221.

[Cover] [Contents] [Index]

Page 297

Figure 7.5 Evidence consensus sequence and their corresponding weighting parameters.

for SIR-C L band HH and HV images is only 36.75%. Clearly, the classification results based on each data source separately are unsatisfactory.

These class-conditional probabilities are first subjected to the extension of Bayesian theory for multisource consensus experiment. Prior probabilities are not considered at this stage. The genetic algorithm is used for detecting the optimal weighting parameters for each data source. Each string in the GA is 21 bits in length, i.e. 7 bits for each weighting parameter, which is equivalent to dividing the weighting value within the range [0, 1] into 2⁷=128 choices. The fitness function is based on the average producer’s accuracy. The initial population is set as 100 with a crossover rate of 0.6 and a mutation rate of 0.001. After 3,000 search iterations, the average producer’s accuracy is improved to 67.85% (see the confusion matrix in Table 7.6a), which is around 10% improvement. The corresponding weighting parameters are shown in Table 7.4, and the classification patterns are shown in Plate 5b.

In approaches using evidential reasoning, GA is applied in a similar way for detecting optimal weighting parameters. The probabilities derived from data sources (1) and (3) are first chosen to perform an orthogonal sum. After 2,000 search iterations, the average producer’s accuracy is only increased to 58.1%. The resulting evidence is then fused with the probabilities derived from data source (2). After 2,000 iterations, the accuracy is enhanced to 65.63% (Table 7.6b) close to the accuracy level that the extension of Bayesian theory achieved. The corresponding weights and classification pattern are shown in Figure 7.5 and Plate 5c, respectively.

[Cover] [Contents] [Index]