An expanded elemental tracer system for atmospheric aerosol

   I am presently (February 2002) developing and testing an expanded elemental tracer system for atmospheric aerosol that is designed to work for regional sources and on large scales. Ms. Jinghua Guo, a Ph.D. student in atmospheric chemistry at Beijing Normal University who is visiting here for a year or so, is helping significantly, and in fact doing most of the real work on it. I decided to post the outlines of this work so that others could watch as it develops and offer comments as we move along.
    The impetus was my impending trip to Beijing in March 2002 to visit the laboratory of Prof. Tian Weizhi in the Department of Nuclear Physics of the China Institute of Nuclear Energy in Beijing. Prof. Tian has a number of project supported by the International Atomic Energy Agency in Vienna, an arm of the United Nations, and requested an Expert Mission from them to offer advice on sampling aerosol and analyzing it by nuclear techniques. The agency asked if I would do this for them, and I agreed. Prof. Tian then asked me to give two lectures while there. In thinking about possible topics, I decided that this would be a good time to revisit earlier thoughts about developing an elemental tracer system for Chinese aerosol. That led me to see the tracer system in the light of an expanded system that I had begun to develop in 1994 and 1995 but had to put aside for more pressing matters. With Jinghua's eager cooperation, we are pressing forward and creating very positive results. We were able to find more than 40 data sets for Chinese aerosol, and have expanded our total number of data sets to over 150, twice as many as I had before.

A brief history
    Our work in regional elemental tracers goes back to the late 1970s and early 1980s, when we were actively studying Arctic haze. It was an easy matter to show that the Arctic aerosol near the surface in winter was dominated by pollution, as evidenced by enrichment of typical pollution elements like V, Mn, Co, Ni, As, Se, and a host of others. It was much harder to show where the aerosol originated, however. Backward air-mass trajectories didn't work, because the air came to Barrow, Alaska, from over the ice cap where there were nearly no meteorological observations. A few degrees difference in direction from Barrow could create such differences in trajectories that we could not distinguish Asia from Europe from North America.
    The only solution seemed to be to distinguish Asian aerosol from European aerosol from North American aerosol chemically. But no one had tried this before, and there seemed several reasons why it might not work. (See below.) We were encouraged, though, by the fact that aerosol from northeastern North America was more enriched in vanadium than was European aerosol, whereas European aerosol was more enriched in Mn. From these two elements came the Mn/V ratio, and eventually a seven-element pollution-based system (V, Mn, As, Se, Sb, Zn, In) that, in concert with large numbers of measurements of aerosols in possible source regions, showed that systematic differences could be found between regions in North America, Europe, and Asia. By 1985 it was becoming clear that Arctic haze, at least near the surface, came from Eurasian sources rather than from North America or the Far East.
    We later applied the same tracer system to the problem of acid deposition in eastern North America and were able to show that, contrary to the conventional wisdom at the time, the majority of the acidity was not automatically linked to emissions in the Midwest, but rather to the smaller emissions of the Northeast. The effect of those nearby emissions was magnified relative to the distant midwestern sources by omnipresence and lack of dilution.
    To recap, the reasons why a regional tracer system should and should not work can be summarized as follows:

Should not work:
    1. All source regions contain the same kinds of fuels, human activities, and industries.
    2. The bigger the region, the more this is so.

Should work:
    1. There are known regional differences in compositions of fuels and uses of fuels.
    2. There are known regional differences in mixes of industries.
    3. There are known regional differences in soils and the rates that they are lifted into the atmosphere.

Obviously, the "should work" won out over the "should not work" for the seven elements and a few broad source regions.

Reason for expanding the system
    I have wanted to expand the tracer system for fifteen years. The main reason is to try to take advantage of the other elements from pollution, crustal, and marine sources that are measured satisfactorily by neutron activation (our traditional analytical technique) but not made part of the tracer system. Our rule of thumb was always that of the 40 to 45 elements measured by NAA, roughly one-half (20 or so) were measured well enough to quality as tracers. The problems were either that they were pollution elements with too much crustal component (Co or Cr, for example), or else they were coarse-particle crustal and marine elements. I held back from involving these other elements, in site of the fact that a broadened system could in principle deal with more sources more reliably.

