Global-Scale Patterns and Trends in Tropospheric NO 2 Concentrations, 2005–2018

: Nitrogen dioxide (NO 2 ) is an important air pollutant with both environmental and epidemiological e ﬀ ects. The main aim of this study is to analyze spatial patterns and temporal trends in tropospheric NO 2 concentrations globally using data from the satellite-based Ozone Monitoring Instrument (OMI). Additional aims are to compare the satellite data with ground-based observations, and to ﬁnd the timing and magnitude of greatest breakpoints in tropospheric NO 2 concentrations for the time period 2005–2018. The OMI NO 2 concentrations showed strong relationships with the ground-based observations, and inter-annual patterns were especially well reproduced. Eastern USA, Western Europe, India, China and Japan were identiﬁed as hotspot areas with high concentrations of NO 2 . The global average trend indicated slightly increasing NO 2 concentrations (0.004 × 10 15 molecules cm − 2 y − 1 ) in 2005–2018. The contribution of di ﬀ erent regions to this global trend showed substantial regional di ﬀ erences. Negative trends were observed for most of Eastern USA, Western Europe, Japan and for parts of China, whereas strong, positive trends were seen in India, parts of China and in the Middle East. The years 2005 and 2007 had the highest occurrence of negative breakpoints, but the trends thereafter in general reversed, and the highest tropospheric NO 2 concentrations were observed for the years 2017–2018. This indicates that the anthropogenic contribution to air pollution is still a major issue and that further actions are necessary to reduce this contribution, having a substantial impact on human and environmental health.


