222.

[Cover] [Contents] [Index]

Page 298

In this multisource classification exercise, which includes both contextual prior probabilities and Markov random fields, the GA is used for two purposes: first, finding optimal values for the weighting parameters and, secondly, for estimating the potential parameters. Only pair-wise potential parameters β are used. For simplicity, we have also used the isotropic assumption (i.e., rotation invariant, single β is used), and both the fitness function and classification algorithm are coded using the iterated conditional mode (ICM) algorithm that is described in Chapter 6. Each string in the G A contains 28 bits. The first 21 bits are used to determine the weighting parameters, while the last 7 bits are used for searching for an optimum value of the potential parameter β_j. The range of β is defined as [0, 3]. After 4000 iterations, a classification accuracy of 79% was achieved (Table 7.6c). This represents an improvement of close to 20% improvement using the TM images alone, and an improvement of more than 10% in comparison with the classification without contextual information. The corresponding weights are listed in Table 7.5, and the classified image is shown in Plate 5d. It is apparent that the patterns shown on the classified image are more patch-like, which gives a more realistic result.

[Cover] [Contents] [Index]