Exploring local spatial features in hyperspectral image

We propose a methodological framework to extract spatial features in hyperspectral imaging data and establish a link between these features and the spectral regions, capturing the observed structural patterns. The proposed approach consists of five main steps: i) two dimensional Stationary Wavelet Transform (2D-SWT) is applied to a hyperspectral data cube, decomposing each single-channel image with a selected wavelet filter up to the maximum decomposition level; ii) a grey-level co-occurrence matrix is calculated for every 2D-SWT image resulting from stage i); iii) distinctive spatial features are determined by computing morphological descriptors from each grey-level co-occurrence matrix; iv) the morphological descriptors are rearranged in a two dimensional data array; v) this data matrix is subjected to Principal Component Analysis (PCA) for exploring the variability of the aforementioned descriptors across spectral channels. As a result, groups of spectral wavelengths associated to specific spatial features can be pointed out yielding a better understanding and interpretation of the data. In principle, this information can also be further exploited, e.g. to improve the separation of pure spectral profiles in a multivariate curve resolution context.