Trends in EC/OC Ratio and Carbon Fractions

Author:  Robert Eldred, Crocker Nuclear Laboratory,

Date:  October 26, 2001

Background:  During the Bend Visibility Conference in October 2-5, 2001, Warren White presented a paper that showed a trend in the EC/OC ratio at some sites.  The question is whether these trends are real or caused by analytical changes. 

Summary and Conclusions:  This report will examine the trends at all sites with a period of record from 1988 to 1999 and eastern sites with a period of record from 1992 to 1999.  In order to shed light on EC and OC, the behavior of the 8 carbon fractions is also examined.  The methodology and implications of artifact subtraction are discussed in Appendix 1.  Although a few artifacts do show trends, there is no indication that the artifacts are significantly involved in any trends in the concentrations.  The conclusion of the report is that there are no general pattern to trends in the EC/OC ratio for the total network.  In the West, as many sites have increasing EC/OC as decreasing.  In the East, however, there is a pattern of decreasing EC/OC.  For all remote East sites, the decrease was about 2% per year.  Five of the sites (16%) in both East and West have statistically significant trends, 4 decreasing and 1 increasing.  At 4 of the 5 sites, the change appears to be associated more with a change in EC than with a change in OC.  There is no evidence that there was an analytical change in the EC-OC split.  In fact, the evidence suggests that the changes are real rather than analytical. 

The TOR Methodology:  A punch is heated in 7 temperature steps and the evolved CO₂ is measured continuously.  The blackness of the filter is monitored by the reflectance of a laser beam.  The first four steps (O1, O2, O3, O4) are in an oxygen-free atmosphere, while the last 3 steps (E1, E2, E3) have added oxygen.  The O4 and E1 steps are at 550°C; the difference is only the added oxygen.  During the first four steps, 98% of the filters blacken as some of the organic pyrolyzes to elemental carbon instead of evolving as CO₂.  When oxygen is added, this pyrolized organic and the ambient elemental carbon will evolve.  The assumption is that the amount of carbon that evolves before the reflectance returns to the initial value is equal to the amount of pyrolized organic.  This is reported as a separate fraction OP, which is a subset of E1 and E2.  Normally the reflectance returns to the original value during the E1 step, but for 15% of the samples it is in the E2 step. 

To have 8 separate fractions, the elemental fractions can be redefined.  The fractions E1 and E2 fractions are replaced by E1P and E2P, which exclude OP.  The equations are:

E1P = max(0, E1-OP)

E2P = max(0, E1+E2-OP-E1P)

OP is very rarely larger than E1 + E2, so there is no need to calculate an E3P. 

The total organic carbon (OC) is all carbon evolved without oxygen plus the pyrolized organic fraction determined by the reflectance measurement:

OC = O1 + O2 + O3 + O4 + OP.

The total elemental carbon (also called light-absorbing carbon) is all carbon evolved in an oxygen atmosphere after the reflectance returns to its initial value:

EC = E1 + E2 + E3 - OP = E1P + E2P + E3

Note that the separation between OC and EC has only a minor dependence on temperature. 

Methodology:  The discussion of methodology will focus on the EC/OC ratio for a single site, Acadia, which showed one of the largest changes in the ratio.  The left plot in Figure 1 shows the EC/OC ratios for each sample over the 12-year period, along with the regression line.  This shows both a seasonal variation and large variation from sample to sample.  Although the regression coefficient is small (r²=0.07), the large number of independent points (1190) means that the uncertainty in the slope is small.  The uncertainty in the slope is given by

                                                                      Equation 1

where b is the slope and n is the number of points.  In this case, the uncertainty is only 10% of the slope.

The right plot shows four possible methods to smooth the data. 

