If your dataset is big enough (100-200+ pictures per category), classical data augmentation should be more than sufficient.
With smaller datasets, a GAN may prove useful. However, the output images must mimic the disease of the leaves accurately. This could prove to be tricky with high resolution images. Cheers, Raoul
GAN seems more artificial and classical data augmentation is natural, I am not sure how the machine takes this, but I think using GAN is something like we are depending on more artificial data to create an artificial machine.