Consistency of long-term elemental carbon trends from thermal and optical measurements in the IMPROVE network

Consistency of long-term elemental carbon trends from thermal and optical measurements in the IMPROVE network L.-W. A. Chen, J. C. Chow, J. G. Watson, and B. A. Schichtel Division of Atmospheric Sciences, Desert Research Institute, 2215 Raggio Parkway, Reno, Nevada, 89512, USA State Key Laboratory of Loess and Quaternary Geology, Institute of Earth Environment, Chinese Academy of Sciences, 10 Fenghui South Road, Xi’an, 710075, China Cooperative Institute for Research in the Atmosphere, Colorado State University, 1375 Campus Delivery, Fort Collins, Colorado, 80523, USA


Introduction
Elemental carbon (EC), a light-absorbing carbon (LAC) component as determined by thermal/optical methods, is the dominant aerosol fraction that absorbs visible radiation in the troposphere (Andreae and Gelencsér, 2006).This fraction is often termed "black carbon" (BC) if quantified by optical or photoacoustic methods (Moosmüller et al., 2009).EC aerosols from incomplete fuel combustion are non-spherical and internally mixed with organic carbon (OC) (Chakrabarty et al., 2006a, b;Chen et al., 2010).Jacobson (2009) estimates the 100-yr global-warming potential (GWP) of EC + OC from fossil-and bio-fuel combustion to be 800-1300 relative to carbon dioxide (CO 2 ).Reducing EC emissions could be a short-term and cost-effective method for slowing global warming (Jacobson, 2002;Bond and Sun, 2005), as well as providing co-benefits for public health, visibility, and material damage (Chow and Watson, 2011).
Long-term monitoring of aerosol chemical composition in the US Interagency Monitoring of PROtected Visual Environments (IMPROVE) network (Watson, 2002) reveals a decreasing trend in average EC concentrations by over 25 % from 1990 to 2004 for the entire country (Murphy et al., 2011) as well as decreases in EC of 40-60 % for urban and non-urban California sites from 1988 to 2007 (Bahadur et al., 2011a, b;Schichtel et al., 2011).These trends are consistent with emission reduction measures implemented to attain PM 2.5 and PM 10 National Ambient Air Quality Standards for engine exhaust (Lloyd and Cackette, 2001), residential wood combustion (Hough and Kowalczyk, 1983;Butler, 1988;Hough et al., 1988), and prescribed burning (Riebau and Fox, 2001;Tian et al., 2008).Even though IMPROVE data are available through 2009, Murphy et al. (2011) chose to exclude data from 2005 onward owing to potential biases that might have been caused by an upgrade in IMPROVE carbon instruments beginning in 2005.Chow et al. (2007)  demonstrated equivalence between measurements made with the original (Chow et al., 1993) and upgraded (Chow et al., 2007(Chow et al., , 2011) ) instruments for hundreds of samples from a variety of environments.However, average EC concentrations and EC/total carbon (TC) ratios increased at some (but not all) IMPROVE sites from 2004 to 2005, as illustrated in Fig. 1.The objective of this study is to investigate the recent (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009) trends in IMPROVE EC along with those of filter reflectance, which serves as an independent surrogate for EC.
The IMPROVE thermal/optical reflectance (TOR) analysis protocol separates EC from OC on filter samples by temperature-dependent volatilization and oxidation.EC is defined as carbon that does not evolve at ∼ 580 • C in an inert helium (He) atmosphere and is subsequently oxidized to CO 2 with the introduction of oxygen (2 %) at higher temperatures, up to 840 • C. A fraction of OC chars in the inert atmosphere, as evidenced by decreases in light (632.8nm He-neon (Ne) laser) reflected from the aerosol deposit on the filter surface during the analysis (Fig. 2).Pyrolyzed OC (POC) is defined as the carbon evolved after oxygen is introduced and before the reflected light intensity returns to its original value (i.e., the reflectance crossover).POC is subtracted from apparent EC measurement to yield "native" EC concentration in the sampled air.When all of the carbon has evolved, the remaining filter is usually white, similar to the appearance of a blank filter.Non-white filters are occasionally found during dust events, and these are flagged as part of the IMPROVE protocol.
