Chemical differentiation of Bolivian Cedrela species as a tool to trace illegal timber trade

Chemical differentiation of Bolivian Cedrela species as a tool to trace illegal timber trade Abstract Combating illegal timber trade requires the ability to identify species and verify geographic origin of timber. Forensic techniques that independently verify the declared species and geographic origin are needed, as current legality procedures are based on certificates and documents that can be falsified. Timber from the genus Cedrela is among the most economically valued tropical timbers worldwide. Three Cedrela species are included in the Appendix III of CITES: C. fissilis, C. odorata and C. angustifolia (listed as C. lilloi). Cedrela timber is currently traded with false origin declarations and under a different species name, but tools to verify this are lacking. We used Direct Analysis in Real Time Time-of-Flight Mass Spectrometry (DART-TOFMS) to chemically identify Cedrela species and sites of origin. Heartwood samples from six Cedrela species (the three CITES-listed species plus C. balansae, C. montana and C. saltensis) were collected at 11 sites throughout Bolivia. Mass spectra detected by DART-TOFMS comprised 1062 compounds; their relative intensities were analysed using Principal Component Analyses, Kernel Discriminant Analysis (KDA) and Random Forest analyses to check discrimination potential among species and sites. Species were identified with a mean discrimination error of 15–19 per cent, with substantial variation in discrimination accuracy among species. The lowest error was observed in C. fissilis (mean = 4.4 per cent). Site discrimination error was considerably higher: 43–54 per cent for C. fissilis and 42–48 per cent for C. odorata. These results provide good prospects to differentiate C. fissilis from other species, but at present there is no scope to do so for other tested species. Thus, discrimination is highly species specific. Our findings for tests of geographic origin suggest no potential to discriminate at the studied scale and for the studied species. Cross-checking results from different methods (KDA and Random Forest) reduced discrimination errors. In all, the DART-TOFMS technique allows independent verification of claimed identity of certain Cedrela species in timber trade. Introduction Illegal trade in timber is a worldwide environmental problem, resulting in damage of natural resources and economic loss. It has been estimated that 10–80 per cent of the total timber trade is illegal (Seneca Creek Associates, 2004) and in some countries, such as Papua New Guinea, Liberia, and the Amazon countries (Stark and Pang Cheung, 2006; Lawson and MacFaul, 2010; Wit et al., 2010), this percentage has been as high as 80–90 per cent of all logging operations. The most common type of fraud concerns false declarations of species and geographic origin, as current legal procedures are generally based on certificates and documents which can be falsified. Most legislative measures focus at combatting international illegal trade but a high proportion (70–90 per cent) of illegal tropical timber is traded in domestic markets (Cerutti and Lescuyer, 2011; Kishor and Lescuyer, 2012; Lescuyer et al., 2014). Clearly, there is a need for forensic techniques to independently verify the origin of traded timber in both domestic and international markets. The genus Cedrela (Meliaceae) delivers one of the most important tropical timbers (tropical cedar), but illegal logging of Cedrela has resulted in CITES-listing of several species in this genus (Compt and Christy, 2008). As a result, timber from these species can be traded internationally only if the appropriate permits have been obtained and presented for clearance at the port of entry or exit (CITES, 2017). The problem is that CITES-listed and non-listed Cedrela species are harvested and traded under the same name (Moya et al., 2013) and are often confused due to wood-anatomical similarities (Gasson, 2011; Gasson et al., 2011; Moya et al., 2013). For authorities enforcing CITES, methods to differentiate Cedrela species are needed. Bolivia harbours as many as six Cedrela species, in different climatic zones, from moist to dry tropical forests, and from low to high altitudes (Mostacedo et al., 2003; Navarro, 2011; Navarro-Cerrillo et al., 2013): Cedrela angustifolia Sessé & Moc. Ex DC., Cedrela balansae C. DC., Cedrela fissilis Vell., Cedrela montana Moritz ex Turcz., Cedrela odorata L. and Cedrela saltensis M.A. Zapater & del Castillo. Cedrela species are highly valued locally (Mostacedo and Fredericksen, 1999) and used in carpentry, fine furniture, doors, windows, joinery, musical instruments, carvings, coatings and plywood (Toledo et al., 2008). However, Cedrela populations have declined considerably in recent years due to overexploitation (Mostacedo and Fredericksen, 1999, 2001). As a result, out of the six species, three are currently listed in Appendix III of CITES: C. odorata, C. fissilis and C. angustifolia (listed as C. lilloi C. DC.) (CITES, 2017). Despite legal harvesting limitations, these species remain at high risk because of continued illegal logging and timber trade (ABT, 2017). The high incidence of illegal trade indicates that control systems have limited effectiveness and methods for independent verification of species and legal origin are needed. Chemical analysis tools, such as mass spectrometry (Fidelis et al., 2012), near-infrared spectroscopy (Braga et al., 2011; Bergo et al., 2016) and stable isotopes (Kagawa and Leavitt, 2010; Förstel et al., 2011; Vlam et al., 2018), can be used to discriminate species and verify the geographical origin of traded timber. For example, previous studies used a specific mass spectrometer to discriminate species that cannot be identified based on wood anatomy in the Americas (Espinoza et al., 2015), Africa (Deklerck et al., 2017), and Asia (McClure et al., 2015). In this study, we focus on chemical characterization by Direct Analysis in Real Time (DART) coupled with Time-of-Flight Mass Spectrometry (TOFMS). This technique has the potential to assist in enforcing protection of Cedrela species as it cannot be falsified, in contrast to current certificates used for declaration of species origin. In DART analysis, the mass spectrometer quickly identifies the chemical components by the differing mass to charge (m/z) of ions/compounds from specimens, without the need for sample preparation. The resulting chemical spectra can be used as a reference database for species identifications. Because this methodology has a high potential to identify species and locations, our aim is to test its applicability to differentiate Cedrela timber obtained from different species and geographic provenances. We answer the following research questions: (1) To what extent can Bolivian Cedrela species be differentiated based on wood chemical composition? (2) To what extent can chemical composition help to differentiate timber sourced from different sites in Bolivia? (3) What is the accuracy for identification of each Cedrela species and site of origin within Bolivia based on their chemical profiles? As the geographical sites of the collected samples may have different environmental conditions, we expect to find distribution patterns of the wood composition that mirror these conditions (Zobel and van Buijtenen, 1989; Wilkins and Stamp, 1990; Mosedale and Ford, 1996; Moya and Calvo-Alvarado, 2012). We also expect that each Cedrela species will present specific chemicals that distinguish it from others (Chatterjee et al., 1971; Cordeiro et al., 2012; Eason and Setzer, 2007; Lago et al., 2004; Maia et al., 2000). Methods Study site and species We studied heartwood samples from six Cedrela species in Bolivia, from 11 sites. In total we sampled 127 trees. Altitude of the sites ranged from 145 m.a.s.l. (meters above sea level) in Riberalta to 2022 m.a.s.l. in Postrervalle (Table 1). We selected sites taking into account the distribution of the study species and we maintained a minimum of 70 km distance between all site pairs to maximize the sampling coverage across the country (Table 1 and Figure 1). The maximum distance between pairs of sampled sites was 1300 km (Cobija-Roboré). We used these samples to perform two types of tests: differentiation of species and differentiation of geographic origin. In the species identification analyses, we included all Cedrela species in the sample collection to analyse cross-species discrimination. For the geographic origin analysis, we only included the two species with the largest sample sizes that we had sampled at multiple sites: C. odorata from three sites and C. fissilis from six sites. The maximum distance between pairs of sites was 80 km for C. fissilis (Espejos-Yapacaní) and 425 km for C. odorata between Ribertalta and Rurrenabaque. Minimum distances between pairs of sites were 70 km (Concepción-Guarayos) for C. fissilis and 285 km (Ribertalta-Cobija) for C. odorata. We performed a stratified random sampling: in each of the Cedrela populations found, trees of diameter ≥10 cm were randomly selected with a minimum distance among trees of at least 50 m in order to obtain a homogeneous sampling in each site and to reduce genetic noise and confounding impact of sampling relatives on site (Gillies et al., 1999). This random selection of samples covered different types of forest strata. Table 1 Cedrela species and sites included in the study. Sample size refers to the number of trees sampled; botanical samples to the number of trees from which botanical samples were obtained for verification of identification by taxonomists. Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 *No botanical samples were collected, but identification was based on previous collections. Table 1 Cedrela species and sites included in the study. Sample size refers to the number of trees sampled; botanical samples to the number of trees from which botanical samples were obtained for verification of identification by taxonomists. Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 *No botanical samples were collected, but identification was based on previous collections. Figure 1 View largeDownload slide Locations of sampled trees belonging to six Cedrela species in Bolivia. Forest cover, Autoridad de Bosques y Tierra, 2015. Figure 1 View largeDownload slide Locations of sampled trees belonging to six Cedrela species in Bolivia. Forest cover, Autoridad de Bosques y Tierra, 2015. Preliminary analyses of sapwood and heartwood showed a wider variation of compounds in heartwood (70.0629–1086.567 m/z) compared with sapwood with a dominance of sugars and starch (69.0285–958.4909 m/z) that were not species-specific. Based on these results, we decided to only include heartwood samples in our analyses. A single heartwood sample was collected from each tree using a 5 mm diameter increment borer (Haglöf) at 50–100 cm stem height. Species were morphologically identified in situ with the help of local guides. In addition, botanical samples were collected for species confirmation when identification in the field was not possible. This was done for 53 per cent of the sampled trees. The voucher preparation and confirmation of the species based on herbarium collections were carried out by an experienced botanist, A. Araujo Murakami at the Museo de Historia Natural Noel Kempff Mercado (Bolivia). Chemical analysis We used Direct Analysis in Real Time Time-of-Flight-Mass Spectrometry (DART-TOFMS) to differentiate Cedrela species and to explore if geographical origin could be determined based on chemical composition of heartwood. The DART source consists of an ionization technique that occurs at atmospheric pressure and is discussed by Cody et al. (2005). Once the molecules from the sample are ionized, they are directed towards the time-of-flight mass spectrometer (TOFMS) (Cody et al., 2005). The mass spectrometer will then characterize the molecules from the sample by determining the mass to charge (m/z) of the ions in their protonated forms. The principal ionization mechanisms for DART-TOFMS have been thoroughly discussed and it has been used to identify timber species with an accuracy of 70–95 per cent (Lancaster and Espinoza, 2012; Evans et al., 2017). To describe the chemotaxonomic relationship of our Cedrela samples, mass spectra were acquired using a DART ion source (IonSense, Saugus, MA, USA) coupled to a JEOL AccuTOF time-of-flight mass spectrometer (JEOL USA, Peabody, MA, USA) in positive ion mode. To check if preparation of wood was needed, we tested the maximum number of compounds by soaking wood in methanol versus using wood with no previous treatment. We did not observe any enhancement with previous preparation of wood samples (data not shown). Hence, we decided to use untreated heartwood samples. We cut slivers of heartwood no wider than 4 mm from each sample with an scalpel. These slivers were held in the DART helium gas stream for 8 s. A mass calibration standard of polyethylene glycol 600 (Ultra, Kingstown, RI, USA) was run between each five samples. The DART source parameters were: needle voltage, 3.5 kV; electrode 1 voltage, 150 V; electrode 2 voltage, 250 V; and gas heater temperature, 350°C. The mass spectrometer settings included: rings lens voltage, 5 V; orifice 1 voltage, 20 V; orifice 2, 5 V; cone temperature, 120°C; peaks voltage, 600 V; ion guide bias, 28 V; focus lens voltage, −120 V; reflectron voltage, 870 V; pusher voltage, 778 V; pulling voltage, −778 V; suppression voltage, 0.00 V; flight tube