Multi-Genus Model Improves Timber Origin Tracing Accuracy

Researchers have developed a statistical model that improves the accuracy of identifying timber harvest locations by combining chemical data from multiple tree genera. Published in the Annals of Forest Science, the study demonstrates that integrating data from different tree families allows for more precise origin predictions, particularly in regions where reference samples for specific species are scarce or unavailable. The approach addresses a significant limitation in current chemical tracing techniques, which often rely on fragmented and small reference datasets that hinder their effectiveness in combating illegal timber trade.

The study, led by Jakub Truszkowski and colleagues, introduces a multi-genus Gaussian process model that leverages shared patterns of spatial variation in stable isotope ratios and trace element concentrations across different tree genera. Traditional methods typically analyze species or genera in isolation, which limits their utility when reference data is missing for the specific tree type being traced. By accounting for statistical differences between genera while utilizing their common environmental signals, the new model can predict harvest locations even when relevant data is absent in certain areas, effectively lowering the need for extensive new field sampling.

Chemical tracing techniques, including stable isotope ratio analysis and trace element analysis, are increasingly used to verify the origin of timber in compliance with regulations such as the European Union’s Regulation on deforestation-free products. These laws mandate precise geolocation reporting and allow enforcement agencies to use scientific means to validate origin claims. However, the practical application of these techniques has been constrained by the high cost and logistical difficulty of collecting reference samples from all relevant species across all areas of interest, especially in regions affected by political instability or conflict.

To evaluate the new approach, the researchers extended a previous harvest location model to jointly model isotope ratios and trace element concentrations in wood from four economically important genera: Betula, Fagus, Pinus, and Quercus. The evaluation utilized a dataset of 1,020 tree samples collected across 12 Eastern European countries. This region was selected as a case study because it illustrates scenarios where areas of interest for international sanctions cannot be sampled for political or security reasons, making the ability to use existing data from other genera particularly valuable.

The model employs a new covariance function that accounts for shared patterns of spatial variation between genera. This function represents each genus as a vector, where the angle between vectors indicates the degree of similarity in spatial patterns of isotope ratios or trace elements. By estimating these vectors from training data, the model learns the similarities between different genera, allowing it to interpolate between reference data points and utilize data from different genera to make predictions for a sample of interest.

In low-data settings, where samples from the target genus were scarce or entirely absent from a focal testing area, the multi-genus model substantially outperformed the single-genus model. For the determination task, which identifies the harvest location, the multi-genus model achieved considerably lower average mode distance, representing the distance between the true and predicted harvest location. The improvement was most pronounced for Quercus, with the average mode distance reduced by 269 kilometers, or 38 percent, relative to the single-genus model. For Pinus, the reduction was 68 kilometers, or 10 percent, while Fagus and Betula saw reductions of 185 kilometers and 143 kilometers, respectively.

For the verification task, which assesses whether a specific harvest location claim is true, the multi-genus model demonstrated higher sensitivity than the single-genus model for all genera. Sensitivity refers to the percentage of incorrect origin claims that are detected by the model. The largest increase in sensitivity was observed for Pinus at 49 percent, followed by Betula at 18 percent, Quercus at 17 percent, and Fagus at 9 percent. While the specificity, or the percentage of correct claims not marked as incorrect, was similar for most genera, the single-genus model achieved higher specificity for Pinus at low sampling densities.

When data from all genera were available across the study area, corresponding to a high-data regime with an average sampling density of 2.0 trees per 10,000 square kilometers, the multi-genus model achieved performance similar to the single-genus model. In this scenario, the differences in accuracy between the two models were less than 14 kilometers for all genera. The accuracy of the multi-genus model also depended on the dimension of the genus vectors used; increasing the dimension from one to two decreased the mode distance by between 12 and 64 kilometers, depending on the genus, while further increases yielded only modest improvements.

The study highlights that the multi-genus approach can lower reference sampling costs by interpolating between existing data points and utilizing information from other genera. This is particularly relevant given that trace element analysis currently costs between 20 and 120 US dollars per sample, while stable isotope ratio analysis costs between 200 and 500 US dollars per sample. By effectively replacing some physical samples with statistical inference, the model could make chemical tracing more practical and cost-effective for routine use in supply chain compliance.

The researchers note that the approach is especially useful for regions where collecting new reference samples is infeasible due to political, logistical, or security-related challenges. For example, following the invasion of Ukraine, sanctions on Russian timber raised concerns about illegal logging and misdeclaration of origin. The multi-genus model could alleviate this problem by utilizing historical samples collected before the conflict, even if they do not belong to the specific genera of interest, thereby enabling tracing in situations where new samples cannot be collected.