The 2005 carbon instrument upgrade led to a transition from the IMPROVE to IMPROVE A thermal/optical analysis protocol (Chow et al., 1993(Chow et al., , 2007)).The new protocol did not change the temperatures plateaus but rather reflected "actual" analysis temperatures that had been implemented since the inception of the IMPROVE network (Chow et al., 2005).With improved electronics and sealing, the upgraded instrument allows for more precise temperature control, flexible data acquisition, a higher intensity laser light beam, and lower trace oxygen levels in the inert He atmosphere than did the original instrument (Chow et al., 2011).It also enables simultaneous monitoring of filter reflectance and transmittance without changing the reflectance measurement configuration (Fig. 2).Since 2005, reflectance as well as transmittance crossover has been used for charring correction.Thermal/optical transmittance (TOT) often reports higher POC and lower EC than TOR.Chen et al. (2004) and Chow et al. (2004) attributed this to charring of organic vapors adsorbed within the filter (Watson et al., 2009;Chow et al., 2010) which attenuate transmittance substantially but have a minor effect on reflectance from the surface deposit.EC hereafter refers to TOR EC.
Optical measurements designed for charring correction provide alternatives for quantifying EC or BC abundances on filters.Filter attenuation using reflected light (τ R ) or transmitted light (τ T ) is defined as (1) where R 0 and T 0 are reflectance and transmittance, i.e., the reflected and transmitted light intensity, of blank filters, respectively, while R and T are reflected and transmitted light  intensities of particle-laden filters (prior to carbon analysis), respectively.τ R or τ T can be a practically linear function of the light absorption coefficient (b abs ) for filter samples (Lindberg et al., 1999;Quincey, 2007).The widely deployed aethalometer (Hansen et al., 1984) and particle-soot absorption photometer (PSAP; Bond et al., 1999) estimate b abs from τ T that is then converted to BC loading using assumed mass absorption efficiencies derived from simultaneous EC measurements (Watson et al., 2005 and references therein).b abs and BC based on τ R are also reported (e.g., Edwards et al., 1983;Janssen et al., 2011).τ R could be more variable in estimating b abs than τ T since the angular distribution of reflectance is more sensitive to the chemical composition of particle deposits (Kopp et al., 1999;Petzold and Schönlinner, 2004).Nonlinearity among b abs (or BC), τ R , and τ T increases with higher sample loading (Arnott et al., 2005) though it was shown in Chen et al. (2004) that the linear relationship between reflectance and transmittance holds up to an EC loading equivalent to ∼ 20 µg cm −2 on a filter or ∼ 2 µg m −3 in ambient air for IMPROVE network samples (32.7 m 3 of air sampled through a 3.53 cm 2 filter area).Since τ R , a measurement of the darkness of the filter deposit, was recorded for every IMPROVE sample before, during, and after the instrument upgrade and is independent of the evolved carbon quantification, it can be used as an independent indicator of EC.Investigating the EC and τ R relationship before and after the instrument upgrade is essential.This relationship could be site-, and possibly season-specific, considering the diverse environments represented by IM-PROVE samples.Determining τ R trends provides additional weight of evidence for observed EC trends.

