Rule of thumb for selecting ROIs for supervised classification

Question

I'd like to perform a supervised classification using the Semi-Automatic Classification Plugin and a maximum likelihood classification algorithm. I read about a rule of thumb which says "if training data are being extracted from n bands, then a minimum of >10n pixels of training data are collected for each class" (Jensen). I'm using a Landsat image with six bands, so I assume that I have to select at least 60 training pixel per class. Somehow, I'm confused because I also heard people saying that you need 10times training pixels as you have classes.

Is there a general rule of thumb for selecting ROIs? Do you count the training pixels by drawing the polygons around them?

Answer 1

This actually depends on the classifier you are using. Different assumptions about data distribution or different optimization strategies may require more or less training data. It is very true that for growing dimensionality of the input space, you require more training samples to fill the space ("the curse of dimensionality"). This is particularly true for distance based functions, since Euclidean distance loses its meaning in undersampled high-dimensional spaces. Note that also Gaussian distribution rely on Euclidean / Mahalanobis distances, so they won't perform well in these situations. I would say that, as a rule of thumb, the more you have the better it is. Which kind of classifier are you planning to use?

However, instead of focusing everything on the number of examples, you should also focus on the amount of variability you cover for your semantic class with your training data. For instance, if you want to classify a forest class which has a lot of spatial variability of the signal, you may want to select ROIs which actually represent these changes well, to improve and guarantee some sort of generalization. For a class "water", the signal variability in space is usually low, and a classifier may generalize well even with few training samples.