2012/09/28 by Elkebir Sarhrouni, Sarhrouni, Elkebir, Ahmed Hammouch +3
Engineering · #Remote-Sensing Image Classification
paper · pdf · doi:10.48550/arxiv.1210.0528
Hyperspectral image is a substitution of more than a hundred images, called\nbands, of the same region. They are taken at juxtaposed frequencies. The\nreference image of the region is called Ground Truth map (GT). the problematic\nis how to find the good bands to classify the pixels of regions; because the\nbands can be not only redundant, but a source of confusion, and decreasing so\nthe accuracy of classification. Some methods use Mutual Information (MI) and\nthreshold, to select relevant bands. Recently there's an algorithm selection\nbased on mutual information, using bandwidth rejection and a threshold to\ncontrol and eliminate redundancy. The band top ranking the MI is selected, and\nif its neighbors have sensibly the same MI with the GT, they will be considered\nredundant and so discarded. This is the most inconvenient of this method,\nbecause this avoids the advantage of hyperspectral images: some precious\ninformation can be discarded. In this paper we'll make difference between\nuseful and useless redundancy. A band contains useful redundancy if it\ncontributes to decreasing error probability. According to this scheme, we\nintroduce new algorithm using also mutual information, but it retains only the\nbands minimizing the error probability of classification. To control\nredundancy, we introduce a complementary threshold. So the good band candidate\nmust contribute to decrease the last error probability augmented by the\nthreshold. This process is a wrapper strategy; it gets high performance of\nclassification accuracy but it is expensive than filter strategy.\n