James Burger

Spectral pre-treatments of hyperspectral near infrared images: analysis of diffuse reflectance scattering

Scattering effects are often encountered when measuring diffuse reflectance near infrared (NIR) spectra of solid and semi-solid materials. How does this phenomenon effect hyperspectral imaging of powders? A series of hyperspectral NIR images of particle size fractions of commercial grade salt and sugar were acquired. Spectral pre-processing techniques, including Kubelka–Munk, standard normal variate and absorbance transforms, unit length or unit area normalisation, first and second derivative transforms, and several variants of multiplicative scatter corrections (MSC) were applied to the images and examined for their effectiveness at reducing or eliminating scatter effects. Principal component analysis (PCA) scoreplots produced expected results: derivative transforms reduced variance, but did not eliminate the particle size dependencies; piecewise MSC transforms reduced the data to two clusters, one for salt and one for sugar. Partial least squares (PLS) regression was applied to examine the impact of the pre-processing transforms on prediction of particle size. RMSEP values between 10 and 50 μm were determined for particle fractions ranging between 140 and 315 μm for all transforms except the piecewise MSC; in spite of the reduction in additive and multiplicative effects, enough correlated variance remained after application of the pre-processing transforms to allow prediction of particle size ranges from PLS models. Additional scatter effect information was obtained by examining particle size distribution histograms and spatial particle size mappings facilitated by the hyperspectral images.

Time series hyperspectral chemical imaging data: Challenges, solutions and applications

Hyperspectral chemical imaging (HCI) integrates imaging and spectroscopy resulting in three-dimensional data structures, hypercubes, with two spatial and one wavelength dimension. Each spatial image pixel in a hypercube contains a spectrum with >100 datapoints. While HCI facilitates enhanced monitoring of multi-component systems; time series HCI offers the possibility of a more comprehensive understanding of the dynamics of such systems and processes. This implies a need for modeling strategies that can cope with the large multivariate data structures generated in time series HCI experiments. The challenges posed by such data include dimensionality reduction, temporal morphological variation of samples and instrumental drift. This article presents potential solutions to these challenges, including multiway analysis, object tracking, multivariate curve resolution and non-linear regression. Several real world examples of time series HCI data are presented to illustrate the proposed solutions.

Bad pixel detection in hyperspectral staring camera systems

Introduction N IR chemical imaging based on hyperspectral images or hypercubes has been made possible by the advent of InGaAs, MCT (HgCdTe) and InSb focal plane array detectors. These sensor arrays (cameras) permit the acquisition of tens of thousands of spectra within seconds. Coupled with chemo metric analysis tools, this enables the spatial determination of both quantitative and classification information of individual chemical constituents, thereby providing sample uniformity and local purity information. But there is an old computer adage that simply states: garbage in = garbage out. With the massive quantity of data produced by a hyperspectral imaging system, how can one assure that the quality of the acquired image data is optimum—free of garbage data? This article discusses one of the issues effecting hyperspectral image data quality: the detection of bad pixels in staring focal plane detectors. Subsequent articles will discuss additional aspects of improving image quality: detection of bad pixels in pushbroom camera systems, bad pixel replacement techniques and the implications of reflectance transforms and other spectral pre-processing treatments specific to hyperspectral imaging. The current high-tech world surrounds us with a barrage of digital imaging devices: LCD devices for computers and TV, digital phones and of course digital colour cameras. In spite of these devices containing millions of pixels, humans are very adept at identifying even individual bad pixels— those pesky white or black dots at the most inappropriate spot of an image. Within the context of a simple visual image these bad pixels may be acceptable, but there may be serious consequences if, for example, hyperspectral image calibration spectra for a classical PLS model include bad pixel data. A single bad pixel can have a very large leverage effect on a simple regression model. But just what is a bad pixel and how do we detect it? All sensors exhibit a baseline noise level dependent on random fluctuations in the photon flux, quantum efficiency and readout electronics. Whether we use a CCD or CMoS device, we expect to see a uniformly continuous increase in digital signal value associated with an increase in illumination or photon flux. There may be nonuniformities across the sensor device, but this phenomenon is addressed with nonuniformity correction (NUC) which also corrects for system optics throughput and spatial illumination variations. Bad pixels are pixels which behave substantially differently based on anomaly detection or filter algorithms. There are several types or classes of bad pixels. The easiest bad pixels to detect are those which have no response to light (dead ), are continuously saturated (hot ) or output a nearly constant intermediate value (stuck). Some pixels may have a significantly non-linear response to light intensity, while others may be excessively noisy. Bad pixels may produce corrupted data continuously, or only intermittently, sometimes referred to as blinkers or speckles. Hyperspectral images or hypercubes contain data representing three dimensions: two dimensions forming the spatial image space and a third spectral dimension representing wavelength variation. Twodimensional focal plane detectors can be used in two different orientations to acquire hyperspectral images. A series of twodimensional spatial–spatial images can be acquired while changing a bandpass filter in front of the camera. This is referred to as a staring system, with the camera staring at the object or scene space. An alternative configuration places the camera orthogonal to this position so that it acquires image frames with spatial–spectral dimensions. The second hyperspectral spatial dimension is obtained by physically scanning across the sample of interest. This acquisition mode is referred to as a pushbroom mode. The approach to bad pixel detection in these two different acquisition configurations varies significantly. only staring imaging systems will be discussed in the current article.

