This non-parametric method based on rank (Gilbert, 1987; Sirois, 1998) is the most appropriate test to compute optical property trends because it can be applied regardless of missing values, statistical distribution and presence of negatives or below detection limit values in the dataset. The MK test determines whether a monotonic increasing or decreasing long-term trend exists; the slope and the confidence limits are then computed by Sen's slope estimator that is based on the median of the slopes calculated from all possible data pairs. For this study, the MK test was applied on daily medians.

The MK test is designed for serially independent data and is, consequently, influenced by autocorrelation in the time series, leading to inflated type 1 error; that is, there is increased probability of rejecting the no-trend hypothesis (i.e., a false positive). Several correction schemes for the MK test were proposed to correctly handle autocorrelated datasets, and the problems induced by autocorrelation and its various corrections have been clearly described (Wang and Swail, 2001; Yue et al., 2002; Zhang and Zwiers, 2004; Bayazit and Önöz, 2007; Blain, 2013; Wang et al., 2015). A new method (Collaud Coen et al., 2020) has been used for this study that tends to minimize the type 1 and 2 errors (type 2 error is non-rejection of a false null hypothesis, i.e., a false negative) as well as issues with the modification of the slope due to pre-whitening procedures by the application of three pre-whitening (PW) methods. The standard pre-whitening by removing the first lag autocorrelation (von Storch, 1995) has a very low type 1 error but also a low test power, whereas the so-called trend-free pre-whitening procedure published by Yue et al. (2002) (called TFPW-Y in Collaud Coen et al., 2020a) restores the test power at the expense of the type 1 error. Both these pre-whitening procedures were applied prior to the MK test to assess the statistical significance of the trend. A trend was then considered to be ss only if both PW and TFPW-Y were ss at the 95 % confidence level or if PW is ss but not TFPW-Y (false negative). Among the trends of all parameters at all stations calculated for this paper, none was ss for the PW but not for the TFPW-Y, meaning that the PW procedure was always powerful enough. In contrast, many trends were not ss when PW was applied, but were ss with the TFPW-Y procedure, leading to false positives and showing that the TFPW-Y rejection rate of the no-trend hypothesis is too high.

After having determined the statistical significance, a third pre-whitening procedure, the variance-corrected trend-free pre-whitening procedure (VCTFPW) allowing an increase in the slope accuracy (Wang et al., 2015), was applied prior to Sen's slope estimation. The confidence limits of Sen's slope were computed at the 90 % confidence level.

Since many of the time series exhibited clear seasonal cycles, the modified seasonal MK test (Hirsch et al., 1982) was always applied to the four meteorological seasons. The annual trends were considered only if the slopes of the four seasons were homogeneous at the 90 % confidence level (Gilbert, 1987; Sirois, 1998).

Figure 2 presents three examples of seasonal MK results and Sen's slopes of σsp. At JFJ, σsp has ss negative annual trends for all of the analyzed periods, with the most recent 10-year period with a larger negative slope than the longer periods. Spring and fall are the seasons at JFJ with the strongest ss trends; winter has tiny ss negative trends. MRN also exhibits σsp annual negative trends for all of the analyzed periods, but only the 15-, 20- and 25-year trends are ss, and their slopes are more negative for the longest periods. At MRN, summer and fall are the seasons with the largest trends, and that is true for all the trend periods (10–25 years), while spring and winter have more scattered and less significant slopes. Finally, MLO has annual trends that are ss negative for the last 10 years, not ss for the last 15 years, and ss positive for the longest periods (20, 25 and 30 years). The spring season at MLO exhibits a not ss negative trend for the last 10 years and positive trends for the longest periods, with only 25- and 30-year trends being ss.

Following the Weatherhead procedure (Weatherhead et al., 2000), the trend is estimated by fitting the following frequently used statistical model for monthly data with an LMS approximation: 1Yt=m+Ct+ρ⋅(t/12)+Mt,t=1…n, where m is a constant term, Ct is a seasonal component, and ρ is the magnitude of the trend per year. The unexplained noise term Mt is modeled as an [AR(1)] process Mt=ϕ⋅Mt-1+ϵ, where ϕ is the autocorrelation coefficient of the data noise. For this study, either the logarithm of the monthly medians or the monthly medians were taken for all the parameters. Due to the non-normal distribution of the studied parameters, the LMS method applied on the logarithm is considered the standard method according to previous trend analyses (CC2013 and Asmi et al., 2013). A trend is considered to be ss at the 95 % confidence level if |ρ/σp|>2, σp being the standard deviation of the slope. Figure 3a and c show the LMS trends and statistics for MLO σsp, respectively. The LMS results are similar to the MK analysis, the last 10-year trend is negative but ss at only the 90 % confidence level, the 15- and 20-year trends are not ss and the 25- and 30-year trends are ss positive. The normal probability plot of the residue (Fig. 3c) shows that the use of the logarithm of the data results in normally distributed residues as required by this statistical tool.

