JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY DECEMBER 1991 VOL. 6 637 Determination of Geographic Origin of Agricultural Products by Multivariate Analysis of Trace Element Composition* Robert S. Schwartz and Le 1. Hecking US Customs Service Research Division Office of Laboratories and Scientific Services 1301 Constitution Avenue Room 71 13 Washington DC 20229 USA Samples of orange juice pistachio and macadamia nuts were analysed for selected elements using inductively coupled plasma atomic emission spectrometry and atomic absorption spectrometry. Discriminant analysis was used to form mathematical models for predicting the geographic origins. The accuracy of these models was assessed using re-substitution and cross-validation analysis for the orange juice and macadamia nut samples and by the calibration set-prediction set method for the pistachio nut samples.Perfect prediction results were achieved for the pistachio nut samples. Re-substitution analysis results for the orange juice and macadamia nut samples indicated prediction accuracies of 96 and 98% respectively while cross-validation results indicated 88 and 78'30 respectively. Separations were visualized by performing canonical discriminant analysis involving the same elements as in the discriminant analysis and plotting canonical scores for the first two or three canonical functions. Complete resolution of all samples by geographic origin was achieved for all three commodities. Keywords Geographic origin; orange juice; pistachio and macadamia nuts; multivariate data analysis; trace element Determination of the geographic origin of imported mer- chandise is an analytically challenging problem that is currently the focus of much attention within the US Customs Service.The Customs Service enforces trade related laws rules and regulations many of which specify treatment based on geographic origin. The three commodi- ties discussed in this manuscript orange juice pistachio and macadamia nuts all require just such origin-based treatment. For orange juice the relevant government programme is the Caribbean Basin Initiative (CBI) which eliminates the tariff on products from any one of 22 designated countries in the Caribbean Basin. This situation provides obvious economic incentive for fraud by trans- shipment of non-CBI merchandise through CBI countries thus evading payment of duty.Orange juice a high-volume item carrying a duty of $0.35 per gallon of single-strength juice is of particular concern due primarily to the prox- imity of Brazil a non-CBI country which exports a great deal of this commodity to the Caribbean Basin. Of major concern therefore is the ability to differentiate CBI orange juice from that of other countries particularly Brazil. For pistachio nuts the relevant issue was the embargo on products from Iran pistachio nuts being one of the more important Iranian exports. The other major producers are Turkey and the USA particularly California so it is important to be able to distinguish pistachio nuts from these three sources. It should be noted that a significant portion of pistachio nuts imported into the USA are Californian nuts which have been processed abroad.Finally Customs interest in macadamia nuts arose from information that South African macadamia nuts were being trans-shipped to the USA through nearby African countries in violation of the ban on the importation of South African agricultural products one of the sanctions specified in the 1986 Anti-Apartheid Act. A powerful method for the determination of the geo- graphic origin of agricultural products is multivariate statistical analysis of the data provided by analytical instruments such as chromatographs and spectrometers *Presented in part at the 1990 Winter Conference on Plasma Spectrochemistry St. Petersburg FL USA 8th- 13th January 1990. which have the ability to determine more than one component at a time in a sample.If these components have sufficient discriminatory power the set of their concentra- tions will form a characteristic pattern or 'fingerprint' relating to the geographic origin of the sample. Multivariate data analysis provides the ability to detect these patterns and is essentially helpful when the number of components necessary to differentiate samples from different geographic origins increases. This methodology has been used to determine the geographic origins of honey,' olive oil,2 orange juice3 and and has also been used to determine adulteration in orange juice.