Fuzzy Land Cover Change Detection and Validation: A Comparison of Fuzzy and Boolean analyses in Tripoli City, Libya

This research extends fuzzy methods to consider the fuzzy validation of fuzzy land cover data at the sub-pixel level. The study analyses the relationships between fuzzy memberships generated by field survey and those generated from the classification of remotely sensed data. In so doing it examines the variations in the relationship between observed and predicted fuzzy land cover classes. This research applies three land cover classification techniques: Fuzzy sets, Fuzzy c-means and Boolean classification, and develops three models to determine fuzzy land cover change. The first model is dependent on fuzzy object change. The second model depends on the sub-pixel change through a fuzzy change matrix, for both fuzzy sets and fuzzy c-means, to compute the fuzzy change, fuzzy loss and fuzzy gain. The third model is a Boolean change model which evaluates change on a pixel-by-pixel basis. The results show that using a fuzzy change analysis presents a subtle way of mapping a heterogeneous area with common mixed pixels. Furthermore, the results show that the fuzzy change matrix gives more detail and information about land cover change and is more appropriate than fuzzy object change because it deals with sub-pixel change. Finally the research has found that a fuzzy error matrix is more suitable than an error matrix for soft classification validation because it can compare the membership from the field with the classified image. From this research there arise some important points: • Fuzzy methodologies have the ability to define the uncertainties associated with describing the phenomenon itself and the ability to take into consideration the effect of mixed pixels. • This research compared fuzzy sets and fuzzy c-means, and found the fuzzy set is more suit-able than fuzzy c-means, because the latter suffers from some disadvantages, chiefly that the sum of membership values of a data point in all the clusters must be one, so the algorithm has difficulty in handling outlying points. • This research validates fuzzy classifications by determining the fuzzy memberships in the field and comparing them with the memberships derived from the classified image.