Methodology
Digital thermograms (which record one second values for temperature, reflectance, and carbon content) for > 83 000 IMPROVE samples acquired by 24-h sampling on every third day from CY2000 through CY2009 were reprocessed to obtain the initial (dark aerosol deposit) and final (white filter) reflectance values.Data recovery varied by site; typically exceeding 92 % for 2005-2009, but ≤ 80 % for 2000-2004 due to deteriorating storage media (floppy disks and CD-ROMs; it was not practical to recover data from the paper documentation).The 65 sites with the longest records and highest data recovery rates are listed in Table 1 and used for subsequent analysis.Each of these sites contains 80-120 samples per year (20-30 samples per season).They represent 25 US geographic regions as described in Table 1 (see Fig. 3 for the site locations).τ R was calculated per Eq. ( 1) from a ten-second average of the initial and final reflectance for each sample.
The final reflectance represents effective R 0 as all EC has been removed from the filter.Pre-and post-upgrade τ R at a particular IMPROVE site are related to EC through a linear model: (3)  3) and ( 4) are nested into where I and O are unit and zero column vectors and c and b represents c + − c − and b + − b − , respectively.Meaningful changes in c and b would lead to c and b that differ from zero at a statistically significant level (Gujarati, 1970a, b).A robust least-squares regression method that lowers the influence of outliers was applied to determine the coefficients and respective standard errors and p-values in Eq. ( 5).This is achieved by Matlab ® robustfit function with the Huber iterative reweighting algorithm (Dutter and Huber, 1981).
Statistical consistency of c and b pre-and post-2005 (i.e., non-significant c and b) result from relatively small c and b or large standard errors.The latter suggests an insufficient correlation between EC and τ R for τ R to be a good predictor for EC.Therefore, it is important to examine the regression's correlation coefficient as well as the fractional changes in b and c, e.g., b/b − and c/EC −med (EC −med : median EC − concentration).c/EC −med is a better evaluation of changes in c than c/c − since c − is usually small to near zero.Lower and Thompson (1988) show that EC + can be related to EC − by solving Eqs. ( 3) and ( 4) after c and b are determined.This relationship would be the best estimate for the relationship between EC + and EC − , given that a direct regression is not possible.
EC and τ R trends were further assessed using a nonparametric Mann-Kendall (M-K) test (Kendall, 1975;Yue et al., 2002), which examines the sign of slopes for all possible data pairs and determines trend significance from the difference in positive and negative signs.All data acquired in the same year are considered as concurrent measurements (ties) in the test to minimize influence of intra-annual trends such as seasonal variations (Salas, 1993).M-K statistics yield Sen's slope (Sen, 1968;Burn and Hag Elnur, 2002), which is the median slope across all possible data pairs, and its p-value and confidence intervals.Sen's slope provides a more quantitative estimate of the trends.M-K statistics were calculated with Matlab ® code provided by Burkey (2009).

