Gerry Downey

Near infrared hyperspectral image regression : on the use of prediction maps as a tool for detecting model overfitting

Calibration models developed from hyperspectral imaging data may be applied at the pixel level to generate prediction maps that estimate the spatial distribution of components in a sample. Such prediction maps facilitate direct visual interpretation of model performance, and performance indicators can be extracted from them. These maps can be used as a tool to evaluate calibration models developed on hyperspectral imaging data. This paper presents a method for calibration model evaluation based on information obtained from prediction maps and demonstrates its usefulness for preventing overfitting. Partial least-squares regression was used for model calibration in this study, although in principle the proposed method may be used to evaluate other multivariate calibration methods, e.g. ridge regression and principal-components regression.

Guest Editorial: Enhancing near Infrared Spectroscopy with an Added Spatial Dimension

Hyperspectral imaging is a technique which combines spectral and spatial information from an object under investigation.1 It was initially developed in the 1980s for remote sensing applications2 and it was in this context that the term was first used. From that time, developments in equipment and applications, together with a wider realisation of its potential in both industrial and laboratory studies, have resulted in its spread to scientific fields as diverse as astronomy, agriculture, pharmaceuticals and medicine over the last 30 years. In the 1990s, examples of near infrared (NIR) hyperspectral imaging applications began to appear, but it was not until the early years of the present century that it moved into the mainstream of analytical techniques in the agriculture and food domains. Initial applications mainly focused on food safety issues,3-5 but this was rapidly followed by many other studies which have been the subject of recent extensive reviews.1,6–9 There are two basic approaches to the acquisition of hyperspectral images. The first of these involves collecting consecutive whole images at each wavelength available using, generally, a tuneable filter—an acousto-optic tuneable filter (AOTF) or liquid crystal tuneable filter (LCTF). Such devices are referred to as “staring imager” configurations and have the advantages of good spatial resolution and collection of the entire image at once. A disadvantage involves the possibility of changes in the image during the time necessary to record all the wavelengths. The alternative approach is to record complete spectral information from one line of the sample at a time; this requires the sample to move relative to the instrument and is normally referred to as a “pushbroom” system. Many systems now use this approach if for no other reason than because it simulates movement of samples along a conveyor belt in industrial situations and may thus produce information more closely-related to such applications than the “staring imager” approach. For example, the pharmaceutical industry has mainly used staring imaging technology while applications in food and agriculture have tended to be focused on pushbroom imaging technology.1 Traditionally, quality measurements on food focussed on proximate composition and overall appearance. Given considerable impetus by labelling requirements, particularly in the case of processed foodstuffs, this type of approach fails to address problems which are not discernible to the naked eye or even to cope with the increasing speed of analysis demanded by the modern food industry. NIR spectroscopy partially solved the issue of analytical speed and accuracy but spectrophotometers only provide an average spectrum of a sample and require high sample homogeneity to ensure representative measurement of a specific quality property. Depending on the uniformity of the distribution of the specific property being measured, repeated measurements might be required to obtain a representative average spectrum. NIR hyperspectral imaging, on the other hand, results in an image consisting of hundreds of single-channel images stacked on top of each other with each image representing a single wavelength in the NIR region, i.e. each pixel represents a single spectrum which is theoretically different to its neighbours. It is this availability of spectral information for each pixel in the NIR image that makes NIR hyperspectral imaging ideal for heterogeGuest editorial: Enhancing near infrared spectroscopy with an added spatial dimension

Hyperspectral imaging for mushroom (agaricus bisporus) quality monitoring

A method for mushroom quality grading based on hyperspectral image analysis in the wavelength range 400-1000 nm is presented. Different spectral and spatial pretreatments were investigated to reduce the effect of sample curvature on hyperspectral data. Algorithms based on chemometric techniques (Principal Component Analysis and Partial Least Squares Discriminant Analysis) and image processing methods (masking, thresholding, morphological operations) were developed for pixel classification in hyperspectral images.

