Vol.
ger
rnegie
cht,
Yunham Dong
A SAR SPECKLE FILTERING ALGORITHM TOWARDS EDGE SHARPENING
Yunhan Dong*, Anthony K Milne**, and Bruce C Forster*
*School of Geomatic Engineering, **Office of Postgraduate Studies
The University of New South Wales
Sydney 2052, Australia
y.dong @unsw.edu.au, t.milne @unsw.edu.au, b.forster @unsw.edu.au
Working Group VII/6
KEY WORDS: Radar, Speckle Reduction, Texture, Image Processing.
ABSTRACT
One of the major difficulties when classifying synthetic aperture radar (SAR) images is the existence of speckle due to
coherent processing. Existing speckle filtering algorithms can effectively reduce the speckle effect but unfortunately
also smear edges and blur images. It is realized that fluctuations in an image can be due to either local oscillations or
edge crossings. An appropriate filtering algorithm should react differently to these two types of fluctuations: to smooth
local oscillations to reduce the speckle level and to enhance edge crossings to avoid blurring. Such a goal is achieved
via two steps. First edge crossings are detected using the second order derivative of the Gaussian function with a proper
dilation factor as the wavelet transform function; and then the traditional mean filtering algorithm using a moving
window is applied only within the region in which no edge crossings exist. As a result uniform areas are smoothed and
edges are effectively sharpened and enhanced. The algorithm can be used in conjunction with other popular SAR
speckle filters in order to sharpen smeared edges.
1 INTRODUCTION
Speckle reduction is becoming a routine process in synthetic aperture radar (SAR) image applications for terrain
classifications. It is well known that speckle is multiplicative noise. In consequence, a number of filtering algorithms
dealing with multiplicative noise have been proposed and widely used in the SAR image processing. The most Notable
ones include Lee's filter (1986), Frost's filter (Frost et al, 1982), Kuan's filter (Kuan et al, 1987), the polarimetric
whitening filter (Novak and Burl, 1990). Filters designed for polarimetric SAR data have also proposed, such as the
minimum mean square error (MMSE) filter (Goze and Lopes, 1993) for one-look data, and filters given by Lopes and
Sery (1997), Lee and Grunes (1999), for multi-look data. In addition, filters such as median filter (Tukey, 1977),
geometric (Crimmins, 1986), morphological (Dougherty, 1992), wavelet transform (Dong et al, 1998) filters have also
found to be useful in suppressing speckle level. Comparisons and evaluations of these various filters have been reported
elsewhere (Lee et al, 1994, Dong et al, 1998).
Quantitative evaluation of a filter includes several criteria such as equivalent number looks, mean bias, edge
preservation, and texture preservation. Among these the most important are:
1. Preservation of the mean;
2. Reduction of the standard deviation;
3. Preservation of edges; and
4. Texture preservation (restoration).
Existing speckle filtering algorithms can effectively reduce the speckle level. These algorithms however also, more or
less, smear edges and blur images. Adaptive filters, such as Lee's, Kuan's, Frost's and MMSE filters, take account of a
speckle distribution model, compute local statistics in a moving window, and assign pixels’ values accordingly, often
leading to better results compared to non-adaptive filters. However smoothing uniform areas and at the same time
preserving and enhancing edges is difficult to accomplish, because the former requires abandonment of high frequency
components, while the latter needs the preservation of high frequency components as much as possible. One method to
minimizing the deficiency of smearing edges, for instance, is to use edge-directed windows, rather than the traditional
square windows (Lee, 1981).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 89