Introduction
Air pollution is one of the main threats to human health, ecosystems and climate on a global scale [1,2]. The global population is growing substantially, and more than half of the world's population now live in urban areas. Large urban areas and high population densities are hotspots for air pollution [1,3]. According to the World Health Organization (WHO), about 3 million people die annually due to ambient air pollution, mainly in low-and middle-income countries, and about 90% of the world's population are exposed to air that exceeds the WHO air quality guidelines [4].
Nitrogen dioxide (NO 2 ) is one of the most important air pollutants in the atmosphere [5] and linked to a number of both environmental and epidemiological effects [2,6]. It is formed in processes where nitrogen reacts with oxygen in high temperatures, e.g., through lightning and the combustion of fuels [7]. The main anthropogenic sources of NO 2 emissions are transport, industry processes and energy production [8]. Some of the main environmental effects linked to high NO 2 concentrations are acidification, eutrophication and photochemical formation of ozone (O 3 ) [6,7,9]. NO 2 also modifies the radiative balance in the atmosphere and influences the atmospheric lifetime of greenhouse gases [10,11]. NO 2 is toxic at high concentrations, and the epidemiological effects include respiratory illnesses such as lung cancer, asthma exacerbations and cardiopulmonary mortality [2,5,7,12]. NO 2 has a short atmospheric lifetime, on average 3.8 ± 1.0 h (mean ± 1 standard deviation) [8] as it reacts with sunlight, which triggers the production of hydroxyl radical OH [13]. Therefore, high concentrations of tropospheric NO 2 are mainly confined to its emission sources, which in general are urban and industrialized areas [2,5].
Monitoring of NO 2 concentrations can be done with ground-based monitoring stations. However, monitoring stations tend to be clustered in city centers, have a small spatial coverage and are often lacking in developing countries [2,14]. Ground-based air quality monitoring is thereby unevenly distributed, and large areas are under-represented [14,15]. An alternative approach to monitor air pollution is the usage of remotely sensed satellite data that increase the spatial coverage. Major advances have been made over the past decades to use satellite sensors to monitor atmospheric pollutants [1]. Satellite monitoring of NO 2 started in 1995 with the Global Ozone Monitoring Experiment (GOME) instrument [3]. Since then, other satellite instruments have been used to monitor tropospheric NO 2 , such as GOME-2, the SCanning Imaging Absorption spectroMeter for Atmospheric CHartographY (SCIAMACHY), the Ozone Monitoring Instrument (OMI) and the recent TROPOspheric Monitoring Instrument (TROPOMI) aboard Sentinel-5 Precursor. Out of these instruments, OMI offers the longest continuous monitoring record (ongoing since 2004) and has a relatively high spatial resolution (13 × 24 km 2 at nadir) [6,7]. Potential errors in estimating NO 2 concentrations from satellite data include uncertainties in surface albedo, aerosols, cloud parameters, slant column density and air mass factor calculations [2,6,16]. Therefore, for satellite-based products to be trustworthy, the data need to be compared against other observations of NO 2 concentrations, such as from ground-based monitoring stations [17].
Studies of long-term trends in air pollution provide information about likely changes and distribution patterns that are useful for assessing the effects of emission mitigation efforts [18][19][20]. Such studies investigating NO 2 trends using OMI data and validating derived results against ground-based measurements have been performed previously. For instance, there are studies on NO 2 trends over USA [2,15,21], over China [22], Russia [23], in eight European cities [1] and in cities around the globe [24]. These studies have reported declining NO 2 trends in their respective study areas and relationship between OMI and ground-based measurements with correlation coefficients ranging between 0.3 and 0.93. NO 2 trend studies on a global scale have also been performed previously using various satellite sensors, but these studies have overall found both negative and positive trends [3,5,19,25,26].
For trend analysis, one of the most widely used methods is the ordinary least-squares (OLS) linear regression, as performed in most of the above-mentioned studies. These simple linear models only provide partial insights on the mechanism essential for an appropriate attribution of drivers of changes. Actual changes can abruptly occur caused by climatic extreme events, anthropogenic mitigations efforts or changes in contributing factors to air pollution. These changes may only be visible for a short period in time, despite having long-lasting effects, and will therefore remain undetected using such traditional linear trend models [27][28][29].
Recent advances in time-series and breakpoint analysis open new possibilities for studying tropospheric NO 2 concentrations observed by Earth observation satellites, as they allow for the detection of nonlinear trends and turning points in the concentrations. Nonlinear trend models (e.g., PolyTrend) can separate trends into linear and nonlinear trend types [30]. Piecewise linear models, such as Break For Additive Season and Trend (BFAST) [27] or Detecting Breakpoints and Estimating Segments in Trend (DBEST) [29], allow for separating time-series into individual segments, capturing dynamics in specific explanatory variables [28,[31][32][33]. By using these methods, dynamics in Remote Sens. 2020, 12, 3526 3 of 18 tropospheric NO 2 concentrations may be better characterized by capturing specific atmospheric conditions and stages of pollution development through time.
Hence, the main aim of this paper is to analyze global and regional patterns and trends in tropospheric NO 2 concentrations using a continuous time-series of tropospheric NO 2 concentrations from the OMI instrument from 2005 to 2018 with novel methods within time-series and breakpoint analyses. Specifically, we aim at (1) comparing the OMI data against NO 2 concentrations from ground-based monitoring stations, (2) analyzing spatial patterns and temporal (nonlinear) trends, (3) investigating whether regional differences can be found in global NO 2 concentrations and (4) spatially explicitly detecting major breakpoints in NO 2 concentrations and estimating their timing and magnitude at global scale.

Satellite-Based NO 2 Dataset
Aura is one of the National Aeronautics and Space Administration's (NASA) Earth Observing System (EOS) satellites. It was launched in 2004 with the mission to collect data of global air pollution and to monitor the chemistry and dynamics of Earth's atmosphere on a daily basis [34]. Aboard Aura there are four instruments, one of which is OMI [34,35]. OMI is a nadir-looking push broom hyperspectral imaging spectrometer that measures reflected solar radiation in the ultraviolet and the visible light (UV/VIS) channels of the electromagnetic spectrum (wavelength range of 264-504 nm) with a spectral resolution of 0.42-0.63 nm [36,37].
We used the OMI/Aura level 3 NO 2 (OMNO2d) standard product (the cloud screened subset 4) downloaded from NASA's Earth Observation data collection [38]. The OMNO2d product contains composites of daily total tropospheric column NO 2 data with a spatial resolution of 0.25 • × 0.25 • . In this study, we used OMI data from 1 January 2005 until 31 December 2018 (in total, 5092 daily OMNO2d files considering 21 gaps in the daily data files). We also excluded all pixels with less than 50 days of data per year, in order to minimize influences of errors in the retrieval process.

