Skip to main content
NCBI home page
Springer Nature - PMC COVID-19 Collection logoLink to Springer Nature - PMC COVID-19 Collection
. 2023 Jan 13;16(4):745–764. doi: 10.1007/s11869-023-01303-6

Surface, satellite ozone variations in Northern South America during low anthropogenic emission conditions: a machine learning approach

Alejandro Casallas 1,2,✉, Maria Paula Castillo-Camacho 3, Edwin Ricardo Sanchez 3, Yuri González 3, Nathalia Celis 4, Juan Felipe Mendez-Espinosa 5, Luis Carlos Belalcazar 3, Camilo Ferro 2
PMCID: PMC9839215  PMID: 36687138

Abstract

2020 presented the ideal conditions for studying the air quality response to several emission reductions due to the COVID-19 lockdowns. Numerous studies found that the tropospheric ozone increased even in lockdown conditions, but its reasons are not entirely understood. This research aims to better understand the ozone variations in Northern South America. Satellite and reanalysis data were used to analyze regional ozone variations. An analysis of two of the most polluted Colombian cities was performed by quantifying the changes of ozone and its precursors and by doing a machine learning decomposition to disentangle the contributions that precursors and meteorology made to form O3. The results indicated that regional ozone increased in most areas, especially where wildfires are present. Meteorology is associated with favorable conditions to promote wildfires in Colombia and Venezuela. Regarding the local analysis, the machine learning ensemble shows that the decreased titration process associated with the NO plummeting owing to mobility reduction is the main contributor to the O3 increase (≈50%). These tools lead to conclude that (i) the increase in O3 produced by the reduction of the titration process that would be associated with an improvement in mobile sources technology has to be considered in the new air quality policies, (ii) a boost in international cooperation is essential to control wildfires since an event that occurs in one country can affect others and (iii) a machine learning decomposition approach coupled with sensitivity experiments can help us explain and understand the physicochemical mechanism that drives ozone formation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11869-023-01303-6.

Keywords: Tropospheric ozone, Machine learning, Satellite/in situ observations, Precursors decomposition, Wildfires

Introduction

In January 2020, the coronavirus disease (COVID-19) was declared a public health emergency of international concern by the World Health Organization (WHO); this disease was characterized as a global pandemic in March 2020 (WHO 2020). COVID-19 spread to several countries, including Colombia in March 2020, where the rapid increase in infected people pushed the government to declare a state of emergency, decreeing a mandatory preventive lockdown for the entire country on March 25. This lockdown measure restricted non-basic activities, promoting emission abatement to historically low levels (e.g., Venter et al. 2020; Miyazaki et al. 2021). Considering the negative economic impact and the worldwide experience, the full lockdown was partially relaxed on April 27 and extended until May 10, allowing the reactivation of specific sectors (e.g., construction, manufacturing, and some commercial activities). However, the strict and partially relaxed lockdowns handled non-significant differences in air quality (Mendez-Espinosa et al. 2020). For this, both periods were analyzed jointly and named full lockdown (March 25–May 10).

Several studies related to the COVID-19 air quality impact have been carried out worldwide and subsequently collected and analyzed (e.g., Keller et al. 2021; Addas and Maghrabi 2021; Gkatzelis et al. 2021; Chossière Guillaume et al. 2021; Cooper et al. 2022). The results indicate reductions in the concentrations of particulate matter, nitrogen dioxide, sulfur dioxide, and carbon monoxide. At the same time, soaring levels of tropospheric ozone were reported in most of the studies. The studies about air quality were performed mainly in Asia and Europe (≈83%) and to a lesser extent in South American countries (≈5%). At the same time, the results showed the importance of either utilizing emerging data sets (i.e., including the analysis of a post-lockdown period) or models accounting for the effects of meteorology, emission trends, and secondary pollutants as the tropospheric ozone. Thus, this study contributes to filling this gap in South America.

Not only have been carried out a few studies in Latin American countries like the ones in Northern South America but also those studies have not considered the analysis after the lockdown (i.e., post-lockdown, to our knowledge) (Mendez-Espinosa et al. 2020; Arregocés et al. 2021; Henao et al. 2021; Sokhi et al. 2021). Mendez-Espinosa et al. (2020) found reductions in NO2 (28–65%), PM10 (37–66%), and PM2.5 (45–76%) related to lockdown restrictions. Moreover, it was found that pollution events throughout the lockdown period were associated with biomass burning (contributing ≈ 20 μg m−3 to PM2.5), and Saharan dust intrusion (contributing between 104 and 168 μg m−3 to PM10). This analysis was executed using surface measurements, satellite images, and modeled data. Arregocés et al. (2021) used five PM2.5 monitoring stations across Colombia and satellite observation to retrieve aerosol optical depth (AOD) information, finding a reduction in PM2.5 (13–86%), whereas the AOD increased up to 59% during the months of lockdown probably due to dust intrusion from the Sahara Desert. The precipitation analysis included in this research did not show a clear pattern and association with PM2.5.

Henao et al. (2021) identified that in a single city in Colombia, large average reductions in PM2.5 (50–63%), PM10 (59–64%), NO (43–47%), NO2 (43–47%), and CO (40–47%), and increments in O3 (19–22%), were presented. Their comparisons of lockdown with pre-lockdown periods used multivariate regression analysis methods, and machine learning (random forest) to untangle the effects of the lockdown on pollution levels. This last method was useful to account for the effects of meteorology, as other researchers have shown (e.g., Grange et al. 2018; Grange and Carslaw 2019; Vu et al. 2019). Sokhi et al. (2021) made a global analysis including several countries in Latin America and also took into consideration all the atmospheric pollutants measured by the local authorities. For Bogotá-Colombia, their results agree with previous studies, and also add numerous findings: (i) tropospheric ozone increase ≈70% which the authors hypothesized is associated with the titration effect; this effect is defined as the removal of O3 through its reaction with NO (Seinfeld and Pandis 2016); (ii) NOx decreased due to the mobility reduction; (iii) CO was reduced considerably, probably due to changes in industrial activities and mobility restrictions. Despite the above, no single study has included either the analysis of a post-lockdown period or non-common pollutants (e.g., tropospheric ozone, formaldehyde) with a regional view. Moreover, none of them have used a machine learning (ML) ensemble to quantify the contribution of the precursors, markers, and meteorological variables in ozone formation. In fact, to our knowledge, this is the first paper that uses a ML approach to understand and quantify the changes related to ozone.

This work aims to better understand the ozone variations in Northern South America before, during, and after the lockdown, considering the meteorology and two types of ML sensitivity experiments. For this, a set of tools, including satellite and surface observations, reanalysis datasets, and a novel machine learning method/analysis were employed. Ghahremanloo et al. (2021) is the only paper (to our knowledge) that has done something similar in the sense of trying to understand changes in a pollutant (in their case PM2.5) by using a ML model. Nevertheless, several aspects are different since we include a decomposition analysis and also several sensitivity experiments were performed. These sensitivity analyses have been principally performed using Eulerian models (e.g., Community Multiscale Air Quality Model—CMAQ, Weather Research and Forecasting coupled with chemistry—WRF-CHEM) that need a lot of computational power, and large data storage which is not easily available in developing countries such as Colombia, so a ML learning approach could solve these difficulties and still provide important discoveries and analyses.

Method

Area of study and lockdown periods

Colombia, located in Northern South America, is the second most populated country in the region (≈50.88 million inhabitants). The country has the Andes Mountains all over the territory and a high range of altitudes (0 to 5000 m.a.s.l) as can be seen in the Online Resource (hereafter OR) Fig. OR1. Due to these mountain ranges, the country has very specific convective dynamics (see Hernandez-Deckers (2021) and reference therein for a detailed description of the convective dynamics of the region). For example, Colombia’s precipitation seems to be related to local convective patterns which increase in frequency depending on the position of the Inter-tropical Convergence Zone (ITCZ), which means that the precipitation patterns could have a strong spatial variability inside the country. These convective patterns change a lot due to the mountain ranges that have profound impacts on the wind behavior, and their convergence, which in turn are associated with the location and development of convection. For instance, at the north of the central mountain range, convection develops principally in the morning, while convection on the northeast and the Pacific coast of the country develops at night, due to the mountain’s interaction with the wind. On the other hand, at the center of the country, where Bogotá, the capital, is located, convection develops at noon, due to strong radiative forcing (Casallas et al. 2021a). This former convective activity is similar to the one presented in the South of the country (Amazon region). All these regional variabilities could have very important consequences for air pollution.

Another important element to be mentioned is that these mountain ranges not only affect the behavior of convection but also difficult the implementation of well-suited ground-base stations (Casallas et al. 2022c). Colombia is formed by 1122 municipalities and has an extension of 1143 million km2. Near 81% of the population in 2020 was projected to live in urban areas (United Nations Population Division 2018). Due to the relevance of atmospheric pollution to public health in these areas, it has become important to have continuous air quality vigilance systems. Thereby, Colombia has made an effort to improve its air quality monitoring networks over the country, currently available nearly 175 air quality monitoring stations and 24 Air Quality Vigilance Systems for large metropolitan areas comprising several surrounding municipalities (IDEAM 2019).

This study examines data from all the territory, and also analyzes promptly two capital cities, i.e., Bogotá and Medellín (markers in Fig. OR1), located in the center and the northwest regions of Colombia, respectively, for the COVID-19 lockdown periods, which are defined as follows:

  • Pre-lockdown (January 1st–March 24th): a period with no COVID-19 cases registered in Colombia; therefore, no measures for the pandemic.