Results and discussion
The majority of correlation coefficients (r) of EC versus τ R from Eq. ( 5) exceed 0.8 (Table S1 in the Supplement).Lower r is found for Urban, Appalachia, and Ohio River Valley sites with high EC concentrations, especially Washington D.C. (U1 in Fig. 3; r = 0.59) and James River Face Wilderness, Appalachia (A1, r = 0.67).Thirty-six of the 65 sites show no changes in regression slope prior to and after 2005 at the 5 % significance level (i.e., p( b) > 0.05).Thirty-four of the 36 sites, including all Appalachian sites, show no significant changes in regression intercept prior to and after 2005 (i.e., p( c) > 0.05).p( c) are < 0.05 but > 0.01 (1 % significance level) for the remaining two sites (Cape Romain NWR (Southeast, SE3) and Canyonlands NP (Colorado Plateau, CP6), see Table 1 and Fig. 3).The absolute values of b and c for these 36 sites are small, generally within 10 % of b − and EC −med , respectively (Group I in Fig. 4).There is no evidence that the instrument upgrade had an effect on EC measurements for samples taken at these sites.
The other 29 sites are separated into two groups according to Fig. 4. Group II (17 sites) exhibits negative b along with positive c.Six Group II sites have both b and c that are significantly different from zero (p < 0.05), including Brigantine NWR (E1), Washington DC (U1), Lostwood (NP3), UL Bend (NP6), Glacier NP (NR1), and Denali NP (AK1).These sites are located in eastern (E1, U1), northern, and northwestern states (NP3, NP6, NR1, AK1).Group III (12 sites) exhibits positive b and mostly negative c. Eight out of 12 Group III sites contain both b and c significantly different from zero (p < 0.05), including White Pass (NW4), Three Sisters Wilderness (ON4), Mount Hood (ON5), Bliss SP (SN3), Death Valley (D1), Great Basin (G1), Hance Camp at Grand Canyon NP (CP3), and Bridger Wilderness (NR4), all of which are located in the Western Cordillera area of the continental US (Fig. 3). Figure 5 shows examples of EC-τ R scatter from these three groups.
The POC fraction generally increased for samples analyzed since the beginning of 2005 due to higher purity of the inert He atmosphere and more rigorous quality control of He purity (Chow et al., 2007(Chow et al., , 2011)).Even with the reflectance correction, some POC can be misclassified as EC, thereby increasing the EC fraction.This is more evident when EC/POC ratios are low and would likely move the EC-τ R regression towards a higher intercept and lower-to-unchanged slope.Figure 4 is not consistent with this effect being dominant, except possibly at a few Group II sites including the Brigantine NWR site (E1; exemplified in Fig. 5b).
For Group III samples, low EC values tend to be even lower beginning in 2005 for the same τ R (e.g., Fig. 5c).The reason for this is unclear, though it might be related to different sensitivities of reflectance splits between the original and upgraded instruments for low EC levels.With an improved signal-to-noise ratio of the reflectance measurement, the upgraded instruments possibly trigger the split (crossover) later  Sen's slope (2000Sen's slope ( -2009)).The blue bar indicates the 95 % confidence interval of the trend.
than the original instruments, leading to lower EC fractions.τ R quantification is little influenced by the noise, as both R and R 0 are averaged over 15 s before and after the thermal analysis.The opposite effects apparent for Groups II and III could occur simultaneously and offset each other to some extent.
The regression analysis was also carried out by season.However, such seasonal segregation does not reduce scatter around the best-fit lines (Fig. S1 in the Supplement).This suggests daily variability (due to changes in chemical composition and/or measurement uncertainty) comparable to seasonal variability in the EC-τ R relationship and that yearround regression analyses are reasonably representative of all cases.To test whether extreme EC values due to special events such as wildfires can bias the robust regression, regressions were also calculated excluding EC > 15 µg cm −2 .This test resulted in only minor changes in regression intercepts and slopes and did not influence the grouping of the 65 sites.
Since the regression slopes increase or decrease while intercepts decrease or increase (i.e., change in opposite direction), EC + may shift higher or lower compared to EC − depending on site and EC loading.Figure 6 shows, by site, the characteristic EC + vs. EC − relationships between the 10th and 90th EC − concentration percentiles, which contains 80 % of the samples.The linear relationships were derived from Eqs. ( 3) and (4) by eliminating the common variable τ R , as suggested by Lower and Thompson (1988).EC + is shown to be within ±10 % of EC − , for the most part.Larger deviations, e.g., 10-20 % or −10 to −20 %, are seen for EC − ≤ 3 µg cm −2 .Two extreme outliers are the Washington, DC (U1) and Denali NP (AK1) sites, which experience the highest and lowest EC concentrations, respectively.There seems to be more variability in the EC responses between the original and upgraded instruments for the high and low extremes.
The robust M-K test confirms decreasing trends of EC from 2000 through 2009 (Fig. 7), with the largest and smallest changes observed at one Appalachian (Sipsy Wilderness; A2: −0.021 µg m −3 yr −1 ) and one Central Rockies (Great Sand Dunes, New Mexico; CR2: −0.003 µg m −3 yr −1 ) site, respectively.The trends are statistically significant for all 65 sites at the 5 % significance level.This implies 1.3-8.3% reduction of ambient EC concentrations each year (scaled to EC −med as 2000-2004 represents the IMPROVE network baseline period).The national average trend, calculated from the percentage trends weighted by EC −med at each site, would be −4.5 % per year.With an unweighted ordinary linear regression, Fig. S2 (Supplement) shows median EC decreasing at 3-5 % per year from 2000-2009.Murphy et al. (2011) reported a lower value, ∼ −2.2 % EC per year, for March 1990-February 2004 for average, rather than median, EC concentrations.
Figure 7 also shows significant decreasing trends (p < 0.05) for τ R at all except one site in the Northwest (White Pass, Washington; NW4) where the p-value is 0.051 for the negative τ R trend (−0.099Mm −1 yr −1 ).The EC and τ R trends are highly correlated, at r 2 = 0.9 and slope = 10 m 2 g −1 (Fig. 8).Washington, DC (U1 site), the only urban site in this dataset, is an outlier where EC + seems much higher than EC − based on reflectance (Fig. 6), leading to a smaller EC trend than expected from the τ R trend.The EC trend at the U1 site contains a large uncertainty, and this may also be the case for other urban sites.The national average τ R trend, as scaled to τ R−med is −4.1 % each year, also consistent with the national EC trend.
Although subtle changes are found in EC-τ R relationships between the pre-and post-2005 periods, the consistency between recent EC and τ R trends for the majority of IMPROVE sites do not support that such changes have introduced a major or common bias for the EC trends.Environmental changes, probably due to changing EC emissions and year-to-year meteorological variability, are of larger influence than measurement uncertainties.EC concentrations appear to continue decreasing beyond the 1990-2004 period examined by Murphy et al. (2011) at an average rate of 4.1-4.5 % per year.The Regional Haze Rule (US EPA, 1999) has set the goal of returning visibility to natural conditions by 2064.For EC, the natural concentrations are estimated to be ∼ 10 % of the 2000-2004 baseline period.At the current rate of progress, this goal should be met by the 2064 deadline.