Ground-Based NO 2 Dataset
The ground-based data are annual averages (n = 6093) of daily observations of atmospheric NO 2 concentrations (n = 1,706,830) from monitoring stations in the USA between the years 2005 to 2018, provided by the United States Environmental Protection Agency (US EPA) [39]. The reference method used by the US EPA for collection of ambient NO 2 is chemiluminescence analysis [40] based on the reaction of nitric oxide (NO) with ozone (O 3 ). The principle of the method is that a sample of ambient gas enters a reaction chamber where NO molecules react with O 3 to form NO 2 . The reaction produces a quantity of light, a phenomenon known as chemiluminescence. The intensity of the light, which is proportional to the concentration of NO 2 , is then measured to determine the concentration of NO 2 [40,41].

Comparison against Ground Observations
The daily OMI NO 2 data were first averaged monthly, and thereafter annually. Annual averages were used since this study focuses on long-term trends, and it is therefore the inter-annual variability that must be validated. The annual averages were then compared to corresponding ground-based NO 2 data in order to verify the validity of OMI NO 2 product. Since the two datasets use different units (10 15 molecules cm −2 for the satellite-based data and part per billion (ppb) for the ground-based data), we calculated z-scores using the z-statistic ((data value − average)/standard deviation) for both datasets. The relationships between the two datasets were quantified using the root-mean-square error (RMSE), and by goodness-of-fit when fitting the ordinary least-squares linear regression on the z-scores for the two datasets.

Analysis of Spatial Patterns and Temporal Trends
The spatial patterns were analyzed by averaging all OMI NO 2 data pixel-wise over the study period. For analyzing the temporal trends, time-series of annual mean NO 2 concentration were first calculated. Then we applied PolyTrend to analyze and classify trends in the annual NO 2 time-series 2005-2018. We also applied the DBEST program to detect the greatest significant breakpoints in the annual NO 2 time-series and estimate their timing and magnitude. The PolyTrend and DBEST analyses were both performed at pixel level having a statistical significance threshold (α) of 0.05. Pixels with an absolute value of annual average tropospheric NO 2 concentration below the OMI detection limit (0.5 × 10 15 molec.cm −2 ) [42] were excluded from the trend analyses.

Nonlinear Trend Analysis with PolyTrend
PolyTrend is an automated method with an algorithm that accounts for nonlinear change in a trend [30]. It uses a polynomial fitting-based scheme that divides trends into linear and nonlinear trend behaviors and then subdivides the nonlinear trends into classes of cubic, quadratic, and concealed trend types. The linear trend type means that the trend line has a uniform direction over the study period (either increasing or decreasing). The quadratic trend type is a trend line with one bend in its curve, implying that the cell has experienced one direction-change in its trend line over the study period (i.e., first positive and then negative trend, or vice versa). The cubic trend type means that the trend line has two bends, implying that corresponding cell has experienced more than one change in the trend direction over the study period (i.e., first decreasing followed by increasing and then again decreasing change, or vice versa). The concealed trend type consists of cells with either cubic or quadratic trend types, but with no significant net change in tropospheric NO 2 concentrations over the study period. We refer to Jamali et al. [30] for more details.

