Classified data often manifest a salt-and-pepper appearance due to the inherent spectral variability encountered by a classification when applied on a pixel-by pixel basis (Lillesand and Kiefer, 1994). It is often desirable to “smooth” the classified output to show only the dominant (presumably correct) classification.
One means of classification smoothing involves the application of a majority filter. In such operations a moving windows is pass through the classified pixel in the window is not the majority class, its identity is changed to the majority class. If there is no majority class in the window, the identity of the center pixel is not changed. As the windows progresses through the data set, the original class code are continually used, not the labels as modified from the previous window position. (Eastman, 1995)
Majority filters can also incorporate some from of class and/or spatial weighting function. Data may also be smoothed mote than once. Certain algorithms can preserve the boundaries between land cover regions and also involve a user-specified minimum area for any given land cover type that will be maintained in the smooth output (Lillesand and Kiefer, 1994).
Ground truth or field survey is done in order to observe and collect information about the actual condition on the ground at a test site and determine the relationship between remotely sensed data and the object to be observed. It is recommended to have a ground truth at the same time of data acquisition, or at least within the time that the environmental condition does not change.
Classification accuracy assessment is a general term for comparing the classification to geographical data that are assumed to be true to determine the accuracy of the classification process. Usually, the assumed true data are derived from ground truth. It is usually not practical to ground truth or otherwise test every pixel of a classified image. Therefore a set of reference pixels is usually used. Reference pixels are points on the classified image for which actual data are(will be) known. The reference pixel are randomly select.(Congalton , 1991)