Fraud involving protected-origin foods, medicines and timber is usually caught by laboratory assays. Examples are stable isotope ratios, which record local water and climate, and trace-element profiles, which record local soil. Both are hard to fake, but both require grinding or acid-digesting the sample. Fast non-destructive scans such as hyperspectral imaging (reflectance at hundreds of wavelengths from the intact product) or near-infrared spectroscopy can run at a border or warehouse. They are cheaper and weaker. This note proposes that the scan should not try to name the region directly. Instead it should predict what the lab would find: a probability distribution over the isotope and element values. Separately, a climate-and-geology model predicts what those values should be for the origin printed on the label. When the two distributions clearly overlap, the shipment is accepted. When they clearly do not, it is flagged. Only when they partly overlap is a sample sent for destructive testing.
The reason to route everything through predicted lab chemistry is that this shared, physically meaningful, low-dimensional space solves three problems at once. Current origin classifiers handle each of them poorly. First, the space gives a natural way to combine a several-hundred-band spectrum with a few dozen tabular climate variables without the spectrum swamping them. Second, it allows verification of claimed origins for which no spectral reference samples exist, because the expected chemistry for a region comes from its climate and geology rather than from a trained class label. Third, it turns "how many samples must we destroy?" into an explicit, tunable budget.
Core Claim
Published origin classifiers already reach near-perfect accuracy when every modality is measured on every sample and the candidate regions are fixed in advance. On rice from six Jilin regions, concatenating laser-induced breakdown spectra with hyperspectral spectra gives 99.85% accuracy with a support vector machine, against 88.07% for hyperspectral alone and 93.06% for the destructive laser spectra alone (Liu et al.). On salmon, principal-component fusion of lipid and element profiles reaches 100% on retail samples, where single platforms reach 65–82% (Hong et al.). Chasing another point of full-information, closed-set accuracy is therefore pointless.
The unsolved operational problem is different. It has two parts: how accurate verification can be when only a small fraction of shipments is destructively tested, and how well it holds up when the claimed origin was never in the spectral training set. Our claim is that verification through predicted assay chemistry dominates the direct alternatives on the curve of accuracy against lab-testing fraction, for both seen and unseen origins. The direct alternatives are a cheap-scan classifier with confidence-based deferral, and the same classifier distilled from a fully fused teacher.
Research Program
The system has two learned parts that meet in assay space. We write $z$ for the vector of destructive measurements, meaning isotope ratios and a chosen panel of elements.
The first part is a scan-to-chemistry model $q_\phi(z \mid x_s)$. It maps a non-destructive spectrum $x_s$ to a Gaussian over $z$ with mean and covariance. It is trained only on paired samples, where both the scan and the lab assay exist. Because its output is a low-dimensional physical quantity rather than a region label, it can be a small 1D convolutional network, like those used on citrus near-infrared spectra (Geographical Origin Identification of Citrus Fruits…, 98.39% with noise augmentation). It can also be a partial-least-squares regressor. The regression target regularizes the network far more strongly than a region label on a few hundred samples would.
The second part is a claimed-origin chemistry model $g_\theta(z \mid e_c)$. It predicts the expected assay distribution from environmental covariates $e_c$ of the claimed location, such as long-run precipitation, vapour pressure, elevation, and soil pH and mineralogy. We reuse the method of the timber work (Chasing the Timber Trail). That work uses gradient-boosted feature selection over 19 NASA atmospheric variables feeding a multi-task Gaussian process, which reports $R^2$ of 0.899 for oxygen and 0.858 for hydrogen isotopes. We extend it with soil covariates to predict elements as well. This model is trained on lab data alone, which is the most plentiful kind of reference data.
The verification score is the overlap of the two predictions:
$$S(x_s, c) = \int q_\phi(z \mid x_s)\, g_\theta(z \mid e_c)\, dz,$$
This integral has a closed form for Gaussians. Its log is a Mahalanobis-type distance that is widened by both models' uncertainty. Two thresholds on $S$ divide shipments into accept, reject and send to lab. Sweeping the width of the middle band traces the budget curve. When a lab result arrives, $z$ is observed directly and compared with $g_\theta$ alone.
