Estimation of the costs of bad classification Tl'/l.

Estimation of the Costs of Bad Classification T_l'/l

The aim is to find the good class for a pixel. Of course, if the class of assignment is the true class of membership of the pixel, there is no cost: T_l_/l=0.

On the other hand, each time there is an error, it is necessary to penalize the bad choice. As there is no reason to penalize a class l' ( $l' \neq l$ ) rather than a class l'' ( $l'' \neq l$ and $l'' \neq l'$ ) (l being the true class), the cost of one error will be the same for all the errors:

T_l'/l= 1

IRIT-UPS