A similar GLS method based on the minimization of the least square errors similar to ordinary least squares fitting (including similar sensitivity to outliers), but taking into account the autocorrelation in the covariance matrix, was also used in this study. The GLS uses an autoregressive bootstrapping algorithm (ARB) to evaluate the potential differences in the GLS trends arising from the noise terms (Asmi et al., 2013). The ARB methodology was used to produce 1000 realizations of the original time series, with randomized noise terms, and the resulting set of trends was used to determine the 5th to 95th percentile confidence intervals (ARB CLs) of the GLS trends. If the ARB CLs did not include a zero trend, we considered the GLS trend to be ss. The GLS and ARB methodologies were adapted from Mudelsee (2010) and applied to both daily and monthly medians. The previous trend analyses (CC2013 and Asmi et al., 2013) used daily medians.

Figure 3b and f show the GLS/ARB results for MLO σsp for daily and monthly medians. Here again the results are similar to the MK analysis, where the 10-year trend is positive but not ss, the 15-year trend is not ss, while the longer periods exhibit ss positive trends. As with many other stations included in this study, the use of daily or monthly medians did not result in normally distributed residues (Fig. 3f), and, in fact, the residues of the daily and monthly medians appeared to represent different types of distributions. It is also obvious that the seasonality fits (fits from monthly and daily medians in red and orange in Fig. 3b) are different for the two time granularities, with similar shape but higher absolute trend values if fitted from daily medians. The timing of the winter minima is also more precisely defined with the daily data.

Long-term trend analysis of σsp has been performed on 37 datasets. Since some nephelometers only measure σsp (Optec and Radiance Research nephelometers) and σbsp was determined to be unusable for several other sites due to various discontinuities (see Sect. 2.4), the hemispheric backscattering coefficient trends were computed on only 28 datasets. The detailed results of MK trend analyses are given in Table 2, while the overall picture for σsp is presented in Fig. 4. The results for σbsp are very similar to those for σsp for sites where both measurements existed; corresponding figures for σbsp can be found in the Supplement (Figs. S1, S2 and S7).

The σsp ss trends are predominantly negative: 20 stations have ss negative 10-year trends, 5 stations ss positive trends and 12 stations no ss trends dispersed across all continents. Eight (nine) stations with time series longer than 10 years have ss negative 15-year (20-year) trends and none (two) of the 15-year (20-year) trends are ss positive. The MK slopes range between -2.45 and +0.39 Mm-1 yr-1 with a mean of -2.19 Mm-1 yr-1. The main results are as follows.

Over North America, all the σsp trends for periods longer than 10 years are ss negative, and the most recent 10-year trends are generally ss negative. Three stations do not have ss trends. (1) EGB's 9-year time series does not allow for a ss trend (too short), but was included as one of only two Canadian sites. (2) MRN is an IMPROVE station on the western coast of the USA with very high humidity, leading to condensation that can disturb the humidity measurement. This makes it difficult to know whether the ss positive RH 10-year trend (Table S4) is real or due to measurement artifacts and uncertainties. If the ss positive RH trend is real, it could mask a decreasing σsp trend, resulting in a not ss trend. The time coverage for the dry σsp (σsp restricted to RH < 50 %) for MRN is too low to be representative for trend analysis. It should be mentioned that the 10-year trends for MRN ending in 2014–2018 are all not ss (see Sect. 3.2.1), so that the absence of σsp trends seems to be a real phenomenon. (3) GLR is also an IMPROVE station with high humidity. The RH trends at GLR are also not ss, and the dry σsp has a ss negative trend, similar to other stations in its vicinity.

In the previous decadal trend paper (CC2013), the trends in scattering for the arid state of Arizona were not consistent (ss positive: IBB, ss negative: SIA, PAZ, not ss: HGC, SCN). Four of the five Arizona sites (IBB, PAZ, SIA and SCB) were closed in 2010 and HGC now exhibits a ss decreasing scattering trend. MZW, the other IMPROVE station with ss positive scattering trends in 2010, also closed in 2010.

The seasonal MK results for σsp are presented in Fig. 5. Spring is the season with the largest number of ss decreasing trends and winter the season with the lowest. ZEP and PAL exhibit ss positive trends only in summer and BRW has ss negative trends only between December and May. The SPO annual trend is ss positive, whereas it is not for NMY. Both Antarctic stations exhibit, however, a coherent seasonality with ss positive trends only in spring. While the 25- and 30-year trends at MLO are all ss positive, with the largest slope in spring when MLO is influenced by Asian long-range transport (CC2013), the most recent 10–20-year trends are not ss for the individual seasons.

The analysis of σap long-term trends has been performed on 33 datasets (see Fig. 6 and Tables 2 and 3). The long-term trends are ss decreasing (21 stations) or not ss (12 stations) for all stations around the world, leading to a mean decreasing trend of -3.05 Mm-1 yr-1. No ss σap positive trends are measured for any of the stations. The other main results are the following.

In North America the number of σap datasets is much lower than the number of σsp datasets (IMPROVE sites do measure aerosol absorption, but with a different instrumental setup; White et al., 2016). From the five sites with long-term aerosol absorption, APP and BND, two continental rural sites, and the marine Caribbean island (CPR) station have ss negative trends. The other three stations representing the continental rural USA (SGP, EGB) and marine western coast of the USA (THD) exhibit not ss trends in σap.

