3. What is Classification?

Classification is the process of allocating objects to a class within a discrete set of classes on the basis of data about that object and in accordance with a set of rules that create decision surfaces between the classes. It is usual that the classes are physically significant. Thus colour may be used to discriminate between oranges, lemons and apples on a process line, or waveband data may be used to create a land cover map from satellite image data.

Typical scattergrams for 4 band satellite image data.

Classification involves defining classes, usually by either defining training areas for each class, or automatically from the data itself. Then it involves setting the rules that define the decision surfaces that enables the classifier to decide on what data values belong to each class and finally it involves the actual conduct of the classification.

The process of classification

Classes are defined either by identifying typical areas for each class and then using training areas to get class statistics, or by using a clustering algorithm to identify clumps or clusters in the data, and then deriving statistics for each of these clusters. The first method is called supervised classification and the second is called unsupervised classification.

1 - Define the classes by deriving class statistics, 2 - Select the Probability Density Function (PDF) to be used, so that the probabilities can be used to define the class decision surfaces, 3 - Use training data to set the PDF parameters and in this way define the decision surfaces and thus the class domains.

There are many ways to do this; we will focus here on parametric methods that assume that the data for a class obeys a standard model; usually the Normal Distribution.

The physical meaning of the land cover classes depicted in one of the scattergrams depicted on this page. The location of the land covers in each of the scattergrams depends on the typical reflectance of the cover type in the wavebands depicted in the scattergram.