voltage, −7000 V; and detector voltage, 2000 V. Spectra covered the mass range of 70–1100 mass-to-charge ratios (m/z) and were obtained at 1 scan per second. The helium flow rate for the DART source was 2.0 mL s−1. The resolving power of the mass spectrometer, as stated by the manufacturer, was ±2.0 millimass units (mmu). The diagnostic compounds for spectrum classification were selected with 250 mmu and 1 per cent threshold (Deklerck et al., 2017). TSS Unity, a mass-spec data-processing software (Shrader Software Solutions, Inc., Grosse Pointe Park, MI, USA), was used to export the data as text files for further analysis. Statistical analysis Our analysis of the masses (m/z) detected and relative intensities obtained from the DART TOFMS consisted of several steps. To evaluate if chemotypes can enable differentiation of Cedrela species (research question 1: species identification) we first evaluated the existence of species specific compounds, reduced the sample-compound data matrix using Principal Component Analysis (PCA) and finally performed a discriminant analysis to classify the species, determine the importance of each compound and predict sample assignment. For these analyses we used Kernel Discriminant Analyses with package ks 1.10.5 (Duong, 2007, 2017), and Random Forest model with package randomForest 4.6.12 (Liaw and Wiener, 2002) and dplyr 0.7.4 (Wickham et al., 2017) in R version 3.3.3 (R Development Core Team, 2017). We used PCA in order to reduce the number of variables (compounds) into principal components that can then be used as input for a first type of discriminant analysis (KDA). A second discriminant analysis (Random Forest) was used to identify the most important compounds that can differentiate between species. Both discriminant analyses were based on randomized samples and variables in every run. The classification results allowed us to assess the classification success by evaluating frequencies of correct and erroneous identifications. For the analysis of geographic origin (research question 2: geographic origin identification) for C. fissilis and C. odorata (Table 1), we followed the same steps. Based on the classifications, cross validation errors were estimated for species and site assignments (research question 3: identification accuracy). In detail the method involved four main steps. First, we produced a heat-map graph to visualize the chemical profiles (or chemotypes) of the specimens and to verify whether heartwood samples of a particular species contain diagnostic molecules (expressed as mass-to-charge ratio: m/z) that allow it to be distinguished from other species. The heatmap is a graphical representation of the raw mass spectra measured by DART-TOFMS and is created using the Mass Mountaineer software (RBC Software, Peabody, MA, USA). It illustrates the mass-to-charge ratio (m/z) of the detected compounds and their intensities in a spectrum. Second, to reduce the large data matrix into a set of variables so that the variation within each set is maximized (Gotelli and Ellison, 2004), a PCA was necessary for the set which consisted of 125 samples and 1062 compounds. PCA aims to find the linear combinations of variables by using the covariance matrix of data. The first axis reflects the linear fit capturing most of the variation and the successive orthogonal axes reflect the linear capturing of remaining variation not captured in each of the previous components. We extracted six principal components from the sample-molecule matrix, reflecting the greatest variation in the data matrix. The loadings of all 125 samples on the first six axes were retained and this new matrix was used as input in the discriminant analysis. We excluded C. montana due to its small sample size (two individuals). Third, we performed Kernel Discriminant Analysis (KDA) to test species identification and geographic origin. As KDA cannot cope with more than six variables, we performed the PCA analysis described above, and used the first six PCA axes. KDA separates the samples based on an a priori classification assignment (to species and sites classes) and looks for the optimal non-linear combination of variables (here the six component loadings) for maximal separation of the samples in the six dimensional space (Baudat and Anouar, 2000). KDA’s learning algorithm uses Bayes discriminant rule which allocates a point x in the sample space to one (and only one) of the sampled populations. Each population is associated to a kernel density which was estimated implementing a diagonal data-driven (constrained, symmetric and positive-definite) bandwidth matrix (Duong, 2007). This learning algorithm needs to be trained in order to assess the discrimination power of KDA. Therefore, our data were split in two sets: 80 per cent for training and 20 per cent for testing the model. The pre-smoothed data were then applied to estimate a Smoothed Cross Validation (SCV) error (Duong, 2007) as a different procedure to test correctness of the assignment tests. This delivers the classification error (%) which is the probability that samples are incorrectly assigned to a provenance. A cross validation error of 0 per cent indicates that all the samples were correctly assigned. Finally, we used Random Forest analysis to generate a sample classification model in which splits are based on just one chemical compound. One Random Forest run creates 500 ‘Random Forests’ which are used to obtain a final model (Breiman, 2001; Liaw and Wiener, 2002). As with KDA, the algorithm uses 80 per cent of the dataset for training and 20 per cent for model validation. Every run of Random Forest uses a different training set and may lead to different results. Therefore, we ran Random Forest 100 times and averaged the results. In this way, a total of 50 000 Random Forests (100 runs × 500 Random Forests) were built. For each run, the model provided a list of compounds, with their value of importance. We selected the most important compounds that occurred in >40 per cent of the runs and calculated their frequency. These tentative assignments were based on 351 molecules described either for the Cedrela genera or the Meliaceae family (Afendi et al., 2012). Chemical composition in wood can vary not only among species but also for a given tree species or even a given tree (Pettersen, 1984), but heartwood extractive and exudates can also be species specific (Hillis, 1987). Therefore we used the list of the most important compounds to check if any species indicative compound was present. Random Forest analysis allowed us to identify specific chemical compounds that separate one species or site from the other. The Out-of-Bag (OOB, take one out) error rate and species class error were estimated for each of the 100 runs and used to calculate the standard deviation (SD) of these estimates. The OOB error estimate is equivalent to the SCV error of the KDA analysis. Both KDA and Random Forest analyses generated confusion matrices showing the frequency at which each species/site was wrongly classified. In addition, the total of samples tested for each species after 100 randomization runs allowed us to check with what species a single sample could be confused. Finally, the mean errors per species for site identification across the 100 runs were obtained together with their corresponding standard deviation. Results A total of 1062 ions were characterized and their respective intensities were described, in six Cedrela species from 11 sites across Bolivia, from the DART-TOMFS spectra. The results were analysed for species and sites identification separately. A first inspection of chemical data in the heatmap (Figure 2) suggests species-specific patterns in the chemical profiles. Further cross-checking with the actual mass spectra confirmed that the C. odorata samples had a higher intensity of compounds with molecular masses around m/z 212 and 480, C. fissilis had higher intensities for compounds in the m/z 484–502 range, C. balansae at m/z 478 and 680, and C. angustifolia showed high intensities for compounds at m/z 212 and 400. Although the samples of C. montana showed distinctive ions at m/z 275 and 398, this species was excluded from further analyses due to small sample size. Figure 2 View largeDownload slide Heatmap of the output of the DART-TOMFS for six Cedrela species in Bolivia. Each row represents one sample (one tree). Each column represents a specific mass-to-charge ratio (m/z) of an ion. Colour gradient represents relative compound intensity (relative to the most abundant compound). Figure 2 View largeDownload slide Heatmap of the output of the DART-TOMFS for six Cedrela species in Bolivia. Each row represents one sample (one tree). Each column represents a specific mass-to-charge ratio (m/z) of an ion. Colour gradient represents relative compound intensity (relative to the most abundant compound). Identification of species The analysis for species differentiation included five species: C. angustifolia, C. balansae, C. fissilis, C. odorata and C. saltensis (Table 1). The PCA analysis showed that the six most important components together explained 72.1 per cent of the variation across the samples and that the samples were reasonably well separated in the PCA space (Figure 3). The variances explained by the six principal components (PCs) were: 24 per cent, 20 per cent, 10 per cent, 8 per cent, 6 per cent and 4 per cent for PC1–6, correspondingly. These six components were used as input for the KDA. The KDA (of the 80 per cent sample) resulted in a clear separation of the species (Table 2). Figure 3 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for species. Scatterplots combining (a) PC1 and PC2, and (b) PC1 and PC3. Figure 3 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for species. Scatterplots combining (a) PC1 and PC2, and (b) PC1 and PC3. Table 2 Error classification for species. Mean Smoothed Cross Validation (SCV) error, mean Out-of-Bag (OOB) error for classification and their corresponding standard deviations (SD) were estimated after 100 runs for KDA and Random Forest, respectively. KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 View Large Table 2 Error classification for species. Mean Smoothed Cross Validation (SCV) error, mean Out-of-Bag (OOB) error for classification and their corresponding standard deviations (SD) were estimated after 100 runs for KDA and Random Forest, respectively. KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 View Large The KDA had a total mean error of 19 per cent for the SCV test (Table 2). Species-specific errors differed strongly, from 8.7 per cent for C. fissilis to 46.1 per cent for C. balansae. The mean error per species (OOB) from the 100 Random Forest analyses was 15 per cent, representing a mean identification accuracy of 85 per cent (Table 2; Supplementary Data Figure A.1). Again, these errors differed substantially between species with the lowest value of 4.4 per cent for C. fissilis and highest error of 42.4 per cent for C. balansae (Supplementary Data Figure A.1). In the KDA analysis, identification errors for C. angustifolia and C. balansae included wrong assignments to all the other species. C. fissilis was wrongly identified as all the species except as C. saltensis. C. odorata was mostly identified as C. fissilis (118 samples out of 562) and in some cases wrongly identified as C. saltensis (11 samples out of 562). It was rarely classified as C. angustifolia (1 samples out of 562) and never as C. balansae. C. saltensis was mostly confused with C. odorata (33 samples out of 188), in some cases with C. angustifolia (10 samples out of 188), rarely as C. balansae (2 samples out of 188) but never as C. fissilis (Supplementary Data Table A.1). From the Random Forest analyses, the most important compounds for species discrimination were selected (Supplementary Data Table A.3). In total, 15 compounds were most frequent in over 58 per cent of the runs (100 runs). For some compounds we were able to infer the molecular formula and make tentative assignments. Figure 4 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for geographic origin. Scatterplots combining (a) PC1 and PC2, (b) PC1 and PC3 for C. odorata and (c) PC1 and PC2 and d) PC1 and PC3 for C. fissilis. Figure 4 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for geographic origin. Scatterplots combining (a) PC1 and PC2, (b) PC1 and PC3 for C. odorata and (c) PC1 and PC2 and d) PC1 and PC3 for C. fissilis. In most cases of the Random Forest analyses, each species was confused with three other species (Supplementary Data Table A.2): C. angustifolia was mostly classified as C. fissilis or C. saltensis and on one occasion as C. balansae. C. balansae was confused with all species except for C. saltensis. A similar pattern holds for C. fissilis, although this species was mostly confused with C. odorata. Vice versa, C. odorata was mostly confused with C fissilis, in addition to two samples that were mistakenly identified as C. saltensis. Finally, C. saltensis was confused with all species, except for C. balansae. Identification of geographic origin The analysis for geographic origin was done for C. fissilis and C. odorata separately. Classification performance was higher for Random Forest compared with Kernel Discriminant analysis. Furthermore, Random Forest showed similar error rates for both species while Kernel Discriminant showed a difference of 6.3 per cent between C. fissilis and C. odorata (Table 3a and b). The error rate for site identification was highly variable for both methods. The first