197.

[Cover] [Contents] [Index]

Page 275

They also compared these results with those achieved by a neural network stacked-vector approach. Both approaches reveal different advantages and disadvantages. More recently, Schistad et al. (1994) tested an extended Bayesian classification approach for classifying Landsat TM and SAR images, and their results were better than those achieved by conventional single-source classification. A similar experiment was conducted by Kim and Swain (1995), who used the evidential reasoning approach to manage multisource data and achieve an acceptably good classification result.

Both the extension of Bayesian classification theory and evidential reasoning methods regard each data source as fully independent. Hence, one has to generate probability (or evidence) measures for each information class using each source separately, and then obtain the classification result in terms of probability (or evidence) consensus. The general steps are shown in Figure 7.2 and are explained as follows.

7.3.1.1 Feature extraction

This is the preparation stage of the classification process, the aim of which is to determine what kind of input features (e.g. pixel grey values, texture measures) should be used in the classification process. The selection of suitable input features can enhance classification accuracy. Certainly, the choice of input features not only depends on the way in which the image was formed (e.g. optical or microwave sensor), but is also determined by the scale relationship between the ground objects of interest and image resolution. If, for example, image resolution is around 30 m then, in the case of lithological classification, textural information may contribute more significantly to interclass variation than tonal information. In the case of classifying crop types in an agricultural area, textural and tonal information may make no significant difference (Tso, 1997) because of the scale and the size of the object being classified (refer to Chapter 5 for a discussion of texture). Where radar images are used, textural features may be more useful than image tonal information (Ulaby et al., 1986b).

7.3.1.2 Probability or evidence generation

This stage involves the definition of a probability density function or some alternative methodology (e.g. by the use of a feed-forward neural network) to generate class-associated probabilities or evidence in which the user has high confidence. If a statistical model is used for generating the probability or statistical evidence then the choice of probability density function should be source-dependent (Kim and Swain, 1995). For example, optical images can be modelled in terms of a Gaussian p.d.f, while radar images may more suitably be modelled as a Gamma distribution (Chapter 1). This step is an

[Cover] [Contents] [Index]