  • Full lockdown (March 25th–May 10th): a period where the national government decreed full measures for the COVID-19 pandemic, including a standstill of non-essential activities, with a reduction in air pollutant emissions because of the stopping of industrial and traffic activities.

  • Post-lockdown (May 11th–December 31st): a period where government measures were relaxed enough to be closer to normal economic conditions (i.e., business as usual—BAU)

Data sources

Regional analysis

Hourly ERA5 reanalysis data from January 1, 2015, to December 31, 2020, with a 0.25° (10 m surface wind velocity, 2 m air temperature, 2 m dew-point temperature, total precipitation, mean surface downward short-wave radiation flux–radiation, boundary layer height—BLH, and total cloud cover) spatial resolution for single levels (Hersbach et al. 2018) and a 0.25° (850hPa wind velocity) spatial resolution for pressure levels (Muñoz-Sabater 2019), were downloaded over Colombia. Relative humidity (RH) was calculated using the 2 m air temperature (T2) and the 2 m dew-point temperature. Data from ERA5 has been previously validated in Colombia by many authors (e.g., Devis-Morales et al. 2021; Gil Ruiz et al. 2021; Casallas et al. 2021a) to study atmospheric convection, wind energy potential, and air quality, among others. Using this data, an anomaly analysis was performed for every lockdown period, by subtracting from the 2020 mean of each period, the multiannual (2015–2019) mean (our BAU) for each lockdown period, for all the variables except for total precipitation which was daily-aggregated (not averaged) and then subtracted. To be clearer about the anomaly calculation, here we show as bullets all the steps to calculate it:

  • Calculate the 2020 mean of the pre-, full-, and post-lockdown period. If the variable is the precipitation, first a daily accumulation is performed and then from those daily values, the mean of the period is calculated. The result would be three values, one for each period.

  • Calculate the 2015–2019 mean of the pre-, full-, and post-lockdown dates, taking into account that the precipitation has to be daily-accumulated first. The result would be three values, one for each period.

  • Subtract from the 2020 mean of each period it’s 2015–2019 analogous. In other words, and as an example, from the 2020 pre-lockdown period mean, subtract the 2015–2019 pre-lockdown dates mean. The result would be 3 values, one for each lockdown period.

For the pollutants’ behavior analysis in a regional context, the TROPOspheric Monitoring Instrument (TROPOMI) was used (Veefkind et al. 2012), which spatial resolution is 3.5 × 7 km2 for NO2, and CH2O, and 28 × 21 km2 for O3 and daily temporal resolution (from 3-day running average). The data was retrieved from offline mode at level 3, which indicates that the images have already been refined leaving only the pixels with optimal quality, of 75% for NO2, and 50% for CH2O and O3. A collection of images of density tropospheric column NO2, CH2O, and O3 was obtained for the study area, then the collection of images for each established quarantine period was filtered and averaged. Finally, the anomalies are calculated by subtracting from the 2020 mean of each period, the mean of its analogous dates of 2019, not from the whole 2015–2019 period, due to data availability.

Additionally, the Hybrid Single-Particle Lagrangian Integrated Trajectory model—HYSPLIT (Rolph et al. 2017) was used to analyze the air mass trajectories that arrived at Bogotá and Medellín in each lockdown period. Information from the Global Data Assimilation System—GDAS with 1° resolution is used as input to calculate 8 daily arrivals; 72 h per trajectory at 500 m above the surface were retrieved, following the procedure described in Mendez-Espinosa et al. (2020). Data related to hotspots were also analyzed for each study period, downloaded from the Fire Information for Resource Management System—FIRMS (NASA, 2021), with a 1 km2 resolution per pixel and a quality larger than 75%. Both data, from HYSPLIT and FIRMS, were analyzed together to find the importance of long-range transport (LRT) to increase pollutant concentrations (Casallas et al. 2022a). To have a more robust analysis, we follow the method of Casallas et al. (2022a) which is based on Mendez-Espinosa et al. (2020) and consist of applying an algorithm that identifies the active wildfires that are causally related to the back trajectory and the air quality at the selected cities. The algorithm creates a buffer zone of a 100 km radius around each of the 72 hourly locations that are defined for at least one of the trajectories reaching the study area. These allow the creation of a daily time series of potential fires related to air quality for the 2020th year. Then, a Spearman correlation (Rho) is used to see the statistical correlation between air pollution and wildfires. To see a detailed explanation of this methodology, the reader is referred to Mendez-Espinosa et al. (2020) and Casallas et al. (2022a).

Local analysis

For the local evaluation, hourly data from background surface measurements of O3, NO, NO2, CO, PM2.5, PM10, radiation, T2, RH, wind speed, and direction, between 2015 and 2020 in Bogotá and Medellín, were retrieved from each local air quality monitoring network, managed by the local environmental authorities. Bogotá’s air quality monitoring network (AQMNB) has 20 stations located throughout the city and divided by background (9), industrial (2), residential (3), and traffic (6) station types (SDA 2020). Alternatively, Medellín has 18 monitoring stations, being part of the 38 monitoring stations belonging to the Early Warning System of Medellín and the Aburra Valley (SIATA), and divided by traffic (3), industrial (1), and background (14) (SIATA 2019). We took these stations and applied Chauvenet’s criterion, which consists of eliminating the error data, and considering the interquartile deviation method (see a detailed explanation in Lishu et al., 2015). Although this criterion is normally used for parametric data, it can be used to determine the hourly outliers. These outliers are evaluated with the closest station to the station that produces the outlier at the same hour, to determine if the outlier is due to sensor error or due to a non-common phenomenon. Then the stations with more than 75% of representation between 2015 and 2019 and in 2020 are selected to be used in the analysis (Casallas et al. 2022a).

Ground-based data from background stations in both cities, 8 for Bogotá and 14 for Medellín, were analyzed for the period between 2015 and 2019 (referred to as BAU), and for 2020. Tables OR1–OR2 describe in detail the monitoring stations chosen for this research. This analysis is divided into two parts: (i) we calculate the daily average of each pollutant (for ozone, the 8-h maximum daily mean, as in Sokhi et al. 2021) for 2020 and the period between 2015 and 2019 (BAU). Then, the multiyear daily means (or daily 8-h mean for ozone) and standard deviation are calculated for the BAU period, to determine if the 2020 pollutant concentrations are significantly different than the BAU values, for each lockdown period. (ii) We calculate the percentage of change between the 2020 pollutant concentrations and the BAU pollutant concentrations for each of the three lockdown periods. In this sense, we show how much the pollutants change in each lockdown period, and also if these changes are significant or not.

Machine learning ensemble

Feature selection

Ozone variations can be largely explained by variations in its precursors (i.e., NO, NO2, CO), in some markers (i.e., PM2.5) which have been used as a proxy for transported pollution (e.g., from wildfires) (Brancher 2021). Also, PM2.5 could change the ozone through secondary organic aerosols (SOA). The particulate matter coarse fraction (PMC) can produce variations in O3 due to the absorption/reflection of radiation (Zoran et al. 2020) and radiation could also play an important role in ozone formation (e.g., Deroubaix et al. 2021; Liu et al. 2021). We also include the T2, the RH, wind speed, and wind components (u and v) to account for the wind direction, following Li et al. (2019) and considering our regional results (see section “Ozone—regional analysis”). We select these eleven variables to perform the variance inflation factor (VIF) (e.g., Ghahremanloo et al. 2021) using the data from Bogotá (8 stations) and Medellín (14 stations) air quality monitoring networks, and exclude the input variables with multicollinearity larger than 5 as proposed by Kline (2015) and Ghahremanloo et al. (2022), since several studies (e.g., Wei et al. 2019) have shown that multicollinearity could reduce model performance.

Input/output matrix description

After the VIF method is performed (Table OR3), nine variables (NO, NO2, CO, PMC, PM2.5, radiation, wind speed, and the wind components) are selected (hereafter precursors when talking about the ensemble) as inputs for the ML techniques designed. Each variable in the training set has an hourly time resolution (70,000 data points) and is downloaded from Bogotá (8 stations) and Medellín (14 stations) air quality monitoring networks. To create the X-matrix (input), the information from all the background stations is averaged (hourly). In the case that no background station has data from one variable, another type of urban station is selected to be part of the X-matrix. The missing values are imputed following Mogollon-Sotelo et al. (2021), if the imputation is done with a neural network or a linear regression taking into account the four closest background stations, as made by Celis et al. (2022), the results did not show considerable differences. Thus, we follow Mogollon-Sotelo et al. (2021) as it needs fewer computational resources. The Y-vector (output) is created by hourly averaging the ozone concentrations from the background stations of each city. It is important to mention that other variables (e.g., formaldehyde, BLH) can have a high impact on ozone formation, but due to a lack of data availability, they were not included.

