In remote sensing, when producing land use maps from satellite imagery products using supervised classification, specific procedures are used. Among the popular choices, Maximum Likelihood (ML) seems to be quite popular, or favoured. Mahalanobis Distance (MD) in its definition is quite close to ML, and is often used or mentionned.
I came across some referenced comparing both, but it appears that, ultimately, ML performs best than MD.
My point is, are there some cases where MD would perform quite better than ML ? If yes, what could be the possible reasons ? Would it be related to the spectral signature of training samples areas ?
I am actually working on a specific case where it looks like MD is better than ML, though I cannot really pinpoint why.
Thanks for reading.