Monitoring Trends of CO, NO 2 , SO 2 , and O 3 Pollutants Using Time-Series Sentinel-5 Images Based on Google Earth Engine

: Air pollution (AP) is a signiﬁcant risk factor for public health, and its impact is becoming increasingly concerning in developing countries where it is causing a growing number of health issues. It is therefore essential to map and monitor AP sources in order to facilitate local action against them. This study aims at assessing the suitability of Sentinel-5 AP products based on Google Earth Engine (GEE) to monitor air pollutants, including CO, NO 2 , SO 2 , and O 3 in Arak city, Iran from 2018 to 2019. Our process involved feeding satellite images to a cloud-free GEE platform that identiﬁed pollutant-affected areas monthly, seasonally, and annually. By coding in the JavaScript language in the GEE, four pollution parameters of Sentinel-5 satellite images were obtained. Following that, images with clouds were ﬁltered by deﬁning cloud ﬁlters, and average maps were extracted by deﬁning average ﬁlters for both years. The employed model, which solely used Sentinel-5 AP products, was tested and assessed using ground data collected from the Environmental Organization of Central Province. Our ﬁndings revealed that annual CO, NO 2 , SO 2 , and O 3 were estimated with RMSE of 0.13, 2.58, 4.62, and 2.36, respectively, for the year 2018. The annual CO, NO 2 , SO 2 , and O 3 for the year 2019 were also calculated with RMSE of 0.17, 2.41, 4.31, and 4.6, respectively. The results demonstrated that seasonal AP was estimated with RMSE of 0.09, 5.39, 0.70, and 7.81 for CO, NO 2 , SO 2 , and O 3 , respectively, for the year 2018. Seasonal AP was also estimated with RMSE of 0.12, 4.99, 1.33, and 1.27 for CO, NO 2 , SO 2 , and O 3 , respectively, for the year 2019. The results of this study revealed that Sentinel-5 data combined with automated-based approaches, such as GEE, can perform better than traditional approaches (e.g., pollution measuring stations) for AP mapping and monitoring since they are capable of providing spatially distributed data that is sufﬁciently accurate.


Introduction
In today's world, air pollution (AP) is one of the most destructive challenges to the quality of life, especially in developing countries [1]. With the rapid expansion of urbanization and the development of cities, along with rapid population growth, industrialization, and the indiscriminate use of fossil fuels, pollution has increased and exceeded the capacity of the environment to tolerate it [2,3]. As a result, citizens are more likely to suffer from respiratory diseases and suffer from worsened heart and lung conditions. Furthermore, environmental damage, including damage caused by air pollution, costs billions of dollars every year in financial credits, human labor, and other resources [4][5][6].
In recent years, AP has become a leading cause of death in both developing and developed countries [7]. In this regard, the amount of air pollutants in many cities in Sentinel-5 images for CO, NO 2 , SO 2 , and O 3 pollutants monitoring, and (3) to explore the efficiency of GEE for AP retrieval.

Location of Study Area
Arak city, the capital of the Central Province is located in the central part of Iran ( Figure 1). As one of Iran's major manufacturing cities, Arak is widely regarded as one of the four economic poles of the country. This city is the industrial capital of Iran due to the presence of mother industries, the production of 80% of the country's energy equipment, and the presence of the biggest aluminum factory, the biggest manufacturer of heavy machinery in the Middle East, the largest gasoline producer, and the largest mineral industries in the country. As a result, it is one of the most polluted cities in Iran (https://markazi.doe.ir, accessed on 1 February 2022). According to the output factors described in Table 1, the fuel consumption of large industries in Arak city and the air pollution caused by this have been estimated based on the available information. Because Arak is one of the most important industrial cores in Iran, fuel consumption in various seasons does not differ significantly from each other [41].  Table 2 shows that other motor vehicles and heating sources contributed less pollution than industries. As shown in Table 1, industries produced 99% of Arak's air pollution, and the number of pollutants was high as well.

