An unsupervised segmentation-based coder for multispectral images

To fully exploit the capabilities of satellite-borne multi/hyperspectral sensors, some form of image compression is required. The Gelli-Poggi coder[1], based on segmentation and class-based transform coding, has a very competitive performance, but requires some a-priori knowledge which is not available on-board. In this paper we propose a new version of the Gelli-Poggi coder which is fully unsupervised, and therefore suited for use on-board a satellite, and presents a better performance than the original. Numerical experiments on test multispectral images validate the proposed technique.