	0.64	3.4	1.9	0.24
	Cubatão-V.Parisi	11.51	9.25	54.1	23.2	0.38		5.49	2.77	15.0	7.5	0.59
	Santos-Ponta da Praia	6.85	9.35	14.9	12.0	0.66		4.90	2.46	14.9	5.4	0.64
10	Sorocaba	2.90	1.86	9.4	3.6	0.71		2.99	2.66	9.6	5.7	0.41
	Tatuí	1.52	0.90	6.3	3.7	0.28		3.08	1.72	9.5	6.3	0.56
13	Araraquara	1.39	0.71	5.5	2.4	0.3		1.62	1.15	18.6	6.1	0.45
	Bauru	1.33	0.84	5.0	2.6	-0.04		3.97	4.32	10.3	10.3	0.67
	Jaú	1.18	0.81	4.6	2.0	-0.04		6.06	2.84	17.3	6.9	0.56
15	Catanduva	1.38	0.82	4.4	2.4	0.26		1.32	0.85	6.1	3.6	0.28
	São José do Rio Preto	1.43	0.77	5.6	2.9	0.06		2.27	1.68	8.8	4.0	0.54
21	Marília	1.16	0.73	6.3	3.0	-0.04		1.18	0.85	4.9	2.6	0.44
22	Presidente Prudente	1.09	0.73	3.6	1.7	-0.18		1.43	0.80	5.3	3.2	0.38

Temporally, the tropospheric NO₂ vertical column density showed similar behavior to that observed for the NO₂ measured at ground level (Fig. 8 ). It is observed that, for 2020, the behavior of concentrations at ground level is better represented by tropospheric NO₂ vertical column density. In addition, looking at the weekly NO₂ concentrations for all the data available in the satellite overpass, even without data for the satellite (grey line in Fig. 8), it is clear how the lack of data affects the average calculations.

Fig. 8.

Weekly dynamics of tropospheric NO₂ VCD derived from TROPOMI (S—5P) and ground-based measurements of NO₂ for 2019 (A) and 2020 (B). The vertical dashed black line indicates the 76th day of 2019 and 2020, when the first restrictions started in 2020. The averages represent data obtained through the smallest extraction radius (3 km). The relative difference was calculated using variable averages for 2019 as a comparison. The blue lines indicate weekly concentrations of NO₂ that have matching satellite data. The red lines indicate the weekly concentrations of NO₂ for all data available in the satellite overpass, even without data for the satellite. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Spatially, the highest correlations were observed for HUWRM 2 (R: 0.57), 6 (R: 0.51), 7 (R: 0.68), and 10 (R: 0.55), all with characteristics of industrial economic vocation (Table S8). Although it has an industrial economic vocation, HUWRM 5 was the only one in this category to show low correlations (R: 0.14). For this region, low correlations were observed especially for the Campinas-Taquaral and Piracicaba stations, which did not present significant (α > 0.05) correlations. Both stations are in the transitions between urban and rural areas (peri-urban stations, Table S7), which presented the worst correlation because they are far from urban centers and, therefore, receive less localized pollution than stations close to major traffic routes. Only HUWRM 13 and 22 did not present statistically significant correlations (α > 0.05). HUWRM 22 has an economic vocation for Farming and Livestock, with a smaller contribution from the pollutant when compared to other regions (Rudke et al., 2021). The HUWRM 13 has an economic vocation characterized as “in the process of industrialization”, but it presents NO₂ concentrations similar to those observed for regions with Farming and Livestock economic vocation (Table 2), which may be related to the type of industry present in the local, mainly agro-industrial (Gonçalves et al., 2017). Fig. 9 shows the distribution of correlations obtained by aggregating all the data evaluated (2019 and 2020) and may shed some light on the observed behavior. The figure shows that the highest correlations were obtained in stations located in intensely inhabited regions (MASP), which present a greater flow of vehicles. In addition, a good correlation was obtained close to the industrial hub of Cubatão (UGRHI 7). This region showed increases in the concentrations of several pollutants, even during the mobility restrictions imposed by the COVID-19 outbreak (Rudke et al., 2021). In these regions, due to the greater smudge of pollution, NO₂ can be constricted and become trapped within the Planetary Boundary Layer, as mentioned by Cersosimo et al. (2020).