• Seasonal Mean:  The first method is to calculate the ratio of the mean EC / mean OC for each season.  This eliminates sample-to sample variation but retains the seasonal variation.  (The use of the ratio of means rather than the mean of the ratios is done to minimize problems when OC becomes small or negative.  In addition, there is an asymmetry in using the mean of ratios—the mean of EC/OC is not the inverse of OC/EC.) 
• Annual Mean:  Because of potential problems with missing data, the average of the four seasons is better than a single annual mean—whenever there is a significant seasonal variation, missing data during the high or low season will have a greater influence on the single mean rather than mean of the seasonal means.  The annual mean eliminates both sample-to-sample and seasonal variations. 
• Annual Median:  The third method is to calculate 12 annual medians—the results are very similar to those with 12 annual means.  The advantage of the median is that it minimizes the effect of outliers.  It also eliminates the need for special treatment when OC is small. 
• Running Annual Median:  The running annual median was the method used by White.  This appears to retain a small amount of the seasonal variation.  The problem with the running mean methodology is that the points are not independent.  Is a regression line justifiable?  Certainly, Equation 1 is inappropriate with interdependent points--it is not clear how to calculate the uncertainty in the slope for this method. 

The slopes and uncertainties for the five methods are listed in Table 1.  The important conclusion is that all of the five methods give equivalent results.  The raw data give the lowest uncertainty, but graphically the trend is more difficult to see.  For a table, the regression line for the raw concentrations may be best, but for figures, smoothing the data is preferred.  The annual mean or annual median is the best visually and allows the uncertainty to be calculated.  Which is better?  The annual median is simpler to justify than the mean, especially when examining ratios.  However, the mean is slightly easier to calculate for the way the data is arranged in the data set and was used for the preliminary data processing, so it will be used in the remainder of the report. 

All years in this report are seasonal rather than calendar, from March to February.

Analysis:  Annual means for sites with long records.  This analysis is done in two parts.  The first combines the sites into three geographic regions.  The second looks at the individual sites.  In all cases, the data for 1988 are excluded because of greater noise. 

First, the sites are separated into three regions, Southwest, Northwest, and East.  No differences were observed between southern and northern sites in the east.  The sites in each region are: 

Southwest: Bandelier, Big Bend, Bryce Canyon, Canyonlands, Chiricahua, Grand Canyon, Great Sand Dunes, Guadalupe Mtns, Mesa Verde, Petrified Forest, San Gorgonio, Tonto, Weminuche

Northwest: Badlands, Bridger, Crater Lake, Denali, Glacier, Jarbidge, Lassen Volcanic, Mount Rainier, Pinnacles, Point Reyes, Redwood, Rocky Mountain, Yellowstone, Yosemite

East: Acadia, Great Smoky Mtns, Shenandoah, Mammoth Cave, Brigantine, Lye Brook, Dolly Sods, Okefenokee, Sipsey, Upper Buffalo, Washington DC

Data prior to 1992 are excluded for the East, since only Acadia, Great Smoky Mtns, Shenandoah, and Washington DC were operating a full year before 1992. 

Figure 2 examines the two major composite variables, OC and EC, and the ratio EC/OC.  The slopes divided by the mean values are listed in Table 2.  The OC plot shows no trends for any region.  EC shows a small downward trend in all regions, but with marginal or no significance.  Because of the EC trends, the EC/OC ratio also shows a downward trend in all regions.  Again the significance is marginal at best, with uncertainties that are nearly as large at the slopes. 

The second approach is to examine the trends for individual sites.  The results are shown in Table 3.  The statistics for individual carbon fractions are also included.  Overall, there is no consistent pattern in the EC/OC trends.  Of the 38 sites, 23 have negative trends and 15 have positive trends.  In the Northwest, 8 of the 14 trends are negative.  In the Southwest, 6 of the 13 trends are negative.  In the East, 9 of the 11 trends are negative.  

The trend was considered significant if the slope is greater than 3 times the uncertainty in the slope, which is given by Equation 1.  Five sites of the 38 sites show significant trends in the EC/OC ratio:  Petrified Forest, Jarbidge, Yosemite, Acadia, and Washington DC.  Of the five sites with significant trends in EO/OC, 4 are negative and 1 (Jarbidge) is positive.  These five sites were examined more closely. 

