Data processing without the bruises

NIR hyperspectral imaging has already demonstrated its great potential for food analysis and testing. By producing images in which every pixel contains spectral data, NIR hyperspectral imaging can highlight how the texture or chemical composition varies across a food stuff. This has allowed it to identify tough bits of a prawn (see No more dodgy prawns) and determine peanut contamination of wheat flour (see Near infrared hyperspectral imaging detects peanut contamination).

The only real problem with NIR hyperspectral imaging is that producing images in which every pixel contains spectral data tends to result in a lot of data, with a single hyperspectral image often taking up more than 50MB. Processing and analysing all the data produced by NIR hyperspectral imaging can thus be very time consuming, even with powerful computers.

Now, scientists from the University of Modena and Reggio Emilia in Italy, led by Alessandro Ulrici, have come up with a method for reducing the size of the data produced by hyperspectral imaging, while still retaining all its pertinent features. Their method involves converting each hyperspectral image into a one-dimensional signal that can still convey spatial and spectral information, which Ulrici likens to a fingerprint of the image data. This signal, which they term a hyperspectrogram, should be much easier to analyse by statistical techniques such as principal component analysis (PCA) and partial least squares-discriminant analysis (PLS-DA).

As a first test of this method, they tried using it to identify bruised apples, an application for which NIR hyperspectral imaging is ideally suited. They analysed 80 apples from two different varieties, Golden Delicious and Pink Ladies, producing 800 images. Converting these 800 images into hyperspectrograms reduced the size of the data by more than half, from 12GB to 4.7GB.

As reported in Chemometrics and Intelligent Laboratory Systems, they then analysed this data by PCA, finding that it split the apples into two groups representing the two varieties. When they analysed these two varieties by PCA, it split them into two broad groups representing pristine and bruised apples. Next, they used PLS-DA to produce a model that could accurately identify bruised apples from the hyperspectrograms.

By taking advantage of the spectral features picked out by the PLS model to identify the bruised apples, Ulrici and his colleagues were also able to convert the hyperspectrograms into images that showed the location of the bruises. They could do this even when the bruises weren’t yet clearly visible on the actual apples.