Datasets
The main objective of this study was to monitor trends in CO, NO 2 , SO 2 , and O 3 pollutants in Arak city using time-series Sentinel-5 images derived from GEE. To this end, Sentinel-5 AP were employed to monitor CO, NO 2 , SO 2 , and O 3 pollutants based on GEE from 2018 to 2019. We also used monthly, seasonal (spring (April, May, and June), summer (July, August, and September), fall (October, November, and December), and winter (January, February, and March)), and annual data on atmospheric pollutants for AP monitoring, collected from the Environmental Organization of Central Province for 2018 and 2019 (Table 3). Figure 1(3) shows the location of pollution monitoring stations in Arak city.

Methodology
To extract pollution maps (CO, NO 2 , SO 2 , and O 3 ), the following steps were taken: in the first step, all satellite-based datasets were preprocessed and prepared. Second, results for air pollutants were obtained through JavaScript coding using GEE. As a third step, GEE and Sentinel-5 results were validated based on data from pollution measuring stations. Figure 2 provides a brief review of the methodology used for AP retrieval.

Google Earth Engine (GEE) for AP Retrieval
Google launched GEE in 2010 to store and process Earth observation data in a more reliable and time-efficient way [35]. This platform is valuable when the goal is to process open-access Earth observation data over a large area or long-time interval and in a timely manner [42,43].
To retrieve CO, NO 2 , SO 2 , and O 3 , the Sentinel-5 images were first converted from level 2 to level 3 through the harpconvert tool by the bin spatial operation. Then, after applying the spatial and temporal filters, CO, NO 2 , SO 2 , and O 3 products from the study area were generated.
Two types of output were then generated, including maps and statistical reports. The results were verified using in situ data obtained from ground-based air pollution stations.

CO Retrieval
The silent killer, CO, is a poisonous and dangerous gas that is odorless, tasteless, and invisible [44]. CO results from the incomplete burning of carbon. As part of Sentinel 5, the CO product is available from 22 November 2018. The features of this product are given in Table 4. Apex angle of the satellite is the angle away from the vertical.

NO 2 Retrieval
As a result of human activities, especially the burning of fossil fuels, millions of tons of NO 2 are produced every year [45]. In the GEE, NO 2 is one of the Sentinel-5 products that provides offline and in-live high-resolution images. The data on this product can be accessed from 7 October 2018. The features of this product are given in Table 5.  As presented in Table 6, the GEE has provided a product for the analysis of SO 2 . NO 2 data is accessible on the GEE platform from 10 July 2018.
The role of O 3 in the thermal structure of the Earth and the balance of solar radiation is critical as it prevents ultraviolet radiation from reaching the Earth's surface. However, O 3 is considered a pollutant when its concentration in the lower atmosphere exceeds the air quality standard threshold [46]. The GEE platform has provided a product to monitor and review this critical issue, which provides a set of high-resolution images in real time. This product's data can be accessed from 7 October 2018, whose features are listed in Table 7.

Accuracy Assessment
Analyzing the accuracy of a retrieval by inversion compared to a standard assumed to be correct is an important step in image analysis. In this regard, median absolute deviation (MAD) [47], mean square error (MSE) [48], root mean square error (RMSE) [49], and mean absolute percentage error (MAPE) [50] statistical analyses were applied to evaluate the accuracy of results for AP retrieval. Equations (1)-(4) describe the MAD, MSE, RMSE, and MAPE operators.
where n is the number of the AP station in the study area, A t is AP recorded by station, and F t is AP obtained using the Sentinel-5.