The role of each prior method is as follows. The timber work supplies the climate-to-chemistry method. Its known collapse to roughly 60% detection within 500 km motivates adding spectral evidence of soil minerals. We treat that collapse as a constraint on what our method can achieve, not as a problem it automatically solves. Data-valuation work (Optimizing Product Provenance Verification using Data Valuation Methods) shows that removing low-value reference points cut error by 0.8025 RMSE, against 0.1893 for random removal. We reuse this to clean the reference set for $g_\theta$.
The medical fusion literature mostly constrains the design. TFormer found that adding a modality by naive late concatenation can lower accuracy (TFormer). Its deep models are also fragile below a few hundred samples. Our design answers this by fusing in a fixed physical space instead of learning a joint embedding. TabMixer (TabMixer) conditions image features on tabular data. CA-MLIF (CA-MLIF) adds cross-attention with an alignment penalty. Both learn this joint space end to end, and we keep CA-MLIF-style fusion as a competitor.
Data is the main practical risk. The file lists no public provenance benchmark. Every cohort is custom. We therefore plan two tracks. The first uses the existing cohorts where they can be obtained from the authors. These are the Jilin rice set (900 spectral sets, with the laser spectra as the destructive modality, and county coordinates that allow climate covariates to be attached), the salmon set (522 samples), and the oak isotope sets (491 global and 287 European samples). The second track is a new collection on one commodity, rice. It would pair hyperspectral scans with isotope and element assays at roughly twenty or more georeferenced sites across at least two harvest years. This is the only way to test unseen origins, distance-stratified detection and year-to-year drift.
Baselines
The naive reference points are majority-class prediction and the hyperspectral-only support vector machine. The latter has a reported 88.07% on Jilin rice (Liu et al.). The full-information ceiling is the concatenated laser-plus-hyperspectral support vector machine at 99.85% on the same data. It is the strongest reported result in the file, but it requires destroying every sample. We reproduce it with the paper's protocol of a 70/30 split, 50 repetitions, and wavelet and scatter-correction preprocessing. Our method should approach it at a small lab fraction. At a lab fraction of 1.0 our method reduces to lab chemistry plus the claimed-origin model, and it should match the ceiling to within noise. A system that beats this baseline must reach at least 99% accuracy with no more than roughly 20% of samples tested destructively.
The strongest budget-aware competitor is a hyperspectral-only classifier with calibrated confidence deferral. Uncertain samples go to the lab and are then classified by the fused model. It gets the same network family, the same augmentation (noise and generative-adversarial augmentation as in the citrus work) and the same hyperparameter search budget as our scan-to-chemistry model. We also train a distilled student, a hyperspectral network trained to mimic the full fused model's outputs, with the same deferral rule. For classical fusion, principal-component mid-level fusion with partial-least-squares discriminant analysis (Hong et al.) and LightGBM on concatenated features (Yu et al., 97.67% on Euryales Semen) are run under the same deferral protocol. For deep fusion, we include a CA-MLIF-style cross-attention model trained end to end on spectra plus climate covariates.
Three ablations isolate the key ingredient. The first removes the claimed-origin model and compares spectra to empirical per-region chemistry, which tests whether the climate prior matters. The second replaces the chemistry bottleneck with a direct fused classifier on the same inputs, which tests whether routing through assay space matters. The third drops the predicted covariance, which tests whether uncertainty-aware overlap matters for deferral.
Why not just…?
Why not just threshold the confidence of a scan-only classifier and send uncertain samples to the lab? This is the real threat in the closed-set setting, and it may match us there. Our method has two structural advantages the classifier lacks. A classifier cannot output a meaningful score for a claimed region with no training class. Our method can, because $g_\theta$ needs only that region's covariates. A classifier's confidence also measures distance to learned region boundaries, not agreement with the region's expected chemistry. Laundering cases, such as briefly holding pond crabs in a prestige lake (Jiang et al.), are designed to look confidently authentic on surface signals. The leave-sites-out experiment below settles the first point. If our method does not beat this baseline on seen origins, we will report that its value lies only in open-origin verification.
