196.

[Cover] [Contents] [Index]

Page 274

consistency is achieved. This method has a drawback because it needs considerable computational resources due to its iterative nature. Strahler et al. (1978, 1980) include topographic information via a stratification approach. The data are first divided into several subsets, within which variation is minimised, based on the values of the dividing variables. Each subdivision is then analysed separately. The a priori probabilities derived from elevation data are also incorporated.

Franklin et al. (1986), Jones et al. (1988), Peddle and Franklin (1991) and Solaiman et al. (1998) describe similar approaches using stratification procedures. Hutchinson (1982) and Hoffer et al. (1979) use a confusion-reduction concept. The classification is first derived from one or several sources. After analysing the classification results, other information sources such as topographic data are introduced in an attempt to reduce the remaining between-class ambiguity. Kim and Swain (1989) note that classification results based on a confusion-reduction method will be different if the ordering of sources is changed. This drawback can be avoided by using the evidential reasoning approach, as described in the following sections.

The methods described above rely on the user’s knowledge, and are strongly affected by user interaction during the classification process. Two alternative approaches, namely, evidential reasoning based on the Dempster-Shafer theory (Shafer, 1979, 1987; Shafer and Logan, 1987) and the extension of Bayesian classification theory, are discussed next. These methods are more general in application and less dependent on user interaction.

7.3 The extension of Bayesian classification theory

Interest in techniques of multisource classification based on the extension of Bayesian theory was triggered by the studies of Lee et al. (1987) and Benediktsson et al. (1990). The model proposed by Benediktsson et al. (1990) is a refinement of the method due to Lee et al. (1987), in which a more complete statistical mechanism was derived. These two models are introduced next, together with a derivation using the results obtained by Benediktsson et al. (1990) to incorporate the Markov random field (MRF) concept in order to achieve multisource MRF-MAP classification.

7.3.1 An overview

The extension of Bayesian classification theory was compared with the evidential reasoning approach based on the Dempster-Shafer theory by Lee et al. (1987). Both showed satisfactory results for performing multisource data classification, although the method based on the extension of Bayesian classification theory performed slightly better than the evidential reasoning method (Lee et al., 1987). Benediktsson et al. (1990) adopted extended Bayesian classification theory to classify multisource remotely sensed data.

[Cover] [Contents] [Index]