Features on the seabed can be mapped from remote sensing multi/hyperspectral imagery, provided that their effects on the measured reflectance spectrum can be made independent of those produced by the atmosphere and water column. The non-linear effect of water column light attenuation can then be corrected to obtain the absolute reflectance of the seabed. Light attenuation by the water column and bathymetry are both determined from the satellite image. The water column attenuation is then removed in order to apply an automated supervised classification, whatever the depth is. We have compared the results obtained with and without the correction of water column attenuation, for two different statistical measures: Euclidean (ED) and spectral angle mapper (SAM) distances. We have applied this methodology to MERIS images acquired on the lagoon of New Caledonia. The best overall accuracy (79%), as compared to in situ data, is obtained with the corrected image and the SAM distance.