As described under Sect. 2.4, ω0 trends have to be considered with greater caution since the σap absolute values suffer from a certain uncertainty related to filter-based absorption photometer artifacts.

The ω0 trends depend directly on both the magnitude and the sign of the σsp and σap trends. If expressed in % yr-1, a σap trend larger (smaller) than the σsp trend will result in an increasing (decreasing) ω0 trend, respectively (see Fig. S8 and the related estimation of ω0 uncertainty due to measurement and Cref errors). The ω0 trends are consequently much more diverse than the σsp and σap trends with 52 % of ss positive (relatively more scattering), 22 % of ss negative (relatively more absorption) and 26 % not ss trends (see Fig. 7 and Table 2). One peculiarity is that all ω0 ss negative trends are found between latitudes 30 and 50, but this is perhaps due to the low spatial coverage outside of North America and Europe. The main results are the following.

The ω0 is decreasing at three stations in North America (BND, SGP and THD), whereas APP and CPR exhibit ss positive ω0 trends. The CPR ω0 increasing trend can perhaps be related to increased Saharan dust load. The seasonal ω0 trends at CPR are, however, ss not only in summer when Saharan influence is greatest, but for every season except spring (Fig. 8). EGB has no ss trend.

The present-day trends for the backscatter fraction b are mostly ss positive (65 %) across all regions (Fig. 9 and Table 2). This suggests a shift in the size distribution towards smaller accumulation-mode aerosol. The two stations with ss negative trends are CPR in Puerto Rico and BEO, located on a summit in the Balkan range. Not ss trends are mostly found in eastern and northern Europe (KPS, MEL, BIR and PAL), in Antarctica (NMY), as well as at BND and AMY for the last 10 years. The Arctic sites (ALT, BRW and ZEP) all exhibit ss positive b trends. CPR's seasonal trend is ss negative only in fall; trends in b for the other seasons at CPR are not ss (see Fig. S4). Similarly, the BEO b seasonal trend is ss negative only in summer, and not ss otherwise. PAL has ss positive b trends in spring and summer and ss negative b trends for fall, leading to an annual not ss trend.

The scattering Ångström exponent (åsp) trends exhibit a higher variability than the trends in other parameters, with 33 % of ss positive trends, 37 % of ss negative trends and 30 % of not ss trends (Fig. 10 and Table 2). There are ss positive and negative trends in North America, Europe and the polar regions, and the various trends cannot be attributed to specific regions or environments. It should be recalled, however, that åsp is affected by higher uncertainties (see Sect. 2.3) that may contribute to the larger observed variability. The seasonal results also exhibit high variability, with summer being the season with the least number of ss åsp trends (10 out of 26 sites), while spring and fall are the seasons with the largest number of ss positive and negative trends in åsp (8 out of 26 sites), respectively (see Fig. S5).

The number of stations with long-term åap measurement is low, with only 14 time series available. Seven stations situated in various geographical regimes exhibit ss positive trends (Fig. 11 and Table 2): polar regions (ALT and ZEP), a Caribbean coastal station (CPR), a rural continental North America station (SGP), and high-altitude stations in the remote Pacific (MLO), in continental (WLG) and in coastal (LLN) Asia. SMR and JFJ, two stations in Europe but with very different environmental footprints and altitudes, exhibit ss decreasing åap trends; the six other stations, consisting of three coastal and two continental sites, have no ss trends.

While CPR and SGP åap trends are ss positive and JFJ ss negative for all four seasons, the other stations exhibit higher variability as a function of the meteorological seasons (see Fig. S6). The absorption Ångström exponent is principally a function of the particle chemical composition and material properties, but its assignment to an aerosol type is not uniquely defined and also depends on the particle size, with larger particles corresponding to lower åap values (Liu et al., 2016; Schmeisser et al., 2017). For example, åap>2 corresponds to mineral dust in the case of big particles and to brown carbon in the case of small particles. In contrast, åap<1 corresponds to large particles with small absorption like sea-salt-dominated aerosol in the case of big particles and to black carbon (BC)-dominated aerosol in the case of small particles. Following these observational constraints, the JFJ and SMR aerosol tends to represent the category “mixed BC/BrC” according to Schmeisser et al. (2018). CPR absorption has a strong contribution from mineral dust and sea salt, whereas at MLO, SGP, ALT and ZEP contributions to absorption are from mixed sources, including various light-absorbing carbon species and dust. Ideally, direct chemical composition measurements would provide more precise information on the aerosol type, but the necessary chemical composition measurements are not yet readily available at many sites.