Fig. 9.

Spatial distribution of correlations between tropospheric NO₂ VCD and ground-based measurements obtained for the first semester of 2019 and 2020. Dots with a black border indicate non-significant correlations (α > 0.05).

Overall, the tropospheric NO₂ data obtained through the TROPOMI have shown good correlations with the ground-based NO₂ acquired from air quality monitoring stations (Bassani et al., 2021; Cersosimo et al., 2020; Fioletov et al., 2022; Goldberg et al., 2021; Kim et al., 2021; Stratoulias and Nuthammachot, 2020; Vîrghileanu et al., 2020; Virta et al., 2021; Zheng et al., 2019). However, some studies also found lower correlations between these measurements for rural areas (Bassani et al., 2021; Jeong and Hong, 2021; Virta et al., 2021). These studies indicate that, in more remote regions, due to their very low NO₂ concentrations, the minimum detectable limit of the sensor may be reached (Bassani et al., 2021), generating unrealistic data. Assessing tropospheric NO₂ variations over remote locations is a challenge, since these areas are generally far away from the primary sources of NO₂ and thus present concentrations much lower than those over urban areas. In São Paulo, the stations with the lowest correlation located in regions with a rural vocation also showed lower NO₂ levels (mean: 5.1 μg·m^-3, SD: 2.7 μg·m^-3, and maximum: 16 μg·m^-3), when compared to the urban stations with the highest correlations (mean: 15.5 μg·m^-3, SD: 10.4 μg·m^-3, and maximum: 108 μg·m^-3). From comparisons to ground-based total and tropospheric column measurements, validation studies demonstrate that TROPOMI tends to overestimate data at rural and remote locations due to lower concentrations of NO₂ (Ialongo et al., 2020; Marais et al., 2021; Verhoelst et al., 2021; Zhao et al., 2020). Due to the much lower NO₂ levels over rural areas, the estimates of tropospheric NO₂ may not follow the slight daily variations observed by the ground-based stations. This behavior may be related to a possible overestimation of the stratospheric column fraction. An example of such assertation is the Bauru region (HUWRM 13), which exhibited non-significant correlations in our estimates and showed a negative Bias between stratospheric NO₂ measured by the monitoring station (Zenith-Scattered-Light DOAS, ZSL-DOAS) and by TROPOMI (-1.2695 ± 0.00726 Pmolec/cm², data available at: https://mpc-vdaf-server.tropomi.eu/no2/no2-offl-zsl-doas-sunset/bauru#Baseline) (Verhoelst et al., 2021). In this case, the underestimation of stratospheric NO₂ is possibly followed by an overestimation of the tropospheric column, since separating these columns in areas with low NO₂ levels remains challenging. (van Geffen et al., 2019) indicate that lower NO₂ levels can lead to greater difficulty in separating the stratospheric and tropospheric NO₂ columns, due to uncertainties in the NO₂ profile. In addition to these uncertainties, a more significant impact of the instrument noise is expected to affect the estimates over rural areas, as NO₂ levels are much lower.

Thus, the results found for the state of São Paulo demonstrate that the TROPOMI sensor is promising for mapping the variability of tropospheric NO₂ in densely populated regions, which present a source of expressive emission of the pollutant. However, this accuracy was not maintained for less populated areas, with economic characteristics focused on agriculture and agribusiness. It is also worth noting that the measurements of air quality stations (punctual data) were compared with 5.5 × 3.5 or 7.0 × 3.5 km² areas (size of TROPOMI pixel), which can reduce the correlations in locations that have very localized sources, once this effect will be diluted along the pixel on the satellite.

Finally, the findings of this study demonstrate the potential of the TROPOMI data and indicate the need for further studies regarding the relationship between AOD and PM. In the second case, studies that seek to identify PM_2.5 from the satellite are important as a function of its high potential for damage to human health, especially in regions with a scarce monitoring network, such as Brazil. As NO_X is an important precursor of secondary PM_2.5, high correlations between NO₂ and PM_2.5 are expected in places where pollutant sources are strongly associated (Wu et al., 2016). In the State of São Paulo, low to moderate correlations were observed (R:0.34-0.44), depending on the period analyzed (Fig. S7). This behavior was also maintained when comparing concentrations of PM_2.5 with tropospheric NO₂ densities from the TROPOMI sensor (R:0.28 to 0.38). This demonstrates that TROPOMI data can also capture this relationship between NO₂ and PM_2.5 at ground level. In addition, the correlations between concentrations of PM_2.5 and tropospheric NO₂ densities were greater than those observed for AOD data.