Breakpoint Analysis with DBEST
DBEST was developed for analyzing time-series of satellite sensor data, and it uses a segmentation method for two main algorithms of trend generalization and change detection [29]. We used its change detection algorithm in order to detect breakpoints with greatest change in tropospheric NO 2 concentrations. Our input data in DBEST were the pixel-wise time-series of the annual average NO 2 concentrations data.
First, DBEST tests for the occurrence of discontinuities, in this case of tropospheric NO 2 concentrations, by analyzing the absolute differences between consecutive data points and comparing this to the first level-shift-threshold set by the user (Table 1). If the difference is greater than the first level-shift-threshold, then it tests whether or not this difference caused a significant shift in the mean level of tropospheric NO 2 concentrations and persisted over the duration-threshold. If the mean level before and after this identified discontinuity is greater than the second level-shift-threshold, DBEST considers this a level-shift point. DBEST then repeats this process for all data points, sorts them into descending order based on the absolute value of tropospheric NO 2 concentrations difference, and tests if the spacing between a data point and an identified level-shift point is at least the duration-threshold. The trend component of the time-series is then segmented using a peak/valley detector function and a method that draws a straight line through detected peak/valley points and compares perpendicular distances to the non-peak and non-valley points between them with the distance-threshold parameter. If the distance is greater than this threshold, these points are added to the set of detected peak/valley points and level-shift points, all of which are called turning points. Detected turning points are then fit to the tropospheric NO 2 concentrations trend using piecewise linear modelling, and those turning points that minimize the Bayesian Information Criterion (BIC) [43] are considered breakpoints. Here, we used the change detection algorithm of DBEST with a set value (2) for the number of significant breakpoints of interest for detection (Table 1), and as such, DBEST identifies a final set of greatest significant breakpoints as requested by user. The results of Remote Sens. 2020, 12, 3526 5 of 18 the change detection algorithm include the starting time of the breakpoints (break date); the change duration, or the temporal period over which this change occurred; the change value, or the amount of change that occurred over this time period; the change type, whether the change is abrupt (level-shift) or non-abrupt; the change significance, based on the statistical significance level (α = 0.05). Here, the annual average tropospheric NO 2 concentrations time-series data were set as non-cyclical type ( Table 1). The first level-shift-threshold was set to 0.1 × 10 15 molecules cm −2 and the second level-shift-threshold to 0.5 × 10 15 molecules cm-2 . It is recommended that the first level-shift-threshold be set to a smaller value than that for the second level-shift-threshold [29]. Therefore, if a detected change was quick (between two consecutive observations/years) and large enough (0.1 × 10 15 molecules cm −2 ) to shift the mean over the user-set duration (2 years) by 0.5 × 10 15 molecules cm −2 before and after the point, it was characterized as an abrupt change, otherwise it was considered a non-abrupt change, provided that it was a significant breakpoint. The distance-threshold is normally set to be a default that is derived internally by DBEST.

Data Comparison against Ground Observations
The comparison of OMI data against ground-based observations showed a strong relationship (Pearson's correlation coefficient R = 0.65) that was statistically significant (p-value < 0.01) ( Figure 1). The relationship was equally strong (R = 0.65) when separating the analysis into a comparison of how well OMI captured the spatial variability (data averaged site-wise; Figure 1b). The OMI data were most successful at reproducing the inter-annual variability (data averaged annually), for which the observations were in a very close relationship with the ground-based observations (R = 0.99) ( Figure 1c). The ordinary least-squares linear trend in annual averages of the z-scores in OMI NO 2 concentrations (−0.220 ± 0.027 z-scores y −1 ; R 2 = 0.85) was very similar to the corresponding trend in the ground-based NO 2 concentrations (−0.218 ± 0.022 z-scores y −1 ; R 2 = 0.83).

Spatial Patterns
There is a distinct difference in the NO2 concentration distribution between the northern and southern hemispheres, where the higher concentrations are almost exclusively found in the northern hemisphere ( Figure 2a Table 2). The maximum NO2 concentration was for China (28.24 × 10 15 molecules cm −2 ) followed by Japan (14.28 × 10 15 molecules cm −2 ), Italy (11.84 × 10 15 molecules cm −2 ),