Figure 12 presents the σsp 10-year trend timelines as a function of the longitude of the stations. The σbsp 10-year trend timelines are similar to the σsp trend timelines (see Fig. S7). The polar stations (in both the Arctic and Antarctic) have been gathered to the bottom of the figure just after the two Pacific stations of MLO and CGO. The main result is that the sites in eastern and central North America (longitude between -68 and -112∘) have ss negative σsp 10-year trends ending after 2009–2012 regardless of their altitude (200–2200 m a.s.l.) and their environments. This is a clear signature of continental-scale modification due to air quality regulations, and this very clear feature relates to the sulfate-dominated aerosol in the eastern USA and to large SO4 reductions in power plant emissions (Hand et al., 2014; McClure and Jaffe, 2018). Almost all the σsp 10-year trends in the southwestern USA (MZW, SCN and HGC) ending before 2011 are ss positive, as published in the previous trend analysis (CC2013). MLO also exhibits ss positive trends for the same period. These four stations are also high-altitude sites (2000–3400 m), so that it is possible that all of them were influenced by long-range transport of highly polluted air masses from Asia (CC2013). It is further interesting to note that the high-altitude site JFJ (3580 m) in Europe also exhibited a ss positive 10-year trend ending in 2005–2008.

The evolution of the European σsp 10-year trends does not show a clear time for trend modification as is seen in North America, probably due to variable timing in implementation of abatement policies in each individual country. Apart from PAL (which can be considered, to some extent, to be a polar station), the σsp 10-year trends in Europe ending after 2008 are all ss negative or not ss. The four stations in Asia do not have ss trends ending in the last 5 years. The two African stations exhibit no ss trend. For polar sites, BRW, ALT and NMY have mostly not ss trends, whereas SPO exhibits alternating ss positive and negative trends, with the oldest 10-year trends being not ss. In contrast, ZEP exhibits positive trends for all three 10-year periods, which is similar to the 10-year trends at PAL for the same time periods. Due to the very low aerosol concentrations at these sites and, thus, larger measurement uncertainty, it is difficult to interpret the evolving σsp polar trends. They could be related to increased influence from the boreal forest and/or changed circulation patterns modifying the sea/ice influence.

The GSN dataset only covers 8 years, with some missing periods due to the destruction of the station by a typhoon. Due to the very low number of long-term measurements in Asia, GSN was included in this study. While GSN σsp summer trends are not reliable (low data coverage and issues in humidity control), the ss negative winter–spring trends corresponding to the dry season are valid and in line with the PM10 decreasing trends in Korea (Kim and Lee, 2018; Nam et al., 2018).

The lengths of the σap time series are much shorter than for σsp (Fig. 13). This means that the oldest 10-year trends cover the period 1998–2007 (BRW and BND), followed by MLO (2001–2010) and JFJ (2002–2011). For these four stations, the most recent 10-year trends are either not ss or ss negative. The present-day (i.e., trends covering 2009–2018) ss negative 10-year trends (JFJ, BND, MLO and BRW) are preceded by not ss trends. The σap 10-year trend evolution of each station is usually homogeneous with either ss negative or not ss 10-year trends in Asia, Europe, Africa and North America. ALT σap 10 years ending in 2017 is ss positive. The BRW polar station and MLO high-altitude station exhibit also some ss positive 10-year trends ending between 2010 and 2014. Unfortunately, only MLO has a long enough σap time series to compare with the ss positive σsp 10-year trends at high-altitude sites (Fig. 12). At MLO, the series of ss positive σsp 10-year trends ended in 2008, while the series of ss positive σap 10-year trends occurred for the period ending 2009–2013.

Figures 12 and 13 suggest that mid-latitude σsp and σap sequential 10-year trends were ss positive for some periods between 2000 and 2013, followed by not ss trends and ending in the present day with ss negative trends. The evolution from increasing to decreasing σsp and σap trends appears to be not simultaneous, with the σsp inflection points occurring some years before those for the σap trends. The sparse number of stations with long enough time series does not allow generalization of this result.

Because it is limited by the length of σap time series, the ω0 10-year trend evolution also only covers the last decade. The following results can be seen in Fig. 14.

All stations at longitude > 10∘ have ss positive ω0 10-year trends except for AMY, which exhibits a ss negative 10-year trend ending in 2018, and MUK with a not ss 10-year trend ending in 2013. For European sites, ss positive ω0 10-year trends exist for all stations at latitude > 46.8∘, apart from BIR, which has a not ss trend ending in 2018. This suggests that the decreasing σap trends in Asia and in eastern and northern Europe are proportionally larger than the decreasing σsp trends.

Both the b and åsp 10-year trends in Asia and Africa exhibit similar 10-year trend patterns that are either ss positive or not ss (Fig. 15). In this context, similar means that the b and åsp trends are never ss when opposite signs of the slope are observed. These results suggest that particle average size tends to decrease at the Asian and African sites. In Europe, b and åsp 10-year trends have a majority of ss negative or not ss trends in the northeast (longitude > 10∘). At lower European longitudes, there is a discrepancy between b and åsp 10-year trends, with IPR, JFJ and PUY having opposite ss trends for b and åsp, b trends being often ss positive and åsp trends often ss negative. The discrepancy in the signs of the trends for b and åsp may be related to shifts in both the fine and coarse modes of the aerosol size distribution – this is discussed more below (see Sect. 4.2).

In North America, the b 10-year trends ending after 2012 are almost all ss positive, whereas the previous 10-year trends are ss negative at BND and MLO. As in western America, one can see discrepancies in the sign of the slope for b and åsp 10-year trends, with APP, CPR, MLO, SGP and THD having at least one 10-year period with opposite signed ss trends. Also, as in western Europe, the b trends are usually ss positive and the åsp trends ss negative. BND records four 10-year periods with opposite ss trends. In contrast to the other stations, BND åsp trends are ss positive, while BND b exhibits ss negative trends.