3.4. Spatiotemporal distribution of the air pollution monitored by satellite

To demonstrate how mobility restrictions during the COVID-19 outbreak have affected air quality over the state of São Palo, the spatial pattern of reduction/increase in tropospheric NO₂ was evaluated. In this section, only tropospheric NO₂ was considered, since these data presented better correlations with the data measured at ground level. Fig. 10 shows the average for the 30 days after the onset of mobility restrictions, considering tropospheric NO₂ observed by satellite and estimated by Eq. 1. Satellite data considered for the estimates used averages of 30 days before and after the beginning of the restrictions adopted in 2020 in comparison to the same period of 2019 (this dataset can be seen in the Supplementary Material - Fig. S8). As mentioned in section 2.4, using Eq.1 to estimate tropospheric NO₂ for 2020 assumes that the meteorological scenario has a similar average impact as 2019. On the other hand, slightly different atmospheric conditions were observed, with rainfall having less effect on wet deposition and a higher number of fire outbreaks during the mobility restrictions. With the prevailing atmospheric conditions in 2020, air pollutant concentrations were expected to be higher than those observed for the same period in 2019. Thus, considering this scenario, it is possible to infer that the values estimated by Eq. 1 are possibly underestimated; the density of tropospheric NO₂ expected for 2020 would be undoubtedly higher than those shown in Fig. 10-B. This scenario observed for 2020 also prevented higher air quality improvements during mobility restrictions, as mentioned by (Rudke et al., 2021, Rudke et al., 2022).

Fig. 10.

Change in air quality over the state of São Paulo for 30 days after the beginning of mobility restrictions (March 16, 2020). Observed values for tropospheric NO₂ VCD (A), expected values estimated by Eq.1 (B), relative difference between A and B (C), and population density (D) for the state of São Paulo.

The satellite data presented in Fig. 10 shows that the areas with the highest amount of pollutants are closely related to the places with the highest population density, as shown in (Rudke et al., 2021, Rudke et al., 2022). This is essentially due to the larger vehicle fleet associated with population size and the industrial hub already established in these regions. In addition, it is observed that there is no homogeneous behavior in the state, with increases registered in the interior, especially in places with an economic vocation for Farming and Livestock, and reductions observed mainly in Industrial areas or in the process of industrialization. Assessing the entire state of São Paulo, 57% of the area showed some level of reduction in Tropospheric NO₂. Reductions in industrial areas or in the process of industrialization reflect the decrease in economic activities and the adoption of mobility control practices in the 30 days. The reduction in mobility is the main driver of the decrease in MASP (HUWRM 6). This area has the highest density of vehicles per inhabitant in Brazil and many other places in the world. For MASP, the results presented in Fig. 10 show a substantial reduction in Tropospheric NO₂, in which the average drops reached 51%.

Unlike industrial or in industrialization areas, there is an increase in Tropospheric NO₂ in several cities in the interior of the State, where the main economic vocation is Farming and Livestock. The most substantial increases were observed for HUWRMs 9 (16.6%), 17 (15.3%), and 22 (87.0%). The increases in Tropospheric NO₂ occurred in areas with lower concentrations of pollutants (Fig. 10); thus, the rises may be related to seasonal pollutants variability or lower data quality for these areas, as shown in the previous section.

Studies developed based on data acquired through air quality monitoring stations demonstrate similar behavior for MASP, with greater or lesser reductions according to the evaluated period (Connerton et al., 2020; Debone et al., 2020; Nakada and Urban, 2020; Noda et al., 2021; Rudke et al., 2021; Siciliano et al., 2020a; Tadano et al., 2021). Regarding the interior of the state, Rudke et al. (2021) demonstrated great spatial variability in the concentrations of pollutants when analyzing one semester. However, the study shows that, even in interior cities, most of the monitored stations showed significant reductions in NO₂ concentrations compared to the same period in 2019. For particulate matter, an inverse behavior was observed; most stations located in the interior did not show significant variation and, when significant, they showed an increasing tendency. It is important to point out that air quality monitoring stations in the state are located, in general, in medium or large urban areas. Increases in Tropospheric NO₂ values (Fig. 10), on the other hand, were mainly observed in areas with lower population density.