Philosophy of the new system
    Eventually I decided that I had to try putting new elements into the system. Because nearly all the candidates were either coarse-particle crustal or marine, or pollution elements with a sizeable crustal or marine component, the new system had to be able to deal with coarse particles in addition to the original fine particles. the only way seemed to be to insert "floating" crustal and marine signatures that could account for those components of crustal/pollution and marine/pollution elements, and could compensate for the more-rapid removal of coarse aerosol during long-range transport to remote areas (by taking on negative coefficients). We could not predict how well floating signatures would actually work; we could only try them.
    The operating principles of this new system differed from the old system in several basic ways. (1) It uses up to 22 elements instead of the previous 7. (2) It restricts itself to total elemental concentrations, and no longer calculates noncrustal V and Mn. (3) It uses individual sites more than regions. (4) It uses simple averages for sites rather than trying to isolate their local contributions (too hard). (5) It uses signatures of certain important types of sources (crust, sea, coal, oil, etc.) in addition to places. (6) It allows negative contributions for some type-sources (crust and sea) to correct for greater fallout of coarse particles (soil and sea salt) during transport. (7) It uses coarse and fine elements, as opposed to only fine before. (8) It begins with all possible signatures (now exceeding 150) and narrows them down progressively. (9) It uses a simplified least-squares fitting and weighting process that I developed, to make it more accessible to anyone who wants to try it.

    At one point, the system used 25 elements: Na, Mg, Al, Cl, Ca, Sc, Ti, V, Cr, Mn, Fe, Co, Zn, As, Se, Br, In, Sb, I, Cs, La, Ce, Sm, Hf, and Th.  These were the ones from the 40 to 45 that were judged to have the best potential as tracers. But the halogens Cl, Br, and I proved to be unreliable, for good chemical reasons, and were dropped, giving a maximum of 22 that still remains the practical limit.

Signatures of sites
   Here is our list of current place-signatures, grouped into eight regions. We believe that data on more sites remain to be found and incorporated.

North America Europe and Russia East Asia Japan and Korea Pacific Coastal Islands
SF Bay Wraymires Ze Dang (Tibet) Sapporo Okushiri (Oku) 
Columbia, MO Lerwick Urumuqi Kawasaki Oku spring (MAM) 
St. Louis, Mo. Chilton Lanzhou Niigata Oku winter (JF) 
Oak Ridge, TN Plynlimon Baotou Sendai Cheju 
Shenandoah Trebanos Xian rural Tokyo Okinawa (Oki) 
NW Indiana Sutton Xian urban Nagoya Oki spring (FMAM) 
Akron, OH Leiston Chengdu Osaka Oki summer (JJAS)
McArthur, OH Styrrup Chongqing Amagasaki Oki fall (ONDJ)
Allegheny Mtn., PA Gresham Jinghong Matsue  
Washington, D.C. Collafirth Beijing urban (Zhuang) Kurashiki Mid-Pacific Islands
Pr. George Co., VA Petten Beijing urban Ube Midway (Mid)
Wye, MD Arran/Bute Beijing suburbs Omuta Mid spring (MAM)
Elms, MD North Sea Beijing April 1989 Nopporo Mid fall (ASONDJ)
Underhill, VT Ghent Beijing May 1989 Nonodake Oahu
West Point, NY Belgium Tianjin Kyoto Hachiman Oahu spring (MAM)
Narragansett, RI Paris Qingdao April 1989 Chikugo-Ogori Oahu fall (SON)
New York City Jungfraujoch Qingdao May 1989 Seoul average  
Portland, ME Sweden Zhuzhou Seoul spring Arctic
Sudbury (Ontario) Finland LeChang Seoul summer Barrow winter
  K-Puszta Guangzhou Seoul fall Laptev Sea
Atlantic Katowice Shanghai Seoul winter Ny lesund
Bermuda (BDA) Bug Linan Mallipo April 1989 Bear Island
BDA summer (JJA) Transported Europe Hong Kong (HK) average Mallipo May 1989 Northern Norway
BDA fall (SON) Moscow HK summer (MJJ)   Norilsk
BDA winter (DJF) W. Siberia HK winter (OND)   Nikel (Kola Peninsula)
BDA spring (MAM) Karasuk Kenting    
Barbados Samarkand Yellow Sea April 1989    
Izana (Canary Is.) Irkutsk Yellow Sea October 1989    
Mace Head (Ireland) Baikal Vladivostok    
    Ho Chi Minh City    

