MSc Dissertation: Mark Gebhardt

Citation:

Gebhardt, Mark. Speckle Reduction in SAR Imagery. MSc Dissertation. Department of Electrical Engineering, University of Cape Town, 1995.

 

Abstract:

Synthetic Aperture Radar (SAR) is a popular tool for airborne and spaceborne remote sensing. Inherent to SAR imagery is a type of multiplicative noise known as speckle. There are a number of different approaches which may be taken in order to reduce the amount of speckle noise in SAR imagery. One of the approaches is termed post image formation processing and this is the main concern of this thesis.

Background theory relevant to the speckle reduction problem is presented. The physical processes which lead to the formation of speckle are investigated in order to understand the nature of speckle noise. Various statistical properties of speckle noise in different types of SAR images are presented. These include Probability Distribution Functions as well as means and standard deviations. Speckle is considered as a multiplicative noise and a general model is discussed. The last section of this chapter deals with the various approaches to speckle reduction.

Chapter three contains a review of the literature pertaining to speckle reduction. Multiple look methods are covered briefly and then the various classes of post image formation processing are reviewed. A number of non-adaptive, adaptive and segmentation-based techniques are reviewed. Other classes of techniques which are reviewed include Morphological filtering, Homomorphic processing and Transform domain methods. From this review, insights can be gained as to the advantages and disadvantages of various methods. A number of filtering algorithms which are either promising or are representative of a class of techniques, are chosen for implementation and analysis.

The chosen filters are implemented and a discussion of their algorithms is presented. The theory and operation of each of the filters is explained. The filters which are presented are the Mean, Median, Lorentzian, K Nearest Neighbour (KNN), Hirosawa, Maxium a Posteriori, Frost and Maximum Homogeneous Region filters. The filters all operate on the principle of a two dimensional window which is shifted across the image, one pixel at a time. The pixels covered by the window are used to determine the new value of the pixel at the centre of the window. For certain of the filters the local mean and standard deviation (the local statistics) are used to modify the filter response in the presence of edges or point targets. Detailed listings of the source code for all of the filters is given in the appendices.

The chosen filters are used to filter three test images, i.e. one and four look ESAR images and a simulated four look image. After filtering, both qualitative and quantitative assessments of filter performance are made. In order to measure the trade-off between geomettic and radiometric resolutions two quantities are calculated from the filtered and unfiltered images. These quantities are the Equivalent Number of Looks (an indication of radiometric resolution) and an edge measure, which represents the geometric resolution of the image. The local statistics filters (Frost and MAP) are found to produce the best geometric resolution, but only a slight reduction in the amount of speckle. Good edge preservation is also provided by the Median filter. The Mean filter is found to provide the best speckle reduction, but it causes degradation of the geometric resolution. Two filters which achieve a compromise between speckle reduction and edge preservation are the Hirosawa and KNN filters. The point is made that all filters are dependent on the selection of parameters. It is possible to change the performance of the filters by changing the number of iterations, the window size, or other parameters. The results presented in this section are therefore not absolute and merely serve to provide information on the typical performance of different filters.

The choice of filtering algorithm and its parameters is seen to be closely related to the purpose for which the final image will be used. Filters should be chosen according to whether large or small scale features are of interest. The work presented in this thesis provides valuable insights into the potential of post image formation speckle reduction methods. These algorithms can be used in addition to, or in lieu of, multiple look methods in order to reduce the speckle in SAR images.

Further research into post image formation techniques, as well as multiple frequency and multiple polarization methods, is suggested. This further comparison would provide valuable information about the potential for further reducing speckle in SAR imagery.

 

 

Leave a Reply