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.

Replacement of hyperspectral image bad pixels

Introduction d igital imaging sensor arrays, either CCd or CMoS devices, typically contain individual sensor array elements that may routinely generate erroneous data values. Individual sensor readout values may be stuck, hot or cold, noisy, or they may drift. When digital imaging devices are used for hyperspectral image acquisition (3d datasets with spatial × spatial × spectral dimensions), the 2d sensor arrays may be aligned parallel to the sample surface in a staredown configuration (spatial × spatial) or they may be perpendicular to the sample surface in a pushbroom configuration (spatial × spectral). The orientation of the sensor array with respect to the sample has significant implications on the perturbations of the image dataset or hypercube caused by bad pixels. When a single bad pixel occurs in a staredown system, a single dataset spectrum is perturbed at all wavelengths, whereas in a pushbroom system, a set of spectra is perturbed, each at a single wavelength point. These differences in dataset error structure impacts the choice of processing algorithms used for both the detection of bad pixels, and the creation of replacement data values. The detection of hyperspectral image bad pixels has been discussed in two previous articles. This current paper discusses different approaches to hyperspectral image bad pixel replacement, primarily with regards to pushbroom imaging systems.

Classification and Prediction Methods

Quantification or classification is a key objective of hyperspectral chemical imaging (HCI). In this chapter, steps involved in the development of classification and prediction models from HCI data are explained, from initial data pretreatment, through to calibration set development and construction of prediction maps using partial least squares discriminant analysis (PLS-DA). As a case study, we demonstrate the prediction of seven classes in an example near infrared (NIR) HCI image containing objects of different chemical composition. The varied effects of spectral pretreatments, sampling procedures, threshold estimation methods and image post-processing are compared in terms of classification model performance and via visual interpretation of prediction maps. The use of graphical data representations such as histograms and color coded prediction maps is presented for further understanding and aiding the development of multivariate model classification models.

