214.

[Cover] [Contents] [Index]

Page 290

reduced or enhanced in proportion to the weight. It can be seen that the results of the classification process depend both on the information provided by the source and on the source weighting factor.

For example, if two sources, denoted by A and B, respectively, are used for labelling a pixel of interest, source A may suggest that the pixel belongs to class a with probability 0.8 and to class b with probability 0.2. However, source B indicates that the pixel belongs to class b with probability 1.0 and assigns zero probability to the pixel belonging to class a. If both sources have been determined as fully reliable, i.e. their associated weights are unity, the result based on both statistical Bayesian or evidential reasoning approaches for source combination will label the pixel as class b. If we have determined that the weight of source B is 0.5, and that of source A is unity, then the pixel will be labelled as class a. Several possible methods for measuring the source weighting parameters are described below.

7.5.1 Using classification accuracy

The derivation of data source weighting parameters from classification accuracy measurements is an instinctive approach. A data source should be assigned a higher weight if the resulting classification accuracy is high. However, if the classification derived from the data source is unsatisfactory, the data are considered to be relatively unreliable, and should be assigned a lower weight. For example, if the classification accuracy using data source A is 80%, and using data source B is only 50%, then the probability measure or evidence derived from data source A should be assigned a higher weight. This method is very easy to apply, but one should note that the 80 and 50% classification accuracies do not necessarily mean that the source weighting parameters are equal to 0.8 and 0.5, respectively, since different methods can be used for analysing classification accuracy (Chapter 2), and these accuracy analysis tools are likely to provide different measures. The reliability measure in terms of classification accuracy can thus only give us a rough guide.

7.5.2 Use of class separability

The second approach to the measurement of source weighting uses the concept of the separability of the information classes. One can assign a higher weight to a source that contributes more to the separability of the information classes, while sources that contribute little to interclass separability can be assigned low weights. The estimates of the weighting parameter values are therefore based on the contribution to the statistical separability of each data source. Several approaches to the estimation of separability, using distance, divergence, or separability functions, are proposed by Fukunaga (1972), Whitsitt and Landgrebe (1977), Swain and

[Cover] [Contents] [Index]