The implementation of fast, economic, and sustainable analytical methods plays a key role for technological advances in the agri-food sector. Since both visual aspect and chemical composition of food has a decisive impact on its quality evaluation, the applications of imaging and spectroscopic techniques in contexts like precision agriculture and real-time food monitoring are rapidly increasing, allowing an objective, fast and automatic characterization of food samples. The use of RGB digital imaging is now favoured thanks to technically sophisticated, user-friendly and affordable systems, like common smartphones, whose sensors can capture the intensity of the visible region of the electromagnetic spectrum in the red, green and blue channels. Colour and texture are the main features that can be derived from RGB images and related to quality aspects in a wide variety of food products. According to the application at hand, food quality could depend on some properties related to regions of the electromagnetic spectrum other than the visible. In this context, NIR spectroscopy is a fast, non-destructive and versatile technique allowing to determine many chemical constituents in most food products. Some food-related applications could be even successfully handled through cheap, portable and commercially available devices working in narrow spectral regions. Imaging and spectroscopic techniques tend to generate large amounts of data in a very short time, and the resulting datasets can be covered by the definition of Big Data. Indeed, Big Data refers to extremely large and complex datasets that cannot be elaborated with traditional data processing tools but require a multivariate statistical approach to extract useful information. The main topic of this PhD project consists in the development and application of advanced analytical methods based on multivariate statistics and machine learning algorithms to analyse NIR spectra and RGB image data for the characterization of different food matrices. More in detail, this PhD thesis is focused on the evaluation of chemometric strategies able to effectively manage Big Data resulting from imaging and spectroscopic techniques, in order to maximize the amount of information that can be acquired, while minimizing the effort required to acquire that information. The thesis investigates the application of data dimensionality reduction algorithms to extract, codify and relate the colour- and texture- related information in RGB images of red grape samples with the corresponding anthocyanins content through PLS calibration models. Images were acquired using a common smartphone by means of a user-friendly interface and a portable device. A large dataset of images was acquired and converted into matrices of signals named colourgrams, subsequently analysed using multivariate approaches. Research activities focused also on the development of an alternative method to extract and codify the texture properties of RGB images into matrices of signals named texturegrams. The effectiveness of this approach has been tested on a subset of red grape sample images, by comparing the results obtained from different feature extraction methods and investigating the possible advantages of combining colour and texture information through different data fusion approaches. Another part of the research work was devoted to evaluating the effectiveness of NIR spectroscopy coupled with chemometrics to address distinct challenges in the agri-food sector and considering also different acquisition devices, i.e., benchtop and portable instruments.
L’implementazione di metodi analitici veloci, economici e sostenibili svolge un ruolo cardine nel progresso tecnologico del settore agroalimentare. Siccome sia l'aspetto esteriore che la composizione chimica degli alimenti sono decisivi nel valutarne la qualità, le applicazioni di tecniche di imaging e spettroscopia nell’agricoltura di precisione e per il monitoraggio in tempo reale sono in aumento, consentendo una caratterizzazione oggettiva, rapida e automatica dei campioni alimentari. L’utilizzo dell’imaging digitale RGB è oggi favorito grazie alla diffusione di sistemi tecnologicamente avanzati, facili da usare ed economici, ad esempio uno smartphone, in grado di catturare l’intensità della regione visibile dello spettro elettromagnetico nei canali del rosso, verde e blu. Colore e texture sono fra le principali proprietà che possono essere ricavate da un’immagine RGB e legate a aspetti qualitativi in diversi prodotti alimentari. In base al tipo di applicazione, la qualità alimentare può dipendere da alcune proprietà esplorabili attraverso regioni dello spettro elettromagnetico diverse da quella visibile. A questo riguardo, la spettroscopia NIR è una tecnica veloce, non distruttiva e versatile che consente di determinare più costituenti chimici nella maggioranza dei prodotti alimentari. È inoltre possibile affrontare con successo alcune sfide in ambito alimentare tramite dispositivi economici e portatili già in commercio che operano in regioni spettrali ristrette. Le tecniche di imaging e spettroscopiche tendono a generare in poco tempo grandi quantità di dati, e i dataset risultanti possono rientrare nella definizione di Big Data, che non possono essere elaborati con strumenti tradizionali di analisi dei dati, ma che richiedono un approccio statistico multivariato per estrarre l’informazione utile. Il filo conduttore di questo progetto di dottorato è costituito dallo sviluppo e applicazione di metodi analitici avanzati basati su statistica multivariata e algoritmi di machine learning per analizzare spettri NIR e immagini RGB per la caratterizzazione