In the polar regions, the b 10-year ss trends ending after 2014 are all ss positive, whereas the older trends are primarily ss negative. Here again, the discrepancy between b and åsp 10-year trends is large, with all the 10-year b and åsp trends for ZEP and ALT being ss with opposite signs. In contrast, BRW exhibits trends with the same sign for both parameters that can be interpreted as an increase in average particle size for early years followed by a decrease after 2014.

The åap time series are not very long, because the first generation of absorption photometers used either white light or only one wavelength (Fig. 16). The longest time series of åap begins in 2002 at JFJ and exhibits a continuous ss åap decrease. Similar to the results for JFJ, most of the stations have consistently ss negative (JFJ, SMR), ss positive (ALT, ZEP, IPR, SGP, WLG, LLN, MLO and CPR) or not ss 10-year trends (TIK, BRW, THD, APP, GSN, MUK and CPT). The not ss trends of TIK and GSN may be due to datasets shorter than 10 years.

A comparison of the present-day trends derived here to our previous trend ending in 2010 (CC2013) demonstrates that the larger number of stations, particularly in Europe, permits a more detailed view of regional trends. The current wide coverage across continental Europe shows decreasing present-day trends. Decreasing σsp, σbsp and σap trends were confirmed for individual stations (e.g., SMR, Luoma et al., 2019; PAL, Lihavainen et al., 2015b; ARN, Sorribas et al., 2019), as well as at ACTRIS sites including JFJ, HPB, IPR, IZO, PAL, PUY, SMR and UGR (Pandolfi et al., 2018). There are some discrepancies in the trends between our current study and Pandolfi et al. (2018) that seem to be principally due to differences in the analyzed periods. Three additional years of data were included in this study and some older periods included in Pandolfi et al. (2018) were invalidated following the evaluations described in Sect. 2.4. The European b and åsp trends computed by Pandolfi et al. (2018) are similar to the results of this study for most of the stations, in that they also found a general ss increase in b and variable åsp trends. In North America the ss decreasing trends in aerosol extensive properties observed in CC2013 are found to have continued in this work with the extended datasets. These results are confirmed by the two other trend studies for in situ aerosol optical properties in North America. While the methodology and time period of Sherman et al. (2015) were different, the sign and ss of their σsp, b, and åsp trends for BND and SGP were the same as reported here. White et al. (2016) found a decreasing trend in absorption coefficient (estimated from light transmittance measurements on 24 h filter samples) at 110 IMPROVE stations for the 2003–2014 period. SPO σsp, b and åsp trends for the 1979–2014 period (Sheridan et al., 2016) do agree with CC2013 results, whereas the 1979–2018 trends reported in this study suggest an evolution towards more ss positive trends. The very low aerosol concentrations in Antarctica and the difference in the MK algorithm could however also explain the differences amongst these three analyses.

There have been multiple trend studies on carbon species (also referred to as BC, elemental carbon, equivalent black carbon, brown carbon or other terms) which are closely related to aerosol absorption. A decreasing trend in BC concentration is found in Europe (Singh et al., 2018; Kutzner et al., 2018; Grange et al., 2020), related primarily to traffic emission decreases rather than changes in wood burning and/or industrial emissions. Similarly, Lyamani et al. (2011) noted a decrease in BC in southern Spain due to the 2008 economic crisis. In contrast, Davuliene et al. (2019) reported an increasing trend in equivalent black carbon (eBC) for the Arctic site of TIK. In North America, White et al. (2016) found that the decreasing elemental carbon trend at IMPROVE sites was larger than the aerosol absorption trend at the same sites due to the impact of Fe content in mineral dust. BC trends in the Arctic have been extensively studied (e.g., AMAP, 2015; Sharma et al., 2019, and references therein) and suggest a decreasing trend. This is consistent with our general trend in absorption for the polar regions (Table 4), although for individual stations most trends were statistically insignificant.

Particulate mass (PM) and visibility are other metrics for atmospheric aerosol loading that can be most readily compared with our trends in aerosol scattering. Tørseth et al. (2012) detailed decreases in PM across Europe, while Hand et al. (2014, 2019) report significant decreases in PM2.5 mass across the USA, with larger trends in the eastern than in the western USA. Both these trends were also confirmed by the PM trend analysis in Mortier et al. (2020) and are consistent with our reported scattering trends. Li et al. (2016) used visibility to assess trends in atmospheric haze and aerosol extinction coefficient around the world. The time delay in when the trends switch sign between North America (late 1970s), Europe (early 1980s) and China (mid 2000s) correlates with SO4 trends, and the trend differences between the eastern and western parts of the USA and Europe are consistent with what is presented in our study.