Results
Using GEE, pollution parameter maps (CO, NO 2 , SO 2 , and O 3 ) were extracted. By coding in JavaScript language in the GEE, four pollution parameters of Sentinel-5 satellite images were called. Using filters, the study years (2018 and 2019) and the location (Arak city) were defined. Following that, images with clouds were filtered by defining cloud filters, and average maps were extracted by defining average filters for both years. The results of the spatiotemporal distribution of CO, NO 2 , SO 2 , and O 3 in Arak are presented in Figures 3-8 monthly, seasonally, and annually, respectively.
As we can see in Figure 3, for the year 2018, the highest amount of CO was recorded in July and August (0.029 ppm). November was the month with the lowest amount of CO. According to Figure 3, while the highest amount of NO 2 was recorded in December (22.06 ppm), the lowest was recorded in May (12.39 ppm). For SO 2 , the highest and lowest amounts were related to July (33.54 ppm) and May (25.14 ppm), respectively, as shown in Figure 3. As we can see in Figure 3, the highest amount of O 3 was recorded in February (0.148 ppm), while the lowest concentrations of O 3 occurred in September and October (0.122 ppm).
According to Figure 4, for the year 2019, the highest amount of CO was recorded in January (0.030 ppm). March was the month with the lowest amount of CO. According to Figure 4, while the highest amount of NO 2 was recorded in January (30.19 ppm), the lowest was recorded in April (11.08 ppm). For SO 2 , the highest and lowest amounts were related to October (45.38 ppm), as shown in Figure 4. As we can see in Figure 4, the highest amount of O 3 was recorded in April (0.147 ppm), while the lowest amount of O 3 belonged to November (0.117 ppm).
As we can see in Figure 5, for the year 2018, the seasonal highest amount of CO was recorded in spring and summer (0.028 ppm). According to Figure 5, the highest amount of NO 2 was found in fall (17.39 ppm), and the lowest was found in spring (12.18 ppm). For SO 2 , the highest and lowest amounts were related to winter (102.12 ppm) and summer (29.05 ppm), respectively, as shown in Figure 5. As we can see in Figure 5, the highest amount of O 3 was recorded in winter (0.142 ppm), while the lowest amount of O 3 belonged to summer (0.124 ppm).   According to Figure 6, for the year 2019, the seasonal highest amount of CO was recorded in spring, summer and winter (0.028 ppm). According to Figure 6, while the highest amount of NO 2 was recorded in winter (21.96 ppm), the lowest amount was recorded in spring (11.66 ppm). For SO 2 , the highest and lowest amounts were related to winter (101.25 ppm) and summer (28.01 ppm), respectively, as shown in Figure 6. As we can see in Figure 6, the highest amount of O 3 was recorded in spring (0.136 ppm), while the lowest amount of O 3 was associated with summer and fall (0.122 ppm).
As we can see in Figures 7  Tables 8 and 9 also show the results of the accuracy assessment for annual and seasonal AP retrieval. Our findings revealed the efficiency of Sentinel-5 AP products based on GEE for mapping and monitoring CO, NO 2 , SO 2 , and O 3 . According to  Table 8.
As we can see in Table 8, the annual CO was estimated with MAD, MSE, RMSE, and MAPE of 0.16, 0.03, 0.17, and 6.75, respectively, for the year 2019. Additionally, the annual NO 2 was calculated with MAD of 2.03, MSE of 5.83, RMSE of 2.41, and MAPE of 9.2, as shown in Table 8. As seen in Table 8 Table 9 also shows the results of the accuracy assessment for CO, NO 2 , SO 2 , and O 3 retrieval. According to Table 9, for spring 2018, the seasonal CO, NO 2 , SO 2 , and O 3 were estimated with RMSE of 0.09, 5.39, 0.70, and 7.81, respectively. In addition, for summer 2018, the seasonal CO, NO 2 , SO 2 , and O 3 were calculated with RMSE of 0.1, 5.17, 1.17, and 2.83, respectively, as shown in Table 9. As seen in Table 9, the seasonal CO, NO 2 , SO 2 , and O 3 were estimated with RMSE of 0.29, 4.47, 2.61, and 2.58, respectively, for fall 2018. Finally, for winter 2018, the seasonal CO, NO 2 , SO 2 , and O 3 were calculated with RMSE of 0.77, 4.80, 3.009, and 2.60, respectively, as shown in Table 9.
According to Table 9, for spring 2019, the seasonal CO, NO 2 , SO 2 , and O 3 were estimated with RMSE of 0.12, 4.99, 1.33, and 1.27, respectively. In addition, for summer 2019, the seasonal CO, NO 2 , SO 2 , and O 3 were calculated with RMSE of 0.11, 5.20, 0.67, and 10.35, respectively, as shown in Table 9. As seen in Table 9, the seasonal CO, NO 2 , SO 2 , and O 3 were estimated with RMSE of 0.069, 1.44, 1.17, and 0.26, respectively, for fall 2019. Finally, for winter 2019, the seasonal CO, NO 2 , SO 2 , and O 3 were calculated with RMSE of 0.055, 5.23, 0.89, and 1.45, respectively, as shown in Table 9.