six PCs were selected from the PCA analysis (Figure 4) as they explained the highest variance: 78.9 per cent in the case of C. fissilis and 86.2 per cent for C. odorata. Table 3 Error classification for sites of C. fissilis (a) and C. odorata (b) based on KDA with six PCs, and Random Forest analyses. Mean classification and standard deviation (SD) were estimated using the classification error per site after 100 runs with different training and testing sets. KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 Table 3 Error classification for sites of C. fissilis (a) and C. odorata (b) based on KDA with six PCs, and Random Forest analyses. Mean classification and standard deviation (SD) were estimated using the classification error per site after 100 runs with different training and testing sets. KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 KDA classification errors for C. fissilis samples were on average 53.9 per cent (range: 39.7–80.9 per cent), while those for C. odorata averaged 47.7 per cent (range: 38.4–48.5 per cent, Table 3). Roboré and Yapacaní showed the highest total mean error for sites discrimination of C. fissilis samples (Table 3a, Supplementary Data Figure A.1c), and Concepción and Guarayos the lowest. Rurrenabaque showed the highest mean error and Riberalta the lowest error for C. odorata sample classification (Table 3b, Supplementary Data Figure A.1d). There was misclassification between three and four other sites of origin (Supplementary Data Table A.4) with the trained algorithm in KDA. However, some sites showed chemical characteristics clearly distinct from other sites. For example, samples from Roboré and Bajo Paraguá were distinct from Espejos but this site was often confused with Concepción and Guarayos. Samples from Bajo Paraguá and Espejos were distinct from each other but wrongly assigned to Concepción and Yapacaní. Guarayos and Espejos were distinct from Bajo Paraguá but were wrongly assigned to, Concepción and Yapacaní. Similarly to KDA, there was misclassification between two and three other sites of origin in the Random Forest analyses (Supplementary Data Table A.5). For example, Bajo Paraguá was distinct from three sites: Espejos, Concepción and Guarayos but some samples were wrongly assigned to Roboré and Yapacaní. Roboré samples had a higher chance of being wrongly assigned to Bajo Paraguá compared with Yapacaní. Samples from Espejos were wrongly assigned to all sites except Bajo Paraguá. The highest error for the identification of C. fissilis sites using Random Forest was observed in Espejos followed by Roboré and Yapacaní while Concepción showed the best performance with the lowest error rate (Supplementary Data Figure A.1e and f) (Table 3). On the other hand, C. odorata showed the highest classification error for Cobija and lowest error for Rurrenabaque. Samples from Cobija were confused with samples from Rurrenabaque and Riberalta (Supplementary Data Table A.6). However, Rurrenabaque samples were mostly assigned to Riberalta followed by Cobija. Riberalta was confused by the other two sites but it had the highest number of correct assignments. Although Random Forest included a higher number of samples from different sites compared with KDA (24 samples), it performed similarly in error rates and assignments. Samples from Cobija were mostly wrongly assigned to Rurrenabaque and to a lesser extent to Riberalta (Supplementary Data Table A.7). Samples from Rurrenabaque were wrongly assigned to Riberalta and Cobija, at roughly equal frequencies. With this method, Rurrenabaque showed the highest number of correct assignments. In each of the 100 Random Forest analyses, the most important compounds for site discrimination were elected (Supplementary Data Table A.8). Discussion To combat the illegal trade in timber, independent methods to identify species and verify geographical origin need to be developed. In this study, we assessed the effectiveness of DART-TOFMS spectra followed by multivariate statistical analysis to determine the potential for differentiating Cedrela species and geographic origin of Cedrela timber. Overall species differentiation error was 15–19 per cent (range for two statistical methods), while that for geographic origin was significantly higher (42–54 per cent). These discrimination errors are higher compared with previous studies that applied DART-TOFMS, which reached discrimination errors of less than 10 per cent for species discrimination (Lancaster and Espinoza, 2012; Musah et al., 2015; Evans et al., 2017) and of ~30 per cent in distinguishing between sites of origin (Finch et al., 2017). We also found strong differences in discrimination error between species. Possible explanations for these differences include (1) low sample sizes for some species, (2) variation within species, (3) misidentification by the curator or (4) variation across the sites where the species are found (e.g. some species are found together as C. fissilis and C. balansae). We will discuss these possible causes below. Low sample size can lead to higher error rates. This is exemplified by C. montana, of which only two samples were collected. Including this species in the analyses increased the error of species identification from 15 to 30 per cent. Yet other studies that applied DART-TOFMS in species with small sample sizes have successfully discriminated between species (Lancaster and Espinoza, 2012; McClure et al., 2015; Wiemann and Espinoza, 2017). This discrepancy depends on the degree of chemical variation which is much smaller in some species than in others. This variability was evidenced by C. fissilis and C. odorata which showed the lowest error rates in the species discrimination analysis compared with C. angustifolia and C. balansae which showed the highest error rates. This indicates that the accuracy of discrimination is highly species specific which thwarts extrapolating these results to other species and sites. Nevertheless, a more accurate conclusion can be reached by identifying representative chemical compounds in a heatmap. This graphical overview facilitates the discovery of particular trends, such as species-specific chemicals. Another possible source of error is misidentification by the curator. This possible observer bias could be solved by having multiple curators identify and compare herbarium samples before further analysis. In this study, the samples identified were based on a large herbarium collection and previous identifications of Cedrela samples throughout Bolivia. The low accuracy of site discrimination may also be caused by local conditions such as climate, soil characteristics and nutrient availability which seem to affect tree performance and composition (Gentry et al., 1995; Medina 1995; Oliveira‐Filho et al., 1998; Toledo et al., 2011). In the Meliaceae family, Noldt et al. (2001) found that some species were more sensitive to environmental conditions due to root systems in the upper soil layers. The Cedrela samples in our study also showed superficial tree roots and site-specific growth variation (Paredes-Villanueva et al., 2016) which indicates that these trees display site-specific characteristics that may have played an important role in wood formation. Such site characteristics vary from large scale, e.g. ecosystem under different climatic regimes to small scale e.g. the micro site factors that contribute to tree development (Reifsnyder et al., 1971). The scale variation of site identification may have played a role in our discrimination among sites: C. odorata sites were more distant than C. fissilis sites. This was confirmed when only Bajo Paraguá, Roboré and Yapacaní (the most distant sites of C. fissilis) were analysed: the accuracy remained similar with Random Forest (57 per cent) and increased to 53 per cent with KDA (data not shown). These results suggest that discriminating between more distant regions or locations may result in higher accuracies than discriminating among neighboring sites. Apart from these external factors that influence discrimination error, the two statistical analyses we used (Random Forest and KDA) also resulted in different error rates. These errors can be reduced by comparing the probabilities of being assigned to another group. Therefore, KDA and Random Forest would best be used alongside each other as triangulation methods. Comparing and cross-checking results between groups and statistical methods will increase the certainty in identifying species and site of origin. In addition, results should also be complemented by other independent statistical tools. Consistent results of these statistical methods could increase the confidence of correct identification when analysing the spectra generated by DART-TOFMS. Decision Trees (Kamiński et al., 2018; Therneau et al., 2018) or other machine learning algorithms that would also provide information of the less abundant chemicals could also be used as multiple approximation methods. Finally, DART-TOFMS is a qualitative analysis; in order to investigate the role of distance, rainfall, altitude and the chemical composition of Cedrela trees in predicting the likelihood of belonging to the conditions of a specific site, it is necessary to apply a quantitative chemical approach. Such an analysis, in which the effects of sample size and time on the detection accuracy of the chemical signals are measured, will allow us to interpret the resulting molecular mass spectra across different spatial and temporal scales. The within-the-tree variation and among-site differentiation of the chemical compounds of the same species represents a great potential for more specific characterization. All samples in this study were collected in Bolivia, a country that is severely impacted by illegal trade in timber, including Cedrela species. The methods used in this study showed the high potential of mass spectrometry for use in Cedrela species identification in Bolivia, with the highest confidence in identifying C. fissilis. DART TOFMS analysis can easily separate Cedrela genus trees from the other look-alike species, like Swietenia macrophylla King and Carapa guianensis Aubl. (Braga et al., 2011; Bergo et al., 2016), and this would help when false declarations and documents are being used. Previous studies also found that most of the difficulties of Cedrela identifications were at the species level rather than at the genus level (Gasson, 2011). Also, the accuracy of identification between samples from the genera Dalbergia and Machaerium was >95 per cent (Espinoza et al., 2015; Lancaster and Espinoza, 2012). This suggests that DART TOFMS analysis may perform better in distinguishing between Cedrela and other look-alike genera, but suffers in species specific assignment within the taxa. Conclusion Cedrela species belong to a timber genus that has been overexploited in the last couple of years. The regulation of their trading has presented many challenges, given that the identification of those species that belong to the CITES list is difficult because of similar wood anatomical characteristics. Our approach offers a strategy for improving identification certainty of Cedrela species by using a complementary approach contributing to their proper forest management and conservation. DART-TOFMS offers an alternative for identification and chemical discrimination among such species. There are several statistical methods to analyse the data generated by DART-TOFMS. Consistent results of two statistical methods (discriminant analyses: KDA and Random Forest) were found in this study, and applying both methods on the same dataset is recommended. Our results reveal potential for Cedrela species assignment (81–85 per cent accuracy), particularly for C. fissilis (95.6 per cent). Our results also show that discrimination of geographical origin is not possible due to low assignment (with accuracies of 46–57 per cent for C. fissilis and 52–58 per cent for C. odorata). Thus, the mass spectrometric approach used here can help to identify species provenance of certain Bolivian Cedrela timbers, but not geographic provenance within the country. Supplementary data Supplementary data are available at Forestry online. Acknowledgements We would like to thank the Museo de Historia Natural Noel Kempff Mercado for their support on the species identification. This research was financed by the NFP/Nuffic fellowship (The Netherlands). Fieldwork was logistically supported by Universidad Autónoma Gabriel René Moreno and financed by Alberta Mennega Stichting and The Rufford Foundation. We are also grateful to Christoph Ruttkies and Tarn Duong for their recommendations and assistance on the statistical analysis. We are also thankful to Pieter Baas and Nicolien Sol for their help during the collection of samples and analysis of data. Conflict of interest statement None declared. Funding This work was supported by the Dutch organization for internationalization in education (NUFFIC) [NFP-PhD.14/61]. Fieldwork was financed by Alberta Mennega Stichting; and The Rufford Foundation [18 670–1]. References ABT . 2017 Audiencia pública inicial y parcial de rendición de cuentas (de enero hasta agosto, gestión 2017). Autoridad de Fiscalización y Control Social de Bosques y Tierra (ABT), Santa Cruz, Bolivia. Afendi , F.M. , Okada , T. , Yamazaki , M. , Hirai-Morita , A. , Nakamura , Y. , Nakamura , K. , et al. 2012 KNApSAcK family databases: integrated metabolite–plant species databases for multifaceted plant research . Plant Cell Physiol. 