Suppressing sample morphology effects in near infrared spectral imaging using chemometric data pre-treatments C. Esquerre a, b , A.A. Gowen b, ⁎, J. Burger c , G. Downey a, b , C.P. O'Donnell b a Biosystems Engineering, School of Agriculture, Food Science and Veterinary Medicine, University College Dublin, Dublin 4, Ireland b Teagasc, Food Research Centre, Ashtown, Dublin 15, Ireland c BurgerMetrics SIA, Jelgava, Latvia abstract article info Article history: Received 5 August 2011 Received in revised form 21 December 2011 Accepted 8 February 2012 Available online 20 February 2012 Keywords: Spectral imaging Near infrared Chemometric pre-treatment Hyperspectral imaging Natural variability in the morphology (shape and size) of samples presents a difficulty in the use of NIR spec- tral imaging for their quality assessment, since the spectral variability introduced can often overshadow the variability arising due to differences in quality. In this study, combinations of chemometric pre-treatments were studied for suppression of sample morphology effects in the spectral domain, as an alternative to geo- metrical corrections. Asymmetric least squares (AsLs) baseline correction of logarithmic linearised reflec- tance log(1/R) was found to be the optimal pre-treatment in the comparative study. Logarithmic linearisation highlighted the differences in absorption bands while ASL was able to compensate for spectral offset and nonlinear baseline features, thereby enhancing chemical and physical differences between samples while the effect of the morphology of the sample on the spectra was attenuated. © 2012 Elsevier B.V. All rights reserved. 1. Introduction In recent years, near infrared (NIR) spectral imaging, also known as hyperspectral imaging, has become a key research tool in the develop- ment of in- or on-line technologies that may be used to assess critical process parameters or quality attributes of food and pharmaceutical products with the goals of reducing over-processing, enhancing consis- tency and minimising rejects [1,2]. NIR spectral imaging systems consist mainly of an illumination source, lens, spectrograph or tuneable filter, NIR camera and a computer [1–3]. This technology, which has its roots in remote sensing, provides both spatial and spectral information, there- by allowing quantification of spectral distribution related to sample ho- mogeneity, identification of local defects such as mould/bacterial contamination, bruises or off-colour spots [1–8]. Natural variability in the morphology (shape and size) of some prod- ucts, most notably those of agricultural origin, presents a difficulty in the use of NIR spectral imaging for their quality assessment [9–12]. Spectral variability due to spatial position in non-flat samples (the morphological effect) in addition to systematic variations due to light scattering and dif- ferences in the effective path length inside solids may be related to three key factors: (a) the distance and angle between the sample surface to both the illuminating source and detector [13] (b) shadows, specular re- flection and other features such as uneven illumination and reflection between samples and (c) the detector, which cannot be focussed simulta- neously for multiple image depths and wavelengths. A specific selection, orientation and position of the components of a spectral imaging system according to the size and shape of the sample can reduce the effect of sample morphology on recorded spectra, possibly making any additional correction of the spectra redundant. However, such systems would re- quire modifications in the configuration and position of their elements to work properly with different sample sizes or shapes. Recently, various geometrical corrections have been proposed to re- duce morphological effects in hyperspectral images while preserving the chemical/physical information of interest [9,14]. These methods as- sume that the sample reflects incident radiation as a Lambertian body (i.e. the intensity reflected in any direction is proportional to the cosine of the angle with the normal to the surface of the object) or as a function of the height of the sample. These methods successfully correct the mor- phology effect in the spectra of individual and relatively simple shapes (i.e. spheres and ellipsoids). Another approach is to use chemometric pre-treatments to compensate for the morphology effect. Chemometric pre-treatments work on the spectral dimension of hyperspectral data and can correct spectra for baseline shift (detrend, derivatives), multipli- cative effects (standard normal variate [SNV], multiplicative scatter cor- rection [MSC], and normalisation) either individually or in combination. In this study, a selection of chemometric pre-treatments have been eval- uated for morphological correction of spectra obtained in hyperspectral imaging experiments involving model (paper and cardboard) samples. 2. Spectral pre-treatments After de-noising spectra, an effective data pre-treatment should contribute to improvement of any subsequent exploratory analysis, Chemometrics and Intelligent Laboratory Systems 117 (2012) 129–137 ⁎ Corresponding author. Tel.: + 353 17167413; fax: + 353 17166104. E-mail address: aoife.gowen@ucd.ie (A.A. Gowen). 0169-7439/$ – see front matter © 2012 Elsevier B.V. All rights reserved. doi:10.1016/j.chemolab.2012.02.006 Contents lists available at SciVerse ScienceDirect Chemometrics and Intelligent Laboratory Systems journal homepage: www.elsevier.com/locate/chemolab