19.8 Conclusions

Hyperspectral data compression has been considered as a crucial step in preprocessing of hyperspectral data. Instead of focusing on design and development of 3D compression algorithms as most of current efforts are devoted to hyperspectral data compression, this chapter takes a rather different approach by addressing and investigating two important and crucial issues arising in hyperspectral data compression, subpixels and mixed pixels analysis. In particular, it shows and demonstrates via experiments several important overlooked issues that have been proven crucial in hyperspectral data compression and need to be addressed. For a hyperspectral data compression to be effective, hyperspectral data compression must be conducted on an exploitation basis and a blind use of data compression technique generally results in inappropriate interpretation. A direct application of 3D lossy compression techniques to hyperspectral imagery may cause significant loss of crucial information provided by subpixels and/or mixed pixels. Secondly, SNR and MSE have been shown inappropriate to be used as compression criteria for subpixel and mixed pixel analysis when the compression ratio is high. In other words, higher SNR or lower MSE does not guarantee better compression performance in terms of information extraction and vice versa when compression rate at low bits. Thirdly, to address the issues of subpixels and mixed pixels, spectral/spatial compression techniques are shown to be always better and more effective than 3D compression techniques. Fourthly, in order to spectral/spatial compression techniques, a newly developed concept of VD is proposed for hyperspectral image compression to estimate number of principal components needed to be retained for dimensionality reduction. Over the past years, this number has been assumed a priori on a trial and error basis. Using spectral/spatial compression in conjunction with VD for hyperspectral data compression is a new approach. Lastly, despite that we did not include spectral/2D compression techniques in this chapter, the results in Ramakishna (2004), Ramakishna et al. (2005), and Ramakishna et al. (2006) have shown that the spectral/3D compression always performed better than the spectral/2D spatial compression techniques where the latter have been extensively used in spectral/spatial compression in the literature, while the former has not received attention in the past. The reason that we believe is that many researchers may have thought that since spectral compression has done its task to de-correlate spectral information among spectral bands, there is no need of using 3D compression and instead, 2D spatial compression may be sufficient. Unfortunately, this generally is not true as demonstrated by experiments in this chapter. As noted, since the goal of this chapter is not to develop new compression algorithms, two well-known 3D compression techniques, 3D SPIHT and JPEG2000 Multicomponent, are used for benchmark compression.