53 , e1(1–12) . Google Scholar CrossRef Search ADS Baudat , G. and Anouar , F. 2000 Generalized discriminant analysis using a Kernel Approach . Neural Comput. 12 ( 10 ), 2385 – 2404 . Google Scholar CrossRef Search ADS PubMed Bergo , M.C.J. , Pastore , T.C.M. , Coradin , V.T.R. , Wiedenhoeft , A.C. and Braga , J.W.B. 2016 NIRS identification of Swietenia macrophylla is robust across specimens from 27 countries . IAWA J. 37 ( 3 ), 420 – 430 . Google Scholar CrossRef Search ADS Braga , J.W.B. , Pastore , T.C.M. , Coradin , V.T.R. , Camargos , J.A.A. and da Silva , A.R. 2011 The use of near Infrared Spectroscopy to Identify solid wood Specimens of (Cites Appendix II) . IAWA J. 32 ( 2 ), 285 – 296 . Google Scholar CrossRef Search ADS Breiman , L. 2001 Random forests . Mach. Learn. 45 ( 1 ), 5 – 32 . Google Scholar CrossRef Search ADS Cerutti , P.O. and Lescuyer , G. 2011 The Domestic Market for Small-scale Chainsaw Milling in Cameroon: Present Situation, Opportunities and Challenges . Center for International Forestry Research (CIFOR) . Chatterjee , A. , Chakrabortty , T. and Chandrasekharan , S. 1971 Chemical investigation of Cedrela toona . Phytochemistry 10 ( 10 ), 2533 – 2535 . Google Scholar CrossRef Search ADS CITES . 2017 Convention on International Trade in Endangered Species of Wild Fauna and Flora. Appendices I, II and III. https://cites.org/eng/app/appendices.php (accessed on 09 December, 2017). Cody , R.B. , Laramée , J.A. and Durst , H.D. 2005 Versatile new ion source for the analysis of materials in open air under ambient conditions . Anal. Chem. 77 ( 8 ), 2297 – 2302 . Google Scholar CrossRef Search ADS PubMed Compt , J. and Christy , T. 2008 The 14th meeting of the Conference of the Parties to CITES . Traffic Bull. 21 ( 3 ), 101 . Cordeiro , J.R. , Martinez , M.I.V. , Li , R.W.C. , Cardoso , A.P. , Nunes , L.C. , Krug , F.J. , et al. 2012 Identification of four wood species by an electronic nose and by LIBS . Int. J. Electrochem. 2012 , 5 . Google Scholar CrossRef Search ADS Deklerck , V. , Finch , K. , Gasson , P. , Van den Bulcke , J. , Van Acker , J. , Beeckman , H. , et al. 2017 Comparison of species classification models of mass spectrometry data: Kernel Discriminant Analysis vs Random Forest; A case study of Afrormosia (Pericopsis elata (Harms) Meeuwen) . Rapid Commun. Mass Spectrom . 31 ( 19 ), 1582 – 1588 . Google Scholar CrossRef Search ADS PubMed Duong , T. 2007 ks: Kernel density estimation and kernel discriminant analysis for multivariate data in R . J. Stat. Softw. 21 ( 7 ), 1 – 16 . Google Scholar CrossRef Search ADS Duong , T. 2017 ks: Kernel Smoothing. https://cran.r-project.org/web/packages/ks/ Eason , H.M. and Setzer , W.N. 2007 Bark essential oil composition of Cedrela tonduzii C. DC. (Meliaceae) from Monteverde, Costa Rica . Rec. Nat. Prod. 1 ( 2–3 ), 24 – 27 . Espinoza , E.O. , Wiemann , M.C. , Barajas-Morales , J. , Chavarria , G.D. and McClure , P.J. 2015 Forensic analysis of CITES-protected Dalbergia timber from the Americas . IAWA J. 36 ( 3 ), 311 – 325 . Google Scholar CrossRef Search ADS Evans , P.D. , Mundo , I.A. , Wiemann , M.C. , Chavarria , G.D. , McClure , P.J. , Voin , D. , et al. 2017 Identification of selected CITES-protected Araucariaceae using DART TOFMS . IAWA J. 38 ( 2 ), 266 – S263 . Google Scholar CrossRef Search ADS Fidelis , C.H.V. , Augusto , F. , Sampaio , P.T.B. , Krainovic , P.M. and Barata , L.E.S. 2012 Chemical characterization of rosewood (Aniba rosaeodora Ducke) leaf essential oil by comprehensive two-dimensional gas chromatography coupled with quadrupole mass spectrometry . J. Essent. Oil Res. 24 ( 3 ), 245 – 251 . Google Scholar CrossRef Search ADS Finch , K. , Espinoza , E. , Jones , F.A. and Cronn , R. 2017 Source identification of western Oregon Douglas-fir wood cores using mass spectrometry and random forest classification . Appl. Plant Sci. 5 ( 5 ), 1600158 . Google Scholar CrossRef Search ADS Förstel , H. , Boner , M. , Höltken , A. , Fladung , M. , Degen , B. and Zahnen , J. 2011 Fighting Illegal Logging Through the Introduction of a Combination of the Isotope Method for Identifying the Origins of Timber and DNA Analysis for Differentiation of Tree Species . WWF Germany . Gasson , P. 2011 How precise can wood identification be? Wood anatomy’s role in support of the legal timber trade, especially cites . IAWA J. 32 ( 2 ), 137 – 154 . Google Scholar CrossRef Search ADS Gasson , P. , Baas , P. and Wheeler , E. 2011 Wood anatomy of Cites-listed tree species . IAWA J. 32 ( 2 ), 155 – 198 . Google Scholar CrossRef Search ADS Gentry , A.H. , Bullock , S.H. , Mooney , H.A. and Medina , E. 1995 Seasonally dry tropical forests. Gillies , A. , Navarro , C. , Lowe , A. , Newton , A.C. , Hernandez , M. , Wilson , J. , et al. 1999 Genetic diversity in Mesoamerican populations of mahogany (Swietenia macrophylla), assessed using RAPDs . Heredity 83 ( 6 ), 722 – 732 . Google Scholar CrossRef Search ADS PubMed Gotelli , N. and Ellison , A. 2004 A Primer of Ecological Statistics . Sinauer Associates Inc , p. 614 . Hillis , W.E. 1987 Heartwood and Tree Exudates . Springer-Verlag . Google Scholar CrossRef Search ADS Kagawa , A. and Leavitt , S.W. 2010 Stable carbon isotopes of tree rings as a tool to pinpoint the geographic origin of timber . J. Wood Sci. 56 ( 3 ), 175 – 183 . Google Scholar CrossRef Search ADS Kamiński , B. , Jakubczyk , M. and Szufel , P. 2018 A framework for sensitivity analysis of decision trees . Cent. Eur. J. Oper. Res. 26 ( 1 ), 135 – 159 . Google Scholar CrossRef Search ADS PubMed Kishor , N. and Lescuyer , G. 2012 Controlling illegal logging in domestic and international markets by harnessing multi-level governance opportunities . Int. J. Commons 6 , 255 – 270 . Google Scholar CrossRef Search ADS Lago , J.H.G. , de Ávila , P. , de Aquino , E.M. , Moreno , P.R.H. , Ohara , M.T. , Limberger , R.P. , et al. 2004 Volatile oils from leaves and stem barks of Cedrela fissilis (Meliaceae): chemical composition and antibacterial activities . Flavour Fragr. J. 19 ( 5 ), 448 – 451 . Google Scholar CrossRef Search ADS Lancaster , C. and Espinoza , E. 2012 Analysis of select Dalbergia and trade timber using direct analysis in real time and time‐of‐flight mass spectrometry for CITES enforcement . Rapid Commun. Mass Spectrom. 26 ( 9 ), 1147 – 1156 . Google Scholar CrossRef Search ADS PubMed Lawson , S. and MacFaul , L. 2010 Illegal Logging and Related Trade: Indicators of the Global Response . Chatham House . Lescuyer , G. , Ndotit , S. , Ndong , L.B.B. , Tsanga , R. and Cerutti , P.O. 2014 Policy Options for Improved Integration of Domestic Timber Markets Under the Voluntary Partnership Agreement (VPA) Regime in Gabon . Center for International Forestry Research (CIFOR) , p. 4 . Liaw , A. and Wiener , M. 2002 Classification and regression by randomForest . R News 2 ( 3 ), 18 – 22 . Maia , B.H.L.N.S. , Paula , J.R.d. , Sant’Ana , J. , Silva , M.F.d.G.F.d. , Fernandes , J.B. , Vieira , P.C. , et al. 2000 Essential oils of Toona and Cedrela Species (Meliaceae): taxonomic and ecological implications . J. Braz. Chem. Soc. 11 , 629 – 639 . Google Scholar CrossRef Search ADS McClure , P.J. , Chavarria , G.D. and Espinoza , E. 2015 Metabolic chemotypes of CITES protected Dalbergia timbers from Africa, Madagascar, and Asia . Rapid Commun. Mass Spectrom. 29 ( 9 ), 783 – 788 . Google Scholar CrossRef Search ADS PubMed Medina , E. 1995 Diversity of life forms of higher plants in neotropical dry forests. In Seasonally Dry Tropical Forests . E. Medina, H.A. Mooney and S.H. Bullock (eds.). Cambridge University Press, Cambridge , pp. 221 – 242 . Google Scholar CrossRef Search ADS Mosedale , J.R. and Ford , A. 1996 Variation of the flavour and extractives of European oak wood from two French forests . J. Sci. Food Agric. 70 ( 3 ), 273 – 287 . Google Scholar CrossRef Search ADS Mostacedo , B. and Fredericksen , T.S. 1999 Regeneration status of important tropical forest tree species in Bolivia: assessment and recommendations . For. Ecol. Manage. 124 ( 2 ), 263 – 273 . Google Scholar CrossRef Search ADS Mostacedo , B. and Fredericksen , T. 2001 Regeneración y Silvicultura de Bosques Tropicales en Bolivia . Editora El País . Mostacedo , B. , Justiniano , J. , Toledo , M. and Fredericksen , T. 2003 Guía dendrológica de especies forestales de Bolivia. Segunda Edición edn. Proyecto de Manejo Forestal Sostenible. Moya , R. and Calvo-Alvarado , J. 2012 Variation of wood color parameters of Tectona grandis and its relationship with physical environmental factors . Ann. For. Sci. 69 ( 8 ), 947 – 959 . Google Scholar CrossRef Search ADS Moya , R. , Wiemann , M.C. and Olivares , C. 2013 Identification of endangered or threatened Costa Rican tree species by wood anatomy and fluorescence activity . Rev. Biol. Trop. 61 , 1113 – 1156 . Google Scholar CrossRef Search ADS Musah , R.A. , Espinoza , E.O. , Cody , R.B. , Lesiak , A.D. , Christensen , E.D. , Moore , H.E. , et al. 2015 A high throughput ambient mass spectrometric approach to species identification and classification from chemical fingerprint signatures . Sci. Rep. 5 , 11520 . Google Scholar CrossRef Search ADS PubMed Navarro , G. 2011 Clasificación de la Vegetación de Bolivia . Centro de Ecología Difusión Simón I. Patiño , p. 713 . Navarro-Cerrillo , R.M. , Agote , N. , Pizarro , F. , Ceacero , C.J. and Palacios , G. 2013 Elements for a non-detriment finding of Cedrela spp. in Bolivia—a CITES implementation case study . J. Nat. Conserv. 21 ( 4 ), 241 – 252 . Google Scholar CrossRef Search ADS Noldt , G. , Bauch , J. , Koch , G. and Schmitt , U. 2001 Fine roots of Carapa guianensis Aubl. and Swietenia macrophylla King: cell structure and adaptation to the dry season in Central Amazonia . J. Appl. Bot. 75 ( 3–4 ), 152 – 158 . Oliveira‐Filho , A.T. , Curi , N. , Vilela , E.A. and Carvalho , D.A. 1998 Effects of canopy gaps, topography, and soils on the distribution of woody species in a central Brazilian deciduous dry forest . Biotropica 30 ( 3 ), 362 – 375 . Google Scholar CrossRef Search ADS Paredes-Villanueva , K. , López , L. and Navarro-Cerrillo , R.M. 2016 Regional chronologies of Cedrela fissilis and Cedrela angustifolia in three forest types and their relation to climate . Trees 30 ( 5 ), 1581 – 1593 . Google Scholar CrossRef Search ADS Pettersen , R.C. 1984 The chemical composition of wood. In The Chemistry of Solid Wood . R. Rowell (ed.), American Chemical Society, Washington, DC , pp. 57 – 126 . Google Scholar CrossRef Search ADS R Development Core Team. 2017 R: A language and environment for statistical computing . R Foundation for Statistical Computing, Vienna, Austria . http://www.R-project.org/ Reifsnyder , W.E. , Furnival , G. and Horowitz , J. 1971 Spatial and temporal distribution of solar radiation beneath forest canopies . Agric. Meteorol. 9 , 21 – 37 . Google Scholar CrossRef Search ADS Seneca Creek Associates, L 2004 ‘Illegal’ Logging and Global Wood Markets: The Competitive Impacts on the U.S. Wood Products Industry . American Forest & Paper Association . Stark , T. and Pang Cheung , S. 2006 Sharing the Blame: Global Consumption and China’s Role in Ancient Forest Destruction . Greenpeace China . Therneau , T. , Atkinson , B. and Ripley , B. 2018 rpart: Recursive Partitioning and Regression Trees. https://CRAN.R-project.org/package=rpart Toledo , M. , Chevallier , B. , Villaroel , D. and Mostacedo , B. 2008 Ecología y silvicultura. de especies menos conocidas: Cedro, Cedrela spp . Proyecto BOLFOR II/Instituto Boliviano de Investigación Forestal . Toledo , M. , Poorter , L. , Peña‐Claros , M. , Alarcón , A. , Balcázar , J. , Leaño , C. , et al. 2011 Climate is a stronger driver of tree and forest growth rates than soil and disturbance . J. Ecol. 99 ( 1 ), 254 – 264 . Google Scholar CrossRef Search ADS Vlam , M. , de Groot , G.A. , Boom , A. , Copini , P. , Laros , I. , Veldhuijzen , K. , et al. 2018 Developing forensic tools for an African timber: Regional origin is revealed by genetic characteristics, but not by isotopic signature . Biol. Conserv. 220 , 262 – 271 . Google Scholar CrossRef Search ADS Wickham , H. , Francois , R. , Henry , L. and Müller , K. , RStudio . 2017 dplyr: A Grammar of Data Manipulation. A fast, consistent tool for working with data frame like objects, both in memory and out of memory. https://CRAN.R-project.org/package=dplyr Wiemann , M.C. and Espinoza , E.O. 2017 Species Verification of Dalbergia nigra and Dalbergia spruceana Samples in the Wood Collection at the Forest Products Laboratory. Review Process: Informally Refereed (Peer-Reviewed). Wilkins , A.P. and Stamp , C.M. 1990 Relationship between wood colour, silvicultural treatment and rate of growth in Eucalyptus grandis Hill (Maiden) . Wood Sci. Technol. 24 ( 4 ), 297 – 304 . Google Scholar CrossRef Search ADS Wit , M. , Van Dam , J. , Cerutti , P.O. , Lescuyer , G. , Kerrett , R. and McKeown , J.P. 2010 Chainsaw Milling: Supplier to Local Markets—A Synthesis . ETFRN . Zobel , B.J. and van Buijtenen , J.P. 1989 Wood Variation: Its Causes and Control . Springer . Google Scholar CrossRef Search ADS © Institute of Chartered Foresters, 2018. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Forestry: An International Journal Of Forest Research Oxford University Press

Chemical differentiation of Bolivian Cedrela species as a tool to trace illegal timber trade

Loading next page...
 
/lp/ou_press/chemical-differentiation-of-bolivian-cedrela-species-as-a-tool-to-OZzYG5itFz
Publisher
Institute of Chartered Foresters
Copyright
© Institute of Chartered Foresters, 2018. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com.
ISSN
0015-752X
eISSN
1464-3626
D.O.I.