Spatial Patterns
There is a distinct difference in the NO 2 concentration distribution between the northern and southern hemispheres, where the higher concentrations are almost exclusively found in the northern hemisphere  Table 2). The maximum NO 2 concentration was for China (28.24 × 10 15 molecules cm −2 ) followed by Japan (14.28 × 10 15 molecules cm −2 ), Italy (11.84 × 10 15 molecules cm −2 ), Germany (11.34 × 10 15 molecules cm −2 ), USA (11.25 × 10 15 molecules cm −2 ) and India (9.22 × 10 15 molecules cm −2 ). Due to their high concentrations in tropospheric NO 2 , we selected these areas as focus areas used for further analysis in the remaining part of the study.
Remote Sens. 2020, 12, x FOR PEER REVIEW 7 of 18 Germany (11.34 × 10 15 molecules cm −2 ), USA (11.25 × 10 15 molecules cm −2 ) and India (9.22 × 10 15 molecules cm −2 ). Due to their high concentrations in tropospheric NO2, we selected these areas as focus areas used for further analysis in the remaining part of the study.

Temporal Trends
Significant trends in NO2 concentrations were observed largely over land and to a much lower degree over oceans along boundaries with lands ( Figure 3). With the insignificant no-trends masked out, 79.55% of the remaining cells had positive trend whereas 20.45% had negative trend. The increasing trends were distributed over large parts of land, but the decreasing trends were generally observed over USA (Figure 3a), Western Europe (Figure 3b), Japan and the eastern parts of China ( Figure 3c). The global average trend in 2005-2018 was slightly increasing (0.004 × 10 15 molecules cm −2 y −1 ); however, the regional negative trends were strong enough to compensate for the global rising trend of NO2 concentrations over larger areas. Globally, the strongest negative trend was −0.969 × 10 15 molecules cm −2 y −1 while the strongest positive trend was only 0.363 × 10 15 molecules cm −2 y −1 ( Table  2).

Temporal Trends
Significant trends in NO 2 concentrations were observed largely over land and to a much lower degree over oceans along boundaries with lands ( Figure 3). With the insignificant no-trends masked out, 79.55% of the remaining cells had positive trend whereas 20.45% had negative trend. The increasing trends were distributed over large parts of land, but the decreasing trends were generally observed over USA (Figure 3a), Western Europe (Figure 3b), Japan and the eastern parts of China (Figure 3c). The global average trend in 2005-2018 was slightly increasing (0.004 × 10 15 molecules cm −2 y −1 ); however, the regional negative trends were strong enough to compensate for the global rising trend of NO 2 concentrations over larger areas. Globally, the strongest negative trend was −0.969 × 10 15 molecules cm −2 y −1 while the strongest positive trend was only 0.363 × 10 15 molecules cm −2 y −1 ( Table 2).
Areas with high average NO 2 concentrations, except India and western parts of China (Figure 2), generally showed negative trends ( Figure 3; Table 2).
On average, the strongest negative trends were found in Europe (Belgium: was for China, the average trend for the entire country was just slightly increasing (0.014 × 10 15 molecules cm −2 y −1 ) because strong increasing trends (0.363 × 10 15 molecules cm −2 y −1 ) were observed over large parts of the country as well (Figure 3c).

Trend Types
In a global context, the linear trend was the dominant trend type with a spatial coverage of 61.98%, out of which 54.47% were positive and 7.51% negative (Figure 4a, Table 3). The concealed trend was the second trend type with 21.89% spatial coverage and mainly found over east of China and Southwestern Europe (Figure 4c,d). For the remaining trends, 9.77% were quadratic and 6.36% were cubic, out of which the majority was found over the eastern parts of USA ( Figure 4a) and west of Europe (Spain and Portugal) (Figure 4b,c). In the focus areas, the dominant trend type was different for different areas. In the USA, the nonlinear trends (67.59%) were spatially more than the linear trends (32.41%) (Figure 4a, Table 3). In the focus areas in Europe, the most common trend type was linear (negative), except for Spain where the nonlinear trends, particularly the quadratic negative trends (57.96%), were dominant ( Figure 4c, Table 3). The most common trend type over India was linear (increasing) (84.36%), and over Japan was linear (decreasing) (43.03%). China was the country with the largest proportion of nonlinear concealed trends in NO 2 concentration (45.81%); it was also the second country with the highest proportion of linearly increasing trends (39.19%) after India (84.36%) (Figure 4c,d, Table 3).