* It should be noted that the orange juice study cited in ref. 3 only considered juice from Florida and Brazil. The components determined in these studies havevaried but can be broadly classified as either organic substances or elements.The use of elemental concentrations for the determination of the geographic origin of vegetable matter and foods derived therefrom has a number of advantages when compared with the use of organic substances. Firstly organic substances must be manufactured by the plant and therefore must be at least in part under genetic control so that the magnitude of the influence of geographic area in determining the levels of these substances might tend to be problematic. Elements on the other hand must be absorbed by the plant from the soil in which it is grown. It is known that the levels of elements present in plant tissue is directly dependent on the levels of these elements in their growth media,9 which in most instances is the soil.To the extent that the level of elements in the soil are characteristicofthe region it might reasonably be expected that the levelsfound in plants grown in a particular soil will reflect this. In addition elements do not decompose on storage as do organic substances produced by the plant. In this study the application ofthis methodology to orange juice is extended by considering Caribbean Basin countries in addition to Brazil. Two new applications are also presented the determination of the geographic origin of both pistachio and macadamia nuts. Experimental Instrumentation The elements B Ba Ca Cu Fe Mg Mn P and Zn were determined by inductively coupled plasma atomic emission638 JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY DECEMBER 1 99 1 VOL.6 spectrometry (ICP-AES) and K and Rb were determined by atomic absorption spectrometry (AAS). The ICP-AES mea- surements were obtained using an Instrumentation Labora- tory Model P L 100 inductively coupled plasma atomic emission spectrometer (Thenno Jarrell Ash Franklin MA USA) equipped with a standard Instrumentation Labora- tory cross-flow nebulizer and torch. Atomic absorption measurements were obtained using a Perkin-Elmer Model 5000 atomic absorption spectrometer (Perkin-Elmer Nor- walk CT USA) equipped with a single-slot burner a K hollow cathode lamp (Fisher Scientific Washington DC USA) and a Rb electrodeless discharge lamp (EDL) and EDL power supply (Perkin-Elmer). Operating conditions for the ICP-AES and AAS instruments are listed in Table 1.Reagents Water for the preparation of all solutions had a resistivity of 18 MSZ and was obtained from a reverse osmosis-de- ionization system (Millipore Bedford MA USA); nitric acid was Instra-Analyzed (J.T. Baker Phillipsburg NJ USA); caesium chloride was Hi Pure grade (Spex Edison NJ USA); lanthanum chloride was of a grade suitable for AAS analysis of alkaline earth elements (Fisher Scientific Pittsburgh PA USA); and Triton X-100 was commercially available material (lot number 75 1560 Fisher Scientific). Multi-element working standard solutions for use in ICP and AA analyses were prepared from suitable dilutions of stock solutions containing all of the elements of interest. Stock solutions were prepared from 1000 pg ml- aqueous standard solutions of each element (Fisher Scientific) for analytes with final concentrations of less than 10 pg ml-l (B Ba Cu Fe Mn Rb and Zn) and from high-purity (99.99% or higher) compounds or metals (Spex) for the analytes Table 1 Operating conditions ICP-AES- Plasma power level Gas flow Coolant Auxiliary Nebulizer Boron Barium Calcium Copper Iron Magnesium Manganese Phosphorus Zinc Observed wavelength Observation height AAS- Flame Gas flow Acetylene Air Potassium Rubidium Potassium Rubidium Observed wavelength Slit width 1.2 kW 15 1 min-l 0.5 1 min-' 0.45 1 min-l 249.77 nm 455.40 nm 317.93 nm 324.75 nm 259.94 nm 279.08 nm 257.61 nm 213.62 nm 206.20 nm 10- 14 mm above load coil (optimized for each element) Air-acetylene 2.0 1 min-l 15.5 1 min-l 404.4 nm 780.0 nm 0.7 nm 1.4 nm present at higher concentrations (Ca K Mg and P).All standard solutions were 1 mol dm-3 in HN03. Samples There were 27 orange juice samples studied all of which were obtained as frozen concentrated material having Brix values of approximately 65". (Degrees Brix are used as a measure of the concentration of an orange juice sample and is numerically equal to the % m/m of soluble solids measured as sucrose. lo Trace element concentrations were determined in the concentrates and were then adjusted to the values that would have been obtained if the juices had first been diluted to a Brix value of 1 1.8" which is the value stipulated in US regulations as that required for unconcen- trated orange juice.) These samples were obtained from commercial shipments entering the USA.The geographic origins as indicated by the documentation accompanying these samples were as follows Honduras 1; Belize 4; Brazil 12; Jamaica 8; Mexico 1; and USA 1. There were 33 pistachio nut samples studied with the following origins Afghanistan 2; California 20; Iran 6; Sicily 1; and Turkey 4. Of these 6 Californian 3 Iranian and all 4 Turkish samples had authenticated origins; all other samples were obtained from commercial shipments and origins were accepted as attributed from the accom- panying documentation with the exception of 3 samples which were determined to be Iranian by two independent methods. Finally 40 macadamia nut samples were studied with the following origins Australia 9; Costa Rica 2; Guatamala 10; Malawi 12; South Africa 3; and Zimbabwe 4.Of these 1 Australian 1 Costa Rican 3 Guatamalan 3 Malawian and 3 South African samples had authenticated origins; the remaining samples were obtained from commercial ship ments with origins accepted as attributed from the accom- panying documentation. Sample Preparation All samples were prepared for analysis by microwave- assisted digestion using an MDS-8 1 D microwave digestion system (CEM Matthews NC USA). Samples were weighed into 120 ml Teflon PFA digestion vessels and 