An investigation into the effects of speckle filters on classification

Synthetic aperture radar (SAR) imagery, which suffers from speckle noise, is regularly used in the process of extracting useful information, like geophysical parameters, from remotely sensed imagery. The methodology often requires a speckle filter prior to implementation of some classification scheme. This paper reports on the results of an investigation into the effects of speckle filters on classification accuracy and reliability. Prior to classification, the data set, consisting of four SIR-C bands, is filtered using one of many possible speckle filters. Next, an unsupervised clustering algorithm is used to obtain `training´ data for a supervised classification of the data set. The result of the classification is then assessed by calculating the Jeffries-Matusita (JM) distance and the Mahalanobis (D) distance. The authors use the normalized average JM distance as an indicator of overall class separability. The D distance, after thresholding and manipulation, is an indicator of overall classification reliability. The product of the `separability´ and `reliability´ is used to rank the results