Trend Types
In a global context, the linear trend was the dominant trend type with a spatial coverage of 61.98%, out of which 54.47% were positive and 7.51% negative (Figure 4a, Table 3). The concealed trend was the second trend type with 21.89% spatial coverage and mainly found over east of China and Southwestern Europe (Figure 4c,d). For the remaining trends, 9.77% were quadratic and 6.36% were cubic, out of which the majority was found over the eastern parts of USA ( Figure 4a) and west of Europe (Spain and Portugal) (Figure 4b,c). In the focus areas, the dominant trend type was different for different areas. In the USA, the nonlinear trends (67.59%) were spatially more than the linear trends (32.41%) (Figure 4a, Table 3). In the focus areas in Europe, the most common trend type was linear (negative), except for Spain where the nonlinear trends, particularly the quadratic negative trends (57.96%), were dominant (Figure 4c, Table 3). The most common trend type over India was linear (increasing) (84.36%), and over Japan was linear (decreasing) (43.03%). China was the country with the largest proportion of nonlinear concealed trends in NO2 concentration (45.81%); it was also the second country with the highest proportion of linearly increasing trends (39.19%) after India (84.36%) (Figure 4c,d, Table 3).

Breakpoints in Tropospheric NO2 Concentrations
The global tropospheric NO2 concentrations showed a slightly decreasing trend from 2005 to 2008, followed by a small, positive change (0.03 × 10 15 Figure 6a shows the greatest breakpoint change detected in the annual average NO2 concentrations at pixel level. The spatial patterns of the detected short-term changes were similar to the long-term overall trends observed over lands (Figure 3): positive breakpoints were found over large areas in all continents (79.4%) and negative breakpoints mainly over the focus areas (20.6%). The greatest negative drop was for China (−12.41 × 10 15 molecules cm −2 ), followed by USA (−5.60 ×  Figure 6a shows the greatest breakpoint change detected in the annual average NO 2 concentrations at pixel level. The spatial patterns of the detected short-term changes were similar to the long-term overall trends observed over lands (Figure 3): positive breakpoints were found over large areas in all continents (79.4%) and negative breakpoints mainly over the focus areas (20.6%). The greatest negative drop was for China (−12.41 × 10 15 molecules cm −2 ), followed by USA (−5.60 × 10 15 molecules cm −2 ), Italy (−3.81 × 10 15 molecules cm −2 ) and Japan (−3.78 × 10 15 molecules cm −2 ) and then the other focus countries in Europe (Figure 6a, Table 4). The greatest positive change was also for China (6.65 × 10 15 molecules cm −2 ) followed by India (2.13 × 10 15 molecules cm −2 ). Range of the change values was therefore the highest for China (19.06 × 10 15 molecules cm −2 ) and the least for Netherlands (1.59 × 10 15 molecules cm −2 ) and Belgium (1.75 × 10 15 molecules cm −2 ), where the average changes were high and no positive change was detected at all ( Table 4). The type of majority of the detected greatest changes was non-abrupt, indicating that most of the changes occurred gradually over time, except for Belgium where the changes mainly happened abruptly (56.86%).

Discussion
The relationship between the satellite-based and the ground-based datasets supports previous OMI validation studies. For instance, the Pearson's correlation coefficient R was 0.65, which is within the middle of the range (0.40-0.80) of several other studies [1,2,15,22]. The statistical comparison further indicated that OMI was more successful at estimating the temporal component than the spatial component (Figure 1b,c). This can partially be explained since the ground-based monitoring stations are focused on a certain emission source (e.g., traffic locations), whereas an OMI pixel (13 × 24 km 2 ) covers a larger area with potential emission sources both within and downwind from the pixel [1]. The strong relationships with the ground-based observations still indicate that OMI data are useful giving spatially explicit time-series of tropospheric NO2 concentrations to study global patterns and trends.
The spatial distribution of average NO2 concentrations found in this study ( Figure 2) resembles those in other studies [5,19,25,26], confirming that the focus areas are indeed the main hotspots of tropospheric NO2 concentrations globally. According to Krotkov et al. [25], the highest NO2 concentrations coincide with urban areas with high populations and industrialized regions. NO2

