Classification processes use features, also known as predictor variables, for discriminating between classes. A feature is an individual measurable property or characteristic of a geographic phenomenon being observed. Features in Earth observation include the individual bands of images and further properties derived from the image data. For example, the single band of a panchromatic image represents a feature that allows distinguishing between pixels of darker and lighter reflectance. Multispectral images have more bands and thereby enable the differentiation between classes by more features. This means, if two classes are different from each other in several of their properties, it becomes easier to distinguish them. The set of features used in a particular classification comprise the feature space where each feature represents one space dimension.
With an increased number of (uncorrelated) features it becomes possible to increase the number of classes that can be separated. For example land cover classifications have a large number of classes. For identifying suitable bands for optical EO satellites, the spectral signatures of all the target classes have to be analysed to identify in which bands they are separable from other classes. Classes like soil, water, and vegetation have spectral signatures that differ in particular in the blue, green, red, and infrared bands of the electromagnetic spectrum. These bands are present in virtually all multispectral sensors used for land cover classification.
Geographic phenomena can be differentiated not only by their reflectance in different bands. Beyond multispectral features, the classification may include image derivatives like derived spectral indices, principal components, or filtered bands (convolution layers). Object-based image analysis also uses spatial features, i.e. distance and proximity features, planar geometric features and topological features.