The Application of Hyperspectral Chemical Imaging to Chemometrics

Introduction A pplications of hyperspectral chemical imaging (hCI) have expanded greatly both in number and in scope in recent years. the assimilation of hCI into analytical chemistry has been achieved primarily on the basis of chemometrics. the massive amount of data contained within a single hyperspectral image (thousands of spectra or hundreds of images) requires chemometric techniques to extract meaningful chemical information. In essence, “chemometrics provides an understanding of hCI”. there is a corollary to this trend which we present in this paper: “hCI enables an understanding of chemometrics”. Further to this we demonstrate that hCI data can be a valuable tool for the teaching of chemometrics. When we explore and validate chemometric principles, we rely heavily on visual perception; take, for example, the interpretation of scatter plots and histograms. hCI further extends visual interpretation of chemometric analyses by introducing spatial images. the speed and performance of our visual processing abilities permits rapid analysis of this spatial (image) representation of data and enables, for example, pattern or object recognition, analysis of spatial distributions (texture, homogeneity, particle size) and detection of outliers. We can further utilise this “image advantage” by observing the effects of processing on familiar image scenes. In this paper, we present and explore some of the fundamental properties of some basic chemometric techniques by examining the results of their application to an example image. the image is a near infrared (NIR) hyperspectral image of a powdered soup mix, purposely contaminated with several foreign objects (a string, piece of wood and paperclip). the hCI hypercube has spatial dimensions of 418 pixels × 318 pixels (each 100 μm × 100 μm in size) and 207 wavelength channels (962– 1643 nm), acquired with a BurgerMetrics hyperPro NIR imaging system. Figure 1 shows a median wavelength intensity spatial image with identified features. the soup mix represents a heterogeneous image background containing identifiable features (noodles and dried tomato pieces) while the selected contaminants represent some real world artefacts with differing spectral signatures. hCI hypercubes are three dimensional (spatial × spatial × wavelength) data structures; however, most chemometric techniques are based on operations on two-dimensional structures (matrices). Consequently, the three-dimensional hypercube data must first be rearranged. this is typically achieved by unfolding the three-dimensional hypercube into a two-dimensional matrix of spectra by stacking each pixel spectrum, one on top of the other, as shown in Figure 2. this may be followed by a chemometric technique that favours two-dimensional structures, for example, principal components analysis (PCA.) Processing results (in this case score vectors) can then be refolded back to obtain spatial (score) images. Such unfolding > processing > refolding operations are common throughout many hCI chemometric operations. Figure 1

Bad pixel detection in hyperspectral pushbroom camera systems

10 Introduction H yperspectral images are collections of light intensity measurements represented as threedimensional data hypercubes in which the dimensions correspond to one spectral and two spatial axes. Focal plane array detectors are widely used to collect image data from two of these dimensions; scanning the third dimension results in a series of data frames forming the hypercube. Faulty focal plane sensor elements can result in data values which are abnormally low, high or excessively noisy. How can these bad pixels be systematically identified? The choice of an appropriate detection technique is dependent on the orientation of the focal plane array axes relative to the spatial or spectral hypercube axes. A previous article described implications with the stare-down image acquisition approach, where the focal plane array measures spatial × spatial image data. This was exemplified by hypercubes acquired at the Centre Wallon de Recherches Agronomique (CRA-W) labs using a staredown camera. The current article presents bad pixel detection from the perspective of linescan or pushbroom image acquisition where the focal plane array contains spatial × spectral data. Within a focal plane array, an individual pixel sensor may be labeled bad when its output differs significantly in performance from the measured output of surrounding neighbour pixels. When the surrounding neighbour pixels span both spatial and spectral variations, defining differs significantly can become difficult—neighbouring data values may appear consistent across the spatial axis but vary significantly in the wavelength or spectral axis. How can pixel sets be selected or their data values transformed to provide more uniform neighbourhoods for bad pixel detection? Various bad pixel detection schemes are examined in this article. The topic of computing a replacement value for the bad pixel sensor will be discussed in a subsequent article. experimental NIR hyperspectral images were acquired using a HyperScan camera system from BurgerMetrics (www.burgermetrics.com) which utilised a pushbroom InGaAs camera system. Every individually acquired camera frame (320 × 212 pixels) contained 320 spectra each consisting of 212 wavelength points with 3.3 nm steps spanning the 960–1662 nm wavelength range. For noise reduction purposes, a series of 16 frames was acquired, averaged and saved. Complete hypercubes were obtained by acquiring a sequence of 101 averaged frames while repositioning the target sample in 100 μm steps with a motion direction perpendicular to the 32 mm long linescan spatial axis, thus providing 100 × 100 μm spatial resolution. For non-uniformity correction (NUC) and reflectance transformation purposes, dark and white reference frames were computed as the average of 32 averaged frames obtained while either blocking the lens entrance (dark) or scanning a 99% standard reflectance material target (www. sphereoptics.com—white). The primary sample target was a homogenous piece of white polyethylene plastic. A series of 21 hyperspectral images of this material was acquired while increasing the camera integration time from 5 ms to 105 ms in steps of 5 ms. All numerical processing and detection algorithms were implemented in MATlAB.