Techniques - Multispectral Imaging (MSI)

A camera/detector collects images of a sample from the reflected light of a light source in the visible and near-infrared regions usually in the range 400-1000nm. The sample is scanned typically using 20 different wavelengths intervals of the light source. Hyperspectral cameras compile spectral and spatial information in the form of a datacube. This forms a stack of images, where each image discloses a different wave band of reflected light, i.e., the light striking each pixel of a hyperspectral camera is broken down into its component spectral bands. As is the case in classical UV-Vis spectroscopy, different components or molecular structures will give different spectra, which translate into different colours.

The resulting image has to be analysed to be able to discriminate between the authentic component and the adulterating one. This uses computer software which analyses the spectral signature of all the pixels represented in each layer and builds a model to discriminate between the two classes of the sample.  An example, below, is the addition of 10% common wheat to a sample of premium durum wheat.

Image - Crown Copyright

The wheat grains are visually identical.  Software removes the pixels of the background picture, leaving just the grains. The discrimination model is then applied to every pixel in the image, which grades the durum wheat spectrum as red, and the common wheat spectrum as blue.

MSI has been widely applied to the discrimination of different types of visually-similar grains and seeds.

The technique is non-destructive, non-invasive, instant and cheap.  It is ideal for in-line applications, scanning (for example) raw material intake.  It is ideal when the material is highly standardised, and the question being asked is "Does this batch look different?"

The key to success is in constructing the reference datasets used to train the classification model (as is the case with all techniques reliant on a statistical classification model, generally built using supervised Machine Learning),  The "genuine" samples used to train the model must be of absolute known provenance and must be fully representative of all of the variables (e.g. countries of origin, agricultural inputs, seasons and weather patterns) likely to be encountered amongst the test samples.  The training data must be fully randomised, and a set of known samples which the model has never seen must be reserved as a validation set.

Classification models such as MSI will never give a definitive result and - although the result is probabilistic - it is difficult to put a numeric uncertainty on the classification result (as this would assume normally distributed statistics, rarely the case).  So laboratory results are usually framed in terms of "The sample is consistent with ...." or "The sample is not consistent with ....." the authentic sample type, without giving any indication of the confidence that can be placed on the conclusion.