The sites are shown in the map below. We will add arrows to selected sites later.

   Here is our current list of type-signatures. We try not to depend on them because the system is place-oriented.

Allen steam-coal plant
Ondov coal
NIST reference flyash
Hungarian power plant
Dybzhinski ref. flyash
Polish hard coal flyash
Polish brown coal flyash
Refuse incinerator
Sewage sludge incinerator
Copper smelter
Field burning
Street dust
Average crustal aerosol
Saharan aerosol
North American inland crustal aerosol
North American coastal crustal aerosol
Russian crustal aerosol
Far Eastern crustal aerosol
Chinese desert dust

Simplified least-squares fitting procedure
   We are deliberately keeping this aspect of the tracer system as simple as possible so that anyone can understand it and do it for themselves. The suite of signatures is fit to the sample with a standard least-squares procedure where elements are rows (cases) and sites (or types) are columns (variables). Any standard package can be used. We use JMP 3.1 or 4.0 because of its superior exploratory capabilities, but other packages such as Statistica or Excel can be used just as well. EPA's CMB (Chemical Mass Balance) package can be used, too, for those so inclined.
    We also use a simplified procedure for weighting the elements. Weighting compensates for the different levels of concentration of the various elements, which range over some six orders of magnitude. Weighting is also supposed to consider the uncertainties of the values of the different elements. We prefer to deal with this second part later, by examining how the various elements behave in practice. This approach is necessary because many of the data sets do not have reliable uncertainties attached to them, either because they are averages or because the uncertainties are omitted. To bring the various elements to a standard concentration, we have found it satisfactory simply to create a single weighting factor for each element that is the reciprocal of its concentration in the sample to be apportioned. The sample and signatures are all multiplied by these factors before they are inserted into the least-squares program. Although this might seem to weight sources with higher concentration more than those with lower concentrations, it does not, because higher sites just get lower regression coefficients. Our procedure removes only the inherent concentration scale of the different elements, as it should.

Procedures and philosophy
   There are several major options on the road to apportioning a sample among the 150 or so possible Northern Hemispheric sources. Framed as questions they include:

  1. Should we apportion "blind" (without knowing or considering the location of the receptor site) or should we use this information?
  2. Should we use individual place-signatures or construct regional averages?
  3. If place-signatures, should we restrict the solution to one continent or mix signatures from all over?
  4. Should we enter the floating crustal and sea signatures first or last?
  5. How many other type-signatures should we allow (since this is primarily a regional exercise)?
  6. Which sets of elements should be used (crustal, marine, pollution)?
  7. How far should we force the fit? (How many tiny contributions should we allow at the end?)
  8. Should we allow signatures from only genuine source regions or accept regional similarity (from similar receptor sites nearby)?

    All these questions are important, and can be debated at length. Our provisional answers are as follows:

