The two-year research project INSIDE, funded by the Italian Ministry of Universities and Research – MUR, aims to understand and quantify the penetration of incident near-infrared (NIR) radiation coupled with hyperspectral imaging (HSI), depending on materials analysed and experimental settings, with the belief that a deeper knowledge on the topic would open up for new applications, with the help of advanced chemometrics. To this goal, ad hoc multi-material samples were built employing a 3D printer, in the shape of 1×1×1 cm and 1×1×3 cm (base × height) parallelepipeds or cylinders. The resulting samples, respectively 37 and 8, were composed with strata of different thickness of polymers with distinct spectral signatures (PLA and PETG). The samples were then analysed via a NIR-HSI camera (Specim, Finland) operating in the reflectance mode within the 1000-2500 nm spectral range, and the obtained spectra were submitted to chemometric data analysis. Exploratory analysis through principal component analysis (PCA) was performed. Subsequently, the stratified components in terms of total height were modelled, and the quantification precision was investigated in terms of root mean square error in prediction. Specifically, partial least squares (PLS) regression was initially applied on the parallelepipedal samples and tested on the cylindric ones. The obtained results pointed out the presence of two consequent linear trends with different slopes, which needed to be modelled separately to optimize performances. Non-linearity of light penetration was then confirmed by ordinary least squares (OLS) and multivariate curve resolution-alternating least squares (MCR-ALS), since both estimated the pixel-based composition of samples with limited accuracy. Subsequently, specific algorithms to deal with non-linearity were considered and applied. Then, regression by means of convolutional neural networks (CNN) was performed, with the aim of describing the whole range with a single model. Though CNN correctly dealt with the non-linear behaviour of the data, it confirmed that quantification becomes more challenging for deeper layers, as shown by a sudden increase of within-level standard deviation in the prediction results.
Dealing with non-linearity in the study of near infrared light penetration through matter
Sara Gariglio;Cristina Malegori;Paolo Oliveri;Monica Casale;
2025-01-01
Abstract
The two-year research project INSIDE, funded by the Italian Ministry of Universities and Research – MUR, aims to understand and quantify the penetration of incident near-infrared (NIR) radiation coupled with hyperspectral imaging (HSI), depending on materials analysed and experimental settings, with the belief that a deeper knowledge on the topic would open up for new applications, with the help of advanced chemometrics. To this goal, ad hoc multi-material samples were built employing a 3D printer, in the shape of 1×1×1 cm and 1×1×3 cm (base × height) parallelepipeds or cylinders. The resulting samples, respectively 37 and 8, were composed with strata of different thickness of polymers with distinct spectral signatures (PLA and PETG). The samples were then analysed via a NIR-HSI camera (Specim, Finland) operating in the reflectance mode within the 1000-2500 nm spectral range, and the obtained spectra were submitted to chemometric data analysis. Exploratory analysis through principal component analysis (PCA) was performed. Subsequently, the stratified components in terms of total height were modelled, and the quantification precision was investigated in terms of root mean square error in prediction. Specifically, partial least squares (PLS) regression was initially applied on the parallelepipedal samples and tested on the cylindric ones. The obtained results pointed out the presence of two consequent linear trends with different slopes, which needed to be modelled separately to optimize performances. Non-linearity of light penetration was then confirmed by ordinary least squares (OLS) and multivariate curve resolution-alternating least squares (MCR-ALS), since both estimated the pixel-based composition of samples with limited accuracy. Subsequently, specific algorithms to deal with non-linearity were considered and applied. Then, regression by means of convolutional neural networks (CNN) was performed, with the aim of describing the whole range with a single model. Though CNN correctly dealt with the non-linear behaviour of the data, it confirmed that quantification becomes more challenging for deeper layers, as shown by a sudden increase of within-level standard deviation in the prediction results.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