10.1093/forestry/cpy019
Publisher site
See Article on Publisher Site

Abstract

Abstract Combating illegal timber trade requires the ability to identify species and verify geographic origin of timber. Forensic techniques that independently verify the declared species and geographic origin are needed, as current legality procedures are based on certificates and documents that can be falsified. Timber from the genus Cedrela is among the most economically valued tropical timbers worldwide. Three Cedrela species are included in the Appendix III of CITES: C. fissilis, C. odorata and C. angustifolia (listed as C. lilloi). Cedrela timber is currently traded with false origin declarations and under a different species name, but tools to verify this are lacking. We used Direct Analysis in Real Time Time-of-Flight Mass Spectrometry (DART-TOFMS) to chemically identify Cedrela species and sites of origin. Heartwood samples from six Cedrela species (the three CITES-listed species plus C. balansae, C. montana and C. saltensis) were collected at 11 sites throughout Bolivia. Mass spectra detected by DART-TOFMS comprised 1062 compounds; their relative intensities were analysed using Principal Component Analyses, Kernel Discriminant Analysis (KDA) and Random Forest analyses to check discrimination potential among species and sites. Species were identified with a mean discrimination error of 15–19 per cent, with substantial variation in discrimination accuracy among species. The lowest error was observed in C. fissilis (mean = 4.4 per cent). Site discrimination error was considerably higher: 43–54 per cent for C. fissilis and 42–48 per cent for C. odorata. These results provide good prospects to differentiate C. fissilis from other species, but at present there is no scope to do so for other tested species. Thus, discrimination is highly species specific. Our findings for tests of geographic origin suggest no potential to discriminate at the studied scale and for the studied species. Cross-checking results from different methods (KDA and Random Forest) reduced discrimination errors. In all, the DART-TOFMS technique allows independent verification of claimed identity of certain Cedrela species in timber trade. Introduction Illegal trade in timber is a worldwide environmental problem, resulting in damage of natural resources and economic loss. It has been estimated that 10–80 per cent of the total timber trade is illegal (Seneca Creek Associates, 2004) and in some countries, such as Papua New Guinea, Liberia, and the Amazon countries (Stark and Pang Cheung, 2006; Lawson and MacFaul, 2010; Wit et al., 2010), this percentage has been as high as 80–90 per cent of all logging operations. The most common type of fraud concerns false declarations of species and geographic origin, as current legal procedures are generally based on certificates and documents which can be falsified. Most legislative measures focus at combatting international illegal trade but a high proportion (70–90 per cent) of illegal tropical timber is traded in domestic markets (Cerutti and Lescuyer, 2011; Kishor and Lescuyer, 2012; Lescuyer et al., 2014). Clearly, there is a need for forensic techniques to independently verify the origin of traded timber in both domestic and international markets. The genus Cedrela (Meliaceae) delivers one of the most important tropical timbers (tropical cedar), but illegal logging of Cedrela has resulted in CITES-listing of several species in this genus (Compt and Christy, 2008). As a result, timber from these species can be traded internationally only if the appropriate permits have been obtained and presented for clearance at the port of entry or exit (CITES, 2017). The problem is that CITES-listed and non-listed Cedrela species are harvested and traded under the same name (Moya et al., 2013) and are often confused due to wood-anatomical similarities (Gasson, 2011; Gasson et al., 2011; Moya et al., 2013). For authorities enforcing CITES, methods to differentiate Cedrela species are needed. Bolivia harbours as many as six Cedrela species, in different climatic zones, from moist to dry tropical forests, and from low to high altitudes (Mostacedo et al., 2003; Navarro, 2011; Navarro-Cerrillo et al., 2013): Cedrela angustifolia Sessé & Moc. Ex DC., Cedrela balansae C. DC., Cedrela fissilis Vell., Cedrela montana Moritz ex Turcz., Cedrela odorata L. and Cedrela saltensis M.A. Zapater & del Castillo. Cedrela species are highly valued locally (Mostacedo and Fredericksen, 1999) and used in carpentry, fine furniture, doors, windows, joinery, musical instruments, carvings, coatings and plywood (Toledo et al., 2008). However, Cedrela populations have declined considerably in recent years due to overexploitation (Mostacedo and Fredericksen, 1999, 2001). As a result, out of the six species, three are currently listed in Appendix III of CITES: C. odorata, C. fissilis and C. angustifolia (listed as C. lilloi C. DC.) (CITES, 2017). Despite legal harvesting limitations, these species remain at high risk because of continued illegal logging and timber trade (ABT, 2017). The high incidence of illegal trade indicates that control systems have limited effectiveness and methods for independent verification of species and legal origin are needed. Chemical analysis tools, such as mass spectrometry (Fidelis et al., 2012), near-infrared spectroscopy (Braga et al., 2011; Bergo et al., 2016) and stable isotopes (Kagawa and Leavitt, 2010; Förstel et al., 2011; Vlam et al., 2018), can be used to discriminate species and verify the geographical origin of traded timber. For example, previous studies used a specific mass spectrometer to discriminate species that cannot be identified based on wood anatomy in the Americas (Espinoza et al., 2015), Africa (Deklerck et al., 2017), and Asia (McClure et al., 2015). In this study, we focus on chemical characterization by Direct Analysis in Real Time (DART) coupled with Time-of-Flight Mass Spectrometry (TOFMS). This technique has the potential to assist in enforcing protection of Cedrela species as it cannot be falsified, in contrast to current certificates used for declaration of species origin. In DART analysis, the mass spectrometer quickly identifies the chemical components by the differing mass to charge (m/z) of ions/compounds from specimens, without the need for sample preparation. The resulting chemical spectra can be used as a reference database for species identifications. Because this methodology has a high potential to identify species and locations, our aim is to test its applicability to differentiate Cedrela timber obtained from different species and geographic provenances. We answer the following research questions: (1) To what extent can Bolivian Cedrela species be differentiated based on wood chemical composition? (2) To what extent can chemical composition help to differentiate timber sourced from different sites in Bolivia? (3) What is the accuracy for identification of each Cedrela species and site of origin within Bolivia based on their chemical profiles? As the geographical sites of the collected samples may have different environmental conditions, we expect to find distribution patterns of the wood composition that mirror these conditions (Zobel and van Buijtenen, 1989; Wilkins and Stamp, 1990; Mosedale and Ford, 1996; Moya and Calvo-Alvarado, 2012). We also expect that each Cedrela species will present specific chemicals that distinguish it from others (Chatterjee et al., 1971; Cordeiro et al., 2012; Eason and Setzer, 2007; Lago et al., 2004; Maia et al., 2000). Methods Study site and species We studied heartwood samples from six Cedrela species in Bolivia, from 11 sites. In total we sampled 127 trees. Altitude of the sites ranged from 145 m.a.s.l. (meters above sea level) in Riberalta to 2022 m.a.s.l. in Postrervalle (Table 1). We selected sites taking into account the distribution of the study species and we maintained a minimum of 70 km distance between all site pairs to maximize the sampling coverage across the country (Table 1 and Figure 1). The maximum distance between pairs of sampled sites was 1300 km (Cobija-Roboré). We used these samples to perform two types of tests: differentiation of species and differentiation of geographic origin. In the species identification analyses, we included all Cedrela species in the sample collection to analyse cross-species discrimination. For the geographic origin analysis, we only included the two species with the largest sample sizes that we had sampled at multiple sites: C. odorata from three sites and C. fissilis from six sites. The maximum distance between pairs of sites was 80 km for C. fissilis (Espejos-Yapacaní) and 425 km for C. odorata between Ribertalta and Rurrenabaque. Minimum distances between pairs of sites were 70 km (Concepción-Guarayos) for C. fissilis and 285 km (Ribertalta-Cobija) for C. odorata. We performed a stratified random sampling: in each of the Cedrela populations found, trees of diameter ≥10 cm were randomly selected with a minimum distance among trees of at least 50 m in order to obtain a homogeneous sampling in each site and to reduce genetic noise and confounding impact of sampling relatives on site (Gillies et al., 1999). This random selection of samples covered different types of forest strata. Table 1 Cedrela species and sites included in the study. Sample size refers to the number of trees sampled; botanical samples to the number of trees from which botanical samples were obtained for verification of identification by taxonomists. Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 *No botanical samples were collected, but identification was based on previous collections. Table 1 Cedrela species and sites included in the study. Sample size refers to the number of trees sampled; botanical samples to the number of trees from which botanical samples were obtained for verification of identification by taxonomists. Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 Species Sites Sample size Botanical samples Altitude (m.a.s.l.) C. angustifolia Monteagudo 2 2 1705 Postrervalle 13 12 2022 C. balansae Concepción 10 10 432 C. fissilis Bajo Paraguá 10 * 287 Concepción 13 13 432 Espejos 6 6 553 Guarayos 13 9 260 Roboré 10 * 632 Yapacaní 10 5 318 C. montana Postrervalle 2 2 2022 C. odorata Cobija 10 * 274 Riberalta 10 * 145 Rurrenabaque 10 4 309 C. saltensis Monteagudo 8 2 1705 Total 127 65 *No botanical samples were collected, but identification was based on previous collections. Figure 1 View largeDownload slide Locations of sampled trees belonging to six Cedrela species in Bolivia. Forest cover, Autoridad de Bosques y Tierra, 2015. Figure 1 View largeDownload slide Locations of sampled trees belonging to six Cedrela species in Bolivia. Forest cover, Autoridad de Bosques y Tierra, 2015. Preliminary analyses of sapwood and heartwood showed a wider variation of compounds in heartwood (70.0629–1086.567 m/z) compared with sapwood with a dominance of sugars and starch (69.0285–958.4909 m/z) that were not species-specific. Based on these results, we decided to only include heartwood samples in our analyses. A single heartwood sample was collected from each tree using a 5 mm diameter increment borer (Haglöf) at 50–100 cm stem height. Species were morphologically identified in situ with the help of local guides. In addition, botanical samples were collected for species confirmation when identification in the field was not possible. This was done for 53 per cent of the sampled trees. The voucher preparation and confirmation of the species based on herbarium collections were carried out by an experienced botanist, A. Araujo Murakami at the Museo de Historia Natural Noel Kempff Mercado (Bolivia). Chemical analysis We used Direct Analysis in Real Time Time-of-Flight-Mass Spectrometry (DART-TOFMS) to differentiate Cedrela species and to explore if geographical origin could be determined based on chemical composition of heartwood. The DART source consists of an ionization technique that occurs at atmospheric pressure and is discussed by Cody et al. (2005). Once the molecules from the sample are ionized, they are directed towards the time-of-flight mass spectrometer (TOFMS) (Cody et al., 2005). The mass spectrometer will then characterize the molecules from the sample by determining the mass to charge (m/z) of the ions in their protonated forms. The principal ionization mechanisms for DART-TOFMS have been thoroughly discussed and it has been used to identify timber species with an accuracy of 70–95 per cent (Lancaster and Espinoza, 2012; Evans et al., 2017). To describe the chemotaxonomic relationship of our Cedrela samples, mass spectra were acquired using a DART ion source (IonSense, Saugus, MA, USA) coupled to a JEOL AccuTOF time-of-flight mass spectrometer (JEOL USA, Peabody, MA, USA) in positive ion mode. To check if preparation of wood was needed, we tested the maximum number of compounds by soaking wood in methanol versus using wood with no previous treatment. We did not observe any enhancement with previous preparation of wood samples (data not shown). Hence, we decided to use untreated heartwood samples. We cut slivers of heartwood no wider than 4 mm from each sample with an scalpel. These slivers were held in the DART helium gas stream for 8 s. A mass calibration standard of polyethylene glycol 600 (Ultra, Kingstown, RI, USA) was run between each five samples. The DART source parameters were: needle voltage, 3.5 kV; electrode 1 voltage, 150 V; electrode 2 voltage, 250 V; and gas heater temperature, 350°C. The mass spectrometer settings included: rings lens voltage, 5 V; orifice 1 voltage, 20 V; orifice 2, 5 V; cone temperature, 120°C; peaks voltage, 600 V; ion guide bias, 28 