Why not just use principal-component fusion or LightGBM, which already hit 97–100%? Those numbers assume every modality is measured on every sample and that the regions are fixed. At a lab fraction near zero they degrade to their scan-only performance, which is 65–82% on retail salmon for single platforms (Hong et al.). The budget curve shows directly whether they recover faster than our method as the lab fraction grows.
Why not distil the fully fused model into a scan-only student? Distillation transfers the teacher's region boundaries, so it inherits the closed-set limitation. Our scan-to-chemistry model is effectively distillation into assay space rather than label space. The second ablation, direct fused classifier against chemistry bottleneck, is the head-to-head comparison.
Why not an end-to-end cross-attention fusion model? With a few hundred samples such models are at high risk of overfitting (TFormer). They also learn a latent space with no physical meaning, which prevents plugging in a climate prior for unseen regions. We include the model anyway. If it matches us on seen origins, the case rests on open-origin performance and interpretability.
Proposed Experiments
The primary metric is verification accuracy as a function of the fraction of samples sent to the lab, summarized as area under that curve over fractions from 0 to 0.3. It is reported on Jilin rice, salmon and the new rice collection. A mislabeled shipment is simulated by pairing a real sample with a false claimed origin. Secondary metrics are AUROC for fraud detection, expected calibration error of the verification score, and detection accuracy stratified by the distance between the true and claimed origin in bands below 500 km, 500–2500 km and above 2500 km. These bands mirror the timber work's reported drop from above 80% beyond 2500 km to roughly 60% within 500 km (Chasing the Timber Trail). No reported result was found on Lacuna for distance-stratified detection on food commodities, so we establish that baseline ourselves.
The decisive check is the leave-sites-out test on the new rice collection. All spectral samples from a set of sites are withheld, and claims naming those sites are then verified. Here classifier-based baselines have no valid output, and our method should hold AUROC close to its seen-origin value. If it does not, the climate-to-chemistry model is too coarse and the central claim fails.
The second decisive check is the chemistry-bottleneck ablation on seen origins. If a direct fused classifier matches our budget curve, the bottleneck buys only open-set capability. We would then narrow the claim accordingly.
The third check trains on one harvest year and tests on the next. This measures whether predicting physical chemistry drifts less than predicting labels. Temporal drift is flagged as unstudied in both the Euryales and salmon work (Yu et al.; Hong et al.).
Finally, on the paired cohorts only, we construct deliberately conflicting inputs. These combine a scan whose surface bands have been shifted toward the claimed region with lab chemistry left at the true origin. The purpose is to measure how often each method's deferral rule sends such cases to the lab. We state plainly that this is a stress test, not evidence about real laundering.
References
- Combined laser-induced breakdown spectroscopy and hyperspectral imaging with machine learning for the classification and identification of rice geographical origin
- Data fusion and multivariate analysis for food authenticity analysis
- Geographical Origin Identification of Citrus Fruits Based on Near-Infrared Spectroscopy Combined with Convolutional Neural Network and Data Augmentation
- Chasing the Timber Trail: Machine Learning to Reveal Harvest Location Misrepresentation
- Optimizing Product Provenance Verification using Data Valuation Methods
- TFormer: A throughout fusion transformer for multi-modal skin lesion diagnosis
- TabMixer: Noninvasive Estimation of the Mean Pulmonary Artery Pressure via Imaging and Tabular Data Mixing
- CA-MLIF: Cross-Attention and Multimodal Low-Rank Interaction Fusion Framework for Tumor Prognostic Prediction
- Interpretable AI-driven multidimensional chemical fingerprints for geographical authentication of Euryales Semen
- Taste Profile Characterization of Chinese Mitten Crab (Eriocheir sinensis) Meat Using Electronic Tongue Analysis