Interview: Anna de Juan Capdevila

In this interview, we have departed a little from profiles of mainstream NIR researchers and hear from Dr Anna de Juan Capdevila. Anna will be known to many of our readers as one of the leading chemometricians involved in developing new ways to process and interpret hyperspectral images. I hope you find her profile as interesting as I did.—Ed.

Interview: Celio Pasquini

In the context of our upcoming ICNIRS Conference in Brazil, I thought it would be interesting to learn a bit more about Celio Pasquini, our main host for the event. I think you will find his information interesting and some of the images at least should whet your appetite for a visit to Foz do Iguassu! Ed.

Hyperspectral imaging: shining light on mushroom quality

10 T he common white mushroom, Agaricus bisporus, is the most widely cultivated mushroom variety worldwide. Since its introduction in the 1930s, Irish production of this small fungus has evolved into a multi-million Euro industry with the result that Agaricus bisporus is now the most important horticultural export from Ireland. The first cultivated strain was brown in colour but consumer demand for white mushrooms led to the development of strains of what we know today as the common white mushroom. Although trends in mushroom consumption vary widely around the world, white mushrooms remain the most popular choice in Ireland and the uK today. In consequence, Agaricus bisporus are graded according to their appearance; white and blemish-free are the main criteria for high quality. Colorimetry was first applied to mushroom quality evaluation in the 1970s when significant correlations were found between Hunter l-value and sensory quality. Since then, different optical techniques [e.g. near infrared (NIR) spectroscopy and red-green-blue (RGB) imaging] have been investigated for mushroom quality estimation; however, l-value measurement remains the industry standard. Hyperspectral imaging (HSI) enables the acquisition of images over a range (>100) of contiguous wavebands, typically ranging from the visible region up to the NIR. Hyperspectral images, known as hypercubes, may be represented as three-dimensional blocks of data, comprising two spatial and one wavelength dimension, as illustrated in Figure 2. The hypercube allows for the visualisation of various constituents of a sample, separated into particular areas of the image, since regions of a sample with similar spectral properties tend to have similar chemical composition. There are two conventional ways to construct a hypercube. One method, involves keeping the spatial image field of view fixed and collecting images one wavelength after another by filtering the incoming light. The other configuration requires relative movement between the object and the detector; this pushbroom method is particularly well suited to conveyor belt systems. Our research team, a joint collaboration between university College Dublin, Dublin Institute of Technology and the Ashtown Food Research Centre (Teagasc), received funding from the Irish government in 2006 to investigate the potential of hyperspectral imaging for mushroom quality evaluation. Beginning in 2007, our HSI work to date has focussed mainly on the Vis-SWNIR (400–1000 nm) wavelength range since this region is also suitable for monitoring changes in colour which are highly-relevant to mushroom quality assessment. Our team have demonstrated the potential of pushbroom hyperspectral imaging in the VisSWNIR (Figure 3) for detection of blemishes on the mushroom surface, characterisation of the sources of such blemishes and prediction of quality parameters such as moisture content, texture and enzymatic activity. We have also examined the use of different spectral pre-processing methods to reduce the effect of sample curvature and for prediction of mushroom quality without removing packaging. However, in order to keep within the scope of NIR news, the remainder of this article deals with hyperspectral images of mushrooms obtained in the NIR (1000–1650 nm) region. Hyperspectral imaging poses some significant challenges compared to point NIR spectroscopy. With detector array sizes of 580 × 580 pixels (and above), hypercubes can contain >300,000 spectra! The problem of extracting useful information from such data is currently a topic of active research; numerous supervised and un-supervised model building strategies for analysis of hyperspectral imaging data may be found in the literature. Discriminant analysis methods commonly used in point spectroscopy, such as PlS-DA and lDA, can readily be applied to spectral data contained in hypercubes. Once a suitable classifier has been trained, it can be applied to an entire hypercube resulting in prediction maps in which the class of each pixel can be identified by its colour or greyscale. One major challenge in the development of models for hyperspectral image classification arises in the selection of appropriate data for model caliHyperspectral imaging: shining light on mushroom quality