Training site refinement

Before proceeding with classification of the imagery, the training sites were analysed to ensure that they were not composed of any outlying pixels, whose digital numbers were uncharacteristic of that class. This was achieved by calculating statistics for each training regions and examining the maximum, minimum and standard deviations (see Table A4 of Annex A for details). If outlying pixels were identified, the boundaries were adjusted or the site dropped from the training set.

An evaluation of the data frequency histograms for each training site was conducted to ensure that the training data were normally distributed. Normal distribution of training data was an assumption for the classifier to be used later (Section 0). None of the training sets had bi- or tri-modal distributions.

To determine whether the classes for which training sites were identified were in fact spectrally separable, a supervised classification was performed and the percentage of correctly classified pixels within each training site was assessed. To be acceptable for classification, a level of 90% correctly classified training pixels in each class was set. Training sites not meeting these criteria were examined as described above.

Supervised classification and refinement of classification results
In addition to the six multi-spectral channels of data included in the classification (bands 1-5 and 7), two additional principal component bands were generated from the merged data. Principal component analysis is a technique employed to reduce the correlation between bands of data and enhance features that are unique to each band (Hall 1998). A characteristic of principal component analysis is that information common to all input bands (high correlation between bands) is mapped to the first principal component, whilst subsequent principle components account for progressively less of the total scene variance (Hall 1998). An increase in classification accuracy of vegetation communities using principal component analysis has been reported (Green et al 1998).

The maximum likelihood classifier is used in the classification. This classifier accounts for the mean and covariance of each class by estimating the likelihood of a class at any digital value. This is the most rigorous algorithm provided that certain requirements are adhered to (Huang and Mausal 1994). This classifier assumes that the training data for each class in each band are normally distributed. This assumption was checked in the training refinement stage (Section 0).

A majority filter was included as part of the classification to remove isolated pixels from the classified output. The filter operates by assigning a pixel the same value as the most frequent value neighbouring pixels within a certain distance of the central pixel. In this case a filter distance or window of 3 x 3 pixels was used. Filtering tends to increase classification accuracy as many of the pixels removed are misclassifications caused by pixels containing a mix of features (Lillesand and Kiefer 1994).