  1. Blind? We feel that in most circumstances it is unfair to insist on blind apportionments, which represent the ultimate challenge ("You give me an aerosol sample; I will tell you where it is from."), one that is also unrealistic. Nevertheless, we are trying to find a procedure that works as well for blind tests as in the real world. We think we are close.
  2. Sites or regions? Although we originally used regional signatures, we came to feel awkward about the arbitrariness of deciding where one region ended and the next began. We are now trying to develop a procedure that combines the best of both worlds, that lets the apportionment define the effective regions for each sample, and then choose the most representative site from each region for the actual apportionment. The key to this seems to be displaying the regional properly at each stage of the apportionment. (See discussion on graphics below.)
  3. One continent or anywhere? The relatively short lifetimes of atmospheric aerosols (days to weeks) imply that sources for a continental site would come overwhelmingly from that continent. In the middle of oceans, however, seasonal averages could be affected by two or more continents. At high altitudes, where aerosol is older and therefore more mixed, even single samples might be from more than one continent. Thus the number of continents to invoke depends on the sampling location.
  4. Enter crust/sea signatures first or last? Although one can argue both ways on this question, our experience has shown that the sea signature should always be entered before the place-signatures, and the crustal signature usually right after the sea signature.
  5. How many other type-signatures? It appears that most type-signatures are making up for spatial gaps in place-signatures. Thus the type-signatures should be used sparingly, and only when they contribute significantly to the solution.
  6. Which sets of elements? This question is not yet fully answered. It appears that the three sets of elements (marine, crustal, pollution) are necessary to best delimit the overall location of the receptor site (ocean vs. continent, for example). Subsequent passes can narrow the list of elements toward pollution and crustal/pollution for better resolution of pollution regions. It is possible that three or more passes will ultimately prove most informative.
  7. When to stop adding signatures? This question is not answered with any sense of finality, either. The practical answer is to stop when nonsensical sources start being called, such as minor type-signatures or place signatures from other continents. We need to work a lot more on this question.
  8. Genuine source signatures or similar nearby receptors? Again, both approaches seem to have their place. The receptor site itself can be most efficiently located by allowing receptor sites with similar properties into the solution. Then the ultimate source of the aerosol can be investigated by limiting the signatures to genuine source regions on the appropriate continent.

Basic stepwise regression procedure
    We use JMP 3.1 or 4.0 for our stepwise regressions. The current series of steps is roughly as follows:

  1. Choose the number of elements to be used. The example below uses all 22.
  2. Prepare a table of weighted signatures for the sample to be apportioned, with the weighting factors being the reciprocal of the elemental concentrations in the sample. The signatures appear as columns, the elements as rows. Multiply the sample and all the other signatures by this weighting factor. This gives a series of 1's for the sample. Prepare the table directly in JMP or in Excel, which is then imported to JMP.
  3. In JMP, open "Analyze," then "Fit model." In the resulting box, choose the sample as Y and the desired signatures as "effects." Constrain the intercept to zero. Click "run model."
  4. A "Current Estimates Table" will appear, with columns as shown below for a shortened list of possible sources for Narragansett. "Parameter" means signature (source), "SS" means the reduction in the residual if the source is entered, and the "F-ratio" is, according to the JMP manual, "the traditional test statistic to test that the term effect is zero. It is the square of a t-ratio. It is in quotes because it does not have an F distribution for testing the term because the model was selected as it was fit." "Prob>F" is, again according to the manual, "the significance level associated with the F statistic. Like the "F-ratio," it is in quotes because it is not to be trusted as a real significance probability." The "F-ratios" provide a working measure of the goodness of fit for each signature to the (Narragansett) sample, with the highest values being the best fits. The table below shows that Portland (Maine), Prince George County (Virginia), and West Point (New York) fit Narragansett best. (The plots in a section below show the F-ratios for the entire suite of signatures.)

  1. Before entering place-signatures into the solution (the least-squares fit), enter sea and crust as floating signatures. This changes the F-ratios to appear as shown below:

  1. The new F-ratios above represent the fit from pollution elements only. Examine the list and choose the place-signature with the best pollution fit (the highest F-ratio) and enter it. That place is West Point (New York). The new F-ratios are shown below.

  1. Examine the list of F-ratios after entering the first place-signature, and see if any other place-signatures can be entered. In this case, Prince George County (Virginia) is the next obvious entry. Entering it produces the revised F-ratios shown below.