Many atmospheric aerosols are formed in the atmosphere rather than being directly emitted, so understanding trends in aerosol precursors is also relevant for understanding changes in the atmospheric aerosol. Our study found similar results for scattering to those found for sulfate trends (Aas et al., 2019), i.e., decreasing sulfate trends across Europe and the USA, albeit with the sulfate decrease in Europe beginning before the decrease was observed in the USA. Aas et al. (2019) also describe potential increases in sulfate in India and increases followed by decreases in SE Asia. Vestreng et al. (2007) monitored the sulfur dioxide emission reduction in Europe and concluded that SO4 emission reductions were largest in the 1990s, with a first decrease in western Europe in the 1980s followed by a large decrease in eastern Europe in the 1990s. Similarly, Crippa et al. (2016) simulated a larger impact of policy reduction in western Europe than in eastern Europe for NOx, CO, PM10 and BC between 1970 and 2010. Likewise, Huang et al. (2017) simulated the non-methane volatile organic compound emissions and found a rapid decrease in Europe and in North America since the 1990s, whereas the emissions of Africa and Asia clearly increased between 1970 and 2012.

A significant advantage of many REM platforms is their global coverage. Satellites often provide coverage over both land and ocean, and the major ground-based REM network AERONET (Holben et al., 1998) is more globally representative than the sites used in this study. However, there are some inherent limitations in comparing aerosol optical property trends from REM retrievals with surface in situ trends. Our study used aerosol optical measurements made at low RH (typically RH < 40 %) at the surface, while column aerosol optical retrievals are made at ambient conditions and represent the atmospheric column including layers aloft. Only in the situation of a well-mixed atmosphere will it be reasonable to compare trends in surface in situ optical properties with those obtained by ground-based or satellite retrievals. It has also to be mentioned that satellite measurements are less sensitive to the near-ground layers containing the greatest aerosol load. Thus, while our trends can be compared with those for column aerosol properties, there is no reason to expect them to be in complete agreement. Below we discuss trends in PM, AOD, column σap and column SSA.

Satellites have been used to assess the decreasing PM trends in North America and Europe and also to estimate PM trends in other regions with sparse surface measurements. For example, Nam et al. (2017) evaluated the trend in satellite-derived PM10 over Asia and reported mixed annual trend values depending on the subregion they looked at. Li et al. (2017) found satellite-derived PM2.5 to continuously increase in some parts of Asia (e.g., in India) for the 1989–2013 period – we also find an increasing trend (for aerosol absorption) at the one site we studied in India (MUK). For China, Li et al. (2017) report that the PM2.5 trend transitions from an increasing to a decreasing trend, with the transition occurring in the 2006–2008 time period similarly to the sulfate trend pattern reported by Aas et al. (2019). The in situ measurements from China (WLG) and Taiwan (LLN) used in our study are not long enough to detect this transition.

Multiple ground-based REM studies (e.g., Yoon et al., 2016; Wei et al., 2019; Mortier et al., 2020) report decreasing trends in AOD over the USA and Europe, with larger decreasing trends over Europe than over the USA, which is the case in our study (see Table 4) as well. The lack of measurements in many regions, similar to the lack of representativeness in the surface in situ aerosol sites discussed in this study (Asia, Africa, South America, etc.), is also emphasized. Ningombam et al. (2019) analyze AOD 1995–2018 trends from 53 remote and high-altitude sites, of which 21 had ss negative trends. Regionally, Ningombam found primarily negative trends at sites in the USA, Europe and the polar regions. Their findings for sites in China and India suggested mixed trends, with some being positive and some negative in those regions. Some of the sites in Ningombam et al. (2019) were also involved in our study. The trends they find for AOD at LLN and MLO are similar to ours (i.e., not ss trends) at SPO (i.e., ss increasing) and at SGP (i.e., decreasing (note: they refer to SGP as “car”)). Their results are different for IZO (we found no ss trends for scattering, while they reported ss decreasing AOD) and for BRW and BIR (we found ss decreasing scattering trends, but they found not ss AOD trends).

Satellite retrievals can offer an even more global picture of aerosol trends than the surface-based REM data. Various satellite trend analyses present a picture of trends in aerosol optical depth for different regions of the world that is quite consistent across satellite (and ground-based) AOD datasets. For example, for the satellite literature that we surveyed, all found decreases in AOD over the USA and Europe (e.g., Hsu et al., 2012; Mehta et al., 2016; Zhao et al., 2017; Alfaro-Contreras et al., 2017; Wei et al., 2019) consistent with what we have reported for the AOD from ground-based, REM instruments. As we note above, this is also consistent with surface in situ scattering trends. There are some discrepancies in the various satellite-derived AOD trends over Asia that are likely due to differences in time period of analysis, trend methodology, regional definitions and/or perhaps satellite data product. Nam et al. (2017) found that AOD trends varied depending on what part of Asia was being evaluated. Zhao et al. (2017) reported an increasing then decreasing trend over China, which was also suggested by others (e.g., Sogachova et al., 2019; Alfaro-Contreras et al., 2017). Wei et al. (2019) found a slightly negative but statistically insignificant AOD trend for China. Our study found statistically insignificant trends in aerosol loading for both the high-altitude surface site in China (WLG) and in Taiwan (LLN), perhaps because measurements at both these sites span the AOD increase/decrease periods mentioned by Zhao et al. (2017). Over India, increasing trends in satellite AOD were reported by all the literature we surveyed (e.g., Wei et al., 2019; Mehta et al., 2016; Hsu et al., 2012; Alfara-Contreras et al., 2017). This is consistent with our finding of an increasing trend for aerosol absorption for the one Indian site (MUK) in our study.