Structure, training, and validation

ML techniques have been used for air quality forecasting by many authors (e.g., Huang and Kao 2018; Sayeed et al. 2021a), and in Colombia, they produce high-precision forecasts (e.g., Casallas et al. 2021b; Celis et al. 2022). This paper takes this idea further and is one of the first to use these techniques to quantify the contribution of the precursors, markers, and meteorological variables in the formation of ozone, an idea that is based on the surface fluxes’ decomposition approach of Tompkins and Semie (2021). For the ML techniques, we used (i) a decision tree (DT) algorithm (Breiman et al. 1984), (ii) a random forest algorithm (RF) (Breiman, 2001), and (iii) a neural network (NN) model (Hinton 1989) (Fig. OR2), from the SKlearn package (version 1.0.1) of Python 3.9.7 as ensemble members. We select these three methods after performing tests with six ML models (i.e., NN, RF, DT, polynomial regression—PR, support vector machine—SVM, and stochastic gradient descent—SGD) to test their capability to represent the ozone trends using the chemistry, markers, and meteorology as inputs. We found that the PR does not follow the trends of pollution as well as the others, and in terms of the SVM and the SGD, we found good results, but the models have larger errors compared to the other three to represent the magnitude of the pollutant. This is why we decide to select the three models that produce the smallest errors in both the representation of the magnitude and behavior of the pollutant.

All the aforementioned methods were trained using 90% of the hourly data, with the remaining 10% used for validation. To prevent overfitting, an early stopping method is used for the NN (Prechelt 1998). The validation results show that all the models have a coefficient of determination—R2 ≥0.80, index of agreement—IOA ≥0.86, mean bias—MB ≤ ∣0.5∣ ppb, Pearson’s correlation coefficient—R ≥0.83, and root mean square error—RMSE ≤7 ppb (more validation details are illustrated in Table OR4 and Figs. OR3–OR4). These statistical results and the model evaluation criteria of Mogollon-Sotelo et al. (2021) clearly show that the ensemble can reproduce the magnitude and trends of the observations with high precision. It is important to mention that (i) a randomized search method was used to find the best number of hidden layers/neurons, and depth of the trees. (ii) This method can be used in any city that measures ozone and its related predictors. (iii) To test the sensitivity of the ensemble to the models selected, we repeat our calculations including the SVM and SDG models, and the results produce very small changes, so for the sake of simplicity, we decided to only include the NN, DT, and RF models in the final ensemble.

Results and discussion

Ozone, meteorology, and precursor regional analysis

Ozone—regional analysis

The anomalies for ozone were obtained by comparing 2019–2020 ozone concentrations from TROPOMI during the interest periods, as is explained in the “Method” section. Results show high increases (positive anomalies) in ozone levels during pre-lockdown, full lockdown, and post-lockdown mainly in the center and Northeast of Colombia (Fig. 1). Although major positive anomalies are found in the whole region during the full lockdown, the largest increments were found in the pre-lockdown. Additionally, the ozone concentrations (Fig. OR5) reveal that during the full lockdown, the concentrations were higher than in the pre-lockdown period, especially in the Colombia-Venezuela border, and in the North of Colombia, which is explained by the low cloud fraction that allow the increase of radiation and also by the presence of wildfires in this region (e.g., Ballesteros-González et al. 2020, Mendez-Espinosa et al. 2020), that develop due to favorable meteorology conditions. These wildfires are correlated with ozone formation since the areas with the highest pollutant emissions are highly correlated with the zones where wildfires are developing, as reported by González (2021) and demonstrated by Casallas et al. (2022a). However, a detailed quantification of the contributions of ozone due to wildfires is out of the scope of this paper. In this way, ozone concentrations are linked with meteorology through the presence of favorable conditions for the development of wildfires which are correlated with the highest emission areas, and with its precursors’ trend. These two features will be further explained in the next two sections.

Fig. 1.

Fig. 1

Open in a new tab

Ozone anomalies during each lockdown period: a pre-lockdown, b full lockdown, c post-lockdown. The anomalies are calculated by subtracting from the 2020 mean of each period, the mean of each period of the 2019 year. In other words, and as an example, the anomaly of the lockdown is the subtraction of the ozone average in this period of 2020, minus its analogous period average in 2019

It is important to remark that these analyses do not intend to be compared with the surface measurements, although one difference is worth noting. In Bogotá and Medellín, the positive anomaly is not as significant as is measured by the local stations. When using only 2019 as BAU for the surface stations’ anomalies, ozone increases by 17.2% for Bogotá and 28.7% for Medellín, while TROPOMI anomalies are 9.4% and 8.8%, respectively. This could probably be explained by three main reasons: (i) TROPOMI measures the total tropospheric column of ozone and not only the surface concentrations. (ii) TROPOMI has a bias associated with the interplay between its orbit, cloud coverage, and geophysical variability (Hubert et al. 2021), which have to be taken into account in cities with a cloud cover average of ≈80%. (iii) A signal of a local ozone increase in a city, represented just by one pixel and a daily time resolution (not the 8h-max), could not be fully captured by the coarse resolution of the satellite instrument (Hubert et al. 2021).

Meteorology—regional analysis

Figure 2 shows the anomalies for radiation during each lockdown period, compared to the same dates in the previous years (2015–2019). It can be observed that the highest increments of solar radiation levels occurred during the pre-lockdown and full-lockdown in the Colombia-Venezuela border, where numerous wildfires are usually produced (Ponomarev and Ranson 2016). The radiation anomaly seems to be in part produced by a negative (≈25%) cloud cover anomaly (Fig. 3) in the Colombian-Venezuelan border and at the North of Colombia. This cloud cover anomaly not only could help explain the radiation anomaly but can also be important to understand the increase of ozone in Colombia and especially in Bogotá and Medellin. In both cities, the cloud fraction was reduced by ≈20%, which produce an increase in the radiation reaching the surface and the ozone could, in turn, increase. On the other hand, the radiative positive anomaly could explain the positive air temperature anomaly in the same regions (Ricke et al. 2010), as shown in Fig. OR6. The air temperature has a significant negative correlation with the RH since they are related throughout the Clausius-Clapeyron equation. For this, it is expected that the negative RH anomaly (Fig. OR7) develops in the zones where the temperature has a positive anomaly, as in fact, is shown in our results. The radiation, temperature, and humidity behavior coincide with the anomalies reported by Sokhi et al. (2021) for Northern South America, and they also indicate that the meteorological conditions in the North of Colombia and Venezuela are favorable for the development of wildfires, as was pointed out by Jain et al. (2021). They describe how an increase in temperature and a decrease in RH are largely associated with the number and intensity of wildfires (Fig. 4), especially when there are low values of precipitation (Fig. OR8) as in the case of Northern South America, for the pre- and full-lockdown periods.

Fig. 2.

Fig. 2

Open in a new tab

Solar radiation anomalies (2020–BAU) in Northern South America during each lockdown period: a pre-lockdown, b full lockdown, c post-lockdown. The anomalies are calculated by subtracting from the 2020 mean of each period, the mean of each period for the 2015–2019 years. In other words, and as an example, the anomaly of the lockdown is the subtraction of the solar radiation of the entire period, minus the average of all the lockdown dates of the 2015–2019 period. The plot was constructed using the Cartopy Python package (Met Office 2015)

Fig. 3.

Fig. 3

Open in a new tab

Total cloud cover anomalies (2020–BAU) in Northern South America during each lockdown period: a pre-lockdown, b full lockdown, c post-lockdown. The anomalies are calculated by subtracting from the 2020 mean of each period, the mean of each period for the 2015–2019 years. In other words, and as an example, the anomaly of the lockdown is the subtraction of the total cloud cover of the entire period, minus the average of all the lockdown dates of the 2015–2019 period. The plot was constructed using the Cartopy Python package (Met Office 2015)

Fig. 4.

Fig. 4

Open in a new tab

Back trajectories’ density of air masses calculated with HYSPLIT (black lines) and MODIS fires events (red dots) during each lockdown period: a pre-lockdown, b full lockdown, c post-lockdown

To analyze the wildfire’s behavior and its contribution to the ozone concentration, the wind field and the wind back trajectories (only for Bogotá and Medellín) are determined. Wind speed at the surface (the results are largely similar for winds at 850 hPa) had a low positive anomaly at the North and North East of Colombia (Fig. OR9), and the winds move air masses from this area towards the center of the country (Fig. 4), where the mountain ranges (and both cities) are located. In the northern zones of Colombia, and Venezuela, numerous wildfires were developing as can be seen in Fig. 4, which could help explain the increase in ozone (Ballesteros-González et al., 2020) and some of its precursors (e.g., Brancher 2021) in Bogotá and Medellín during the pre-and full-lockdown periods, as several other studies (e.g., Mendez-Espinosa et al. 2020; Ballesteros-González et al. 2020; Sokhi et al. 2021; Casallas et al. 2022a; Casallas et al. 2022b) have shown. It is important to mention that in the case of Bogotá, the wildfires that produce changes in the ozone and the other pollutants are related to Venezuela’s events; this is shown, since the correlation of the back trajectories and the number of wildfires associated to them with the ozone concentration is 0.54 (Rho) with a p-value of 0.0012.

On the other hand, Medellín is affected by the wildfires that developed in the North of the country, where the Caribbean cities are located (Fig. 4). The Northern wildfires associated with the back trajectories have a Rho of 0.57 with ozone, with a p-value of 0.042. This means that there is statistical significance, and shows that wildfires are able to increase the ozone concentration in the studied cities; nevertheless, the exact mechanism of this is still to be described in detail, but this is out of the scope of this research. Additionally, to have a more robust understanding of the mechanisms that are driving the ozone behavior, the BLH (Fig. OR10) is analyzed; this variable increases in the northern region during the pre- and full lockdown (coinciding with the temperature anomaly). Nevertheless, this variable does not seem to be important for the production of ozone, since meteorology impacts the ozone formation principally by producing wildfires, which in combination with a lack of precipitation, high radiation, and low cloud cover significantly increase ozone pollution in the region.