While the study focuses on temperate forests in Eastern Europe, the authors suggest that the method has potential for broader application. Future work could apply the model to tropical species in regions where illegal logging poses immediate threats to ecosystems, such as Central Africa, where a significant portion of timber production is estimated to be illegal. The shared spatial patterns in wood chemistry across diverse lineages indicate that the model could be adapted to other commodities relevant to deforestation-free regulations, such as soybean and cocoa, by modeling samples collected in different years as distinct but related spatial patterns.

The findings demonstrate that wood chemistry shares common spatial patterns across distantly related genera, driven primarily by differences in climate and soil composition rather than species-specific biology alone. By leveraging these shared patterns, the multi-genus model provides a robust tool for identifying timber harvest locations, supporting both regulated companies and enforcement agencies in ensuring compliance with import regulations and promoting legal and sustainable forest products trade.

The study employed a probabilistic framework based on Gaussian process regression, a statistical technique that quantifies uncertainty in model predictions. This approach extends previous single-genus models by incorporating a new covariance function designed to handle data from multiple tree genera simultaneously. In this model, each genus is represented as a vector, where the angle between two vectors indicates the degree of similarity in their spatial patterns of isotope ratios or trace elements. The coordinates of these vectors are estimated directly from the training data, allowing the model to learn the statistical relationships between different genera. If the vectors for two genera are orthogonal, the model treats their data as uncorrelated; if they are parallel, it indicates that predictions for one genus can be derived from the other through shifting and rescaling. This structure enables the model to account for both the distinct baseline levels of chemical measurements in different genera and the shared patterns of spatial variation caused by environmental factors.

The dataset used for evaluation consisted of 1,020 tree samples from four economically important genera: Betula, Fagus, Pinus, and Quercus, collected across 12 Eastern European countries. For each sample, researchers measured five stable isotope ratios, including carbon, hydrogen, nitrogen, oxygen, and sulfur, as well as the relative abundance of 17 trace elements using X-Ray Fluorescence Spectroscopy. This resulted in 20 distinct chemical measurements per sample, although measurements for barium, manganese, and bromine were excluded from further analysis due to missing data. The study area for each genus was defined as all locations within 300 kilometers of any sample from that genus, provided they fell within the known geographic range of the genus. This spatial definition allowed the model to predict harvest locations across a continuous area, regardless of political borders, which is crucial for verifying origin claims in international trade.

The experimental setup involved two distinct regimes to test the model’s robustness: a high-data regime and a low-data regime. In the high-data regime, all available samples were used to simulate a scenario where reference data is abundant across the study area. In the low-data regime, the researchers simulated situations where samples from the target genus were scarce or entirely missing from specific focal testing areas. This was achieved by down-sampling the training data within these areas to various densities, ranging from zero to one sample per 10,000 square kilometers. The model’s performance was evaluated using two metrics: mode distance for the determination task, which measures the distance between the true and predicted harvest location, and sensitivity and specificity for the verification task, which assesses the model’s ability to detect incorrect origin claims while avoiding false alarms.

Results indicated that the multi-genus model’s advantage over the single-genus model was most significant in low-data settings. As the number of training samples from the target genus increased, the performance gap between the two models diminished. For the determination task, the difference became negligible when the sample density reached 0.2 trees per 10,000 square kilometers for Betula and Pinus. However, for the verification task, the multi-genus model retained a sizeable advantage in sensitivity across most experimental settings, even at higher sample densities. This suggests that while the models converge in accuracy when data is plentiful, the multi-genus approach provides a more robust safety net for detecting fraudulent claims when reference data is limited. The findings underscore the value of integrating diverse chemical data sources to enhance the reliability of timber tracing tools in complex regulatory environments.

Subject of Research: Development of a multi-genus Gaussian process model for improving the accuracy of timber harvest location identification using stable isotope and trace element data.

Article Title: Chemical timber tracing: combining tree-genera information lowers reference data needs and makes harvest location identification more accurate

Article References: Truszkowski, J., Boeschoten, L., Mortier, T., Smith, C., Buliga, B., Chater, C., Janssens, S. B., Saunders, J., Trischler, J., Zuidema, P., Antonelli, A., & Deklerck, V. (2026). Chemical timber tracing: combining tree-genera information lowers reference data needs and makes harvest location identification more accurate. Annals of Forest Science, 83(1), Article 18. https://doi.org/10.1186/s13595-026-01341-x

Image Credits: AI Generated

DOI: 10.1186/s13595-026-01341-x

Keywords: timber tracing, stable isotope analysis, trace element analysis, illegal logging, Gaussian process, forest science, supply chain compliance, Eastern Europe, Chemical, timber, tracing, combining