219.

[Cover] [Contents] [Index]

Page 295

Table 7.4 Optimal consensus weights detected by GA for the extension of Bayesian theory

Source	TM	SIR-C C band HH, HV	SIR-C L band HH, HV
Weights	0.882	0.693	0.142

important result pointed out by Holland is that if P_i denotes the number of individuals in generation i, the number of schema being tested will be P_i³. This result explains why a substantial search improvement can be achieved using a relative small population size, and the concept is therefore termed implicit parallelism.

During the search process, the progress of GA survivals is determined by a user-defined function called the fitness function. The fitness function can be regarded as controlling GA by providing the ability to evaluate and hence select higher quality individuals. Where a GA is used to determine data source weighting parameters, and if the aim of classification is to improve the overall classification accuracy, one can then simply use the overall classification accuracy as the fitness function.

7.6 Experimental results

Three classification approaches (the extension of Bayesian theory, Markov random fields and evidential reasoning) are used in a classification experiment with the aim of evaluating their relative classification performance. Three data sources are used: (1) Landsat TM bands 1–5 and 7; (2) Shuttle Imaging Radar C (SIR-C) C-band HH and HV polarisation images; and (3) SIR-C L-band HH and HV polarisation images. Each image is 256× 256 pixels in size. Examples of these data sources are shown in Plate 4. The study area is located in the Red Hills, Sudan, and the aim is to classify surface lithology into eight types, listed in Table 7.3, which also gives the number of training pixels selected for each class. ‘Ground truth’ is available in the form of a paper map (Plate 5a).

For the six TM bands the pixel values are used as inputs, while for the SIR-C radar images, textural features are extracted, using the multiplicative

Table 7.5 Weights and potential parameter in MRF model for multi-source classification

Source	TM	SIRC C band HH, HV	SIRC L band HH, HV
Weights	0.937	0.661	0.039
β		1.37

[Cover] [Contents] [Index]