In summary, during the pre- and full-lockdown period, radiation had a positive anomaly (due in part to a negative cloud cover anomaly) which increased the air temperature and indirectly decreased RH. This humidity reduction also affects the total precipitation, producing less rain. The negative rain anomaly decreases the pollutant dilution (e.g., Dey et al. 2018; Casallas et al. 2022b) and generates favorable conditions for the development of wildfires in the North of Colombia and Venezuela, which increase pollution (i.e., O3, PM, VOCs) in these areas, as the statistical evaluation demonstrates. The polluted air parcels are transported to Bogotá and Medellín due to the wind conditions, first because the wind is moving from the north and northeast towards the cities and second since the wind speed has a positive anomaly between these fire areas and both cities, facilitating the transport of particles to these sites. It is also important to note that the cloud cover and the radiation could be playing an important role in local and regional ozone changes since less cloud cover implies more available radiation to be used in the formation of ozone.

Precursors—regional analysis

For the aim of this research, NO2 (Figs. OR11–OR12) and formaldehyde as a VOC representative (Fig. OR13) were retrieved from TROPOMI, as ozone precursors. Formaldehyde anomalies are presented at a regional scale in Fig. 5. During the pre- and full-lockdown periods, there are seen significant positive anomalies in formaldehyde concentrations, promptly in the Northeast, Orinoco, and Northern Amazon regions, which may be linked to the development of wildfires and biomass burning, often in these seasons (e.g., Evtyugina et al. 2013, Ballesteros-González et al. 2020). Whereas, during the post-lockdown period, the formaldehyde levels decreased in Colombia, getting anomalies close to zero and even negative, agreeing with the reduction in wildfires that have been pointed out as primary VOCs source (e.g., Yamasoe et al. 2015). Furthermore, NO2 anomalies (Fig. OR12), show a decrease in NO2 concentrations almost all across the country for the pre- and full-lockdown periods previously attributed to the reduction in traffic emissions because of the lockdown measurements (e.g., Sokhi et al. 2021). In the Colombia-Venezuela border, there is a considerable increase during the full lockdown that may be also related to the development of wildfires in these seasons, as shown in the VOCs (Fig. OR13) trend (e.g., Sitnov and Mokhov 2017). For the post-lockdown, NO2 similar to formaldehyde concentrations did not show remarkable changes concerning 2019. The behavior found for formaldehyde and NO2 might explain some of the positive ozone anomalies, mainly due to the role that these two pollutants play as ozone precursors (e.g., Jiang and Fast 2004).

Fig. 5.

Fig. 5

Open in a new tab

Formaldehyde anomalies during each period of confinement: a pre-lockdown, b full lockdown, c post-lockdown, and the analyzed cities (1) Bogotá and (2) Medellín. Notice that only here the cities are included because this pollutant does not have in situ measurements. The anomalies are calculated by subtracting from the 2020 mean of each period, the mean of each period of the 2019 year. In other words, and as an example, the anomaly of the lockdown is the subtraction of the formaldehyde average in this period of 2020, minus its analogous period average in 2019

Since there are no surface VOC measurements, the formaldehyde data measured by TROPOMI is used to evaluate its anomalies in Bogotá and Medellín (Fig. 5 and Fig. OR13 bottom panels). Both cities presented an increase in formaldehyde concentrations during the pre-lockdown period, which can be associated with the masses’ trajectory (e.g., Sitnov and Mokhov 2017) arising from the wildfires developed during the first months of 2020. Subsequently, during the full-lockdown period, a reduction in concentrations is observed in both cities, previously explained by a strong correlation between VOC concentrations and vehicle emissions (e.g., Gkatzelis et al. 2021). This also agreed with the post-lockdown trend, where anomalies remained negative. Although comparing it to the full-lockdown period, the concentrations increased (Fig. OR13) most likely due to the increment in vehicular emissions (e.g., Pakkattil, et al. 2021).

Ozone and precursor local analysis

Ozone—local analysis

Daily 8h-max mean ozone concentrations followed a similar trend to the previous years in both cities, being within the standard deviation (half a sigma) of the BAU (Fig. OR14). However, it is noticeable that between the third and fifth months, these concentrations increased significantly, by ≈69% in Bogotá and ≈42% in Medellín (Fig. 6), compared to the analogous periods in the BAU, being near the maximum values, and even higher than them, concurring with the pre- and full-lockdown (Sokhi et al. 2021). This behavior can be attributed to many factors, such as the pollutants (i.e., PM2.5, PM10, VOCs) concentrations emitted by wildfires that were highly active in the pre-lockdown phase and usually happen over those months, becoming one of the main sources of tropospheric ozone (e.g., Yamasoe et al. 2015; Ballesteros-González et al., 2020). Moreover, the decrease in NOx and PM is attributed to the reduction in traffic and industry emissions during the full-lockdown period (e.g., Guevara et al. 2021). The decrease in NOx could be related to the increase in ozone due to the titration effect, as has been pointed out by other authors (e.g., Guevara et al. 2021; Sokhi et al. 2021). On the other hand, CO reduction could be also significant due to its secondary role in ozone formation, via OH− liberation (Yamasoe et al. 2015). Thus, these factors will be further analyzed in the next section.

Fig. 6.

Fig. 6

Open in a new tab

Percentage changes of O3, NO, NO2, PM2.5, PM10, and CO for the lockdown periods, compared to the mean/8-h max on their analogous periods in 2015–2019 (BAU) in the cities of a Bogotá and b Medellín

Precursors—local analysis

The precursors NO, NO2, PM2.5, PM10, and CO concentrations were also evaluated to find out the possible causes of the O3 trend in 2020. Considering that O3 is a short-lived gas produced by complex photochemical reactions between the sunlight and some precursor gasses (e.g., VOCs, CO) in the presence of NOx (e.g., Seinfeld and Pandis 2016).

Percentage changes for each precursor were calculated as shown in Fig. 6, by each lockdown period; the time comparisons between the 2020 trends and the multiannual average were also calculated for each pollutant and are shown in Figs. OR15–OR16. The findings indicate that nitrogen oxides (NO, NO2) had a significant (larger than a sigma) reduction in Bogotá (≈63% and ≈32%, respectively) and in Medellín (≈67% and ≈50%, respectively) relative to the analogous BAU periods, in the full lockdown, as previously reported by Sokhi et al. (2021). However, in the pre-lockdown, the NO reduction was faster (≈6% in Bogotá and ≈22% in Medellín) than the NO2 decrease, since the last one did not have any significant change, until the beginning of the full lockdown. By the post-lockdown, the NO kept a ≈23% decline in Bogotá but in Medellín increased to ≈7%, in contrast to the NO2 that did not get a significant change in Bogotá, but in Medellín, the reduction remained by ≈12%. The drop in these pollutants for the 2020 full-lockdown scenario has been documented and attributed to the high reduction in traffic emissions due to mobility restrictions (e.g., Sokhi et al. 2021; Deroubaix et al. 2021). Nevertheless, the slow decrease of NO2 is also an indicator of external emissions, such as the ones produced by wildfires (e.g., Betancourt-Odio et al. 2021; Brancher 2021). Thus, as explained by Brancher (2021), the great reduction in NO is one of the main reasons for the O3 increment, since that species is necessary to lead the titration phenomenon that controls the production of tropospheric ozone.

PM2.5 and PM10 showed a similar trend with a reduction in the three lockdown phases, oscillating between 2% and 44%, being greater in the full lockdown, and having similar behavior in both cities. This is because PM is widely related to traffic emissions in urban areas (e.g., Castillo-Camacho et al. 2020; Sokhi et al. 2021). However, its decrease is lower compared with the other pollutants in a low emissions scenario, suggesting that the PM can be a secondary pollutant (Deroubaix et al. 2021), and also because of the LRT produced by wildfires at the start and middle of the full-lockdown (Figs. OR15–OR16). PM behavior is also significant when analyzing the increase of tropospheric ozone, because, as Zoran et al. (2020) reported, the PM reflects the sunlight radiation and retard the photochemical reaction forming ozone, so in the lack of PM, the O3 formation is going to increase, also because it can act as an LRT proxy, and as shown by Ivatt et al. (2022), the ozone could present an aerosol-inhibited regime.

In Bogotá, CO had reductions between 9% and 17% compared to BAU during the pre- and post-lockdown period. Although, at the beginning of January, it had a notable 10-day peak (Fig. OR15), which could be explained by a specific industrial or traffic activity that could have taken place in the early morning (5 to 7h Colombian Local Time) or by a sensor error not identified on the data quality control. Instead, Medellín got a ≈1% increment in the pre-lockdown, a ≈1% reduction in the post-lockdown, and during the full lockdown, both cities showed a decrease of ≈46%. The CO behavior is principally explained by the mobile sources since the pollutant correlation (Rho) with NOx is ≈0.69 (p-value ≤0.05) and with a lesser degree to stationary sources that are more related to SO2, which has a Rho ≈0.35 (p-value ≤0.05) with CO. In terms of its relationship with ozone formation, CO has a negative correlation and can be acting to increase ozone formation, a feature that will be further investigated in the “Ozone formation decomposition” section.

