Band Selection and Classification of Hyperspectral Images using Mutual Information: An algorithm based on minimizing the error probability using the inequality of Fano

摘要

Hyperspectral image is a substitution of more than a hundred images, calledbands, of the same region. They are taken at juxtaposed frequencies. Thereference image of the region is called Ground Truth map (GT). the problematicis how to find the good bands to classify the pixels of regions; because thebands can be not only redundant, but a source of confusion, and decreasing sothe accuracy of classification. Some methods use Mutual Information (MI) andthreshold, to select relevant bands. Recently there's an algorithm selectionbased on mutual information, using bandwidth rejection and a threshold tocontrol and eliminate redundancy. The band top ranking the MI is selected, andif its neighbors have sensibly the same MI with the GT, they will be consideredredundant and so discarded. This is the most inconvenient of this method,because this avoids the advantage of hyperspectral images: some preciousinformation can be discarded. In this paper we'll make difference betweenuseful and useless redundancy. A band contains useful redundancy if itcontributes to decreasing error probability. According to this scheme, weintroduce new algorithm using also mutual information, but it retains only thebands minimizing the error probability of classification. To controlredundancy, we introduce a complementary threshold. So the good band candidatemust contribute to decrease the last error probability augmented by thethreshold. This process is a wrapper strategy; it gets high performance ofclassification accuracy but it is expensive than filter strategy.