The annual mean data are shown for each of the five sites in Figure 3.  Whenever there is a trend in OC or EC, the trend in the components are shown in Figure 4.

Summary of table and figures for sites with significant trend in EC/OC:

Petrified Forest:  The trend is associated with a downward trend in the EC concentration.  Both E1P and E2P are decreasing.

Yosemite:  This is the only site where the trend in EC/OC is associated more with trends in OC than with trends in EC.  There is an upward trend in OC, mostly associated with an increase in 1999. 

Jarbidge:  This is the only site where the EC/OC ratio is significantly increasing.  The trend is associated with an upward trend in the EC concentration.  Both E1P and E2P are increasing.

Acadia and Washington DC:  Both EC and OC are decreasing, but EC is decreasing faster, resulting in a downward trend in the ratio.  The EC decrease is primarily in E1P, the lowest temperature fraction.  All of the OC fractions are decreasing.

Table 3. Trends for sites with long records, expressed as the slope in the regression vs. time divided by the mean.  The units are % per year.  Cases where the slope exceeds 3 times the uncertainty in the slope are shown as bold.  All based on the 11-year record between 1989 and 1999, except for 7 eastern sites labeled by *.  The sites are arranged from north to south.

Appendix 1:  Artifacts

In the IMPROVE network, the artifacts for each of the 8 fractions are determined from secondary filters.  These are subtracted from the measured values in order to determine the concentrations of the fractions.  OC and EC are determined by adding the fractions.  There is some controversy on how the artifacts from secondary filters should be handled in general.  However, based on the explanation below, there can be no doubt that the correct procedure for IMPROVE is to subtract the artifacts. 

The quartz material generally does not adsorb organic gases from the atmosphere.  However, in the process of pre-firing, some foci are created that allow polar organic molecules to be adsorbed.  Tests done at Davis show that the quartz filter will gain an artifact when exposed to the atmosphere.  The artifact will increase for a few days and then level off to a constant value for filters of equal areas.  The artifact reaches a saturation limit in µg/cm².  In IMPROVE, the filters are exposed to enough atmosphere before sampling so that no additional amount is added during sampling.  Thus, the field blanks are roughly similar to secondary filters, as shown in Figure 5.  Organic gases in the inlet air stream or from volatilization of particles collected on the primary filter will not be adsorbed on the secondary filter.  The artifact will be the same for both primary and secondary filters.  Thus, to determine ambient concentrations, it is necessary to subtract this artifact. 

The field blanks and secondary filters in Figure 5 are not precisely equal.  The difference is largest for the most volatile fractions; obviously some of the artifact is volatilized during sampling.  One would expect the vacant foci would be filled by other carbon molecules.  This would mean that the O3 and O4 fractions should be higher for the secondary filters.  This is not observed—the O4 fractions are equal.  Perhaps particles collected on the filter covers these foci making them are inaccessible to atmospheric organic.  In any case, we would expect the artifact on the primary filter to be better represented by the secondary filter than by the field blank. 

The source of the controversy basically is that many studies handle the quartz filters differently than IMPROVE does.  In many special studies, the primary and secondary pre-fired quartz filters are removed from a sealed storage immediately before being loaded into the sampler and beginning sampling.  There is no time for the filters to adsorb the organic artifact before sampling.  During sampling, the primary filter will acquire organic from ambient gases.  The secondary filter will acquire organic from any ambient gases that are not collected on the primary filter, plus any gases derived from volatilization of organic particles originally collected on the primary filter.  If the assumption is made that the filters will not collect artifact from organic gases in the atmosphere, then the primary filter will gain nothing from the atmosphere and the secondary filter will measure only organics volatilized from the primary filter.  In this case, the correct procedure would be to add the secondary to the primary.  However, this assumption is inconsistent with the observation that a quartz filter will acquire artifact simply by exposure to the atmosphere.  If this assumption is dropped, there is absolutely no way that the artifact on the secondary filter will predict the artifact on the primary filter.  Even for special studies, it would be better to saturate the filter by exposure to the atmosphere prior to sampling. 

