Then a rulebased technique using thematic information (DEM, municipal map and water bodies, etc.) was employed to correct previously misclassified land cover categories in ERDAS’s spatial modeler. Although application of the rule-based technique greatly improved the MLC classification, a small amount of misclassification was still found between wetland and cultivated lands. This was mainly attributed to their geographical contiguity. GIS tools such as Area of Intere(AOI) were afterward applied using visual analysis, reference data, and local knowledge to split and
recode these covers into their original classes. It is necessary to mention that ground truth information was also of great value in the refinement process. Applying those techniques substantially improved the result of pre-classification by the supervised algorithm. A 3×3 majority filter finally applied to the classified land cover data to reduce the salt-andpepper effect (Lillesand and Kiefer 1999). To determine the changes in land use/cover at different years, a post classification comparison of change detection was used. Even though this technique presents few limitations (Singh 1989; Coppin et al. 2004), it is the most common approach (Jensen 1996; Mundia and Aniya 2006) to compare data from different sources and dates. The advantage of postclassification comparison is that it bypasses the difficulties associated with the analysis of images acquired at different times of the year and/or by
different sensors (Yuan et al. 2005; Coppin et al. 2004; Alphan 2003). Moreover, the post classification method also answers the amount, location, and nature of change (Howarth and Wickware 1981). A major pitfall, however, is that the accuracy of the change maps depends on the accuracy of individual classifications and subject to error propagation (Yuan et al. 2005; Zhang et al. 2002). A comparison between the classified maps was carried out subsequently on a pixel-by-pixel basis (Jensen and Ramsey 1987).