Reduction and segmentation of hyperspectral data cubes

Reduction and segmentation of multiwavelength images be- come problematic when the number of bands increases. Integral fleld spectroscopy and other instrument designs allowing for enhanced spectral and spatial resolution lead to extremely large hyperspectral data cubes (typically 370 millions of pixels per exposure for the MUSE instrument). New analysis tools exploring jointly spectral and spatial features are re- quired. We propose a new approach, based on the Mean-Shift method (Comaniciu & Meer 2002), to reduce the dimensionality of large data cubes and extract the main spectral patterns. A set of spectra extracted from the cube is used as an initial reference basis. Each spectrum in the observation is projected on this basis, and represented by a vector of pro- jection coe-cients or weights. The Mean-Shift method is then carried out for the whole dataset to flnd the modes in the projection space. These modes are selected for a new projection basis and the algorithm is iter- ated until convergence. The distance between two spectra is deflned as the angle between their related vector coe-cients to increase e-ciency: this speeds up the convergence and gives more weight to the compari- son of spectral patterns, minimizing the efiect of the average intensity of each spectrum. This approach has been tested on simulated data. It is promising, especially for very high spectral resolution data cubes.