I am currently creating land cover maps of a number of national parks across Asia from Landsat 8 imagery and also later from SPOT imagery. I plan to use supervised classifications in ERDAS IMAGINE and we hope to create accurate maps showing individual types of vegetation. I am wondering if it would be better to perform this classification on the original multispectral image or the pan-sharpened image. Obviously, it would be preferential to use the pan-sharpened image as the spatial resolution would be higher, but I'm afraid that there may be some loss of spectral information in the resampling, especially when computing NDVI values. I am not sure whether this loss of information is worth compromising for the increased spatial resolution. I am fairly new to using this technique so just hoped someone could give me some advice!
Best way is to try both and see the influence on the accuracies. If resolution is not critical, pan-sharpening is not a good choice because it alters/reduces the amount of spectral information available, and in the case where you have a time-series, also changes the spectral relation between dates, which tends to lead to lower accuracies.
That being said, if some of the classes are occuring in narrow strips or edges/borders between homogeneous areas, pan-sharpening could help obtain better accuracy on theses classes, from the improved resolution.
There is also a data volume consideration. Pan sharpening from 30 to 15 meters will quadruple the file size, and you may not be able to include as many temporal bands in the classifier, depending on the implementation / memory limitations.