Discussion
The relationship between the satellite-based and the ground-based datasets supports previous OMI validation studies. For instance, the Pearson's correlation coefficient R was 0.65, which is within the middle of the range (0.40-0.80) of several other studies [1,2,15,22]. The statistical comparison further indicated that OMI was more successful at estimating the temporal component than the spatial component (Figure 1b,c). This can partially be explained since the ground-based monitoring stations are focused on a certain emission source (e.g., traffic locations), whereas an OMI pixel (13 × 24 km 2 ) covers a larger area with potential emission sources both within and downwind from the pixel [1]. The strong relationships with the ground-based observations still indicate that OMI data are useful giving spatially explicit time-series of tropospheric NO 2 concentrations to study global patterns and trends.
The spatial distribution of average NO 2 concentrations found in this study ( Figure 2) resembles those in other studies [5,19,25,26], confirming that the focus areas are indeed the main hotspots of tropospheric NO 2 concentrations globally. According to Krotkov et al. [25], the highest NO 2 concentrations coincide with urban areas with high populations and industrialized regions. NO 2 concentrations are generally much lower over oceans than that over land since there are no sources of NO 2 emissions except for passing ships [44]. This indicates that the trends observed along offshore boundaries are possibly caused by atmospheric deposition of NO 2 transported from their source by large-scale circulation [45]. According to Peters et al. [44], satellite instruments have issues with detecting trace gases over oceans because of the low NO 2 concentrations often being below the detection limit of the instruments (0.5 × 10 15 molecules cm −2 ).
The global and regional trends seen ( Figure 3) generally agree with the results from previous studies. Previous studies have shown increasing trends over both India and China [5,19,25,26], where our results show increasing trends over both countries too (Figure 3). The decreasing trend with major drop in NO 2 that we observed over Eastern USA confirms the previous study by Krotkov et al. [46] reporting a dramatic decrease in OMI NO 2 from 2005 to 2015, as a result of both technological improvements and stricter regulation of emissions. In agreement with our trend results derived for Western Europe, recently Wang et al. [47] observed decreasing trends over Netherlands, Belgium, Germany and Italy, as detected in OMI NO 2 concentrations for 2012-2018. The trend results seem to be consistent among studies with data used from different satellite instruments and/or study periods [5,19,25,26].
Decreases of NO 2 concentrations can primarily be attributed to either local-, regionalor country-level environmental regulations, improvements in emission control technology (e.g., power plants and vehicles), or economic changes and the associated effects in energy usage [24,25]. Since the spatial distribution of average concentrations and significant decreasing trends correlate well, this indicates that environmental regulations and technological improvements in the countries with the most severe pollution have had a positive effect on concentrations of NO 2 . However, it should also be noted that the two final years of this study period (2017-2018) were the years with the highest average global concentrations. This clearly shows the importance of continuous satellite-based monitoring of global patterns and trends in NO 2 concentrations, also for assessing the effects of regional environmental regulations and technological improvements to reduce emissions [48].
Linear regression models assume that changes occur linearly and gradually, which is not always the case [30,49]. Here, a polynomial fitting-based scheme (PolyTrend) was used to account for nonlinear trends. This polynomial approach thus helps to detect nonlinear trends in time-series that would not be identified by an ordinary least-squares (i.e., linear model) approach. The linear trend type was the dominant trend type globally (Figure 4; Table 3) as well as for Europe (except Spain), India and Japan, indicating monotonic (non-decreasing or non-increasing) trends over these areas. The nonlinear trends with a significant slope (quadratic and cubic) were mainly found over eastern parts of USA and Spain. Since the curve of these trends has one (quadratic) or two (cubic) bends, this indicates that the NO 2 concentration trends in these areas either started with an increase and then decreased or the opposite started with a decrease and then increased (quadratic), or with even more short-term changes in the direction of the trend (cubic). The latter case is in agreement with the regional trend curve for USA: a cubic trend starting with a short-term downward trend, then an upward trend, and then again another downward trend (Figure 5b). The identified areas with the concealed trends, mainly in the eastern parts of China and south of Spain, are new findings that, up to the best of our knowledge, have not been reported yet. The reason is that the OLS method is often used in trend studies, and such nonlinear trends are not detectable when OLS is applied for the entire studied period. If OLS applies here, no significant trend in 2005-2018 is detected. However, the concealed changes are credible patterns of nonlinear changes such as the greatest breakpoint changes we detected in NO 2 concentrations over China.
The majority of the detected significant breakpoints were non-abrupt indicating that the concentrations of NO 2 changed gradually, possibly due to stricter environmental regulations or economic cycles, as opposed to abrupt changes (e.g., in Belgium and Netherlands), which could be due to power plants or industries that have been either opened or shut down suddenly. The years 2005 to 2009 were by far the years with the highest occurrence of negative breakpoints, and regional-scale reductions of tropospheric NO 2 concentrations were also observed for USA, Europe and Japan during these years ( Figures 5-7). It has also previously been pointed out that 2008 was a year of significant reductions in NO 2 emissions (e.g., [21,22,50,51]) due to the start of the great economic recession [50,51]. This was an event, which caused large-scale economic reductions and affected anthropogenic activity globally, which in turn reduced the associated emissions of air pollution from, for example, vehicles, power plants and industries. According to the results of this study, the largest change magnitudes in NO 2 concentrations during 2005-2008 were found in USA and Japan. The European countries appear to have suffered less, based on the changes in tropospheric NO 2 concentrations ( Figure 6, Table 4). The negative breakpoint we found over Eastern China with a four-year duration (2011-2014) is in general agreement with Li et al.'s [52] study of analyzing global change of tropospheric NO 2 from 2012 to 2017 using data from the Ozone Mapping Profiler Suite (OMPS) Nadir Mapper (NM) onboard the Suomi National Polar Partnership (SNPP). They reported a large decline of NO 2 in Eastern China started in 2013 and was almost entirely driven by wintertime decreases, thus indicating a decrease in anthropogenic emissions over the area. Souri et al. [53] in'their study of analyzing long-term trends of OMI NO 2 concentration 2005-2014 over East Asia, also found downward trends in Japan and more developed Chinese cities such as Guangzhou and Beijing, and upward trends in the majority of northern regions of China in 2010-2013. This supports the concealed trend (upward-downward) we observed for China. Another study by Krotkov et al. [46] also showed similar severe declines of NO 2 in Eastern China in 2011-2014 due to an economic shutdown and government efforts to restrain emissions from the power and industrial sectors. Likewise, the steepest increasing trend we observed was over India, and they reported a fast-growing trend from 2005 to 2015 for India's NO 2 level from coal power plants and smelters.
The time-series analysis methods used in this study (PolyTrend and DBEST) benefit from recent developments, as mentioned earlier, but like many other methods they also have weaknesses. They work on a pixel-by-pixel basis, and they consider each pixel's time-series data as an isolated entity in their trend classification and change detection procedure; the spatial behavior of adjacent areas is not used to improve the robustness of trend/change detection [54]. Thus, the obtained trend and breakpoint results should be interpreted with caution.
Future research could include multiple breakpoint detection analyses using data for pre-and post-pandemic phases of COVID-19 to study impacts of possible changes in anthropogenic sources of NO 2 emissions (e.g., transport, industry processes and energy production) on air pollution and tropospheric NO 2 concentration trends.

Conclusions
This study contributes to the ongoing research regarding spatiotemporal patterns and trends in tropospheric NO 2 concentrations using data from the OMI instrument, and it investigates how the tropospheric concentrations have changed globally and regionally over the period of 2005 through 2018. By applying novel techniques for analysis of time-series and their breakpoints, we quantified long-term nonlinear trends and provided information about distribution patterns in the point in time with the greatest changes.