194.

[Cover] [Contents] [Index]

Page 272

to source weighting parameters. If these weighting parameters are not chosen properly, then the multisource consensus will give disappointing results although it is based on a theoretically robust mechanism.

In this chapter, the principal approaches for dealing with multisource classification are addressed. Several classification methods, based on stacked-layer, incorporation of topographic data, the extension of Bayesian theory (in which a further derivation to incorporate Markov random field (MRF) concept to perform multisource MRF-MAP is also illustrated), and evidential reasoning, respectively, are discussed. We also introduce possible options for source weighting factor assignment, and experimental results are presented in the final section.

7.1 Stacked-vector method

The most straightforward approach to deal with a multisource classification problem is simply to extend the dimension of the data vectors to include each source. This approach is known as the stacked-vector or augmented-vector method. For example, if one has a six-band TM image and a three-band SPOT image, then nine bands can be used together as inputs to the classifier. Although this method is easy to apply, several issues require attention.

The first issue concerns the scale of measurement of each source. Since different data sources are likely to have different measurement scales, it is generally recommended that all data should be mapped into the same scale. Such a process is called normalisation (Figure 7.1). For instance, if data

Figure 7.1 The stacked-vector classification process.

[Cover] [Contents] [Index]