20 ml of concentrated nitric acid was added to each sample. [Teflon PFA (perfluoroalkoxy) is manufactured from tetrafluoroe- thylene with a fully fluorinated alkoxy side chain.] Sample sizes were as follows 2 g of orange juice concentrate; 1 g of pistachio nutmeat; or 0.7 g of macadamia nutmeat.Twelve samples of orange juice or pistachio nuts or six samples of macadamia nuts were processed at one time. Each vessel was then fitted with a pressure relief valve and cap and sealed using the capping station of the MDS-8 1D system. For the orange juice samples the vessels were left unsealed for approximately 30 min to allow a preliminary exother- mic reaction to be completed. The carousel of the mi- crowave oven was then loaded with the vessels and the samples were processed using the appropriate programme as listed in Table 2. At the conclusion of the digestion programme the vessels were allowed to cool to room temperature uncapped and the solutions were reduced by evaporation to a volume of approximately 1 ml in the microwave oven. For the pistachio nuts 2 ml of 3Ooh hydrogen peroxide were added prior to evaporation which was initiated after the initial bubbling had stopped.Liquid residues were taken up in either 1 mol dm-3 HN03 for orange juice and pistachio samples or 1 mol dm-3 HN03-O. 1% m/v Cs (as CsCl) for the macadamia samples transferred into 25 ml calibrated flasks and made up to volume with the same solvent. All ICP analyses were conducted on portions of these solutions that had beenJOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY DECEMBER 199 1 VOL. 6 639 Table 2 Power-time programme for microwave oven digestion of orange juice concentrate pistachio nuts and macadamia nuts. Values given are for time (in min)* Power level (%) Commodity 50 100 Orange juice concentrate 15 10 10 12 - Pistachio nuts 10 10 10 10 - Macadamia nuts 15 10 10 15 15 to cool for 5 min and then vented.*At the conclusion of each time peiod the vessels were allowed made 0.05% v/v in Triton X-100. The determinations of K in orange juice and pistachio nut samples were conducted on 1 0-fold dilutions of the final solution using a 1 mol dm-3 HN03-0.2% m/v La (as LaC13) diluent; determinations of K in macadamia nut samples were conducted on the final solution without modification. The determinations of Rb in all instances were conducted on the final solution without modification. All analyses were conducted versus multi- element standards prepared using the same diluents. Multivariate Data Analysis Multivariate analysis of elemental concentration data was performed on an IBM PC/XT personal computer using SASISTAT version 6 software (SAS Institute Cary NC USA).The two major techniques used in this study were discriminant analysis and canonical discriminant analysis. Discriminant analysis uses a sample set with known group membership called a calibration set to form a mathema- tical model which is used for predicting group membership of unknown samples. A related technique known as stepwise discriminant analysis was used to help select a subset of elements from amongst those determined having good discriminatory ability to form the calibration model. Canonical discriminant analysis is a dimension reduction technique that is used to visualize group separations in 2- dimensional plots. This technique uses linear combinations of the original elemental concentrations to form a new set of variables referred to as canonical discriminant functions.The number of canonical functions is significantly lower than the number of original variables the elements in this instance. Each sample can be represented by its set of scores on these functions and as most of the variance or information content is concentrated in the first two or three functions this information can be simply displayed in 2- dimensional plots of scores for each of the samples on the first two or three canonical functions. In these plots if the elements chosen are sufficiently discriminatory samples from the same area will cluster together whereas samples from different areas will be separated. There are a number of ways to demonstrate the validity of a discriminant analysis.The most straightforward way is to divide the samples into two groups a calibration set and a prediction set. The model formed from the calibration set is then used to predict the group membership of the prediction set which is then compared with actual group membership. To be successful the calibration set should include a statistically sound representation of samples from all classes of interest. Dividing the samples into two groups however reduces the number of samples available to form a statistically sound calibration set. When the number of samples available is limited the statistical soundness of the calibration model can be impaired and other validation techniques which make maximum use of the available samples to form a calibration model can be applied.Two such techniques used in this study are re-substitution analysis and cross-validation analysis. In a re-substitution analysis all the samples are used for the calibration model the predictive ability of the model is then evaluated by using it to ‘predict’ the group membership geographic origin in this instance of the