Neighbouring pixel data augmentation: a simple way to fuse spectral and spatial information for hyperspectral imaging data analysis

Hyperspectral imaging has become an essential tool in analytical chemistry. Indeed, it is used to explore complex and heterogeneous samples providing simultaneously spectral and spatial information in the acquired data cube. Many chemometric tools have been developed and optimised to extract latent features in such data structure, but it has to be said that the spatial information (ie, pixel neighbourhood) is almost not used during the data analysis framework. It is in this sense that we propose the neighbouring pixel data augmentation as a new and simple way to consider all dimensions of the cube. We will share the potential of the concept on the basis of 2 data set explorations using Raman hyperspectral imaging and deep UV synchrotron fluorescenxce microscopy. It shall be first demonstrated that more information can be extracted when neighbouring pixel data augmentation is used. Furthermore, a good behaviour against noise will be also observed.