V; focus lens voltage, −120 V; reflectron voltage, 870 V; pusher voltage, 778 V; pulling voltage, −778 V; suppression voltage, 0.00 V; flight tube voltage, −7000 V; and detector voltage, 2000 V. Spectra covered the mass range of 70–1100 mass-to-charge ratios (m/z) and were obtained at 1 scan per second. The helium flow rate for the DART source was 2.0 mL s−1. The resolving power of the mass spectrometer, as stated by the manufacturer, was ±2.0 millimass units (mmu). The diagnostic compounds for spectrum classification were selected with 250 mmu and 1 per cent threshold (Deklerck et al., 2017). TSS Unity, a mass-spec data-processing software (Shrader Software Solutions, Inc., Grosse Pointe Park, MI, USA), was used to export the data as text files for further analysis. Statistical analysis Our analysis of the masses (m/z) detected and relative intensities obtained from the DART TOFMS consisted of several steps. To evaluate if chemotypes can enable differentiation of Cedrela species (research question 1: species identification) we first evaluated the existence of species specific compounds, reduced the sample-compound data matrix using Principal Component Analysis (PCA) and finally performed a discriminant analysis to classify the species, determine the importance of each compound and predict sample assignment. For these analyses we used Kernel Discriminant Analyses with package ks 1.10.5 (Duong, 2007, 2017), and Random Forest model with package randomForest 4.6.12 (Liaw and Wiener, 2002) and dplyr 0.7.4 (Wickham et al., 2017) in R version 3.3.3 (R Development Core Team, 2017). We used PCA in order to reduce the number of variables (compounds) into principal components that can then be used as input for a first type of discriminant analysis (KDA). A second discriminant analysis (Random Forest) was used to identify the most important compounds that can differentiate between species. Both discriminant analyses were based on randomized samples and variables in every run. The classification results allowed us to assess the classification success by evaluating frequencies of correct and erroneous identifications. For the analysis of geographic origin (research question 2: geographic origin identification) for C. fissilis and C. odorata (Table 1), we followed the same steps. Based on the classifications, cross validation errors were estimated for species and site assignments (research question 3: identification accuracy). In detail the method involved four main steps. First, we produced a heat-map graph to visualize the chemical profiles (or chemotypes) of the specimens and to verify whether heartwood samples of a particular species contain diagnostic molecules (expressed as mass-to-charge ratio: m/z) that allow it to be distinguished from other species. The heatmap is a graphical representation of the raw mass spectra measured by DART-TOFMS and is created using the Mass Mountaineer software (RBC Software, Peabody, MA, USA). It illustrates the mass-to-charge ratio (m/z) of the detected compounds and their intensities in a spectrum. Second, to reduce the large data matrix into a set of variables so that the variation within each set is maximized (Gotelli and Ellison, 2004), a PCA was necessary for the set which consisted of 125 samples and 1062 compounds. PCA aims to find the linear combinations of variables by using the covariance matrix of data. The first axis reflects the linear fit capturing most of the variation and the successive orthogonal axes reflect the linear capturing of remaining variation not captured in each of the previous components. We extracted six principal components from the sample-molecule matrix, reflecting the greatest variation in the data matrix. The loadings of all 125 samples on the first six axes were retained and this new matrix was used as input in the discriminant analysis. We excluded C. montana due to its small sample size (two individuals). Third, we performed Kernel Discriminant Analysis (KDA) to test species identification and geographic origin. As KDA cannot cope with more than six variables, we performed the PCA analysis described above, and used the first six PCA axes. KDA separates the samples based on an a priori classification assignment (to species and sites classes) and looks for the optimal non-linear combination of variables (here the six component loadings) for maximal separation of the samples in the six dimensional space (Baudat and Anouar, 2000). KDA’s learning algorithm uses Bayes discriminant rule which allocates a point x in the sample space to one (and only one) of the sampled populations. Each population is associated to a kernel density which was estimated implementing a diagonal data-driven (constrained, symmetric and positive-definite) bandwidth matrix (Duong, 2007). This learning algorithm needs to be trained in order to assess the discrimination power of KDA. Therefore, our data were split in two sets: 80 per cent for training and 20 per cent for testing the model. The pre-smoothed data were then applied to estimate a Smoothed Cross Validation (SCV) error (Duong, 2007) as a different procedure to test correctness of the assignment tests. This delivers the classification error (%) which is the probability that samples are incorrectly assigned to a provenance. A cross validation error of 0 per cent indicates that all the samples were correctly assigned. Finally, we used Random Forest analysis to generate a sample classification model in which splits are based on just one chemical compound. One Random Forest run creates 500 ‘Random Forests’ which are used to obtain a final model (Breiman, 2001; Liaw and Wiener, 2002). As with KDA, the algorithm uses 80 per cent of the dataset for training and 20 per cent for model validation. Every run of Random Forest uses a different training set and may lead to different results. Therefore, we ran Random Forest 100 times and averaged the results. In this way, a total of 50 000 Random Forests (100 runs × 500 Random Forests) were built. For each run, the model provided a list of compounds, with their value of importance. We selected the most important compounds that occurred in >40 per cent of the runs and calculated their frequency. These tentative assignments were based on 351 molecules described either for the Cedrela genera or the Meliaceae family (Afendi et al., 2012). Chemical composition in wood can vary not only among species but also for a given tree species or even a given tree (Pettersen, 1984), but heartwood extractive and exudates can also be species specific (Hillis, 1987). Therefore we used the list of the most important compounds to check if any species indicative compound was present. Random Forest analysis allowed us to identify specific chemical compounds that separate one species or site from the other. The Out-of-Bag (OOB, take one out) error rate and species class error were estimated for each of the 100 runs and used to calculate the standard deviation (SD) of these estimates. The OOB error estimate is equivalent to the SCV error of the KDA analysis. Both KDA and Random Forest analyses generated confusion matrices showing the frequency at which each species/site was wrongly classified. In addition, the total of samples tested for each species after 100 randomization runs allowed us to check with what species a single sample could be confused. Finally, the mean errors per species for site identification across the 100 runs were obtained together with their corresponding standard deviation. Results A total of 1062 ions were characterized and their respective intensities were described, in six Cedrela species from 11 sites across Bolivia, from the DART-TOMFS spectra. The results were analysed for species and sites identification separately. A first inspection of chemical data in the heatmap (Figure 2) suggests species-specific patterns in the chemical profiles. Further cross-checking with the actual mass spectra confirmed that the C. odorata samples had a higher intensity of compounds with molecular masses around m/z 212 and 480, C. fissilis had higher intensities for compounds in the m/z 484–502 range, C. balansae at m/z 478 and 680, and C. angustifolia showed high intensities for compounds at m/z 212 and 400. Although the samples of C. montana showed distinctive ions at m/z 275 and 398, this species was excluded from further analyses due to small sample size. Figure 2 View largeDownload slide Heatmap of the output of the DART-TOMFS for six Cedrela species in Bolivia. Each row represents one sample (one tree). Each column represents a specific mass-to-charge ratio (m/z) of an ion. Colour gradient represents relative compound intensity (relative to the most abundant compound). Figure 2 View largeDownload slide Heatmap of the output of the DART-TOMFS for six Cedrela species in Bolivia. Each row represents one sample (one tree). Each column represents a specific mass-to-charge ratio (m/z) of an ion. Colour gradient represents relative compound intensity (relative to the most abundant compound). Identification of species The analysis for species differentiation included five species: C. angustifolia, C. balansae, C. fissilis, C. odorata and C. saltensis (Table 1). The PCA analysis showed that the six most important components together explained 72.1 per cent of the variation across the samples and that the samples were reasonably well separated in the PCA space (Figure 3). The variances explained by the six principal components (PCs) were: 24 per cent, 20 per cent, 10 per cent, 8 per cent, 6 per cent and 4 per cent for PC1–6, correspondingly. These six components were used as input for the KDA. The KDA (of the 80 per cent sample) resulted in a clear separation of the species (Table 2). Figure 3 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for species. Scatterplots combining (a) PC1 and PC2, and (b) PC1 and PC3. Figure 3 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for species. Scatterplots combining (a) PC1 and PC2, and (b) PC1 and PC3. Table 2 Error classification for species. Mean Smoothed Cross Validation (SCV) error, mean Out-of-Bag (OOB) error for classification and their corresponding standard deviations (SD) were estimated after 100 runs for KDA and Random Forest, respectively. KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 View Large Table 2 Error classification for species. Mean Smoothed Cross Validation (SCV) error, mean Out-of-Bag (OOB) error for classification and their corresponding standard deviations (SD) were estimated after 100 runs for KDA and Random Forest, respectively. KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 KDA Random Forest Species Mean error (%) SD (%) Mean error (%) SD (%) C. angustifolia 26.5 28.0 33.9 7.9 C. balansae 46.1 36.8 42.4 17.7 C. fissilis 8.7 7.1 4.4 1.8 C. odorata 22.3 17.5 15.8 5.3 C. saltensis 20.2 32.4 29.6 17.2 Mean 18.9 7.0 14.9 2.3 View Large The KDA had a total mean error of 19 per cent for the SCV test (Table 2). Species-specific errors differed strongly, from 8.7 per cent for C. fissilis to 46.1 per cent for C. balansae. The mean error per species (OOB) from the 100 Random Forest analyses was 15 per cent, representing a mean identification accuracy of 85 per cent (Table 2; Supplementary Data Figure A.1). Again, these errors differed substantially between species with the lowest value of 4.4 per cent for C. fissilis and highest error of 42.4 per cent for C. balansae (Supplementary Data Figure A.1). In the KDA analysis, identification errors for C. angustifolia and C. balansae included wrong assignments to all the other species. C. fissilis was wrongly identified as all the species except as C. saltensis. C. odorata was mostly identified as C. fissilis (118 samples out of 562) and in some cases wrongly identified as C. saltensis (11 samples out of 562). It was rarely classified as C. angustifolia (1 samples out of 562) and never as C. balansae. C. saltensis was mostly confused with C. odorata (33 samples out of 188), in some cases with C. angustifolia (10 samples out of 188), rarely as C. balansae (2 samples out of 188) but never as C. fissilis (Supplementary Data Table A.1). From the Random Forest analyses, the most important compounds for species discrimination were selected (Supplementary Data Table A.3). In total, 15 compounds were most frequent in over 58 per cent of the runs (100 runs). For some compounds we were able to infer the molecular formula and make tentative assignments. Figure 4 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for geographic origin. Scatterplots combining (a) PC1 and PC2, (b) PC1 and PC3 for C. odorata and (c) PC1 and PC2 and d) PC1 and PC3 for C. fissilis. Figure 4 View largeDownload slide Results of Principal Component Analysis used for KDA analyses for geographic origin. Scatterplots combining (a) PC1 and PC2, (b) PC1 and PC3 for C. odorata and (c) PC1 and PC2 and d) PC1 and PC3 for C. fissilis. In most cases of the Random Forest analyses, each species was confused with three other species (Supplementary Data Table A.2): C. angustifolia was mostly classified as C. fissilis or C. saltensis and on one occasion as C. balansae. C. balansae was confused with all species except for C. saltensis. A similar pattern holds for C. fissilis, although this species was mostly confused with C. odorata. Vice versa, C. odorata was mostly confused with C fissilis, in addition to two samples that were mistakenly identified as C. saltensis. Finally, C. saltensis was confused with all species, except for C. balansae. Identification of geographic origin The analysis for geographic origin was done for C. fissilis and C. odorata separately. Classification performance was higher for Random Forest compared with Kernel Discriminant analysis. Furthermore, Random Forest