samples. This however gives an optimistic estimate as the same samples contributing to the model are the ones being predicted. A more realistic estimate of predictive ability is obtained from cross- validation in which the model to be used for the classifica- tion of a given sample includes all samples except the one being classified. As none of the samples being classified contributes to the model used for classifying it the positive bias is eliminated.Results and Discussion Orange Juice Originally it was planned to use as a calibration set a group of samples of frozen orange juice concentrate supplied by the Food and Drug Administration (FDA) which had authenticated origins. This set was to be used for the determination of the geographic origin of orange juice concentrates obtained by the US Customs Service from normal commercial shipments. In practice however this worked out poorly. The calibration set formed from the FDA samples had poor predictive ability for the Customs samples especially in the important category of Brazilian juices where all 12 Customs samples were misclassified. It was found that the trace element levels for five of the seven elements used in the multivariate analysis were significantly higher for the Customs samples these were Ba Ca Mn P and Rb.The differences observed for Ba were particularly dramatic the Customs Brazilian samples averaged 0.37 ppm of Ba while the FDA Brazilian samples averaged 0.06 ppm of Ba. As mentioned later Ba is one of the more important elements in determining the geographic origin of orange juice. It was felt that these differences were most likely to be due to differences in processing the FDA samples had been processed in special non-commercial facilities. Although further study of the source of these differences was not pursued it is possible that the differ- ences in processing could have led to the differences observed in the levels of trace elements. It was then decided to use the entire set of 27 Customs samples as a calibration set taking the geographic origin from the accompanying documentation and to assess the accuracy of the calibration model by using re-substitution and cross-validation techniques.A discriminant analysis was performed on this sample set using B Ba Ca K Mn P and Rb. The order of discriminatory power for this group of elements was B> Mn> Ba>Rb>>P>Ca>K. Results of the re-substitu- tion and cross-validation analysis are shown in Table 3. It can be seen that the discriminating power of the model is good and that as expected re-substitution results are somewhat more optimistic than those of the cross-valida- tion technique. As noted in Table 3 the cross-validation Table 3 Discriminant analysis results for orange juice samples Brazilian result7 Correct result* Number False Type of analysis (%I correct positives Re-substitution 96 12 1 Cross-validation$ 88 1 1 1 *From all orange juice samples tested n=27.f n = 12. $Results adjusted for areas represented by only one sample.640 JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY DECEMBER 1991 VOL. 6 5.0 .- c C 3 2.5 c - lu .- s O - 10 results exclude samples that are the sole representatives of a single area; this is because in this type of analysis the sample being classified is excluded from the model making it impossible to classify a single sample representing one area correctly. In order to visualize group separations a canonical discriminant analysis was carried out on this sample set using the data for the same elements from the discriminant analysis.In Fig. 1 a plot of the scores for all samples on the first two canonical functions is shown; good separation is observed for all samples except for the noticeable overlap between Brazilian and Jamaican samples. It should be realized however that these plots are only 2-dimensional representations of multi-dimensional data and do not always contain all the relevant information in the data. In this instance it was found that the Brazilian and Jamaican samples could be separated according to their scores on the third canonical function. This is illustrated in Fig. 2 which - - J J J J L 10.0 1 u I 10 8 - 6 - C .o 4 1 2 - c C 0 C .- e 0 - 0 -0 0 c -2 -4 $ -6 -8 7.5 1 M I I I - T T T T cc - C C C - C - S - J I I I I 1 I -20 -15 -10 -5 0 5 10 15 $ -5.0 -7-5 i J J L L L .LL L JLL L L 8 6 8 8 H -10.0 I I I I I I I I -7.5 -5.0 -2.5 0 2.5 5.0 7.5 10.0 12.5 First canonical function Fig.1 Plot of scores on the first two canonical functions for the orange juice samples. Country codes B Belize; H Honduras; J Jamaica; L Lower Siio Paulo Brazil; M Mexico; and U USA c L 2 L .