The satellite measurements also enable evaluation of aerosol loading changes in regions with few to no long-term surface in situ aerosol optical property measurements. The Middle East exhibited an increasing trend in AOD, while South America exhibited variable trends (e.g., Wei et al., 2019; Metha et al., 2016; Hsu et al., 2012; Alfaro-Contreras et al., 2017). Wei et al. (2019) found a statistically insignificant trend in South America and suggested it was due to complex and changing aerosol sources. Mehta et al. (2016) looked specifically at Brazil and found a decreasing annual AOD trend but an increasing AOD trend in springtime. Decreasing AOD trends were found over central Africa (Wei et al., 2019), over the African deserts (Metha et al., 2016) and on African coasts (Alfaro-Contreras et al., 2017) regardless of whether they are dominated by smoke aerosols (southwest) or dust (northwest).

In addition to AOD, trends for other column aerosol property such as column σap and column SSA can be considered. While there appear to be many investigations focusing on trends in column aerosol properties other than AOD at individual sites, there are only a few papers that take a more global, multi-site approach (e.g., Li et al., 2014; Zhao et al., 2017; Mortier et al., 2020). There have been several studies related to changes in column σap using AERONET REM retrievals. For example, Li et al. (2014) suggest an increase in column σap over the USA and a decrease over Europe and at most sites in Asia. More recently, Mortier et al. (2020) found ss decreasing σap trends in Europe and North America and ss increasing σap trends in Asia and Africa. Zhao et al. (2017) used satellite retrievals and reported decreasing trends of column σap over both the USA and Europe and a not ss column σap trend over China. Nam et al. (2018) suggested there was an increasing trend in the column extinction Ångström exponent over Asia based on satellite observations. These findings are mostly consistent with our results (Table 4) which indicated decreasing åsp trends in the USA and Europe but perhaps an increasing trend in Asia.

Comparisons of in situ and column ω0 trends are more fraught, because, in addition to the above-mentioned caveats related to comparing surface and column measurements, column ω0 can only be obtained from REM techniques under higher aerosol loading conditions. For example, Kahn and Gaitley (2015) indicate that MISR SSA retrieval requires AOD > 0.15–0.2. Similarly, AERONET retrievals require AOD (at 440 nm) > 0.4 (Dubovik et al., 2000). This limits the sites for which column SSA can be retrieved. Andrews et al. (2017) present a plot derived from global model simulations suggesting more than 80 % of the globe has annual AOD values below 0.2, and, indeed, many of the surface in situ sites discussed here are in remote locations with annual AOD consistently below 0.2. Andrews et al. (2017) also suggest there is a systematic variability of SSA with loading that might result in column SSA biases if retrievals are constrained to higher levels of AOD. With these caveats in mind, we can compare our surface ω0 trend results with satellite column ω0 trends. Li et al. (2014) studied 2000–2013 trends in column ω0 at select AERONET sites. Their findings suggest that column ω0 is increasing in the USA, Europe and Asia. However, they noted that the uncertainty in these trends is high because they used level 1.5 data (AOD < 0.4) in order to have enough data points for their analysis. Zhao et al. (2017) utilized satellite retrievals and reported decreasing trends in column ω0 over the eastern USA and Europe and a not statistically significant trend over China for the 2001–2015 period. Their results over the USA and western Europe are consistent with the overall regional ω0 trends reported in this study (i.e., Table 4), although Fig. 7 suggests there is a fair amount of variability in the surface ω0 trends at the individual sites in these two regions. Our study found an increasing trend in ω0 at the surface in Asia (based on three sites), which is consistent with Li et al.'s column ω0 trend but not with the lack of trend in column ω0 over China suggested by Zhao et al. (2017). But, as noted above, remote sensing retrievals of column SSA should be considered with caution and, clearly, further effort in column SSA trend analysis is warranted.

While it is beyond the scope of this effort to explore in depth the causes of the observed trends of aerosol optical properties, some general comments can be made. First, tendencies in regional trends for variables representing aerosol loading (e.g., surface in situ aerosol scattering, PM, and AOD) are generally consistent across multiple datasets. Overall, the main cause of observed decreasing trends in loading is likely a strong reduction of both primary aerosols and precursors of secondary aerosol formation connected to mitigation strategies on regional to continental scales (e.g., Huang et al., 2017; Crippa et al., 2016; Pandolfi et al., 2016; Vestreng et al., 2007). Detailed analysis of PM reductions and composition changes in Europe and the USA has enabled attribution of the trends to changes in source types and emission levels (e.g., Hand et al., 2019; Pandolfi et al., 2016; Ealo et al., 2018). The explanations of the trends based on long-term measurements are supported by modeling efforts. Like many satellite retrievals, model simulations also provide global coverage and, in addition, can be used to investigate reasons for observed changes in aerosol. Model simulations described in Li et al. (2017) suggested that the decrease in PM2.5 in western and central Europe is principally due to sulfate, whereas in eastern Europe decreases in organic aerosol also play a role. The EMEP status report (2019) notes that the difference in emission trends between western and eastern Europe has become more significant since 2010. Further, the EMEP status report suggests that estimated increasing emissions of all pollutants since 2000 in eastern Europe are mainly influenced by emission estimates for the remaining Asian areas in the EMEP modeled domain. Similarly, Zhao et al. (2019) used a model to attribute the AOD, ω0 and åsp decreases in North America and Europe to considerable emission reductions in all major pollutants except in mineral dust and ammonia.