  1. Given the goodness of fit (SSE down to 0.32 from the original 22), the apportionment can be stopped.

Elaborated stepwise regression procedure
    Actual apportionments are often more complicated than the simplified example above. The number of place-signatures is far greater (150 or so), and span three continents rather than just North America, as shown for Narragansett. Type-sources also come into more serious consideration.
    The trickiest part usually has to do with treating the different continents. When should potential sources be restricted to a single continent, and when should they be allowed from anywhere? The answer depends on whether we are apportioning blind. If not blind, we restrict the sources to the continent of the receptor (the sample). If the receptor is in the middle of an ocean or otherwise extremely remote, we may allow sources from multiple continents, however. When working blind and getting indications that sources on different continents may fit the sample similarly well, we often insert a branching point into the procedure. If for example the sources could reasonably come from Europe and North America, we first explore Europe to its full potential (find the best fit with only European sources) and then do the same for North American sources. The continent with the best fit is then considered the source. In the unlikely event that two continents, or even three, fit the sample equivalently, we can draw no firm conclusion about its source. This seems to happen roughly 10% to 20% of the time.

Illustrative apportionments
   As examples of signatures that are fit unambiguously and well, we offer Narragansett (Rhode Island), Elms (Maryland), Wye (Maryland), Lerwick (United Kingdom), Northern Norway, Barrow (Alaska) winter, Hong Kong, Okinawa spring, Amagasaki (Japan), and Midway spring. Because the discussions of each apportionment require considerable space, we have given them separate sections.
    Narragansett. The Narragansett signature is apportioned so well because it is both measured well (several hundred samples) and is surrounded by other well-measured signatures. Although its step 2 (after sea and crust have been entered) is clearly North American, we show trial apportionments with European and Asian sources for illustrative purposes.
    Elms. Elms is located in Maryland, part of the northeastern United States, in the Chesapeake watershed. Its aerosol is apportioned very well to nearby source signatures, and at the same time illustrates the self-correcting nature of the apportionment process.
    Wye. Wye is also in Maryland, near Elms. Its aerosol is likewise apportioned very well to multiple place-sources in the Eastern United States. Threads for Europe and Japan/Korea are followed because they are indicated in step 2, but they do not pan out.
    Lerwick. Lerwick is a site in the Shetland Islands that is one of several sites in the United Kingdom characterized by the Atomic Energy Research Establishment (AERE) in the 1970s and 1980s. As you might expect, it is fit clearly to nearby sources in the UK and Europe. It is best fit to Collafirth, another site in the Shetland Islands. Of course, Collafirth is not really a source, but a parallel receptor to Lerwick. (See point 8 under "Procedures and Philosophy" above.)
    Northern Norway. Northern Norway lies between Europe and the Arctic both geographically and meteorologically, and there is every reason for its aerosol to show characteristics of both those other types. The results of the apportionment confirm this expectation only partway, though: it looks more European than Arctic. To reach this answer, we carefully tested the three possible source areas indicated by step 2 of the apportionments: Europe, Japan/Korea, and the Arctic. Only Europe survived solidly. Along the way, we found that two of the three strands self-corrected themselves out of the answer.
    Barrow winter. The aerosol of Barrow, Alaska, during winter is known as Arctic haze. We now know that it is derived largely from pollution aerosol of midlatitudes. The question is whose midlatitudes? Europe, Asia, North America, or some combination? Our earlier seven-element tracer system showed that it was Eurasian. The extended system has the potential to do better, but for a long time we though it had to be used in a series of steps, a kind of geographic leap-frog, for example Barrow to Norway and Norway to Europe and Russia. The apportionments presented here go directly back to Eurasia for the first time, and confirm the mix of European and Russian sources that we had deduced in the 1980s.
    Hong Kong. Surprisingly, Hong Kong's aerosol is influenced markedly by aerosol from Mainland China, particularly during the dry season of winter. It therefore comes as no surprise that our average Hong Kong signature is fit well by Guangzhou and Kenting (Taiwan), 0.59, in addition to crust and sea, of course. The next two signatures that enter, Cheju (island south of South Korea) and Mallipo (west coast of South Korea) are from farther north, and improve the fit only slightly (to 0.42).
    Okinawa spring. This aerosol is fit simply and directly by two essentially equivalent Asian threads.
    Okinawa fall. Okinawa fall is fit similarly to Okinawa spring, except that the second Asian thread self-corrects to the first Asian thread.
    Amagasaki. This is one of a dozen Japanese cities for which we have good signatures. Its profile is fit nicely to other Japanese cities and nearby East Asian sites.
    Midway spring. This island location in the mid-Pacific experiences large amounts of crustal dust mixed with smaller amounts of pollution aerosol during spring. We have now traced them both back to China.

