Non-Linear Data-Driven Methods for feature reduction in Hyperspectral imaging

Hyperspectral imaging (HSI) enables non-invasive and contact-free analysis of object composition by capturing hundreds of wavelength bands and providing their spatial distribution visualization and characterization. Thank to their high spectral resolution, hyperspectral images (HI) offer finer details about materials and structures that are beyond human vision capabilities. On the other side, the high dimensionality of hyperspectral data poses significant computational challenges, necessitating the use of dimensionality reduction techniques, particularly when integrating HI with advanced computational paradigms such as Machine Learning (ML) and Deep Learning (DL) approaches. Dimensionality reduction is essential for HSI and can be approached through two primary strategies: feature extraction and feature selection, the latter often referred to as band selection [1]. Feature extraction identifies essential spectral features in a transformed domain through non-linear approximation methods in a proper expansion basis. In contrast, feature selection focuses on selecting representative spectral bands directly from the original data employing non-uniform sampling methods. Two distinct strategies for dimensionality reduction of HI are proposed to enhance classification performance while reducing computational demands of resources and time in ML and DL frameworks. The first approach projects data into a transformed space, where relevant components are automatically selected based on the entropic normalized information distance [2]. The second approach utilizes the wavelet transform to perform non-uniform sampling of spectral bands, leveraging its ability to capture the non-stationary characteristics of signals [3]. These adaptive methods automatically determine the number of spectral bands required. Their respective properties are analysed and their effectiveness in ML/DL-based classification tasks is evaluated through comparative studies conducted on benchmark datasets.