To summarize, the development of wildfires in the region played an important role in the high formation of tropospheric ozone, which is confirmed through the fact that pollutants such as PM increase their concentration at the same time as the ozone and when numerous wildfires were active. This became one of the reasons that led to the O3 increment found at the end of the pre-lockdown, and at the beginning and middle of the full-lockdown period. Additionally, the significantly low concentrations of NO and the slow decrease of NO2 produce a reduction in the titration phenomenon, which reduces the O3 consumption and allows it to increase. Taking these results into consideration and to improve our understanding of the role that the precursors are playing in the production of ozone, chemistry-transport models or ML models can be used, as have been suggested by other authors (e.g., Guevara et al. 2021; Deroubaix et al. 2021; Sokhi et al. 2021). In the next section, novel ML experiments are described, as they were used to disentangle the importance of each precursor in ozone production.

Ozone formation decomposition

To understand the surface ozone anomalies in the lockdown period, we broke them into the contribution their precursors made (see the “Method” section for details) for Bogotá and Medellín using an ensemble of three ML models. To understand the contribution of the precursors, we perform two types of experiments for the full-lockdown period/dates:

  1. Precursor homogenization: a set of 10 sensitivity tests are performed, the first allows all the precursors to behave normally, and the other executions homogenize (to their BAU mean) all the precursors except for one. In other words, if we are interested in the changes produced by NO, the other precursors are replaced by its BAU values, and only the NO is allowed to behave as it did in the full-lockdown period.

  2. Precursors modification by year: we run the ML models for the full-lockdown dates of 2017 to 2019 and then perform sensitivity experiments by replacing each variable of the simulated year with its 2020 values. For the sensitivity experiment of the year 2020, each precursor is replaced with its 2019 value.

Precursor homogenization

The ML ensemble can follow with high precision the magnitude and tendencies of the O3 for Bogotá and Medellín (Fig. 7 and Fig. OR17), with higher accuracy and less spread for Bogotá. This high precision is seen when comparing the cyan and black lines. The experiments showed that the increases in ozone are principally associated with the NO decrease which, due to titration, reduces ozone consumption in both cities, as was previously suggested by Sokhi et al. (2021) for Colombia. PM2.5 is also playing an important role at the beginning and between days 23 and 26 of the lockdown period (Fig. 7b–e); this can be explained because PM2.5 acts as a marker for transported pollutants (Zoran et al. 2020), and at these dates, wildfires were developing in the North of Colombia and Venezuela (Fig. 4). Wildfires not only transported aerosols (Mendez-Espinosa et al. 2021), they also produced increases in ozone formation, as was previously pointed out by Ballesteros-González et al. (2020), a feature that is also captured by the ensemble.

Fig. 7.

Fig. 7

Open in a new tab

a–c Bogotá and d–f Medellín ozone daily 8h-max concentration for the ensemble means during the lockdown period when all the precursors except for one are homogenized. The precursors in the legend label represent the one that is not homogenized, and the shades are the ensemble standard deviation

For Bogotá, CO and PMC have no significant (≈5±2 ppb) positive contributions to the O3 formation (Fig. OR17b–c), which was expected since both pollutants present a low correlation (CO - Rho = −0.28 with a p-value of 0.023) and in the case of the PMC, no statistical significance (PMC - Rho = 0.26 with a p-value of 0.15) was found. The NO2 has low (≈1±3.5 ppb) contributions, but conclusions for this pollutant have to be taken with caution since its uncertainty oscillates between positive and negative contributions. Further sensitivity experiments are performed, by artificially increasing/decreasing the NO2 by 5% and 10% of its mean (not shown); this indicates that the changes could be related to the NO2/NO ratio (hereafter NOratio), in which smaller values produce larger ozone inhibition (Seinfeld and Pandis 2016), so in the case of Bogotá when the NO is homogenized, NO2 tend to produce a positive contribution to ozone formation (Betancourt-Odio et al. 2021). It is possible that the addition of more pollutants (e.g., formaldehyde) to the X-matrix, the inclusion of more robust ML techniques (Celis et al. 2022), the development of sensitivity experiments in physical models (Ballesteros-González et al. 2020), or the combination of both techniques as in Sayeed et al. (2021b) could help to reduce the NO2 uncertainties related to this experiment.

The wind speed and its components (Figs. 7c and OR17c) have a very low (≈2 ±1 ppb) contribution to the ozone formation, probably because these variables reduce the concentration of the pollutant by advection and turbulent mixing, so its contribution would depend on the magnitude of the eddies they produce. For Bogota, the wind is normally very slow, so the magnitude of eddies is not able to mix the pollutant and reduce its concentration, as also found by Casallas et al. (2022b). In terms of solar radiation, the model predicts a small (≈3±1.5 ppb) negative contribution, although the radiation values of 2020 have a positive (≈4.2%) anomaly (Fig. 2 and Fig. OR15), it is not enough to produce a positive ozone contribution. Radiation has to be artificially increased (not shown) by a 5% or 12% to have a null contribution or a 5 ppb positive contribution to O3 formation, which means that this variable does not produce large changes to ozone concentration as is expected for a high cloud fraction (≈88±6%) area such as Bogotá (Deroubaix et al. 2021).

In the case of Medellín (Fig. 7d–f and Fig. OR17d–f), the ensemble has larger (but not significant) uncertainties, so these results have to be taken with caution. Nevertheless, some signals allow us to make hypotheses and explain some features of ozone formation. CO and PMC produce small (≈4±2 ppb and ≈7±5 ppb) negative contributions (Fig. OR17d–e), the first for the same reason as in Bogotá. The PMC on the other hand has a negative contribution, which suggests that the VOC oxidation mentioned by Deroubaix et al. (2021) decreases and this produces a negative contribution to ozone formation. This result also seems to indicate that the VOC oxidation is more important than the radiative properties of the PMC for the formation of ozone in Medellín, something that deserves further investigation. PM2.5 has a low (≈3±2 ppb) negative correlation that increases when wildfires (North of Colombia) are present, although this signal is not as strong as in Bogotá (Fig. 4) due to the number and intensity of the wildfires. Solar radiation shows no significant impacts on ozone formation, possibly because Medellín has a cloud fraction percentage of ≈73±4% in the majority of the full-lockdown period (not shown). In the case of NO2, the same sensitivity experiments performed for Bogotá show that the NOratio hypothesis could explain the ozone contributions due to NO2. Medellín NOratio tends to be ≈10% smaller than in Bogotá, which produces a larger O3 inhibition, so a negative contribution is seen, as expected (e.g., Finlayson-Pitts and Pitts 1999; Seinfeld and Pandis 2016). In terms of the wind speed and its components, they have a low negative (≈4 ±3) contribution possibly because the magnitude of the wind is larger than in Bogota, which produce an increase in the magnitude of the eddies and reduce the pollutant by mixing it (e.g., Tang et al. 2016; Casallas et al. 2022b), although this is only a hypothesis, and needs a more detailed investigation.

In general, the ensemble shows that the increase in O3 was principally due to the decrease in NO and its titration effect. It also shows that wildfires are an important contributor to the increase of ozone at the beginning and middle of the lockdown. Another significant result is that for Medellín, PMC seems to be acting through VOC oxidation, which is possibly one critical contributor that is not measured in any air quality station in the country. In fact, if this ML ensemble is run with the Copernicus Atmosphere Monitoring Service—CAMS (Inness et al. 2019) data (for each city) as input (not shown), adding formaldehyde and the BLH. It is found that the major contributor to the ozone formation is the NO, followed by the formaldehyde and that the other precursors behave similarly to the model that uses the surface stations as input. The wind has low negative contributions to the O3 increase, due to its low magnitudes, and probably because the wind produces eddies that mix and reduce the pollutant. The ensemble also shows that there are some significant uncertainties related to radiation, CO, and especially NO2 that need further investigation, although some of them are clarified with the second set of experiments.

Precursor modification by year

The second set of experiments (Figs. 8–9 and Figs. OR18–OR19) coincides with the conclusions made in the previous section. Especially since allowing all the precursors to change gives a better and more realistic idea of the importance of the 2020 values in the ozone increase. In these experiments, three pollutants stand out as the ones that produce the largest contribution to the O3 concentration in both cities, NO, PM2.5, and NO2, in order of their relative importance. 2020 NO values are the ones that have a large positive contribution to the ozone formation in the other years evaluated. Additionally, when replacing the values of 2019 with the ones of 2020 for NO, the ozone concentration decreases considerably to values similar to the 2019 year, which means that the titration effect is essential (and the most important) to explain the ozone increase (Sokhi et al. 2021; Betancourt-Odio et al. 2021), reinforcing the conclusion of the previous section. In Medellín, after day 20, ozone values in the 2019 and 2020 experiments (for NO) get close to the observations due to a decrease of NO in 2019 during the lockdown dates, highlighting the importance of the titration effect. On the other hand, PM2.5 has positive contributions when wildfires are active, at the beginning of the lockdown dates, and at day 20 when the 2020 values are used. When the PM2.5 2019 values are replaced in the 2020 conditions, the wildfire signal also appears but between days 5 and 20 in which an intense wildfire develops at the Colombia-Venezuela border (Mogollon-Sotelo et al. 2021), a feature that Ballesteros-González et al. (2020) reported as essential for the O3 increase.

Fig. 8.