For Asia, modeling by Li et al. (2017) suggests aerosol changes are principally related to increases in organic aerosol and secondary inorganic aerosol, whereas the increases in BC, nitrate and ammonium are comparably moderate. Yoon et al. (2016) use a model to ascribe the observed increases in AOD over India to increases in BC and water-soluble materials – both related to anthropogenic emissions. Over China, Yoon et al. (2016) observe a disconnect between the model chemical composition and the measured AOD, which they explain by noting that the measurement sites they rely on in the region are far from the population centers where most of the emissions occur. Zhao et al. (2019) use a model to attribute the increase in AOD followed by a decrease in AOD to emission increases induced by rapid economic development until 2008–2009 followed by decreases in both anthropogenic primary aerosols and aerosol precursor gases.

Zhao et al. (2017) suggest that the larger reductions in aerosol precursors (e.g., SO4 and NOx emissions) rather than primary aerosols, including mineral dust and black carbon, can explain the decreases in ω0 and åsp observed over Europe and the USA. This is because the secondary aerosols formed from such precursors tend to be primarily scattering, so less secondary aerosol would change the relative balance between scattering and absorption, driving ω0 down. Similarly, secondary aerosol particles tend to be small, so a decreasing trend in secondary aerosol would change the relative contribution of small to large particles in the aerosol size distribution and lead to a decreasing trend in åsp. In contrast, in Asia simultaneous increases in aerosol precursors and BC before 2006 and a simultaneous decrease after 2011 explains the trends ω0 and åsp they observed there. Modifications in emissions of aerosol precursors also impact the atmospheric chemistry, leading to non-linear response of the formation of secondary inorganic aerosol (Banzhaf et al., 2015).

While regional changes in emissions are one driving factor in trends, because of long-range transport, out-of-region changes in sources also have the potential to affect trends. For example, Saharan dust impacts CPR, IZO and UGR (e.g., Denjean et al., 2016; Rodriguez et al., 2011; Garcia et al., 2017; Lyamani et al., 2008), and its emissions may change (decrease) in a warmer world (Evan et al., 2016). Other examples of sites clearly impacted by long-range transport include IZO impacted by northern African pollution due to developing industries (Rodriguez et al., 2011) and the high-altitude station of MLO which is impacted by Asian pollution (e.g., Perry et al., 1999). Mountainous stations can also be affected by modifications of the planetary boundary layer or of the continuous aerosol layer heights responding to ground temperature or mesoscale synoptic weather changes (e.g., Collaud Coen et al., 2018, and references therein). The oscillation in trend sign for several variables at the Arctic sites is potentially caused by the very low aerosol loading, but the Arctic region is changing rapidly and the impact of evolving transport patterns, atmospheric removal processes or local sources cannot be excluded (e.g., Willis et al., 2018) and requires closer study. While both increasing and decreasing levels of aerosol due to changes in anthropogenic emissions have been observed, the role of non-anthropogenic sources may become more important in the future. For example, climate change also affects soil drought, and the positive feedback between drought and wildfires can also affect aerosol optical properties (Hallar et al., 2017; McClure and Jaffe, 2018). The number and intensity of wildfires are increasing in several regions (e.g., Moreira et al., 2020; Turco et al., 2018; Hand et al., 2014). McClure and Jaffe (2018) confirmed an increasing trend of PM2.5 98th percentiles in the northwestern USA due to an increase in wildfires superimposed on the global decrease in anthropogenic emissions. Yoon et al. (2016) also note an increase in extreme AOD events in the western USA, which they hypothesize could be due to wildfires. Another example of potential changes in natural aerosol may take place in the Arctic, where decreases in sea ice coverage might play a role in natural aerosol increases in the region (e.g., Willis et al., 2018) (decreases in sea ice coverage may also lead to enhanced anthropogenic emissions due to increased human activity (e.g., Aliabadi et al., 2015)). Whether such changes in natural aerosol emissions lead to observable changes in overall aerosol trends or trends at the extremes of aerosol loading is something to look for in future trend analyses.

Detailed studies at each station are necessary to discriminate between direct causes like changes in anthropogenic emissions and indirect causes related to general climate changes such as drought, changes in surface albedo, biogenic aerosol concentration, atmospheric chemistry, sea ice coverage or atmospheric circulation patterns. The availability of the homogenized dataset from this study will provide a useful tool for these types of analyses.

In order to get a truly global overview of aerosol trends, surface in situ measurements need to be paired with model simulations and satellite observations. This will enable evaluation of the uncertainty in regional and global trends based on deficiencies in spatial and/or temporal coverage. Satellites and models are able to fill the gaps in coverage from ground-based measurements, but both rely on surface measurements for ground truth.