Sixty–seventy training sites varying in size from 286 to 8,914 pixels (5–46%) were used to locate training pixels on the images. Except for the bare soil/ landfill category, training samples for each class were 5– 12 subclasses. The training samples were then evaluated by using class histogram plots. Training samples were refined, renamed, merged, and deleted after the evaluation of class histogram and statistical parameters. A supervised maximum likelihood classification (MLC) algorithm was subsequently applied to each image which has generally been proven to obtain the best results from remotely-sensed data if each class has a
Gaussian distribution (Bolstad and Lillesand 1991). Misclassification was observed in the classified land cover categories obtained from the MLC classification. For example, certain urban surfaces were misclassified as landfill sites due to their similar spectral characteristics. Likewise, misclassification was also found between the wetland/lowlands category and the cultivated land, water bodies, and lowland/wetland category. It may be noted that initially the wetland category was identified as a separate class but eventually it merged with the lowland class as it was not possible to separate from the lowland category because of their alike spectral properties. Post-classification refinement, therefore, was used to improve the accuracy as it is simple, efficient and easily executable method (Harris and Ventura 1995). It is important to note that misclassification was higher in the MSS image among the datasets. To surmount the difficulty of misclassification, a number of strategies were considered. For
example, thematic information (e.g. water bodies, vegetation, and bare soil) was first extracted from the V-S-W index (Yamagata et al. 1997).