Fig. 1 .
Fig. 1.Annual average elemental carbon (EC), organic carbon (OC), and the ratio of EC to total carbon (TC = OC + EC) for: (a) all IMPROVE data, (b) downtown Washington DC (U1), and (c) Bryce Canyon National Park (CP1) between 1989 and 2009.Data were acquired from the Visibility Information Exchange Web System (VIEWS) website (http://views.cira.colostate.edu/).An EC increase from 2004 to 2005 corresponds with the carbon instrument upgrade for (a) and (b), but this is not observed at every site, as shown in (c).

Fig. 3 .
Fig. 3. Sixty-five IMPROVE sites in 25 regions (seeTable 1 for definitions).Color codes indicate the changes of EC-τ R regression coefficients across the instrumental upgrade in 2005.Red: significant change in slope (p < 0.05); solid edge: significant change in intercept (p < 0.05); green: all other sites without significant changes.See text for details.

Fig. 5 .
Figure 5. 1 where bold italics indicate column vectors of EC or τ R including all pre (−)/post (+) upgrade (on 1 January 2005) data, and c and b are regression coefficients (c: intercept; b: slope).c and b are expected to differ (i.e., c + = c − and/or b + = b − ) only if the instrument upgrade introduced a bias in EC that is larger than typical measurement uncertainties.To examine the changes in c and b, Eqs. (
Table 1 for definitions).Color codes indicate the changes of EC-τ R regression coefficients across the instrumental upgrade in 2005.
Red: significant change in slope (p < 0.05); solid edge: significant change in intercept (p < 0.05); green: all other sites without significant changes.See text for details.

Table 1 .
Region, location, and data completeness (2000-2009)of EC and τ R for 65 IMPROVE sites selected for this study.

California Alaska t R Trend (Mm -1 yr -1 ) t R ×A/V (Mm -1 )
EC by thermal/optical reflectance (TOR) and (b) τ R at 65 IMPROVE sites.See Table1for site details.A and V are nominal filter area (3.53 cm 2 ) and sample volume (32.7 m 3 ), respectively.Medians are those of 2000-2004 baseline period.Trends are based on