General Discussion
From a methodological standpoint, the AP can be computed using two methods: ground-based methods and remote sensing technology. Ground-based methods are the main approach to AP retrieval. These methods, however, are time-consuming and expensive. Additionally, they lack frequent records in dynamic environments, such as cities. Over the last decades, satellite-based models have been employed for AP retrieval. The satellite-based technology is considered an innovative technique that can estimate and monitor AP at dense spatial sampling intervals and large scales. The efficiency of satellitebased data and methods such as the Terra Moderate Resolution Imaging Spectroradiometer (MODIS), the Sentinel-5 Precursor (Sentinel-5P), Global Precipitation Measurement (GPM), Soil Moisture Active and Passive (SMAP), the National Centers for Environmental Prediction (NCEP), Climate Forecast System Reanalysis (CFSR), and the Global Land Data Assimilation System (GLDAS) have proven useful for AP retrieval . This study used Sentinel-5 based on GEE to retrieve monthly, annual and seasonal CO, NO 2 , SO 2 , and O 3 from 2018 to 2019. The results showed that satellite-derived data and applied methods performed well for AP retrieval (Tables 8 and 9). The results of this study revealed that remote sensing technology would make AP mapping and monitoring fast and easier in dynamic areas, such as cities. Our findings also showed a strong correlation coefficient between obtained values from pollution measuring stations and Sentinel-5 (Figures 9-12). According to Figures 5 and 6   According to Figure 9, the correlation coefficients (R 2 ) for CO, NO 2 , SO 2 , and O 3 in spring 2018 were 0.9509, 0.5692, 0.9285, and 0.5627, respectively. The correlation coefficients between the data obtained from Sentinel-5 and the ground data for summer 2018 were 0.5544, 0.7463, 0.9221, and 0.8601 for CO, NO 2 , SO 2 , and O 3 , respectively, as shown in Figure 9. According to Figure 9, strong correlation coefficients of 0.970, 0.4627, 0.9307, and 0.7841 were found for CO, NO 2 , SO 2 , and O 3 , respectively, in fall 2018. In addition, the correlation coefficients between the data obtained from Sentinel-5 and the ground data for winter 2018 were 0.7117, 0.727, 0.9506, and 0.8748 for CO, NO 2 , SO 2 , and O 3 , respectively, as shown in Figure 9. In general, strong correlation coefficients were estimated between data obtained from remote sensing technology and ground-based data in all seasons, in particular winter for the year 2018. The results of this research are in accordance with reports received from the Environmental Organization of Central Province (https://markazi.doe.ir, accessed on 1 February 2022).
According to Figure 10, the correlation coefficients for CO, NO 2 , SO 2 , and O 3 in spring 2019 were 0.8974, 0.6174, 0.8959, and 0.5192, respectively. The correlation coefficients between the data obtained from Sentinel-5 data and the ground data for summer 2019 were 0.92, 0.855, 0.8391, and 0.8298 for CO, NO 2 , SO 2 , and O 3 , respectively, as shown in Figure 10. According to Figure 10, strong correlation coefficients of 0.917, 0.6275, 0.9326, and 0.8197 were found for CO, NO 2 , SO 2 , and O 3 , respectively, in fall 2019. In addition, the correlation coefficients between the data obtained from Sentinel-5 and the ground data for winter 2019 were 0.9009, 0.9471, 0.8933, and 0.8332 for CO, NO 2 , SO 2 , and O 3 , respectively, as shown in Figure 10. In sum, the results of the correlation coefficients between the predicted data from remote sensing technology and actual data received from pollution measuring stations demonstrated a difference (about <0.6) from 2018 to 2019 in all seasons. This difference could be because of some missing data received from pollution measuring stations.
From Figure 11, it can be seen that strong correlation coefficients of 0.9509, 0.9344, 0.9344, and 0.8376 for CO, NO 2 , SO 2 , and O 3 , respectively, were obtained in 2018.

