| dc.description.abstract | The innumerable capabilities of SAR image have contributed to its several applications, topmost
of which is land cover classification. The presence of speckle in SAR images makes the automated
digital image classification a challenging task. Reduction of speckles noise is therefore a very
important component in SAR image processing. This research therefore aimed at investigating the
effect of Gamma Map, Median, Mean and Lee de-speckling filters on the classification accuracy
of land cover maps. Secondary data source i.e., dual polarized (VV+VH), high resolution, ground
range detected (GRD) Sentinel 1A image of Kampala district was acquired and used. The image
was filtered using the aforementioned de-speckling filters while varying the filter window size and
later subjected to unsupervised classification using K-Mean clustering algorithm. The results were
evaluated using quantitative measures, which included Equivalent Number of Look (ENL),
Speckle suppression index (SSI), Speckle suppression and mean preservation index (SMPI), Mean
square error (MSE), Producer classification accuracy, User classification accuracy, Overall
classification accuracy and Kappa coefficient.
It was observed that the aforementioned quantitative evaluation measures cannot achieve the best
performance in speckle suppression and feature preservation simultaneously. De-speckle filters
with good noise removal capabilities often tended to degrade the spatial and radiometric resolution
of the original image thus leads to poor classification accuracy. The classification accuracy of
landcover maps produced by SAR imagery is dependent on the quality of the image which can be
attributed to the different de-speckling techniques and filter window sizes used. The results of the
research showed that Gamma Map Lee and Mean filter provide a good balance in feature
preservation as well as speckle reduction compared to Median filter which has a significant
performance in speckle reduction, but at the cost of feature loss. An overall classification accuracy
of 89.6 %, with 0.87 kappa coefficient value was archived using Gamma Map filter. It was also
observed that majority of the de-speckling filters (Gamma Map, Median and Lee filter) attained
best overall classification accuracies and Kappa Coefficient values with 5x5 kernel size, except
for Mean filter. Thus, concludes 5x5 filter window size and Gamma Map filter to be the best for
land cover classification due to the ability to remove noise and preserve the details of the image. | en_US |