Fig. 8

Open in a new tab

Bogotá’s ozone daily 8h-max concentration ensemble means during the lockdown date of 2017 to 2019 when each precursor is replaced (one at a time) with the values of 2020. The bottom panel shows the ozone concentrations of 2020 when each precursor is replaced (one at a time) by the 2019 values. Precursors in the legend label represent the one replaced, and the shades are the ensemble standard deviation

Fig. 9.

Fig. 9

Open in a new tab

Medellín’s ozone daily 8h-max concentration ensemble means for the lockdown date of 2017 to 2019 when each precursor is replaced (one at a time) with the values of 2020. The bottom panel shows the ozone concentrations of 2020 when each precursor is replaced (one at a time) by the 2019 values. Precursors in the legend label represent the one replaced, and the shades are the ensemble standard deviation

In terms of the NO2, it has a small negative contribution (≈2.2±0.7ppb) when the 2020 values are replaced into the other years, possibly because of the changes in the NOratio in both cities, especially for Medellín. The only difference between both cities is that in the case the 2019 values are replaced in 2020 conditions, the contributions in Bogotá are negative and in Medellín are positive after day 20 (before this day, they are both negatives). This behavior could be explained by the fact that the NOratio in Medellín is 8% larger than in Bogotá at the beginning of the lockdown, so the ozone inhibition in Medellín is weaker (Notario et al. 2012; Seinfeld and Pandis 2016). Medellín O3 concentration after day 23 in the 2020 conditions for the NO2 experiment is similar to the 2019 values, which could also be explained by means of the NOratio, which is ≈5% larger compared with 2019 (Zoran et al. 2020). The other precursors have a low contribution (≥1ppb) to the ozone formation in both cities, the reason for this is explained in the previous section and coincides with the results of this section.

In general, the ensemble results suggest that the precursors’ contribution to the ozone formation during the full lockdown are as follows: (a) the reduction in the titration effect due to the decrease in NO concentrations. (b) Changes in NOratio due to the changes in NOx (notice that the NO2 conclusions have to be further investigated). (c) The LRT, which is produced by wildfires in the North of Colombia for Medellín and Venezuela for Bogotá. (d) Radiation does not have a strong contribution due to the large cloud cover in both cities, especially in Bogotá. (e) CO has a weak contribution to the ozone formation since its correlation to O3 is not significant and (f) the PMC in Bogotá has a low contribution related to the reflection of radiation, but for Medellín, the oxidation of VOCs could be important. (g) The wind speed and its components have a weak negative contribution possibly associated with the magnitude of the eddies acting to mix and reduce the pollutant concentration.

Conclusions

2020 was an ideal setting for studying the air quality response to several emission reductions, due to the COVID-19 pandemic lockdowns. Several global and local studies have identified an increase in tropospheric ozone concentrations; nevertheless, the reasons for this behavior have not been fully understood and described, especially in Northern South America. Thereby, this research focused on the tropospheric ozone behavior and formation, in three lockdown periods. The analysis was done considering a regional scale from satellite observations (TROPOMI and ERA5), a local scale from air quality monitoring networks, and an ozone decomposition technique based on a machine learning ensemble, composed of 3 models. NO, NO2, CO, PM, formaldehyde, and meteorological variables were considered for the analysis, as tropospheric ozone precursors.

Ozone had positive anomalies in all of Northern South America, except in the Caribbean Sea, where low changes were identified. These changes are associated with NO2 and CH2O and especially with the development of wildfires on the Caribbean coasts of Colombia and in the Venezuelan cities close to the Colombian border. The wildfires are developing due to positive anomalies in radiation and temperature, and also due to negative anomalies of humidity and precipitation. In terms of the local changes (Bogotá and Medellín), at the beginning and middle of the full lockdown, the O3 increases are associated with the LRT from the wildfires developed in Venezuela in the case of Bogotá and from the North of Colombia in the case of Medellín, as the statistical analyses showed.

Considering the observation and ensemble results, the most important factor in the behavior of the surface ozone is the titration effect, which is reduced due to the significant decrease in NO produced by the low mobility during this period. It is also found that changes in NOratio due to the decrease in NOx are a critical factor to explain the increase in O3 concentrations, although this needs further investigation. Additionally, the LRT (due to wildfires) can be a factor and the ML ensemble is able to capture it by using PM2.5 as a proxy. In terms of radiation and CO, the first does not have a large contribution due to the cloud cover (≥70%) in both cities, and the second has a weak contribution to the ozone formation since its correlation to O3 is not significant and the titration effect is much stronger. The PMC for Bogotá has a low contribution to ozone, related to the reflection of radiation, which differs from the Medellín results, since the PMC contribution, although low, seems to be related to the organic oxidation of VOCs. These VOC results, summed up by the regional results and the ensemble results (with CAMS and in situ data as input), suggest that it is extremely important to start measuring and analyzing VOCs and their contribution to the ozone formation on a global scale but especially in Northern South America, where no stations measure this pollutant. The wind speed and its components have a negative low contribution, because when their magnitudes are high the mixing increases throughout the eddy’s effects and also due to the amount of pollutant that they can advect.

The results of this paper highlight three essential elements that need to be addressed or used in future policies/research: (i) the improvement of the vehicle fleet will reduce the NO and NO2 concentrations, which due to titration would produce an unwanted increase in surface ozone, so this has to be taken into account when developing policies. (ii) Extra border policies and cooperation is needed to reduce the risk, intensity, and number of wildfires, which although are produced in one country, several others could be affected, as in the case of Colombia and Venezuela, where the PM and the O3 increase considerably when wildfires are active. (iii) ML ensembles are a good alternative to analyze ozone concentration and its precursors; they could also be a tool to better understand other pollutants (i.e., PM), especially in cases where no good emission inventories or computer power is available. To improve our ensemble, the inclusion of more robust ML techniques and/or a hybrid model (between physicochemical models and ML) could be performed.

Supplementary information

ESM 1 (11.5MB, docx)

(DOCX 11755 kb)

Acknowledgements

The authors would like to thank Dr. Cyril Caminade and Dr. Ellie López-Barrera for their fruitful comments, which help to improve the results of this manuscript. We also like to thank Dr. Sebastian K. Muller for his important comments on the ML experiments. We would also like to thank two anonymous reviewers for their constructive comments. They were very important to greatly improve our results and be more precise in the methods and results of this manuscript. A special thanks to the NumPy, Matplotlib, and Pandas developers’ team.

Author contribution

Alejandro Casallas: conceptualization; data curation; methodology; software; validation; formal analysis; investigation; visualization; writing—original draft. Maria Paula Castillo-Camacho: conceptualization; data curation; methodology; formal analysis; software; visualization; writing—original draft. Edwin Ricardo Sanchez and Yuri González: data curation; methodology; software; visualization; writing—original draft. Nathalia Celis: software; validation; formal analysis; writing—original draft. Juan Felipe Mendez-Espinosa: writing—original draft. Luis Carlos Belalcazar: conceptualization; methodology; formal analysis; writing—original draft. Camilo Ferro: investigation; software; validation; formal analysis; writing—original draft.

Funding

This work was supported by the MINISTERIO DE CIENCIA, TECNOLOGÍA E INNOVACIÓN - MINCIENCIAS and Fondo de Investigación en Salud (FIS) of Colombia, through the project “Fortalecimiento de Proyectos en ejecución de ciencia, tecnología e innovación en Ciencias de la Salud con Talento Joven e Impacto Regional” from the announcement 874 of 2020; and MINCIENCIAS Grant No. 905–2019. Nevertheless, the funding sources have no involvement in the design or development of the research.

Data availability

All the data, models, and scripts used for the making of this paper are available upon request to the authors.

