Selection of representative hyperspectral data and image pretreatment for model development in heterogeneous samples: A case study in sliced dry-cured ham

Abstract

Sliced dry-cured ham arranged in ready-to-eat packages is a convenient and widely consumed commodity characterised by heterogeneity in composition not only among different industrial batches but also through their horizontal and vertical profiles, making precise nutrition labelling of the packages a difficult task. Hyperspectral imaging techniques can serve as a steadfast solution not only to predict the overall composition of the major constituents of dry-cured ham but also to visualise their distributions. The main aim of this study was to define the optimal protocol for pretreating hyperspectral images and selecting representative hyperspectral data for developing accurate predictive models in excessively heterogeneous samples, using sliced dry-cured ham as a case study. Hyperspectral images (400–1000 nm) were acquired for heterogeneous sliced dry-cured ham and homogeneous unsliced dry-cured muscles. Partial least squares (PLS) regression models to predict fat, water, salt and protein contents were developed and tested in an independent dataset. The PLS predictive models developed from the whole surface of sliced dry-cured ham were the most accurate ones for predicting fat, water, salt and protein contents with a determination coefficient in prediction () of 0.89, 0.85, 83 and 0.63 and standard error in prediction (SEP) of 1.43, 1.21, 0.51 and 1.57%, respectively. The chemical images resulting from the models gave advantages of hyperspectral imaging technique over traditional chemical methods to visualise the spatial distribution of different constituents within the packaged ham slices.