- E l 2 0 -3 t L L L L M L J R E E 6 H J 1 J J J - 4 1 J J -7.5 -5.0 -2.5 0 2.5 5.0 7.5 10.0 12.5 First canonical function Fig. 2 Plot of scores on the first and third canonical functions for the orange juice samples. Country codes as in Fig. 1 AJOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY DECEMBER 1991 VOL. 6 64 1 the excellent separation achieved for the most important Iranian category. Macadamia Nuts Owing to the limited number of standards available it was decided to combine the samples and standards into one large set rather than divide them into a calibration and a prediction set as had been done with the pistachio samples.A discriminant analysis was performed using the entire 40- sample set as a calibration model using data for the following nine elements Ba Ca Cu Fe K Mg Mn P and Zn. The order of discriminatory power for this set of elements was Zn>Cu>Ca=Mg>P=K>Mn=Fe-Ba. The validity of this nine-element model was checked by performing a re-substitution and a cross-validation analysis on the entire sample set. The results are shown in Table 5. Although excellent results were obtained for the re-substitu- tion analysis the cross-validation results were significantly poorer both in terms of the over-all percentage correctly classified and the classification of the most important South African category.However as mentioned earlier re-substi- tution would be expected to give optimistically biased results the degree of difference observed here between the re-substitution and cross-validation results for the South African samples merits additional comment. This large difference can be ascribed at least in part to there being only three South African samples in the set. In a cross-validation analysis the sample being classified is excluded from the model. This would leave the remaining two South African samples to form the model for the classification of the third. It is unlikely that just two samples will form an accurate model for the samples from a given country and such a model would not be expected to give reliable results.The one false positive was a Guatamalan sample. The estimated probability of group membership for this sample was 0.54 South African and 0.40 Guatamalan with the remainder of the probability associated with other areas. In practice such a sample which has significant probabilities of belonging to more than one country would not have been assigned to any country owing to the uncertainty. This makes the false positive classification a moot point in this instance. Group separations were visualized in a manner similar to that described for orange juice and pistachio nut samples. A canonical discriminant analysis was performed on the entire 40-sample set using the same nine elements as for the discriminant analysis.Fig. 4 shows a plot of the scores for each sample on the first two canonical functions. The samples from each country are well separated except for Malawi and South Africa where there is a small degree of overlap. This overlap is reasonable in view of the geogra- phic proximity of these two countries. While the first two canonical functions do not completely separate the samples from these countries it was found that they were resolved by their scores on the third canonical function. This is illustrated in Fig. 5 which is a plot of the scores on the first ~~ ~ ~ Table 5 Discriminant analysis results for macadamia nut samples South African result? Correct result* Number False Type of analysis (Oh) correct positives Re-substitution 98 3 0 Cross-validation 78 1 1 *From all macadamia nut samples tested n=40.tn = 3. 5 4 3 C 0 3 . 2 w- r ' E 0 f3 .- C rn -1 -2 -3 - 6 - 4 - 2 0 2 4 6 8 First canonical function Fig. 4 Plot of scores on the first two canonical functions for the macadamia samples. Country codes A Australia; C Costa Rica; G Guatamala; M Malawi; S South Africa; and Z Zimbabwe. Asterisks indicate samples with authenticated origins. Perimeters of the regions containing samples from a given country have been outlined for clarity A A G G G G' G G G G' G' G I I t I I t - 6 - 4 - 2 0 2 4 6 8 First canonical function Fig. 5 Plot of scores on the first and third canonical functions for the macadamia samples. See Fig. 4 for details and third canonical functions.It can be seen that the samples from Malawi and South Africa are now completely separated. Conclusions The applicability of multivariate analysis of trace element composition to the determination of the geographic origin of three agricultural products namely orange juice pista- chio nuts and macadamia nuts has been demonstrated. The multivariate techniques known as discriminant analysis642 JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY DECEMBER 199 1 VOL. 6 and canonical discriminant analysis have been shown to be useful for predicting geographic origin and for helping to visualize the separation of products of different geographic origins. This approach should have wide applicability to the products in general. 5 Kwan W. O. and Kowalski B. R. Anal. Chirn. Acta 1980 122 215. 6 Frank I. and Kowalski B. R. Anal. Chirn. Acta 1984 162 241. Lebensm. Unters. Forsch. 1983 177 15. 8 Pane S. W. Food Technol.. 1986.40. 104. determination of the geographic origin of agricultural 7 Borszeki J.9 KoltaY L-9 InczedY J- and Gems E. Z. References Gilbert J. Shepherd M. J. Wallwork M. A. and Hams R. G. J. Apic. Res. 1981 20 125. Forina M. and Armanino C. Ann. Chirn. (Rome) 1982 72 127. Bayer S. McHard J. A. and Winefordner J. D. J. Agric. Food Chem. 1980,28 1306. Kwan W. O. Kowalski B. R. and Skogerboe R. K. J. Agric. Food Chew. 1979,27 1321. 9 KaTbata-Pendias A. and Pendias; H.; Trace Elements in Soils and Plants CRC Press Boca Raton 1984 pp. 51-55. 10 Oficial Methods of Analysis ed. Horwitz W. Association of Official Analytical Chemists Arlington VA 13th edn. 1980 sect. 22.025 p. 363. Paper 1 /00896J Received February 2Sth 1991 Accepted July 30th 1991