di diverse matrici alimentari. In particolare, il focus di questa tesi verte sulla valutazione di strategie chemiometriche in grado di gestire efficacemente i Big Data ottenuti da tecniche di imaging e spettroscopiche, al fine di massimizzare la quantità di informazioni che possono essere acquisite e minimizzare lo sforzo necessario per l'acquisizione delle informazioni stesse. Si valuta l'applicazione di algoritmi di riduzione della dimensionalità dei dati per estrarre e correlare le proprietà relative a colore e texture di immagini RGB di uve rosse con il corrispondente contenuto di antociani tramite modelli di calibrazione multivariata mediante PLS. Le immagini sono state acquisite utilizzando un semplice smartphone tramite un'interfaccia user-friendly e un dispositivo portatile. L’ampio dataset di immagini così ottenuto è stato convertito in una matrice di segnali chiamati colorigrammi, che è stato quindi analizzato con tecniche multivariate. È stato anche sviluppato un metodo alternativo per estrarre e codificare le proprietà di texture di immagini RGB in matrici di segnali denominati texturegrammi. La sua efficacia è stata testata su un subset di immagini di uve rosse, confrontando i risultati ottenuti da diversi metodi di estrazione di features e valutando possibili vantaggi della combinazione dell’informazione legata a colore e texture mediante diversi approcci di data fusion. Un’altra parte del lavoro di ricerca è stato dedicato alla valutazione dell'efficacia della spettroscopia NIR accoppiata alla chemiometria per affrontare sfide caratteristiche nell’ambito agroalimentare, considerando anche diversi dispositivi di acquisizione, da banco e portatili.
Analisi multivariata di Big Data ottenuti attraverso tecniche analitiche non distruttive nel settore agroalimentare
MENOZZI, CAMILLA
2025
Abstract
The implementation of fast, economic, and sustainable analytical methods plays a key role for technological advances in the agri-food sector. Since both visual aspect and chemical composition of food has a decisive impact on its quality evaluation, the applications of imaging and spectroscopic techniques in contexts like precision agriculture and real-time food monitoring are rapidly increasing, allowing an objective, fast and automatic characterization of food samples. The use of RGB digital imaging is now favoured thanks to technically sophisticated, user-friendly and affordable systems, like common smartphones, whose sensors can capture the intensity of the visible region of the electromagnetic spectrum in the red, green and blue channels. Colour and texture are the main features that can be derived from RGB images and related to quality aspects in a wide variety of food products. According to the application at hand, food quality could depend on some properties related to regions of the electromagnetic spectrum other than the visible. In this context, NIR spectroscopy is a fast, non-destructive and versatile technique allowing to determine many chemical constituents in most food products. Some food-related applications could be even successfully handled through cheap, portable and commercially available devices working in narrow spectral regions. Imaging and spectroscopic techniques tend to generate large amounts of data in a very short time, and the resulting datasets can be covered by the definition of Big Data. Indeed, Big Data refers to extremely large and complex datasets that cannot be elaborated with traditional data processing tools but require a multivariate statistical approach to extract useful information. The main topic of this PhD project consists in the development and application of advanced analytical methods based on multivariate statistics and machine learning algorithms to analyse NIR spectra and RGB image data for the characterization of different food matrices. More in detail, this PhD thesis is focused on the evaluation of chemometric strategies able to effectively manage Big Data resulting from imaging and spectroscopic techniques, in order to maximize the amount of information that can be acquired, while minimizing the effort required to acquire that information. The thesis investigates the application of data dimensionality reduction algorithms to extract, codify and relate the colour- and texture- related information in RGB images of red grape samples with the corresponding anthocyanins content through PLS calibration models. Images were acquired using a common smartphone by means of a user-friendly interface and a portable device. A large dataset of images was acquired and converted into matrices of signals named colourgrams, subsequently analysed using multivariate approaches. Research activities focused also on the development of an alternative method to extract and codify the texture properties of RGB images into matrices of signals named texturegrams. The effectiveness of this approach has been tested on a subset of red grape sample images, by comparing the results obtained from different feature extraction methods and investigating the possible advantages of combining colour and texture information through different data fusion approaches. Another part of the research work was devoted to evaluating the effectiveness of NIR spectroscopy coupled with chemometrics to address distinct challenges in the agri-food sector and considering also different acquisition devices, i.e., benchtop and portable instruments.File | Dimensione | Formato | |
---|---|---|---|
PhD_Thesis_Menozzi_Camilla_XXXVII (1).pdf
accesso aperto
Dimensione
11.22 MB
Formato
Adobe PDF
|
11.22 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/202180
URN:NBN:IT:UNIMORE-202180