Environ. Sci. Technol. 2010, 44, 1690–1697
PCDD/F Source Apportionment in the Baltic Sea Using Positive Matrix Factorization K . L . S U N D Q V I S T , * ,† M . T Y S K L I N D , † P. GELADI,‡ P.K. HOPKE,§ AND K. WIBERG†
Received October 6, 2009. Revised manuscript received January 8, 2010. Accepted January 18, 2010.
Positive Matrix Factorization (PMF) was used to identify and apportion candidate sources of polychlorinated dibenzo-p-dioxins and polychlorinated dibenzofurans (PCDD/F) in samples of offshore and coastal surface sediments from the Baltic Sea. Atmospheric deposition was the dominant source in offshore and pristine areas, in agreement with previous studies. Earlier chlorophenol use and a source suggested origins from pulp and paper production and related industries were identified as important coastal sources. A previously presumed major source, chlorine bleaching of pulp, was of only minor importance for modern Baltic surface sediments. The coastal source impacts were mostly local or regional, but pattern variations in offshore samples indicate that coastal sources may have some importance for offshore areas. Differences between subbasins also indicated that local and regional air emissions from incineration or other high-temperature processes are more important in the southern Baltic Sea compared to those in northerly areas. These regional differences demonstrated the importance of including offshore sediments from the Bothnian Bay, Gulf of Finland, and other areas of the Baltic Sea in future studies to better identify the major PCDD/F sources to the Baltic Sea.
Introduction Multivariate receptor modeling uses statistical methods to identify emission sources based on pollution fingerprints and quantifies their contributions. Varying receptor modeling techniques have been used since the 1960s, primarily within the field of air pollution, and several reviews are available (1, 2). Applications of receptor modeling have expanded over the years to include studies of environmental pollutants in sediments, whereby emission sources for metals and organic compounds have been apportioned (3-10). Recently, Positive Matrix Factorization (PMF, ref 11) has emerged as the most commonly used approach. Hitherto, few studies have applied PMF to source apportionment of polychlorinated dibenzop-dioxins, polychlorinated dibenzofurans (PCDD/Fs, refs 12, 13), and polychlorinated biphenyls (PCBs, refs 14-16) in sediments or resuspended sediments. * Corresponding author phone: +46 90 7865000; fax: +46 90 7867655; e-mail:
[email protected]. † Department of Chemistry, Umeå University, SE-90187 Umeå, Sweden. ‡ Unit of Biomass Technology and Chemistry, Swedish University of Agricultural Sciences, SE-90183 Umeå, Sweden. § Department of Chemical and Biomolecular Engineering and Center for Air Resources Engineering and Science, Clarkson University, Potsdam, New York 13699. 1690
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 44, NO. 5, 2010
In the Baltic Sea, a large brackish water inland sea in northern Europe, the sources for elevated PCDD/F in biota are not fully understood. Despite efforts to control PCDD/F pollution in the Baltic environment, the decreasing concentration trend seen in guillemot (Uria aalge) eggs in the 1970s-1980s has leveled off since the mid 1980s (17). On the basis of the levels found in sediment throughout the Baltic Sea (18), the most abundant PCDD/F congener in Baltic herring, 2,3,4,7,8-penta-CDF, has been suggested to be of atmospheric origin. Other studies also suggest atmospheric deposition as the overall most important current source in the region (18, 19). However, comprehensive congener pattern comparisons between offshore surface sediment and atmospheric deposition indicate that it is not the only important source (20). Su and Christensen (8) attempted to quantify PCDD/F sources to the Baltic Sea sediments applying the chemical mass balance model to a single sediment core with source types suggested by Kjeller and Rappe (21). Their model evaluation was based on homologue profiles and suggested that in the mid 1980s, municipal solid waste incineration contributed approximately 45% of the PCDD/ Fs to the sediments, coal combustion contribited 37%, and pentachlorophenol contributed 18%. However, one disadvantage with most chemical mass balance models is that all important sources and their profiles must be identified before beginning the analysis. Sources not included and variability in candidate source profiles or, even worse, biased profiles may lead to inaccurate conclusions. In a previous study, PCDD/F concentrations in surface sediments sampled along the Swedish coast and in offshore areas were determined (20). A principal component analysis (PCA) of the data identified indicator congeners for various sources types. In the coastal areas, pulp and paper production and other wood industry-related activities were found to be of major importance (22). This conclusion was based on a high abundance of indicator congeners as well as similarities between sediment and source candidate PCDD/Fs fingerprints. It was suggested that major sources included chlorophenol preservatives and a source plausibly connected to pulp and paper industries (22). Profiles of atmospheric deposition were similar to the profiles in offshore sediments, supporting the hypothesis of the importance of this source. However, the PCA indicated several source options, e.g., longrange atmospheric transport and regional high-temperature processes, and the relative importance of the various options could not be estimated because PCA results cannot be interpreted quantitatively. The current study uses PMF, a multivariate method that yields quantitative candidate source information. This data analysis tool was applied to the data set in order to estimate the relative importance of various sources. The aim was to identify and apportion the candidate offshore and coastal Baltic Sea sediment PCDD/F sources. The large size of the field data set (62 variables in 143 samples) contrasts to previous PMF studies of PCDD/F sources in sediments (85 variables in 35 samples and 83 variables in 19 samples, respectively, refs 12, 13).
