Explainable timber geographic provenance using flame emission spectroscopy dynamics and a transformer model
Abstract
The misrepresentation of timber provenance undermines global supply chain transparency and sustainable forest management practices. Although reference forensic approaches, such as stable isotope analysis and chemical and genetic approximations, are reliable, their high cost and complexity hinder their widespread adoption. This work evaluates the feasibility of a high-throughput framework for provenance classification that integrates VIS-NIR flame emission spectroscopy dynamics and an end-to-end deep learning model. Eucalyptus globulus samples were harvested from five distinct geographical sectors in Chile and subjected to controlled combustion. Combustion kinetic profiles were recorded using a VIS-NIR spectrophotometer and processed using Transformer architecture, benchmarked against Support Vector Machine (SVM) and XGBoost classifiers. The proposed model achieved a test accuracy and F1-score of 97.5%, significantly outperforming SVM (80.0%) and XGBoost (62.5%) baselines. Traditional models exhibited severe overfitting owing to static feature compression, whereas the Transformer successfully captured the transient temporal patterns of the combustion cycle. Interpretable feature attribution analysis using Shapley Additive exPlanations (SHAP) confirms that the classification relies on specific atomic emission lines (potassium doublet at 766.4 and 770.1 nm), validating the physico-chemical basis of the model. These findings confirm that treating wood provenance as a dynamic pattern recognition task yields superior robustness, providing a data-efficient basis for a compact optical sensor for traceability that complements existing timber tracing frameworks.