307.

[Cover] [Contents] [Index]

Page 75

to be investigated. For a more detailed discussion, the reader is referred to Krishnaparam et al. (1991, 1993, 1995).

2.3.4 Statistical supervised methods of pattern recognition

The supervised approach to pixel labelling requires the user to select representative training data for each of a predefined number of classes. Classification performance is highly dependent on how well the user is able to model the target class distribution. The user’s experience can be very helpful in identifying and locating training areas. Ideally, the training areas should be sites where homogeneous examples of known cover types are found (Townshend, 1981). A supervised statistical classification can be carried out by the following three steps:

1 Define the number and nature of the information classes, and collect sufficient and representative training data for each class

2 Estimate the required statistical parameters from the training data, and

3 Use an appropriate decision rule.

Although the selection of training data may be tedious, a supervised approach is preferred by most researchers because it generally gives more accurate class definitions and higher accuracy than do unsupervised approaches. Three statistical classifiers are in general use. These are the parallelepiped method, minimum distance classifier and the maximum likelihood algorithm.

2.3.4.1 Parallelepiped method

The parallelepiped method is implemented by defining a parallelepiped-like subspace (i.e. a hyper-rectangle) for each class. The boundaries of the parallelepiped, for each feature, can be defined by the minimum and maximum pixel values in the given class, or, alternatively, by a certain number of standard deviations on either side of the mean of the training data for the given class. The decision rule is simply to check whether the point representing a pixel in feature space lies inside any of the parallelepipeds. An example illustrating the specification of the topology of a parallelepiped classifier in the case of a two-dimensional feature space is shown in Figure 2.8.

The parallelepiped method is quick and easy to implement, but errors may arise, particularly when a pixel lies inside more than one parallelepiped or outside all parallelepipeds. These two situations are, in fact, likely to occur, because in the feature space the distribution of pattern vectors is often quite complex. Therefore, it is hard to provide a robust classification performance using this simple method.

[Cover] [Contents] [Index]