Methods Sampling and Chemical Analysis. Surface sediments were collected along the Swedish coast and in offshore areas of the Baltic Sea using gravity corers. The sampling and analysis has been described in detail elsewhere (20). In brief, a total of 146 dried sediments and one settling particulate matter sample were toluene-extracted in a Soxhlet-Dean-Stark setup. Cleanup consisted of multilayer silica column chro10.1021/es9030084
2010 American Chemical Society
Published on Web 02/01/2010
matography and sulfur removal using copper followed by fractionation on a carbon/Celite column. It has been shown that comprehensive congener specific data is more advantageous for source identification than only using the 17 2,3,7,8substituted PCDD/F congeners traditionally analyzed (23). Therefore, a total of 72 chromatographic peaks were resolved on a DB-5 column using gas chromatography-high resolution mass spectrometry (GC-HRMS) (20). The chromatographic peaks correspond to individual congeners or groups of congeners. In figures and tables, congener names are abbreviated as defined in Sundqvist et al. (22): (i) degree of chlorination is expressed as T, Pe, Hx, Hp, and O for tetra-, penta-, hexa-, hepta-, and octa-substituted congeners; (ii) dibenzo-p-dioxins ) D and dibenzofurans ) F; and (iii) commas between numbers are excluded. In addition, chromatographic peaks are represented by only one congener, and addition of an asterisk indicates coelution with one or several congeners (e.g., 1378TF*). Receptor Modeling and Positive Matrix Factorization (PMF). Multivariate receptor modeling includes different statistical tools used to reconstruct plausible source patterns based on environmental data from receptor sites (in this case, the sediment samples) and to calculate their contribution to each receptor sample (1, 2). These techniques are based on the assumption that the pollutants from different sources are linearly additive and that no degradation or other pattern alteration takes place during the transport from source to receptor. Pattern changes can, however, still be considered as long as processed patterns are known or can be estimated. PCA can compress a large number of variables into a few components, identify similarities between samples, and find marker congeners and variables, but the results cannot be interpreted quantitatively. The PMF model describing the contributions from p independent sources is described by eq 1, where xij is the fraction (or concentration) of the ith congener in the jth sample of the original data matrix X, fik is the the fraction of the ith congener in the kth factor, gkj is the contribution of the kth factor on the jth sample, and eij is the model residual for the ith congener in the jth sample.
For missing data, they were set to 4 × average concentration, and for the remaining data points, the uncertainty was estimated to be DL + 0.1 × concentration. This treatment of uncertainties, missing data, and values below DL were in line with recommendations from Polissar et al. (with minor changes (25)). Finally, all data points in the two matrices were divided by the total concentration (normalization to unit concentration) in each sample to eliminate the effect of concentration differences and other sample specific variables such as organic content. On the basis of a previous PCA analysis (22) and data evaluation, some samples and PCDD/F congeners were excluded from the modeling. As in the earlier study, one sediment sample was excluded because of its extreme congener pattern consisting almost entirely of PCDFs. Three samples with more than 40% of the data below the DL (signalto-noise ) 3) were also excluded. Chromatographic peaks, for which more than 2/3 of the data did not pass the positive identification criterion (the level exceeded DL and the isotopic ratios were satisfying) (20), were also excluded. After these adjustments, the data set consisted of 143 observations (142 surface sediments and 1 sample of settling particulate matter) and 62 variables (chromatographic peaks). Diagnostic Tools. The results of the PCA analysis can be used as an indicator of the factor number, although not as an exact number but providing a possible range. Using this range, factor models within that range were calculated. In these models, the stability of the profiles in repeated calculations is a good diagnostic tool (26). Ten repeated calculations with random starting points were used in the present study. Another diagnostic tool often used is the Coefficient of Determination (COD) defined in eq 3, where xij is the measured concentration, yij is the modeled concentration, and xji is the average measured concentration of the ith congener.
SSerr COD ) R2 ) 1 ) SStot
∑ (x
ij
- yij)2
i
∑ (x
ij
(3)
- xj i)2
i
p
xij )
∑f
ik
× gkj + eij
(1)
k)1
The objective function, Q (eq 2), is the weighted sum of squares of the difference between the model output and the original data where sij is the uncertainty of the ith congener in the jth sample in the original data set consisting of m congeners and n samples. Q is minimized in order to obtain an optimal solution.
∑ ∑ [s ] m
Q)
n
i)1 j)1
eij
2
(2)
ij
PMF is a factor model and subject to rotational ambiguity (24). No unique solution can generally be obtained. PMF2 (11) was used for the modeling presented in this work. The source candidates calculated by the model were identified through the presence of marker congener and overall similarities with actual source patterns (subjective examination). No statistical evaluation was performed because too few reports are available of full congener analysis of source pattern. Pretreatment of the Data. Prior to modeling, values below the detection limit (DL, signal-to-noise
