195.

[Cover] [Contents] [Index]

Page 273

from the optical spectrum are used, one may convert them into ground reflectance units. If radar images are used, one may convert these data into backscatter coefficients (σ⁰, Chapter 1). However, where a variety of data sources is used then choice of a data normalisation scheme may be logically difficult, and the classification process may be dominated by data sources that display a larger scale of variation. For instance, if the first data source in a two-source case has a range of [0, 32], and the second has a range of [0, 16384] then, the second data source is more likely to dominate the classification process due to the effects of measurement scale. Some classification procedures require explicit normalisation of the input data (for example, a feed-forward neural network). Other methods, such as maximum likelihood (ML), use a hidden form of normalisation. ML uses estimates of the variance-covariance matrices (S_i) of the classes in order to generate class membership probabilities. The calculation of S_i involves the subtraction of the class mean of each feature from the feature measurements.

A second issue is that the stacked-vector method may not be practical in terms of computational cost when the number of data vectors is large. For example, if one uses the statistical Gaussian maximum likelihood method for classifying N dimensional variables, the resulting computational cost is proportional to N². A way to reduce this problem is to use a feature selection procedure, which involves selecting a subset of data sources. A suitable subset can be selected by using some distance measure such as the divergence index (Singh, 1984) or the B-distance (Haralick and Fu, 1983) as shown in Chapter 2. Another strategy for feature selection is through the use of data transforms, such as principal component analysis (PCA), Tasselled Cap (Crist and Cicone, 1984a), vegetation indices, or the self-organised feature map (SOM) (Chapters 2 and 3). Note that, although feature selection can decrease the computational cost of classification, some valuable information may be contained in the discarded data sources. Such a dilemma always results when feature selection techniques are applied.

A third issue is concerned with the reliability (or uncertainty) of the data. The stacked-vector approach treats each data source as being fully reliable, i.e. each source contributes equally to the classifier in the determination of the location of decision boundaries in feature space. This may not always be the case in practice, and could be a drawback to the achievement of higher classification accuracy.

7.2 Incorporating topographic data

Some studies have directed their interest towards the use of topographic data in the classification process. For example, Richards et al. (1982) combine Landsat images with topographic information and spatial context to iteratively adjust the label assigned to each pixel until between-source class

[Cover] [Contents] [Index]