Multimodal spectral fusion for food authenticity and geographical origin traceability: principles, modeling strategies, and application advances

Food authenticity and geographical origin traceability now affect safety governance, recall management, market surveillance and consumer trust. For high-value foods, mislabeling and adulteration are rarely controlled by one marker; they reflect coupled changes in molecular composition, elemental background, spatial heterogeneity, processing history and supply-chain context. This critical review therefore asks when multimodal spectral fusion adds independent analytical evidence, rather than merely enlarging the variable space. We evaluate this question through physicochemical complementarity, block-matrix formalism, leakage-controlled validation, quantitative performance audits and routine-laboratory constraints. The resulting rule is conditional. Low-level fusion is defensible only for commensurate, strongly preprocessed and transparently weighted blocks with sufficient independent samples. Mid-level fusion is the most reliable default for the small-to-moderate food-authenticity datasets that dominate the field, because it controls redundancy while preserving cross-block interactions. High-level fusion is mainly an operational strategy for asynchronous instruments, missing blocks, confidence-based rejection or fault-tolerant screening. Latent deep fusion should be reserved for large, balanced, blocked and externally tested datasets. Reported examples show that fusion is most persuasive when it improves the same-task single-modality baseline under independent or external validation, as in milk-type classification, salmon authentication and saffron origin studies; however, the salmon retail-set gain is suggestive rather than definitive because the external set is small. Conversely, small random-split improvements are not evidence of a chemometric breakthrough. Future progress depends less on stacking sensors than on defining minimal effective modality combinations, reporting uncertainty, confronting software and graphical-user-interface limitations, validating across batches, years and instruments, and justifying the cost, throughput, maintenance and training overhead of multi-instrument workflows.