It is important to consider that, by October 2020, one-third of the research published on the effects of mobility restriction policies (adopted during the COVID-19 outbreak) on air quality used data monitored by satellite (Gkatzelis et al., 2021). Although this type of information is very relevant for spatial assessments, especially in places with little or no ground-level monitoring network, some considerations should be made regarding its monitoring capacity.

As for the monitoring carried out through remote sensing, one of the main problems in data retrieval by passive sensors is related to the presence of clouds (Cersosimo et al., 2020; Sano et al., 2007). Furthermore, regarding the properties of atmospheric aerosols, an additional challenge concerns the distinction between scattered light by aerosol particles and by atmospheric gases and some types of complex surfaces (e.g., snow/ice, desert, etc.) (Dubovik et al., 2021; Mhawish et al., 2019). For NO₂ retrievals, another issue is related to the presence of aerosols in the atmosphere, which significantly impacts scattering properties (Vasilkov et al., 2020). This type of interference, along with errors and problematic retrievals, substantially reduces the availability of continuous and uniform sampling and coverage for a given region.

Fig. S9 shows the availability of Tropospheric NO₂ data for the state of São Paulo, considering data observed for the 30 days before and after the beginning of the restrictions adopted in 2020, compared to the same period for 2019. The percentage of valid days during the 30 days before the mobility restrictions is approximately 35% for São Paulo [average quantity of data for the state was 8.9 (min 0 – max 23) for 2019 and 12.9 (min 0 – max 28) for 2020]. During the restrictions, this value increases to approximately 50% [2019: 12.4 (min 1 - max 24); 2020: 17.5 (min 2 – max 28)]. This occurs due to the greater presence of clouds during the rainy season (October to March), with increased cloudiness in early spring and maximum peak occurring in summer (Moura et al., 2016).

Locally, data availability can be even more affected. For example, for MASP, the availability of valid days was <30% before and approximately 40% during the restrictions (data available for the period before the restrictions: 2019–7.3 and 2020 - 6.4, during the restrictions: 2019–8.3 and 2020–14.7). This fact is due to the high cloudiness present over the region in all year periods compared to other areas of the state (Fig. S10). In the specific case of MASP, few data were used in the comparison presented in Fig. 10; this fact shows that unrealistic results can be obtained when satellite data is used to track short-term changes, once days with low or high concentrations of pollutants can be considered on the statistics and thus generate under- or overestimated averages. (Prunet et al., 2020) also observed low availability of valid days for Paris (approximately 40%) due to the presence of clouds. The authors also indicate that the lack of valid data for some weeks analyzed for Milan may have contributed to artificially reducing the 2020 averages, probably intensifying the reductions observed during the lockdown period. Therefore, considering the uncertainties related to the lack of data associated with constant presence of clouds worldwide (Fig. S11), the findings must also be associated with a more careful analysis. For instance, the review of Gkatzelis et al. (2021) distinguishes studies based on observations alone and studies that involve models that can account for the weather impacts on the data. Besides, one way to improve satellite data availability is associated with using mathematical and/or statistical methods, such as interpolation and regression methods (Cersosimo et al., 2020; Wu et al., 2021; Zheng et al., 2019).

4. Conclusions

This study examined the capability of satellite products, tropospheric NO₂ column density from the TROPOMI sensor and MAIAC-AOD from MODIS sensors, as predictors of NO₂ and particulate matter (PM₁₀ and PM_2.5) concentrations at ground level for the state of São Paulo during the period of vehicle circulation restrictions as a function of the COVID-19 pandemic. This approach was performed by calculating the correlation between ground-based air quality data and satellite data at spatial windows of 3, 5, and 10 km, and at a temporal window of ±1 h. The pandemic outbreak period was chosen due to the recurrent use of these products to describe the trend in pollutant concentrations, especially in large urban areas. The main findings of this study were:

1) For the State of São Paulo, the distribution of changes in pollutants measured by satellite (relative difference between observed and expected data) presented a very heterogeneous behavior. Overall, about 57% of the state of São Paulo showed some reduction in Tropospheric NO₂ density values. The regions with the greatest reductions were those with industrial economic vocation, such as MASP. On the other hand, increases in tropospheric NO₂ were observed in regions with an economic vocation for Farming and Livestock, of up to 87%.