Monthly Distribution of AP in 2018 and 2019
Detailed zoning maps of Arak in 2018 showed that the most polluted areas were in the city's center (Areas 4 and 5), while the least polluted area was 3, which is in accordance with reports received from the Environmental Organization of Central Province (Figure 3). Using satellite images, it was found that the least pollution was recorded in area 3 due to the presence of vegetation and the surrounding gardens. However, the highest distribution of pollution occurred in area 1, which was due to the location of all small and large industries of Arak city in this area, in accordance with collected data from air pollution measuring stations. According to Figure 3, the highest concentration of CO occurred in July and August (0.029 ppm) for the year 2018. The highest concentration of NO 2 was also estimated for December (22.06 ppm). The highest concentration of SO 2 occurred in August (33.54 ppm). The highest value for O 3 was estimated in February, as shown in Figure 3.
The pollution zoning maps for Arak city in 2019 showed the highest levels of pollution in areas 1 and 2, and the least levels of pollution in area 3 ( Figure 4). Based on the distribution of pollution using satellite images, it was found that area 3 had the least pollution. This was due to the excellent vegetation and the surrounding gardens. However, area 1 had the highest pollution, which was due to the location of all small and large industries of Arak city in this area, which is in accordance with data received from the Environmental Organization of Central Province. As we can see in Figure 4, the highest concentration of CO occurred in January (0.030 ppm) for the year 2019. The highest concentration of NO 2 was also estimated for January (30.19 ppm), while the highest concentration of SO 2 belonged to October (45.38 ppm). The highest value for O 3 was estimated in April (0.147 ppm), as shown in Figure 4.

The Effect of Land Use on Air Pollution in Arak City
Based on an analysis of land use and its effect on air pollution, it was found that the most polluted area was the Arak industrial town, which is located in area 1. It has been reported that the industrial factories of Arak play a significant role in the air pollution of Arak city by producing suspended particles, carbonaceous oxides, nitrogenous oxides, ammonia, etc. [41]. The least air pollution was found in area 3 due to the high density of vegetation, which is in accordance with the identified green areas on the land use map.
Arak is surrounded by hills and the city's highlands are located to the south and the city's industrial area is to the southeast. The prevailing winds in this area are from the west and southwest. Regardless of other natural factors, due to the average speed (between 7 and 10 K/H) of the prevailing winds, pollution from the industries in the east of this city cannot penetrate into the residential areas [51]. Due to the topography, the south and west of Arak city are surrounded by highlands, and, since local winds are typically from the east and northeast, this factor, as well as the fact that January and April are the months with the greatest percentage of still air, contributed to creating an inversion, which is in accordance with reports received from the Environmental Organization of Central Province.

Limitation of the Study
Although satellite images are a promising source of data for generating estimates at high spatial resolution on a local scale as they capture some spatial variability, they are limited to generalizing in certain areas [52,53]. Future work should focus on combining additional datasets that are readily available globally (e.g., additional bands of satellite data, normalized difference vegetation index (NDVI)/enhanced vegetation index (EVI), meteorology, digital elevation model (DEM)) that can be combined with meter-scale satellite images to generate better estimates of air pollutants. In addition, future efforts will need to assess the sensitivity of image-based models to images collected with different temporal aspects, such as time of day and season.

Conclusions
Recent progress in Earth observation technology, and remote sensing in particular, has turned remote sensing into big data technology, which demands efficient, effective, and cost-effective data-driven methods. Therefore, applying different data-driven approaches and comparing their efficiency can be considered as the state of the art for remote sensing sciences which is the object of the current research. This study employed Sentinel-5 images based on GEE to retrieve CO, NO 2 , SO 2 , and O 3 parameters. According to our findings, Sentinel-5 images based on GEE turned out to be the most efficient approach for AP retrieval. This study also confirmed a strong correlation between CO, NO 2 , SO 2 , and O 3 retrieved from Sentinel-5 and air pollution stations.
Our findings confirm that GEE is appropriate to exploit vast amounts of data, and that it can be regarded as a testbed for machine learning algorithms. In the investigation of air pollution in urban areas, satellite images in both time and space can provide optimal management and high accuracy. Considering the fact that most pollution monitoring station data are incomplete due to malfunctioning devices, AP classification from remote sensing images is a challenging task because of the wide range of features that can cause heterogeneity. The present study addresses this complexity by utilizing an automated datadriven platform. This study proves the applicability of Sentinel-5 in GEE for providing a general framework for the monitoring of CO, NO 2 , SO 2 , and O 3 at various levels and scales. The results of this study contribute to monitoring programs for CO, NO 2 , SO 2 , and O 3 changes in dynamic environments, such as cities. The results are readily generalizable to more complex Earth feature monitoring.