showed similar error rates for both species while Kernel Discriminant showed a difference of 6.3 per cent between C. fissilis and C. odorata (Table 3a and b). The error rate for site identification was highly variable for both methods. The first six PCs were selected from the PCA analysis (Figure 4) as they explained the highest variance: 78.9 per cent in the case of C. fissilis and 86.2 per cent for C. odorata. Table 3 Error classification for sites of C. fissilis (a) and C. odorata (b) based on KDA with six PCs, and Random Forest analyses. Mean classification and standard deviation (SD) were estimated using the classification error per site after 100 runs with different training and testing sets. KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 Table 3 Error classification for sites of C. fissilis (a) and C. odorata (b) based on KDA with six PCs, and Random Forest analyses. Mean classification and standard deviation (SD) were estimated using the classification error per site after 100 runs with different training and testing sets. KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 KDA Random Forest Mean error SD (%) Mean error SD (%) (%) (%) (a) C. fissilis sites  Bajo Paraguá 45.8 36.5 38.5 16.9  Espejos 43.3 44.4 86.7 11.7  Concepción 37.5 36.8 23.5 13.3  Guarayos 39.7 32.2 36.9 17.5  Roboré 80.9 33.3 57.3 17.9  Yapacaní 60.9 36.8 48.2 20.7  Mean 53.9 12.5 42.7 4.8 (b) C. odorata sites  Cobija 47.8 38.5 60.7 15.8  Riberalta 38.4 38.6 40.9 19  Rurrenabaque 48.5 38.5 30.4 21.6  Mean 47.7 19.7 42.4 8.6 KDA classification errors for C. fissilis samples were on average 53.9 per cent (range: 39.7–80.9 per cent), while those for C. odorata averaged 47.7 per cent (range: 38.4–48.5 per cent, Table 3). Roboré and Yapacaní showed the highest total mean error for sites discrimination of C. fissilis samples (Table 3a, Supplementary Data Figure A.1c), and Concepción and Guarayos the lowest. Rurrenabaque showed the highest mean error and Riberalta the lowest error for C. odorata sample classification (Table 3b, Supplementary Data Figure A.1d). There was misclassification between three and four other sites of origin (Supplementary Data Table A.4) with the trained algorithm in KDA. However, some sites showed chemical characteristics clearly distinct from other sites. For example, samples from Roboré and Bajo Paraguá were distinct from Espejos but this site was often confused with Concepción and Guarayos. Samples from Bajo Paraguá and Espejos were distinct from each other but wrongly assigned to Concepción and Yapacaní. Guarayos and Espejos were distinct from Bajo Paraguá but were wrongly assigned to, Concepción and Yapacaní. Similarly to KDA, there was misclassification between two and three other sites of origin in the Random Forest analyses (Supplementary Data Table A.5). For example, Bajo Paraguá was distinct from three sites: Espejos, Concepción and Guarayos but some samples were wrongly assigned to Roboré and Yapacaní. Roboré samples had a higher chance of being wrongly assigned to Bajo Paraguá compared with Yapacaní. Samples from Espejos were wrongly assigned to all sites except Bajo Paraguá. The highest error for the identification of C. fissilis sites using Random Forest was observed in Espejos followed by Roboré and Yapacaní while Concepción showed the best performance with the lowest error rate (Supplementary Data Figure A.1e and f) (Table 3). On the other hand, C. odorata showed the highest classification error for Cobija and lowest error for Rurrenabaque. Samples from Cobija were confused with samples from Rurrenabaque and Riberalta (Supplementary Data Table A.6). However, Rurrenabaque samples were mostly assigned to Riberalta followed by Cobija. Riberalta was confused by the other two sites but it had the highest number of correct assignments. Although Random Forest included a higher number of samples from different sites compared with KDA (24 samples), it performed similarly in error rates and assignments. Samples from Cobija were mostly wrongly assigned to Rurrenabaque and to a lesser extent to Riberalta (Supplementary Data Table A.7). Samples from Rurrenabaque were wrongly assigned to Riberalta and Cobija, at roughly equal frequencies. With this method, Rurrenabaque showed the highest number of correct assignments. In each of the 100 Random Forest analyses, the most important compounds for site discrimination were elected (Supplementary Data Table A.8). Discussion To combat the illegal trade in timber, independent methods to identify species and verify geographical origin need to be developed. In this study, we assessed the effectiveness of DART-TOFMS spectra followed by multivariate statistical analysis to determine the potential for differentiating Cedrela species and geographic origin of Cedrela timber. Overall species differentiation error was 15–19 per cent (range for two statistical methods), while that for geographic origin was significantly higher (42–54 per cent). These discrimination errors are higher compared with previous studies that applied DART-TOFMS, which reached discrimination errors of less than 10 per cent for species discrimination (Lancaster and Espinoza, 2012; Musah et al., 2015; Evans et al., 2017) and of ~30 per cent in distinguishing between sites of origin (Finch et al., 2017). We also found strong differences in discrimination error between species. Possible explanations for these differences include (1) low sample sizes for some species, (2) variation within species, (3) misidentification by the curator or (4) variation across the sites where the species are found (e.g. some species are found together as C. fissilis and C. balansae). We will discuss these possible causes below. Low sample size can lead to higher error rates. This is exemplified by C. montana, of which only two samples were collected. Including this species in the analyses increased the error of species identification from 15 to 30 per cent. Yet other studies that applied DART-TOFMS in species with small sample sizes have successfully discriminated between species (Lancaster and Espinoza, 2012; McClure et al., 2015; Wiemann and Espinoza, 2017). This discrepancy depends on the degree of chemical variation which is much smaller in some species than in others. This variability was evidenced by C. fissilis and C. odorata which showed the lowest error rates in the species discrimination analysis compared with C. angustifolia and C. balansae which showed the highest error rates. This indicates that the accuracy of discrimination is highly species specific which thwarts extrapolating these results to other species and sites. Nevertheless, a more accurate conclusion can be reached by identifying representative chemical compounds in a heatmap. This graphical overview facilitates the discovery of particular trends, such as species-specific chemicals. Another possible source of error is misidentification by the curator. This possible observer bias could be solved by having multiple curators identify and compare herbarium samples before further analysis. In this study, the samples identified were based on a large herbarium collection and previous identifications of Cedrela samples throughout Bolivia. The low accuracy of site discrimination may also be caused by local conditions such as climate, soil characteristics and nutrient availability which seem to affect tree performance and composition (Gentry et al., 1995; Medina 1995; Oliveira‐Filho et al., 1998; Toledo et al., 2011). In the Meliaceae family, Noldt et al. (2001) found that some species were more sensitive to environmental conditions due to root systems in the upper soil layers. The Cedrela samples in our study also showed superficial tree roots and site-specific growth variation (Paredes-Villanueva et al., 2016) which indicates that these trees display site-specific characteristics that may have played an important role in wood formation. Such site characteristics vary from large scale, e.g. ecosystem under different climatic regimes to small scale e.g. the micro site factors that contribute to tree development (Reifsnyder et al., 1971). The scale variation of site identification may have played a role in our discrimination among sites: C. odorata sites were more distant than C. fissilis sites. This was confirmed when only Bajo Paraguá, Roboré and Yapacaní (the most distant sites of C. fissilis) were analysed: the accuracy remained similar with Random Forest (57 per cent) and increased to 53 per cent with KDA (data not shown). These results suggest that discriminating between more distant regions or locations may result in higher accuracies than discriminating among neighboring sites. Apart from these external factors that influence discrimination error, the two statistical analyses we used (Random Forest and KDA) also resulted in different error rates. These errors can be reduced by comparing the probabilities of being assigned to another group. Therefore, KDA and Random Forest would best be used alongside each other as triangulation methods. Comparing and cross-checking results between groups and statistical methods will increase the certainty in identifying species and site of origin. In addition, results should also be complemented by other independent statistical tools. Consistent results of these statistical methods could increase the confidence of correct identification when analysing the spectra generated by DART-TOFMS. Decision Trees (Kamiński et al., 2018; Therneau et al., 2018) or other machine learning algorithms that would also provide information of the less abundant chemicals could also be used as multiple approximation methods. Finally, DART-TOFMS is a qualitative analysis; in order to investigate the role of distance, rainfall, altitude and the chemical composition of Cedrela trees in predicting the likelihood of belonging to the conditions of a specific site, it is necessary to apply a quantitative chemical approach. Such an analysis, in which the effects of sample size and time on the detection accuracy of the chemical signals are measured, will allow us to interpret the resulting molecular mass spectra across different spatial and temporal scales. The within-the-tree variation and among-site differentiation of the chemical compounds of the same species represents a great potential for more specific characterization. All samples in this study were collected in Bolivia, a country that is severely impacted by illegal trade in timber, including Cedrela species. The methods used in this study showed the high potential of mass spectrometry for use in Cedrela species identification in Bolivia, with the highest confidence in identifying C. fissilis. DART TOFMS analysis can easily separate Cedrela genus trees from the other look-alike species, like Swietenia macrophylla King and Carapa guianensis Aubl. (Braga et al., 2011; Bergo et al., 2016), and this would help when false declarations and documents are being used. Previous studies also found that most of the difficulties of Cedrela identifications were at the species level rather than at the genus level (Gasson, 2011). Also, the accuracy of identification between samples from the genera Dalbergia and Machaerium was >95 per cent (Espinoza et al., 2015; Lancaster and Espinoza, 2012). This suggests that DART TOFMS analysis may perform better in distinguishing between Cedrela and other look-alike genera, but suffers in species specific assignment within the taxa. Conclusion Cedrela species belong to a timber genus that has been overexploited in the last couple of years. The regulation of their trading has presented many challenges, given that the identification of those species that belong to the CITES list is difficult because of similar wood anatomical characteristics. Our approach offers a strategy for improving identification certainty of Cedrela species by using a complementary approach contributing to their proper forest management and conservation. DART-TOFMS offers an alternative for identification and chemical discrimination among such species. There are several statistical methods to analyse the data generated by DART-TOFMS. Consistent results of two statistical methods (discriminant analyses: KDA and Random Forest) were found in this study, and applying both methods on the same dataset is recommended. Our results reveal potential for Cedrela species assignment (81–85 per cent accuracy), particularly for C. fissilis (95.6 per cent). Our results also show that discrimination of geographical origin is not possible due to low assignment (with accuracies of 46–57 per cent for C. fissilis and 52–58 per cent for C. odorata). Thus, the mass spectrometric approach used here can help to identify species provenance of certain Bolivian Cedrela timbers, but not geographic provenance within the country. Supplementary data Supplementary data are available at Forestry online. Acknowledgements We would like to thank the Museo de Historia Natural Noel Kempff Mercado for their support on the species identification. This research was financed by the NFP/Nuffic fellowship (The Netherlands). Fieldwork was logistically supported by Universidad Autónoma Gabriel René Moreno and financed by Alberta Mennega Stichting and The Rufford Foundation. We are also grateful to Christoph Ruttkies and Tarn Duong for their recommendations and assistance on the statistical analysis. We are also thankful to Pieter Baas and Nicolien Sol for their help during the collection of samples and analysis of data. Conflict of interest statement None declared. Funding This work was supported by the Dutch organization for internationalization in education (NUFFIC) [NFP-PhD.14/61]. Fieldwork was financed by Alberta Mennega Stichting; and The Rufford Foundation [18 670–1]. References ABT . 