Tutorial: Time series hyperspectral image analysis

Abstract

A hyperspectral image is a large dataset in which each pixel corresponds to a spectrum, thus providing high-quality detail of a sample surface. Hyperspectral images are thus characterised by dual information, spectral and spatial, which allows for the acquisition of both qualitative and quantitative information from a sample. A hyperspectral image, commonly known as a "hypercube", comprises two spatial dimensions and one spectral dimension. The data of such a file contain both chemical and physical information. Such files need to be analysed with a computational "chemometric" approach in order to reduce the dimensionality of the data, while retaining the most useful spectral information. Time series hyperspectral imaging data comprise multiple hypercubes, each presenting the sample at a different time point, requiring additional considerations in the data analysis. This paper provides a step-by-step tutorial for time series hyperspectral data analysis, with detailed command line scripts in the Matlab and R computing languages presented in the supplementary data. The example time series data, available for download, are a set of time series hyperspectral images following the setting of a cement-based biomaterial. Starting from spectral pre-processing (image acquisition, background removal, dead pixels and spikes, masking) and pre-treatments, the typical steps encountered in time series hyperspectral image processing are then presented, including unsupervised and supervised chemometric methods. At the end of the tutorial paper, some general guidelines on hyperspectral image processing are proposed.