Thursday 2nd November 2017
Multivariate hyperspectral data visualisation
and classification of tea products
02/11/17 AT 3pm (UK time)
Abstract for webinar:
Hyperspectral (HS) data generated from different optical spectroscopy techniques is often characterised as high dimensional with a large number of correlated dimensions leading to a much lower intrinsic dimensionality. To visualise the structure in the data, the HS data is transformed to a lower dimensionality explaining the data regarding clusters at various scales. When transforming the data from high to low dimensional space, the shape of the data manifold in the higher dimension largely determines the effectiveness of the visualisation method regarding preserving the structure of the data. In cases where the manifold represented by the data points is linear, classical methods like principal component analysis (PCA) and multi-dimensional scaling (MDS) can perform well in preserving the structure. However, when the points in high dimensional space lie very near or in the non-linear manifold, methods like PCA and MDS fail to capture the structure. This is because the aim of these methods is to keep distant points as far as possible and do not utilise information from neighbouring data points. In this case, non-linear methods utilising neighbourhood information can perform better.
In this webinar, different data visualisation techniques will be explored for visualising the near-infrared (NIR) hyperspectral data of different tea products. Furthermore, different multi-class support vector machine classification modelling will be explored for classification of different commercial tea products and tea products of different geographic origin. Different variable selection methods (Filter and wrapper-based) will also be explored to have insight on the success of different classification models.
This webinar will be presented by Puneet Mishra from the University of Strathclyde.
If you wish to register for the webinar please contact Christine.firstname.lastname@example.org