Biased Run-Length Coding of Bilevel Classification Label Maps of Hyperspectral Images

Abstract

For efficient coding of bilevel sources with some dominant symbols often found in classification label maps of hyperspectral images, we proposed a novel biased run-length (BRL) coding method, which codes the most probable symbols separately from other symbols. To determine the conditions in which the BRL coding method would be effective, we conducted an analysis of the method using statistical models. We first analyzed the effect of 2-D blocking of pixels, which were assumed to have generalized Gaussian distributions. The analysis showed that the resulting symbol blocks tended to have lower entropies than the original source without symbol blocking. We then analyzed the BRL coding method applied on the sequence of block symbols characterized by a first-order Markov model. Information-theoretic analysis showed that the BRL coding method tended to generate codewords that have lower entropies than the conventional run-length coding method. Furthermore, numerical simulations on lossless compression of actual data showed improvement of the state of the art. Specifically, end-to-end implementation integrating symbol blocking, BRL, and Huffman coding achieved up to 4.3% higher compression than the JBIG2 standard method and up to 3.2% higher compression than the conventional run-length coding method on classification label maps of the widely used “Indian Pines” dataset.