Feature-based lossy compression of multispectral data

A new lossy compression algorithm is proposed which, unlike most conventional techniques, gives the user control over the distribution of the inherent compression-induced error over various components of the scene. This technique effectively allows features of interest in the image to be coded at a relatively higher precision level (bit rate) than the nonessential features. This approach minimizes the impact of lossy compression on the “final product” domain. The compression technique is based on a modified linear pixel unmixing procedure followed by standard JPEG algorithm. Each spectrum in the data set is modeled as a linear combination of a finite number of spectrally distinct signatures referred to as end-members. The end-members are selected via a modified ISODATA clustering algorithm which uses the spectral angle criterion, instead of the Euclidean distance criterion. The resulting end-members are more likely to correspond to identifiable physical features, i.e. species, on the ground. The concentration of each end-member in an image pixel is determined and mapped to 0-255 range. The ensemble of all pixel concentrations for each end-member forms a compositional map, known as abundance image. The resulting set of abundance images will exhibit a much smaller spectral correlation since each abundance image shows the concentration of only one species in the entire scene. Depending on the relative importance, each of the resulting abundance images are then coded via standard JPEG at preset quality levels. The preliminary results are very encouraging