188.

[Cover] [Contents] [Index]

Page 267

tational cost of MPM is larger than that of ICM, but is much less than that required by SA. The algorithm for MPM estimation is given in Figure 6.21.

6.6 Experimental results

In this section, the results of classification experiments by using the algorithms introduced in Section 4.5.3 are described. The test data set is the same as that used in Chapter 3 and described in Section 3.6.2.

The results of the classification experiments are shown in Plate 3 and corresponding confusion matrices are shown in Table 6.2. For comparison, the result produced by the maximum likelihood method is shown in Plate 3a. The ML method achieves an overall classification accuracy of 55.93% (kappa=0.460). It is clear that ML does not provide a very clean result. The addition of contextual information results in images that are much more patch-like. Plate 3b is the classification result output by the ICM algorithm, which achieved an overall classification accuracy of 63.78% (kappa 0.561). In comparison with the result obtained by the maximum likelihood algorithm, the ICM result shows an improvement of around 8% in terms of overall classification accuracy. The result generated by the SA algorithm is shown in Plate 3c. The initial value of parameter T was set to 3, and N_nner=100 (see Figure 6.19 for a description of the algorithm). The SA algorithm achieved an overall accuracy of 68.21% (kappa=0.610). Compared to the maximum likelihood and ICM results, SA shows an improvement of around 13% and 5%, respectively. Plate 3d shows the results of the MPM algorithm, with T=1, k=50 and n=200 (as described

Figure 6.21 Maximiser of the posterior marginals classification algorithm.

[Cover] [Contents] [Index]