Declarations

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

  1. Addas A, Maghrabi A. The impact of COVID-19 lockdowns on air quality—a global review. Sustainability. 2021;13:10212. doi: 10.3390/SU131810212. [DOI] [Google Scholar]
  2. Arregocés HA, Rojano R, Restrepo G. Impact of lockdown on particulate matter concentrations in Colombia during the COVID-19 pandemic. Sci Total Environ. 2021;764:142874. doi: 10.1016/J.SCITOTENV.2020.142874. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Ballesteros-González K, Sullivan AP, Morales-Betancourt R. Estimating the air quality and health impacts of biomass burning in northern South America using a chemical transport model. Sci Total Environ. 2020;739:139755. doi: 10.1016/j.scitotenv.2020.139755. [DOI] [PubMed] [Google Scholar]
  4. Betancourt-Odio MA, Martínez-de-Ibarreta C, Budría-Rodríguez S, Wirth E. Local analysis of air quality changes in the community of Madrid before and during the COVID-19 induced lockdown. Atmosphere. 2021;12:659. doi: 10.3390/atmos12060659. [DOI] [Google Scholar]
  5. Brancher M. Increased ozone pollution alongside reduced nitrogen dioxide concentrations during Vienna’s first COVID-19 lockdown: significance for air quality management. Environmental Pollution. 2021;284:117153. doi: 10.1016/j.envpol.2021.117153. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Breiman L. Random forests. Machine Learning. 2001;45:5–32. doi: 10.1023/A:1010933404324. [DOI] [Google Scholar]
  7. Breiman L, Friedman JH, Olshen RA, Stone CJ. Classification and regression trees. 2. Belmont: Wadsworth; 1984. [Google Scholar]
  8. Casallas A, Castillo-Camacho MP, Guevara-Luna MA, Gonzalez Y, Sánchez E, Belalcázar LC. Spatio-temporal analysis of PM2.5 and policies in Northwestern South America. Sci. Total Environ. 2022;852:158504. doi: 10.1016/j.scitotenv.2022.158504. [DOI] [PubMed] [Google Scholar]
  9. Casallas A, Córdoba T, Sanchez-Cardenas L, Guevara-Luna MA, Belalcázar LC (2022b) Understanding the atmospheric characteristics of high polluted events in a tropical megacity. Rev EIA Esc Ing. 10.17533/udea.redin.20220682
  10. Casallas A, Ferro C, Celis N, Guevara-Luna M, Mogollon-Sotelo C, Guevara-Luna F, Merchan M. Long short-term memory artificial neural network approach to forecast meteorology and PM2.5 local variables in Bogotá, Colombia. Model Earth Syst Environ. 2021;8:2951–2964. doi: 10.1007/s40808-021-01274-6. [DOI] [Google Scholar]
  11. Casallas A, Hernandez-Deckers D, Mora-Paez H. Understanding convective storms in a tropical, high-altitude location with in-situ meteorological observations and GPS-derived water vapor. Atmósfera. 2021;36(2):225–238. doi: 10.20937/ATM.53051. [DOI] [Google Scholar]
  12. Casallas A, Jiménez-Saenz C, Torres V, Quirama-Aguilar M, Lizcano A, Lopez-Barrera EA, Ferro C, Celis N, Arenas R. Design of a forest fire early alert system through a deep 3D-CNN structure and a WRF-CNN bias correction. Sensors. 2022;22(22):8790. doi: 10.3390/s22228790. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Castillo-Camacho MP, Tunarrosa-Grisales IC, Chacón-Rivera LM, Guevara-Luna MA, Belalcázar-Cerón LC. Personal exposure to PM2.5 in the massive transport system of Bogotá and Medellín, Colombia. Asian J Atmos Environ. 2020;14:210–224. doi: 10.5572/ajae.2020.14.3.210. [DOI] [Google Scholar]
  14. Celis N, Casallas A, Lopez-Barrera EA, Martínez H, Peña-Rincón CA, Arenas R, Ferro C. Design of an early alert system for PM2.5 through a stochastic model and machine learning. Environ Sci Pol. 2022;127:241–252. doi: 10.1016/j.envsci.2021.10.030. [DOI] [Google Scholar]
  15. Chossière Guillaume P, Xu H, Dixit Y, Isaacs S, Eastham Sebastian D, Allroggen F, et al. Air pollution impacts of COVID-19-related containment measures. Sci Adv. 2021;7:eabe1178. doi: 10.1126/sciadv.abe1178. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Cooper MJ, Martin RV, Hammer MS, et al. Global fine-scale changes in ambient NO2 during COVID-19 lockdowns. Nature. 2022;601:380–387. doi: 10.1038/s41586-021-04229-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Deroubaix A, Brasseur GP, Gaubert B, Labuhn I, Menut L, Siour G, Tuccela P. Response of surface ozone concentration to emission reduction and meteorology during the COVID-19 lockdown in Europe. Meteorological Applications. 2021;28:e1990. doi: 10.1002/met.1990. [DOI] [Google Scholar]
  18. Devis-Morales A, Rodríguez-Rubio E, Montoya-Sánchez RA. Modelling the transport of sediment discharged by Colombian rivers to the southern Caribbean Sea. Ocean Dynamics. 2021;71:251–277. doi: 10.1007/s10236-020-01431-y. [DOI] [Google Scholar]
  19. Dey S, Gupta S, Chakraborty A, Sibanda P. Influences of boundary layer phenomena and meteorology on ambient air quality status of an urban area in eastern India. Atmósfera. 2018;31:69–86. doi: 10.20937/atm.2018.31.01.05. [DOI] [Google Scholar]
  20. Evtyugina M, et al. VOC emissions of smouldering combustion from Mediterranean wildfires in central Portugal. Atmos Environ. 2013;64:339–348. doi: 10.1016/j.atmosenv.2012.10.001. [DOI] [Google Scholar]
  21. Finlayson-Pitts BJ, Pitts JN., Jr . Chemistry of the upper and lower atmosphere: theory, experiments, and applications. Elsevier; 1999. [Google Scholar]
  22. Ghahremanloo M, Lops Y, Choi Y, Jung J, Mousevinezhad S, Hammond D (2022) A comprehensive study of the COVID-19 impact on PM2.5 levels over the contiguous United States: a deep learning approach. Atmos Environ 118944. 10.1016/j.atmosenv.2022.118944 [DOI] [PMC free article] [PubMed]
  23. Ghahremanloo M, Lops Y, Choi Y, Yeganeh B. Deep learning estimation of daily ground level NO2 concentrations from remote sensing data. J. Geophys. Res Atmos. 2021;126:e2021JD034925. doi: 10.1029/2021JD034925. [DOI] [Google Scholar]
  24. Gil Ruiz SA, Barriga JEC, Martínez JA. Wind power assessment in the Caribbean region of Colombia, using ten-minute wind observations and ERA5 data. Renew Energy. 2021;172:158–176. doi: 10.1016/j.renene.2021.03.033. [DOI] [Google Scholar]
  25. Gkatzelis GI, et al. The global impacts of COVID-19 lockdowns on urban air pollution: a critical review and recommendations. Elementa: Science of the Anthropocene. 2021;9:00176. doi: 10.1525/ELEMENTA.2021.00176. [DOI] [Google Scholar]
  26. González Y (2021) Evaluación del impacto en la calidad del aire generado por el aislamiento preventivo como medida frente al COVID-19 en tres ciudades de Colombia. Universidad Nacional de Colombia. https://repositorio.unal.edu.co/handle/unal/80998. Accessed 10 Feb 2022
  27. Grange SK, Carslaw DC. Using meteorological normalisation to detect interventions in air quality time series. Sci Total Environ. 2019;653:578–588. doi: 10.1016/J.SCITOTENV.2018.10.344. [DOI] [PubMed] [Google Scholar]
  28. Grange SK, Carslaw DC, Lewis AC, Boleti E, Hueglin C. Random forest meteorological normalisation models for Swiss PM10 trend analysis. Atmos Chem and Phys. 2018;18:6223–6239. doi: 10.5194/ACP-18-6223-2018. [DOI] [Google Scholar]
  29. Guevara M, Jorba O, Soret A, Petetin H, Bowdalo D, Serradell K, Tena C, Denier van der Gon H, Kuenen J, Peuch V, Pérez García-Pando C. Time-resolved emission reductions for atmospheric chemistry modelling in Europe during the COVID-19 lockdowns. Atmos Chem Phys. 2021;21:773–797. doi: 10.5194/acp-21-773-2021. [DOI] [Google Scholar]
  30. Henao JJ, Rendón AM, Hernández KS, Giraldo-Ramirez PA, Robledo V, Posada-Marín JA, Bernal N, Salazar JF, Mejía JF. Differential effects of the COVID-19 lockdown and regional fire on the air quality of Medellín. Colombia. Atmosphere. 2021;12:1137. doi: 10.3390/atmos12091137. [DOI] [Google Scholar]
  31. Hernandez-Deckers D (2021) Features of atmospheric deep convection in northwestern South America obtained from infrared satellite data. Quart J Royal Meteor Soc. 10.1002/qj.4208
  32. Hersbach H, et al. Copernicus Climate Change Service (C3S) Climate Data Store (CDS) 2018. ERA5 hourly data on pressure levels from 1979 to present. [Google Scholar]
  33. Hinton GE. Connectionist learning procedures. Artificial Intelligence. 1989;40:185–234. doi: 10.1016/0004-3702(89)90049-0. [DOI] [Google Scholar]
  34. Huang CJ, Kuo PH. A deep CNN-LSM model for particulate matter (PM2.5) forecasting in smart cities. Sensors. 2018;1:2220. doi: 10.3390/s18072220. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Hubert D, et al. TROPOMI tropospheric ozone column data: geophysical assessment and comparison to ozonesondes, GOME-2B and OMI. Atmospheric Measurement Techniques. 2021;14:7405–7433. doi: 10.5194/amt-14-7405-2021. [DOI] [Google Scholar]
  36. IDEAM (2019) INFORME DEL ESTADO DE LA CALIDAD DEL AIRE EN COLOMBIA. http://documentacion.ideam.gov.co/openbiblio/bvirtual/023898/InformeCalidadAire_2019.pdf
  37. Inness A, et al. The CAMS reanalysis of atmospheric composition. Atmos Chem Phys. 2019;19:3515–3556. doi: 10.5194/acp-19-3515-2019. [DOI] [Google Scholar]
