Carlotta Ferrari

Handling large datasets of hyperspectral images: reducing data size without loss of useful information.

Hyperspectral Imaging (HSI) is gaining increasing interest in the field of analytical chemistry, since this fast and non-destructive technique allows one to easily acquire a large amount of spectral and spatial information on a wide number of samples in very short times. However, the large size of hyperspectral image data often limits the possible uses of this technique, due to the difficulty of evaluating many samples altogether, for example when one needs to consider a representative number of samples for the implementation of on-line applications. In order to solve this problem, we propose a novel chemometric strategy aimed to significantly reduce the dataset size, which allows to analyze in a completely automated way from tens up to hundreds of hyperspectral images altogether, without losing neither spectral nor spatial information. The approach essentially consists in compressing each hyperspectral image into a signal, named hyperspectrogram, which is created by combining several quantities obtained by applying PCA to each single hyperspectral image. Hyperspectrograms can then be used as a compact set of descriptors and subjected to blind analysis techniques. Moreover, a further improvement of both data compression and calibration/classification performances can be achieved by applying proper variable selection methods to the hyperspectrograms. A visual evaluation of the correctness of the choices made by the algorithm can be obtained by representing the selected features back into the original image domain. Likewise, the interpretation of the chemical information underlying the selected regions of the hyperspectrograms related to the loadings is enabled by projecting them in the original spectral domain. Examples of applications of the hyperspectrogram-based approach to hyperspectral images of food samples in the NIR range (1000-1700 nm) and in the vis-NIR range (400-1000 nm), facing a calibration and a defect detection issue respectively, demonstrate the effectiveness of the proposed approach.

Minimisation of instrumental noise in the acquisition of FT-NIR spectra of bread wheat using experimental design and signal processing techniques

Spectral resolution (R) and number of repeated scans (S) have a significant effect on the S/N ratio of Fourier transform-near infrared (FT-NIR) spectra, but the optimal values of these two parameters have to be determined empirically for a specific problem, considering separately both the nature of the analysed matrix and the specific instrumental setup. To achieve this aim, the instrumental noise of replicated FT-NIR spectra of wheat samples was modelled as a function of R and S by means of the Doehlert design. The noise amounts in correspondence to different experimental conditions were estimated by analysing the variance signals derived from replicate measurements with two different signal processing tools, Savitzky–Golay (SG) filtering and fast wavelet transform (FWT), in order to separate the “pure” instrumental noise from other variability sources, which are essentially connected to sample inhomogeneity. Results confirmed that R and S values leading to minimum instrumental noise can vary considerably depending on the type of analysed food matrix and on the different instrumental setups, and helped in the selection of the optimal measuring conditions for the subsequent acquisition of a wide spectral dataset.

The potential of spectral and hyperspectral-imaging techniques for bacterial detection in food: A case study on lactic acid bacteria

Official methods for the detection of bacteria are based on culture techniques. These methods have limitations such as time consumption, cost, detection limits and the impossibility to analyse a large number of samples. For these reasons, the development of rapid, low-cost and non-destructive analytical methods is a task of growing interest. In the present study, the capability of spectral and hyperspectral techniques to detect bacterial surface contamination was investigated preliminarily on gel cultures, and subsequently on sliced cooked ham. In more detail, two species of lactic acid bacteria (LAB) were considered, namely Lactobacillus curvatus and Lactobacillus sakei, both of which are responsible for common alterations in sliced cooked ham. Three techniques were investigated, with different equipment, respectively: a macroscopic hyperspectral scanner operating in the NIR (10,470-5880cm(-1)) region, a FT-NIR spectrophotometer equipped with a transmission arm as the sampling tool, working in the 12,500-5800cm(-1) region, and a FT-MIR microscopy operating in the 4000-675cm(-1) region. Multivariate exploratory data analysis, in particular principal component analysis (PCA), was applied in order to extract useful information from original data and from hyperspectrograms. The results obtained demonstrate that the spectroscopic and imaging techniques investigated can represent an effective and sensitive tool to detect surface bacterial contamination in samples and, in particular, to recognise species to which bacteria belong.

Iodine Value and Fatty Acids Determination on Pig Fat Samples by FT-NIR Spectroscopy: Benefits of Variable Selection in the Perspective of Industrial Applications

In this work, FT-NIR spectroscopy was employed to determine iodine value (IV) and fatty acids (FA) content of pig fat samples, through the combined use of signal preprocessing, multivariate calibration, and variable selection methods. In particular, the main focus was on the use of variable selection methods, both in order to improve the predictive performance of the calibration models, and to identify relevant wavelengths that could be subsequently used for the development of simple, fast, and cheap hand-held devices, able to measure IV and FA content directly on the fat without the need of any sample pretreatment. Firstly, for each property of interest, partial least squares (PLS) multivariate calibration models were calculated considering the whole spectral range and testing different signal preprocessing methods. Then, once chosen the optimal signal preprocessing method, a two-step variable selection procedure was applied. In the first step, the interval-PLS variable selection algorithm was used to calculate a set of calibration models, whose outcomes were considered altogether in the second step, in order to select the optimal calibration model. The variable selection procedure allowed to lower the number of spectral variables retained by the model, and often led to an increase of the performance in prediction of the external test set samples.