Joint Classification and Compression of Hyperspectral Images

In this paper, the problem of compressing hyperspectral images using a classification point of view is examined. The goal is to compress with loss an image in order to obtain interesting compression ratio with the constraint to be near lossless for a classification algorithm. Our proposed method is based on a spectral vector quantization performed by the Kohonen’s Self Organizing Map and an entropy coding. This method gives high compression ratio (up to 100 : 1) and appears to have the same strategy of a Spectral Angle Mapper algorithm. Thus, it is possible, on the first hand, to make classifications into the compressed domain, or on the other hand to classify the dictionary of vector quantization to have its semantic meaning. This algorithm was applied to CASI images with 48 spectral bands acquired over Saint-Michel in France for green seaweed proliferation monitoring. It proved to be very efficient for compressing images while still remaining of excellent quality for monitoring usage.