Statistical Tests for Hyperspectral Coded Data Unsupervised Classification

We propose a novel method for unsupervised classification of coded hyperspectral acquisitions using a DD-CASSI (Double Disperser - Coded Aperture Spectral Snapshot Imager) system, which reduces the number of required acquisitions, typically by an order of magnitude. Leveraging the Separability Assumption (SA) and non-parametric Gaussianity statistical tests, our approach identifies homogeneous regions, which are areas of pixels made of the same material, and determines their unique spectral signatures directly from the coded measurements. By combining these statistical tests with spatial characteristics from panchromatic images, our iterative method effectively classifies regions without reconstructing the entire hyperspectral cube. This approach demonstrates the potential for accurate classification with minimal data, paving the way for optimized hyperspectral data analysis.