If the primary and secondary filters are saturated prior to sampling, then any carbon volatilized from the primary filter will pass through the second filter.  The system will simply not register this loss. 

In the IMPROVE network, the artifact is estimated from the median secondary filter.  Field blanks are collected at all sites and used only as a QA check.  A set of artifacts for each fraction is calculated each season using all secondary filter measurements and used for all sites during that season. 

The early network collected a secondary filter with each sample and analyzed only a small fraction.  This permitted us to monitor any site differences without the cost of analyzing every secondary filter.  The disadvantage was that if a secondary and primary filter were interchanged in downloading the cassettes, and only one filter were analyzed, we would use a secondary value for the ambient value.  Thus, we would incorrectly get near-zero ambient concentrations.  If both filters are analyzed, it is possible to check for this error during data processing.  Since no site differences were observed, we decided that it would be best to collect secondary filters at a few fixed sites and analyze every secondary filter.  Since , secondary filter have been collected at 4 to 8 sites distributed around the country; all are analyzed.  No attempt is made to differentiate between sites.  

In the IMPROVE network, a single set of artifacts is used for all sites in the network.  This methodology requires two assumptions.  The first is that the artifact on the primary filter does not need to be calculated from the artifact on the corresponding secondary filter.  The fact that secondary filters are statistically equal, at least for a given season, indicates the validity of this assumption.  In reality, a composite artifact is better than a one-to-one subtraction because of better statistics.  The second assumption is that the artifacts on secondary filters are independent of the site.  Figure 6 show the comparison between the four sites that had secondary filters for the 5-year period.  In all cases, the medians for the four sites are not statistically different.

The IMPROVE methodology of calculating separate artifacts for each season allows variation from season to season and variation from year to year.  (However, the main reason for using seasonal artifacts is that the data are generated every season.)  Table 4 shows that there is some seasonal variation, with summer somewhat larger than winter for both OC and EC.  The time trends are shown in Figure 7. 

The artifact is subtracted from each measured value before the concentration is calculated.  If the measured fraction is less than the seasonal artifact, the concentration for that fraction will be negative; this is left negative in the data base.  The seasonal artifacts are listed in Table 5.  Note that all values are in µg/filter.  To convert to ng/m³, multiply by 30.  The median artifacts are 50% of the median measured values for O1 and E1, 32%-38% for O2, O3, and E2, 20% for O4, and 5% for E1. 

Trends in the artifacts.  The trends in the artifacts are plotted in Figure 1.  The point at 1990 indicates winter 1989-90.  These values were used for the first two years.  The largest changes have been in O1, the most volatile organic fraction.  There was a decrease in 1991 and an increase in 1995.  Several fractions showed a decrease through 1995, followed by a leveling off. 

One test of the correctness of the chosen artifact is the lower end of the distribution of concentrations.  Figure 8 shows the trend in the concentrations of the 2nd percentile of OC and EC for each year.  (These 2nd percentile is arbitrary; any low percentile would suffice.)  Note that the 2nd percentile concentration is always positive for OC, but is negative for EC during the first half of the network.  If the chosen artifact overestimates the true artifact, the concentration for a given percentile would decrease.  There is no trend in OC, suggesting that the chosen artifact is a consistent representation of the true artifact.  However, there is an upward trend in EC, suggesting a changing relationship between the true and chosen artifact.  If the assumption that the chosen artifact is equal to the true artifact for the years between 1994 and 1999, then the earlier chosen artifact actually overestimated the true artifact.  We might note a confounding factor in this methodology:  some of the negative results in the early years could also reflect larger uncertainties in collection and analysis.  The earlier analysis showed a decrease in the chosen EC artifact over time.  The annual average EC artifact expressed as ng/m³ is also shown.  It is clear that this change is insufficient to account for the lower 2nd percentile concentrations in the early network.