    As examples of signatures that are not presently fit well, we offer Jungfraujoch (Swiss Alps) and Lin'an (southeastern China). The Jungfraujoch case may be a bit misleading, for the obvious solution failed (Japan), whereas a more subtle (hidden) solution that used closer sources (UK and surroundings) did quite well. Some of the difficulties in fitting Jungfraujoch may have arisen because its high altitude (4000 m) mixes aerosols from all over, possibly including the Atlantic, and because it lies well south of any of our European signatures. For example, we do not have signatures from southern France, anywhere in Spain or Portugal, or even the Mediterranean Basin, for that matter. At present, we have no idea why Lin'an does not work.

Summary of step 2 apportionments
    Since most of the final apportionments work out pretty much as indicated at step 2, it is illustrative to compare the step 2 plots for the samples discussed above. They are shown in a separate section.

Why do the signatures work?
    At this time it is impossible to give a full answer as to why the signatures work as well as they do. The explanation seems to lie in systematic differences in proportions among the elements. Although it would be convenient to be able to identify one or two elements as key to making signatures unique, our experience does not support this view. As a result, we know more that they work than how they work.

Aspects of the tracer system to be tested

  1. Which classes of elements are most important (crustal, marine, pollution)?
  2. Are any individual elements more important than others, and if so, why?
  3. Are there any groups of elements that are most important to regional tracing, and why? (Example: nonmetals from the upper right of the periodic table.)
  4. Is any particular analytical technique best suited for extended elemental tracing? What combinations of analytical techniques are best?
  5. Can signatures developed from different particle sizes of aerosol be used simultaneously in a tracer system? (TSP, PM10, PM2.5, etc.)
  6. Is there a minimum number of samples needed to provide a reliable signature?
  7. What is the preferred spacing of samplers when developing a regional signature? In other words, when does information begin to become redundant as sampling sites are moved closer together?
  8. How large is the seasonal effect on place-signatures, that is, the difference in signatures with season? Are some places more stable with season than others, and if so, why?
  9. Do any existing signatures contain analytical artifacts?
  10. Is there a discernible "laboratory effect," that is, where samples from one analytical laboratory tend to cluster together because they all contain the same analytical bias?
  11. Is there a limit to the distance over which regional tracers work?
  12. What is the maximum fraction of missing data that can be filled in and still provide signatures satisfactory for extended tracer systems?
  13. What is the best weighting system for least-squares fitting, or are they all functionally equivalent?
  14. How reliably can the technique be applied to short-term samples, such as daily, as opposed to the long-term averages used in developing this system?
  15. How fast does tracing ability degrade when applied to samples significantly shorter than a day, say an hour or so? How much would such degradation depend on the remoteness of the sampling site?
  16. To what extent can old place-signatures (say 10 years or more) be used with current signatures? (In other words, how fast do typical place-signatures change?)
  17. Can we use this tracer system to validate Prof. Christian Junge's rule of thumb that aerosols are fully mixed above 5 km?

Next steps
    We are currently documenting the extent to which the system works with its imperfect suite of signatures culled from over 30 years of measurements by groups all over the Northern Hemisphere, nearly none of whom had tracer systems in mind as they were generating their data. After we learn just how far the system can be pushed, we will turn to exploring why it works. That in turn should allow us to begin to lay out steps for broadening it (both geographically and in terms of constituents) and improving it (specificity, distance, etc.). All the while, of course, we will be seeking data sets to add to the system. We believe that there may be as many more usable data sets as we have in the system now. Adding data sets will give us the chance to learn how far they can be from the ideal and still add value to the tracer system.

Back to Home Page