Profiling of Food and Agricultural Supply Chain by Analytical Data Modelling and Machine Learning Techniques

The modern agricultural and food supply chain faces growing global demands regarding product quality, safety, authenticity, and environmental sustainability. Traditional quality control protocols rely heavily on off-site, destructive laboratory testing and single-target analytical methods that struggle to resolve the biochemical complexity and high dimensionality of modern food matrices. To address these bottlenecks, “smart” agriculture requires a transition toward rapid, non-invasive, and data-driven analytical strategies. This Ph.D. research establishes an integrated computational framework combining advanced analytical platforms - such as comprehensive two-dimensional gas chromatography, fluorescence spectroscopy, and handheld Raman devices - with chemometrics, Machine Learning, and Deep Learning algorithms across the farm-to-fork landscape. The overarching objectives were to develop portable diagnostic tools for early, pre-symptomatic crop-pathogen detection, establish chemometric signal harmonization procedures for volatile profiling, design interpretable multiway Deep Learning architectures for food authenticity assessment, and integrate microbiological and volatilome profiling to verify biological terroir and geographical provenance. Through these multi-disciplinary workflows, volatilome profiling of commercial craft ciders and regional wines provided novel chemical benchmarks, identifying key volatile organic compounds that correlate strongly with perceived sensory attributes and highlighting aromatic nuances linked to specific production practices. Field-deployable handheld Raman spectroscopy coupled with discriminant chemometric algorithms enabled early, pre-symptomatic diagnosis of viral infections in plant leaves prior to the onset of visible symptoms, as well as rapid taxonomic classification of phytopathogenic bacteria. Furthermore, reinterpreting multidimensional fluorescence data as digital images processed by Variational Autoencoders within a class-modelling framework achieved remarkable efficiency in food adulteration detection, delivering visually interpretable reconstruction error maps that pinpoint specific spectral regions involved in contamination. Finally, multi-omics integration demonstrated that vineyard-specific microbial signatures persist throughout inoculated grape fermentations, directly shaping wine volatilome regionality. Overall, this doctoral research introduces an innovative chemometric-driven analytical paradigm linking physical measurements, advanced signal preprocessing, and interpretable pattern recognition, bridging analytical chemistry and smart agriculture to advance non-destructive quality assurance, crop monitoring, and provenance verification