Uncovering which elements drive timber and soybean traceability models

Abstract. Independent verification of commodity origin is increasingly needed to support deforestation-free and legality-based supply-chain regulation. Multi-element chemical fingerprints are a promising tool for this purpose as plant and wood chemistry integrates information about soil conditions, which carries a spatially distinct signal. However, machine-learning provenance models are often difficult to scrutinize and interpret. The goal of this work is to evaluate whether elemental fingerprints can predict geographic harvest location and produce interpretable and environmentally plausible evidence for commodity tracing. Its novelty lies in combining continuous coordinate prediction with multiple explainable-machine-learning tools to compare predictive performance, robust feature importance, local explanations and reduced element panels across both soybean and timber. The source data comprised 398 soybean (Glycine max) samples from Argentina, Bolivia, Brazil and Paraguay, and 1253 East-European and West-Asian timber samples from Betula, Fagus, Pinus and Quercus. We trained Random Forest, XGBoost and support-vector regression models to predict longitude and latitude and evaluated the models with a Haversine skill score (HSS) relative to a centroid baseline. Feature importance was assessed with permutation importance, conditional permutation importance, SHapley Additive exPlanations (SHAP) values and cumulative feature-addition analyses. Soybean models achieved the strongest coordinate-prediction performance (mean HSS 0.596–0.628), while timber genera showed lower but consistently positive skill scores. Attribution analyses identified relatively small-element suites: soybean predictions were driven mainly by Co, Ba, Cu, Mg, Ni, Se, Ca and Cs. Timber markers were genus-specific, with Sr, Pb and Fe being ranked the highest for Betula; Cl, Fe and Bi for Fagus; Rb, Br and Zn for Pinus; and Fe, Sr and Ni for Quercus. The results show that explainable machine learning can provide a more transparent and auditable basis for elemental-based commodity-origin prediction.