Hyperspectral imaging techniques: an attractive solution for the analysis of biological and agricultural materials.
- Baeten, V. , Dardenne, P. & Fernández Pierna, J.A. (2007). Hyperspectral imaging techniques: an attractive solution for the analysis of biological and agricultural materials. In: Techniques and Applications of Hyperspectral Image Analysis, Hans F. Grahn & Paul Geladi Editors. 289-311.
| Type | Book Section |
| Year | 2007 |
| Chapter title | Hyperspectral imaging techniques: an attractive solution for the analysis of biological and agricultural materials. |
| Book title | Techniques and Applications of Hyperspectral Image Analysis |
| Editor | Hans F. Grahn & Paul Geladi Editors |
| Label | U15-0496 |
| Recnumber | 44 |
| Pages | 289-311 |
| Endnote Keywords | Hyperspectral imaging|Chemometrics|Support Vector Machines|Classification|Screening| |
| Abstract | Hyperspectral imaging data and Support Vector Machines for the trace back of compound feeds J. A. Fernández Pierna1, V. Baeten2, I. Fissiaux2, A. Michotte Renier2 and P. Dardenne2 1 Scientific collaborator F.N.R.S. Statistics and Informatics Department, University of Agronomical Sciences, Avenue de la Faculté 8, 5030 Gembloux, Belgium 2 Walloon Agricultural Research Centre (CRA-W), Quality of Agricultural Products Department, Chaussée de Namur n°24, 5030 Gembloux, Belgium. (baeten@cra.wallonie.be) A complete screening of feedingstuffs, i.e. a quantitative and qualitative determination of vegetal and mineral feed ingredients, is decisive in the support of the Directive 2002/2/EC of the European Parliament on the marketing of compound feeds. In this document, the Commission underlines the labelling previsions of compound feeds in order to facilitate the trace back of feed materials. It implies a compulsory declaration for all the feed materials as well as their amount in compound feeds. For this reason, it was decided to apply our knowledge on hyperspectral NIR imaging in order to perform this screening. Hyperspectral imaging consists in the analysis of several hundreds of particles being the result of the grinding of compound feeds. The major advantages of this technique are that a) the recognition is independent on the expertise of the analyst, b) it is possible to automate the analysis and to analyse a large quantity of samples per unit of time, c) the method is non destructive allowing further analysis of the material by other methods and d) this technique has a low limit of detection. To perform the trace back of the materials, a classification tree was built to sort the particles following a dichotomist way. At each node of this tree, a discrimination model is used to classify the particles. The first step of the process was to create the discriminating models from the hyperspectral databases of each class of raw material. The models will allow to separate the particles of interest (labelled as +1) from the other particles (labelled as -1). Chemometric method SVM - Support Vector Machines- with a RBF Gaussian kernel was used as classification method to construct the discriminating models. Once the classification tree has been constructed, the aim is to find, for a new sample composed of different feed materials, its real composition by testing such a sample in each of the models constituting the tree. For each step of the tree, particles are tested in order to verify whether they belong to class +1 or to class -1. Then, only the spectra considered as -1 are analysed at the next step of the tree. When analysing the results of each model, it is possible to know whether a kind of material is present or not and to estimate the percentage of each material present in the new sample. As general conclusion, hyperspectral data in combination with SVM as chemometric classification technique seems to be a promised methodology for the determination of open formulations. As future work a larger data base has to be constructed in order to cover the whole diversity of materials included in compound feeds. Moreover the discrimination equations could be improved by performing a correct optimisation of the parameters included in the SVM models. |
| Author address | Baeten Vincent, Quality Department of Agro-food Products, Walloon Agricultural Research Centre (CRA-W), Chaussée de Namur, 24, B-5030 Gembloux, baeten@cra.wallonie.be |
| Fichier | |
| Caption | U15-0496-baeten-2007.pdf |
| Authors | Baeten, V., Dardenne, P., Fernández Pierna, J.A. |
Team
