201.

[Cover] [Contents] [Index]

Page 279

By dividing Equation (7.8) by Equation (7.5), the probability ratio is seen to be:

(7.9)

This is the contribution of the (n+1)^th source, and therefore Equation (7.5) can be modified as follows:

(7.10)

For reliability measurements, we can again use the idea of appending exponents to each data source, and Equation (7.6) can be rewritten as:

(7.11)

Clearly, the above expression provides a more satisfactory relationship for dealing with multisource fusion problems in comparison with Equation (7.6). The reason is that when a source is found to be fully unreliable, i.e. α_i=0, both posterior and associated prior probabilities should be rejected simultaneously as shown in Equation (7.11). However, it is worthwhile to note that if prior probability is not used (which is equivalent to regarding P(ω_j) as being uniformly distributed for all information class j), then Equations (7.6) and (7.11) are equivalent.

The sensitivity measure as given by Equation (7.7) is converted to the form:

(7.12)

Another way to understand the behaviour of α_i more clearly is to represent Equation (7.11) in logarithmic form (Benediktsson et al. 1990)

(7.13)

It is now apparent that α_i is a coefficient that corresponds to the weighting parameter associated with data source i.

7.3.4 Multisource classification using Markov random field

In most remote sensing image classification experiments, the prior probability P(w) is rarely used because of the problem of properly modelling

[Cover] [Contents] [Index]