  38. Ivatt PD, Evans MJ, Lewis AC. Suppression of surface ozone by an aerosol-inhibited photochemical ozone regime. Nat. Geosci. 2022;15:536–540. doi: 10.1038/s41561-022-00972-9. [DOI] [Google Scholar]
  39. Jain P, Castellanos-Acuna D, Coogan SCP et al (2021) Observed increases in extreme fire weather driven by atmospheric humidity and temperature. Nat Clim Chang 10.1038/s41558-021-01224-1
  40. Jiang G, Fast JD. Modeling the effects of VOC and NOx emission sources on ozone formation in Houston during the TexAQS 2000 field campaign. Atmos Environ. 2004;38:5071–5085. doi: 10.1016/j.atmosenv.2004.06.012. [DOI] [Google Scholar]
  41. Keller CA, et al. Global impact of COVID-19 restrictions on the surface concentrations of nitrogen dioxide and ozone. Atmos Chem Phys. 2021;21:3555–3592. doi: 10.5194/acp-21-3555-2021. [DOI] [Google Scholar]
  42. Kline RB. Principles and practice of structural equation modeling. Guilford publications; 2015. [Google Scholar]
  43. Li K, Jacob DJ, Liao H, Shen L, Zhang Q, Bates KH. Anthropogenic drivers of 2013-2017 trends in summer surface ozone in China. Proceedings of the National Academy of Sciences. 2019;116:422. doi: 10.1073/pnas.1812168116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Lishu W, Yan L, Qiuyan L, Shoutian D, Lijing T. Greenhouse environmental data collection based on improved Chauvenet’s criterion Trans. Chin. Soc. Agric. Eng. 2015;31:212–217. doi: 10.3969/j.issn.1002-6819.2015.05.030. [DOI] [Google Scholar]
  45. Liu Y, Wang T, Stavrakou T, Elguindi N, Doumbia T, Granier C, Bouarar I, Gaubert B, Brasseur GP. Diverse response of surface ozone to COVID-19 lockdown in China. Sci Total Environ. 2021;789:147739. doi: 10.1016/j.scitotenv.2021.147739. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Mendez-Espinosa JF, Belalcazar LC, Morales Betancourt R. Regional air quality impact of northern South America biomass burning emissions. Atmos Environ. 2020;203:131–140. doi: 10.1016/j.atmosenv.2019.01.042. [DOI] [Google Scholar]
  47. Mendez-Espinosa JF, Rojas NY, Vargas J, Pechon JE, Belalcazar LC, Ramírez O. Air quality variations in Northern South America during the COVID-19 lockdown. Sci Total Environ. 2021;749:141621. doi: 10.1016/j.scitotenv.2020.141621. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Miyazaki K, Bowman K, Sekiya T, Takigawa M, Neu Jessica L, Sudo K. Global tropospheric ozone responses to reduced NOx emissions linked to the COVID-19 worldwide lockdowns. Sci Adv. 2021;7:eabf7460. doi: 10.1126/sciadv.abf7460. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Mogollón-Sotelo C, Casallas A, Vidal S, Celis N, Ferro C, Belalcazar LC. A support vector machine model to forecast ground-level PM2.5 in a highly populated city with a complex terrain. Air Qual Atmos Health. 2021;14:399–409. doi: 10.1007/s11869-020-00945-0. [DOI] [Google Scholar]
  50. Muñoz-Sabater J (2019) ERA5-land hourly data from 1981 to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS). 10.24381/cds.e2161bac (accessed 06 September 2021)
  51. NASA (2021) MODIS Collection 6 Hotspot/Active Fire Detections MCD14ML distributed from NASA FIRMS. https://earthdata.nasa.gov/firms (accessed 20 December 2021)
  52. Notario A, Bravo I, Adame JA, Díaz-de-Mera Y, Aranda A, Rodríguez A, Rodríguez D. Analysis of NO, NO2, NOx, O3 and oxidant (OX = O3 + NO2) levels measured in a metropolitan area in the southwest of Iberian Peninsula. Atmospheric Research. 2012;104:217–226. doi: 10.1016/j.atmosres.2011.10.008. [DOI] [Google Scholar]
  53. Office M. Cartopy: a Cartographic Python Library with Matplotlib Interface. Exeter, UK: Met Office; 2015. [Google Scholar]
  54. Pakkattil A, Muhsin M, Varma MR. COVID-19 lockdown: effects on selected volatile organic compound (VOC) emissions over the major Indian metro cities. Urban Climate. 2021;37:100838. doi: 10.1016/j.uclim.2021.100838. [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Ponomarev E, Kharuk V, Ranson K. Wildfires dynamics in Siberian larch forests. Forests. 2016;7:125. doi: 10.3390/f7060125. [DOI] [Google Scholar]
  56. Prechelt L. Early stopping - but when? In: Orr GB, Müller KR, editors. Neural Networks: Tricks of the Trade. Lecture Notes in Computer Science. 1998. [Google Scholar]
  57. Ricke K, Morgan M, Allen M. Regional climate response to solar-radiation management. Nature Geosci. 2010;3:537–541. doi: 10.1038/ngeo915. [DOI] [Google Scholar]
  58. Rolph G, Stein A, Stunder B. Real-time Environmental Applications and Display sYstem: READY. Environ Model Softw. 2017;95:210–228. doi: 10.1016/j.envsoft.2017.06.025. [DOI] [Google Scholar]
  59. SDA (2020) Reporte de estaciones. http://rmcab.ambientebogota.gov.co/Report/stationreport (accessed 02 October 2021)
  60. Sayeed A, Lops Y, Choi Y, Jung J, Salman AK. Bias correcting and extending the PM forecast by CMAQ up to 7 days using deep convolutional neural networks. Atmos Environ. 2021;253:118376. doi: 10.1016/j.atmosenv.2021.118376. [DOI] [Google Scholar]
  61. Sayeed A, Choi Y, Eslami E, Jung J, Lops Y, Khan Salman A, Lee JB, Park HJ, Choi MH. A novel CMAQ-CNN hybrid model to forecast hourly surface-ozone concentrations 14 days in advance. Sci Rep. 2021;11:10891. doi: 10.1038/s41598-021-90446-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Seinfeld J, Pandis S. Atmospheric chemistry and physics: from air pollution to climate change. 3. Wiley; 2016. [Google Scholar]
  63. SIATA (2019) Air quality stations information. https://siata.gov.co/descarga_siata/index.php/info/aire/ (accessed 4 November 2021)
  64. Sitnov SA, Mokhov II. Formaldehyde and nitrogen dioxide in the atmosphere during summer weather extremes and wildfires in European Russia in 2010 and Western Siberia in 2012. Inter J Remote Sensing. 2017;38:4086–4106. doi: 10.1080/01431161.2017.1312618. [DOI] [Google Scholar]
  65. Sokhi R, et al. A global observational analysis to understand changes in air quality during exceptionally low anthropogenic emission conditions. Enviro Int. 2021;157:106818. doi: 10.1016/j.envint.2021.106818. [DOI] [PubMed] [Google Scholar]
  66. Tang G, et al. Mixing layer height and its implications for air pollution over Beijing, China. Atmos Chem Phys. 2016;16:2459–2475. doi: 10.5194/acp-16-2459-2016. [DOI] [Google Scholar]
  67. Tompkins AM, Semie AG. Impact of a mixed ocean layer and the diurnal cycle on convective aggregation. Journal of Advances in Modeling Earth Systems. 2021;13:e2020MS002186. doi: 10.1029/2020MS002186. [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. United Nations Population Division (2018) World urbanization prospects: 2018 revision. Urban population (% of total population) - Colombia. Urban population (% of total population) - Colombia. World bank web. https://data.worldbank.org/indicator/SP.URB.TOTL.IN.ZS. Accessed 29 Dec 2021
  69. Veefkind JP, et al. TROPOMI on the ESA Sentinel-5 Precursor: a GMES mission for global observations of the atmospheric composition for climate, air quality and ozone layer applications. Remote Sensing of Environment. 2012;120:70–83. doi: 10.1016/j.rse.2011.09.027. [DOI] [Google Scholar]
  70. Venter ZS, Aunan K, Chowdhury S, Lelieveld J. COVID-19 lockdowns cause global air pollution declines. Proceedings of the National Academy of Sciences. 2020;117:18984. doi: 10.1073/pnas.2006853117. [DOI] [PMC free article] [PubMed] [Google Scholar]
  71. Vu T, Shi Z, Cheng J, Zhang Q, He K, Wang SM, Harrison R. Assessing the impact of clean air action on air quality trends in Beijing using a machine learning technique. Atmos Chem Phys. 2019;19:11303–11314. doi: 10.5194/ACP-19-11303-2019. [DOI] [Google Scholar]
  72. Wei J, Huang W, Li Z, Xue W, Peng Y, Sun L, Cribb M. Estimating 1-km-resolution PM2.5 concentrations across China using the space-time random forest approach. Rem Sens Environ. 2019;231:111221. doi: 10.1016/j.rse.2019.111221. [DOI] [Google Scholar]
  73. WHO (2020) COVID-19: cronología de la actuación de la OMS. https://www.who.int/es/news/item/27-04-2020-who-timeline%2D%2D-covid-19
  74. Yamasoe MA, Sauvage B, Thouret V, Nédélec P, Le Flochmoen E, Barret B. Analysis of tropospheric ozone and carbon monoxide profiles over South America based on MOZAIC/IAGOS database and model simulations. Tellus B: Chemical and Physical Meteorology. 2015;67:1. doi: 10.3402/tellusb.v67.27884. [DOI] [Google Scholar]
  75. Zoran MA, Savastru RS, Savastru DM, Tautan MN. Assessing the relationship between ground levels of ozone (O3) and nitrogen dioxide (NO2) with coronavirus (COVID-19) in Milan, Italy. Sci Total Environ. 2020;740:140005. doi: 10.1016/j.scitotenv.2020.140005. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

ESM 1 (11.5MB, docx)

(DOCX 11755 kb)

Data Availability Statement

All the data, models, and scripts used for the making of this paper are available upon request to the authors.

Articles from Air Quality, Atmosphere, & Health are provided here courtesy of Nature Publishing Group

RESOURCES