2017 Audiencia pública inicial y parcial de rendición de cuentas (de enero hasta agosto, gestión 2017). Autoridad de Fiscalización y Control Social de Bosques y Tierra (ABT), Santa Cruz, Bolivia. Afendi , F.M. , Okada , T. , Yamazaki , M. , Hirai-Morita , A. , Nakamura , Y. , Nakamura , K. , et al. 2012 KNApSAcK family databases: integrated metabolite–plant species databases for multifaceted plant research . Plant Cell Physiol. 53 , e1(1–12) . Google Scholar CrossRef Search ADS Baudat , G. and Anouar , F. 2000 Generalized discriminant analysis using a Kernel Approach . Neural Comput. 12 ( 10 ), 2385 – 2404 . Google Scholar CrossRef Search ADS PubMed Bergo , M.C.J. , Pastore , T.C.M. , Coradin , V.T.R. , Wiedenhoeft , A.C. and Braga , J.W.B. 2016 NIRS identification of Swietenia macrophylla is robust across specimens from 27 countries . IAWA J. 37 ( 3 ), 420 – 430 . Google Scholar CrossRef Search ADS Braga , J.W.B. , Pastore , T.C.M. , Coradin , V.T.R. , Camargos , J.A.A. and da Silva , A.R. 2011 The use of near Infrared Spectroscopy to Identify solid wood Specimens of (Cites Appendix II) . IAWA J. 32 ( 2 ), 285 – 296 . Google Scholar CrossRef Search ADS Breiman , L. 2001 Random forests . Mach. Learn. 45 ( 1 ), 5 – 32 . Google Scholar CrossRef Search ADS Cerutti , P.O. and Lescuyer , G. 2011 The Domestic Market for Small-scale Chainsaw Milling in Cameroon: Present Situation, Opportunities and Challenges . Center for International Forestry Research (CIFOR) . Chatterjee , A. , Chakrabortty , T. and Chandrasekharan , S. 1971 Chemical investigation of Cedrela toona . Phytochemistry 10 ( 10 ), 2533 – 2535 . Google Scholar CrossRef Search ADS CITES . 2017 Convention on International Trade in Endangered Species of Wild Fauna and Flora. Appendices I, II and III. https://cites.org/eng/app/appendices.php (accessed on 09 December, 2017). Cody , R.B. , Laramée , J.A. and Durst , H.D. 2005 Versatile new ion source for the analysis of materials in open air under ambient conditions . Anal. Chem. 77 ( 8 ), 2297 – 2302 . Google Scholar CrossRef Search ADS PubMed Compt , J. and Christy , T. 2008 The 14th meeting of the Conference of the Parties to CITES . Traffic Bull. 21 ( 3 ), 101 . Cordeiro , J.R. , Martinez , M.I.V. , Li , R.W.C. , Cardoso , A.P. , Nunes , L.C. , Krug , F.J. , et al. 2012 Identification of four wood species by an electronic nose and by LIBS . Int. J. Electrochem. 2012 , 5 . Google Scholar CrossRef Search ADS Deklerck , V. , Finch , K. , Gasson , P. , Van den Bulcke , J. , Van Acker , J. , Beeckman , H. , et al. 2017 Comparison of species classification models of mass spectrometry data: Kernel Discriminant Analysis vs Random Forest; A case study of Afrormosia (Pericopsis elata (Harms) Meeuwen) . Rapid Commun. Mass Spectrom . 31 ( 19 ), 1582 – 1588 . Google Scholar CrossRef Search ADS PubMed Duong , T. 2007 ks: Kernel density estimation and kernel discriminant analysis for multivariate data in R . J. Stat. Softw. 21 ( 7 ), 1 – 16 . Google Scholar CrossRef Search ADS Duong , T. 2017 ks: Kernel Smoothing. https://cran.r-project.org/web/packages/ks/ Eason , H.M. and Setzer , W.N. 2007 Bark essential oil composition of Cedrela tonduzii C. DC. (Meliaceae) from Monteverde, Costa Rica . Rec. Nat. Prod. 1 ( 2–3 ), 24 – 27 . Espinoza , E.O. , Wiemann , M.C. , Barajas-Morales , J. , Chavarria , G.D. and McClure , P.J. 2015 Forensic analysis of CITES-protected Dalbergia timber from the Americas . IAWA J. 36 ( 3 ), 311 – 325 . Google Scholar CrossRef Search ADS Evans , P.D. , Mundo , I.A. , Wiemann , M.C. , Chavarria , G.D. , McClure , P.J. , Voin , D. , et al. 2017 Identification of selected CITES-protected Araucariaceae using DART TOFMS . IAWA J. 38 ( 2 ), 266 – S263 . Google Scholar CrossRef Search ADS Fidelis , C.H.V. , Augusto , F. , Sampaio , P.T.B. , Krainovic , P.M. and Barata , L.E.S. 2012 Chemical characterization of rosewood (Aniba rosaeodora Ducke) leaf essential oil by comprehensive two-dimensional gas chromatography coupled with quadrupole mass spectrometry . J. Essent. Oil Res. 24 ( 3 ), 245 – 251 . Google Scholar CrossRef Search ADS Finch , K. , Espinoza , E. , Jones , F.A. and Cronn , R. 2017 Source identification of western Oregon Douglas-fir wood cores using mass spectrometry and random forest classification . Appl. Plant Sci. 5 ( 5 ), 1600158 . Google Scholar CrossRef Search ADS Förstel , H. , Boner , M. , Höltken , A. , Fladung , M. , Degen , B. and Zahnen , J. 2011 Fighting Illegal Logging Through the Introduction of a Combination of the Isotope Method for Identifying the Origins of Timber and DNA Analysis for Differentiation of Tree Species . WWF Germany . Gasson , P. 2011 How precise can wood identification be? Wood anatomy’s role in support of the legal timber trade, especially cites . IAWA J. 32 ( 2 ), 137 – 154 . Google Scholar CrossRef Search ADS Gasson , P. , Baas , P. and Wheeler , E. 2011 Wood anatomy of Cites-listed tree species . IAWA J. 32 ( 2 ), 155 – 198 . Google Scholar CrossRef Search ADS Gentry , A.H. , Bullock , S.H. , Mooney , H.A. and Medina , E. 1995 Seasonally dry tropical forests. Gillies , A. , Navarro , C. , Lowe , A. , Newton , A.C. , Hernandez , M. , Wilson , J. , et al. 1999 Genetic diversity in Mesoamerican populations of mahogany (Swietenia macrophylla), assessed using RAPDs . Heredity 83 ( 6 ), 722 – 732 . Google Scholar CrossRef Search ADS PubMed Gotelli , N. and Ellison , A. 2004 A Primer of Ecological Statistics . Sinauer Associates Inc , p. 614 . Hillis , W.E. 1987 Heartwood and Tree Exudates . Springer-Verlag . Google Scholar CrossRef Search ADS Kagawa , A. and Leavitt , S.W. 2010 Stable carbon isotopes of tree rings as a tool to pinpoint the geographic origin of timber . J. Wood Sci. 56 ( 3 ), 175 – 183 . Google Scholar CrossRef Search ADS Kamiński , B. , Jakubczyk , M. and Szufel , P. 2018 A framework for sensitivity analysis of decision trees . Cent. Eur. J. Oper. Res. 26 ( 1 ), 135 – 159 . Google Scholar CrossRef Search ADS PubMed Kishor , N. and Lescuyer , G. 2012 Controlling illegal logging in domestic and international markets by harnessing multi-level governance opportunities . Int. J. Commons 6 , 255 – 270 . Google Scholar CrossRef Search ADS Lago , J.H.G. , de Ávila , P. , de Aquino , E.M. , Moreno , P.R.H. , Ohara , M.T. , Limberger , R.P. , et al. 2004 Volatile oils from leaves and stem barks of Cedrela fissilis (Meliaceae): chemical composition and antibacterial activities . Flavour Fragr. J. 19 ( 5 ), 448 – 451 . Google Scholar CrossRef Search ADS Lancaster , C. and Espinoza , E. 2012 Analysis of select Dalbergia and trade timber using direct analysis in real time and time‐of‐flight mass spectrometry for CITES enforcement . Rapid Commun. Mass Spectrom. 26 ( 9 ), 1147 – 1156 . Google Scholar CrossRef Search ADS PubMed Lawson , S. and MacFaul , L. 2010 Illegal Logging and Related Trade: Indicators of the Global Response . Chatham House . Lescuyer , G. , Ndotit , S. , Ndong , L.B.B. , Tsanga , R. and Cerutti , P.O. 2014 Policy Options for Improved Integration of Domestic Timber Markets Under the Voluntary Partnership Agreement (VPA) Regime in Gabon . Center for International Forestry Research (CIFOR) , p. 4 . Liaw , A. and Wiener , M. 2002 Classification and regression by randomForest . R News 2 ( 3 ), 18 – 22 . Maia , B.H.L.N.S. , Paula , J.R.d. , Sant’Ana , J. , Silva , M.F.d.G.F.d. , Fernandes , J.B. , Vieira , P.C. , et al. 2000 Essential oils of Toona and Cedrela Species (Meliaceae): taxonomic and ecological implications . J. Braz. Chem. Soc. 11 , 629 – 639 . Google Scholar CrossRef Search ADS McClure , P.J. , Chavarria , G.D. and Espinoza , E. 2015 Metabolic chemotypes of CITES protected Dalbergia timbers from Africa, Madagascar, and Asia . Rapid Commun. Mass Spectrom. 29 ( 9 ), 783 – 788 . Google Scholar CrossRef Search ADS PubMed Medina , E. 1995 Diversity of life forms of higher plants in neotropical dry forests. In Seasonally Dry Tropical Forests . E. Medina, H.A. Mooney and S.H. Bullock (eds.). Cambridge University Press, Cambridge , pp. 221 – 242 . Google Scholar CrossRef Search ADS Mosedale , J.R. and Ford , A. 1996 Variation of the flavour and extractives of European oak wood from two French forests . J. Sci. Food Agric. 70 ( 3 ), 273 – 287 . Google Scholar CrossRef Search ADS Mostacedo , B. and Fredericksen , T.S. 1999 Regeneration status of important tropical forest tree species in Bolivia: assessment and recommendations . For. Ecol. Manage. 124 ( 2 ), 263 – 273 . Google Scholar CrossRef Search ADS Mostacedo , B. and Fredericksen , T. 2001 Regeneración y Silvicultura de Bosques Tropicales en Bolivia . Editora El País . Mostacedo , B. , Justiniano , J. , Toledo , M. and Fredericksen , T. 2003 Guía dendrológica de especies forestales de Bolivia. Segunda Edición edn. Proyecto de Manejo Forestal Sostenible. Moya , R. and Calvo-Alvarado , J. 2012 Variation of wood color parameters of Tectona grandis and its relationship with physical environmental factors . Ann. For. Sci. 69 ( 8 ), 947 – 959 . Google Scholar CrossRef Search ADS Moya , R. , Wiemann , M.C. and Olivares , C. 2013 Identification of endangered or threatened Costa Rican tree species by wood anatomy and fluorescence activity . Rev. Biol. Trop. 61 , 1113 – 1156 . Google Scholar CrossRef Search ADS Musah , R.A. , Espinoza , E.O. , Cody , R.B. , Lesiak , A.D. , Christensen , E.D. , Moore , H.E. , et al. 2015 A high throughput ambient mass spectrometric approach to species identification and classification from chemical fingerprint signatures . Sci. Rep. 5 , 11520 . Google Scholar CrossRef Search ADS PubMed Navarro , G. 2011 Clasificación de la Vegetación de Bolivia . Centro de Ecología Difusión Simón I. Patiño , p. 713 . Navarro-Cerrillo , R.M. , Agote , N. , Pizarro , F. , Ceacero , C.J. and Palacios , G. 2013 Elements for a non-detriment finding of Cedrela spp. in Bolivia—a CITES implementation case study . J. Nat. Conserv. 21 ( 4 ), 241 – 252 . Google Scholar CrossRef Search ADS Noldt , G. , Bauch , J. , Koch , G. and Schmitt , U. 2001 Fine roots of Carapa guianensis Aubl. and Swietenia macrophylla King: cell structure and adaptation to the dry season in Central Amazonia . J. Appl. Bot. 75 ( 3–4 ), 152 – 158 . Oliveira‐Filho , A.T. , Curi , N. , Vilela , E.A. and Carvalho , D.A. 1998 Effects of canopy gaps, topography, and soils on the distribution of woody species in a central Brazilian deciduous dry forest . Biotropica 30 ( 3 ), 362 – 375 . Google Scholar CrossRef Search ADS Paredes-Villanueva , K. , López , L. and Navarro-Cerrillo , R.M. 2016 Regional chronologies of Cedrela fissilis and Cedrela angustifolia in three forest types and their relation to climate . Trees 30 ( 5 ), 1581 – 1593 . Google Scholar CrossRef Search ADS Pettersen , R.C. 1984 The chemical composition of wood. In The Chemistry of Solid Wood . R. Rowell (ed.), American Chemical Society, Washington, DC , pp. 57 – 126 . Google Scholar CrossRef Search ADS R Development Core Team. 2017 R: A language and environment for statistical computing . R Foundation for Statistical Computing, Vienna, Austria . http://www.R-project.org/ Reifsnyder , W.E. , Furnival , G. and Horowitz , J. 1971 Spatial and temporal distribution of solar radiation beneath forest canopies . Agric. Meteorol. 9 , 21 – 37 . Google Scholar CrossRef Search ADS Seneca Creek Associates, L 2004 ‘Illegal’ Logging and Global Wood Markets: The Competitive Impacts on the U.S. Wood Products Industry . American Forest & Paper Association . Stark , T. and Pang Cheung , S. 2006 Sharing the Blame: Global Consumption and China’s Role in Ancient Forest Destruction . Greenpeace China . Therneau , T. , Atkinson , B. and Ripley , B. 2018 rpart: Recursive Partitioning and Regression Trees. https://CRAN.R-project.org/package=rpart Toledo , M. , Chevallier , B. , Villaroel , D. and Mostacedo , B. 2008 Ecología y silvicultura. de especies menos conocidas: Cedro, Cedrela spp . Proyecto BOLFOR II/Instituto Boliviano de Investigación Forestal . Toledo , M. , Poorter , L. , Peña‐Claros , M. , Alarcón , A. , Balcázar , J. , Leaño , C. , et al. 2011 Climate is a stronger driver of tree and forest growth rates than soil and disturbance . J. Ecol. 99 ( 1 ), 254 – 264 . Google Scholar CrossRef Search ADS Vlam , M. , de Groot , G.A. , Boom , A. , Copini , P. , Laros , I. , Veldhuijzen , K. , et al. 2018 Developing forensic tools for an African timber: Regional origin is revealed by genetic characteristics, but not by isotopic signature . Biol. Conserv. 220 , 262 – 271 . Google Scholar CrossRef Search ADS Wickham , H. , Francois , R. , Henry , L. and Müller , K. , RStudio . 2017 dplyr: A Grammar of Data Manipulation. A fast, consistent tool for working with data frame like objects, both in memory and out of memory. https://CRAN.R-project.org/package=dplyr Wiemann , M.C. and Espinoza , E.O. 2017 Species Verification of Dalbergia nigra and Dalbergia spruceana Samples in the Wood Collection at the Forest Products Laboratory. Review Process: Informally Refereed (Peer-Reviewed). Wilkins , A.P. and Stamp , C.M. 1990 Relationship between wood colour, silvicultural treatment and rate of growth in Eucalyptus grandis Hill (Maiden) . Wood Sci. Technol. 24 ( 4 ), 297 – 304 . Google Scholar CrossRef Search ADS Wit , M. , Van Dam , J. , Cerutti , P.O. , Lescuyer , G. , Kerrett , R. and McKeown , J.P. 2010 Chainsaw Milling: Supplier to Local Markets—A Synthesis . ETFRN . Zobel , B.J. and van Buijtenen , J.P. 1989 Wood Variation: Its Causes and Control . Springer . Google Scholar CrossRef Search ADS © Institute of Chartered Foresters, 2018. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

Forestry: An International Journal Of Forest ResearchOxford University Press

Published: Jun 7, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off