2) The assessment of temporal data availability demonstrated that, due to factors that make data acquisition difficult by passive sensors - especially clouds, there is a large presence of failures in the data series. Considering the satellite data available for thirty days, the analyses show that retrievals are available <20% of the time, depending on the region of the state. This fact can provide misleading assessments of the real variability of pollutants in the lower atmosphere when satellite information is used as ground truth. This misleading can occur when few data from days with high or low concentrations of pollutants are used, generating over- or underestimated values, respectively.

3) Comparisons between AOD data and particulate matter (PM₁₀ and PM_2.5) showed that AOD has a low potential to explain the spatiotemporal behavior of PM in the study area. Correlations between data were low and frequently not significant, demonstrating the weak direct relationship between data. Furthermore, although the performance of the MAIAC AOD retrievals varied in time and space, PM_2.5 data were better represented by the AOD in some specific stations, reaching correlations of up to 0.61 in a few stations. The low correlations between AOD and PM data observed for most stations are corroborated by previous studies, and they may be related to the type of aerosol present in the atmosphere of the state of São Paulo, which is very different from other regions where the correlations between these data proved to be promising.

4) Tropospheric NO₂ data demonstrated the potential to represent changes in NO₂ concentrations at ground level. The correlations between Tropospheric NO₂ column density and NO₂ concentrations at ground level were >0.60. The highest correlations were observed especially in industrialized regions, which are significant sources of NO₂ and NO. On the other hand, low correlations were observed for regions with an economic vocation for Farming and Livestock. These regions had low tropospheric NO₂ densities due to the low number of sources present in the area, making the correlations generally low and not significant.

Finally, based on the findings of this research, it is possible to conclude that the Tropospheric NO₂ column density product from the TROPOMI sensor is promising for evaluating NO₂ concentration in the lower atmosphere of the state of São Paulo. On the other hand, the MODIS-AOD data retrieved through the MAIAC algorithm showed low potential for directly predicting MP concentrations at ground level. Thus, our findings demonstrate that, although MAIAC-AOD has shown relatively good results as a PM predictor in some areas around the world, its use may be limited to locations with less uncertainty regarding atmospheric parameters and other assumptions and approximations used in AOD retrieval.

Author statement

Conceptualization, A. P. Rudke, T. T. de A. Albuquerque, and J. A. Martins; Data curation, A. P. Rudke; Formal analysis, A. P. Rudke and J. A. Martins; Funding acquisition, T. T. de A. Albuquerque, and J. A. Martins; Investigation, A. P. Rudke, L. D. Martins, R. Hallak, E. D. Freitas, M. F. Andrade, P. Koutrakis, T. T. de A. Albuquerque, and J. A. Martins; Methodology, A. P. Rudke and J. A. Martins; Project administration, T. T. de A. Albuquerque and M. F. Andrade; Resources, T. T. de A. Albuquerque, and J. A. Martins; Software, A. P. Rudke; Supervision, T. T. de A. Albuquerque, L. D. Martins, and J. A. Martins; Visualization, A. P. Rudke, D.S. de Almeida, and A. Beal; Writing—original draft preparation, A. P. Rudke, L. D. Martins, T. T. de A. Albuquerque, D.S. de Almeida, and J. A. Martins; Writing—review and editing, A. P. Rudke, J. A. Martins, D.S. de Almeida, L. D. Martins, A. Beal_, R. Hallak, E. D. Freitas, M. F. Andrade, P. Koutrakis, and T. T. de A. Albuquerque.

Declaration of Competing Interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Taciana Toledo de Almeida Albuquerque reports financial support was provided by Federal University of Minas Gerais. Anderson Paulo Rudke reports financial support was provided by Coordination of Higher Education Personnel Improvement. Maria de Fatima Andrade reports financial support was provided by State of Sao Paulo Research Foundation. Daniela Sanches de Almeida reports financial support was provided by National Council for Scientific and Technological Development.

Edited by Dr. Menghua Wang

Appendix A. Supplementary data

Supplementary material

Data availability

Data will be made available on request.