Near-lossless compression of hyperspectral data through classified spectral prediction

This paper presents a novel scheme for lossless/near-lossless hyperspectral image compression, that exploits a classified spectral prediction. MMSE spectral predictors are calculated for small spatial blocks of each band and are classified (clustered) to yield a user-defined number of prototype predictors for each wavelength, capable of matching the spatial features of different classes of pixel spectra. Unlike most of the literature, the proposed method employs a purely spectral prediction, that is suitable for compressing the data in band-interleaved-by-line (BIL) format, as they are available at the output of the on-board spectrometer. In that case, the training phase, i.e., clustering of predictors for each wavelength, may be moved off-line. Thus, prediction will be slightly less fitting, but the overhead of predictors calculated on-line is saved. Although prediction is purely spectral, hence 1D, spatial correlation is removed by the training phase of predictors, aimed at finding statistically homogeneous spatial classes matching the set of prototype spectral predictors. Experimental results on AVIRIS data show improvements over the most advanced methods in the literature, with a computational complexity far lower than that of analogous methods by other authors.