A Neural Approach to Compression of Hyperspectral Remote Sensing Imagery

This paper presents an original research for hyperspectral satellite image compression using a fully neural system with the following processing stages: (1) a Hebbian network performing the principal component selection; (2) a system of "k" circular self-organizing maps for vector quantization of the previously extracted components. The software implementation of the above system has been trained and tested for a hyperspectral image segment of type AVIRIS with 16 bits/pixel/band (b/p/b). One obtains the peak-signal-toquantization noise ratio of about 50 dB, for a bit rate of 0.07 b/p/b (a compression ratio of 228:1). We also extend the previous model for removal of the spectral redundancy (between the R, G, B channels) of color images as a particular case of multispectral image compression; we consider both the case of color still images and that of color image sequences.