SAR Image Enhancement:Combining Image Filtering and Segmentation

Jean-Marie BeaulieuComputer Science Department

Laval UniversityQuebec City, Quebec, G1K-7P4, Canada

Abstract – SAR images are generally affected by speckle noise due to the interference of reflected waves from manyelementary scatterers. The speckle complicates image interpretation and reduces the effectiveness of automatic imageanalysis techniques. A good filter should maintain the signal mean value and reduce the speckle while preserving theedges. Adaptive filters use an estimation of the variation coefficient. The utilization of fixed size windows produces poorestimates. We examine new ways to use image segmentation techniques to improve image filtering processes. In a complextask, the objectives of the first step should be limited in order to obtain reliable results that will be the base of next steps.We use a region growing process to reduce extremum values as a pre-processing step to SAR image filtering.

Keywords: SAR image, filtering, segmentation, region growing, remote sensing.

1- Introduction

Synthetic aperture radar (SAR) systemsprovide high-resolution images for remotesensing application. The utilization of themicrowave band makes the image acquisitionmostly independent of weather condition. SARimages are generally affected by speckle noisedue to the interference of reflected waves frommany elementary scatterers. The specklecomplicates image interpretation and reduces theeffectiveness of automatic image analysistechniques. The speckle is generally modeled as amultiplicative noise.

A good filter should maintain the signal meanvalue and reduce the speckle while preserving theedges. Several filtering methods have beenproposed such as the Lee filter, the Kuan filterand the Gamma filter [1] [2] [3]. These filters areadaptive and use local statistics on a fixed sizewindow to determine the weighting factor. Thesefilters are better than the Box filter or the medianfilter because they consider the multiplicativenature of the speckle noise. However, they use anestimation of the coefficient of variation of thesignal intensity to make a decision about thehomogeneity of an area. In some cases, thisestimation may not be reliable which producespoor results.

In remote sensing, a segmentation processcould be used to detect land fields and to improvepixel classification. The segmentation of SARimages is greatly complicated by the presence ofcoherent speckle. The complex structure of theSAR images requires the utilization of a compositecriterion for the segmentation. SAR image filteringis often used as a pre-processing step.

We examine new ways to use imagesegmentation techniques to improve image filteringprocesses. In a complex task, the objectives of thefirst step should be limited in order to obtainreliable results that will be the base of next steps.We use a region growing process to reduceextremum values as a pre-processing step to SARimage filtering.

2- Utilization of spatial information

2.1- SAR image model

Radar images are mainly characterized by thepresence of speckle [2]. The radar signal can bemodeled as a random process. The variance of thesignal can be reduced by performing multi-lookprocessing, i.e., by averaging L independentviews of the data. The probability density

function of L-look signal follows a Gammadistribution:

−

Γ= −

RILexpI

R

L

)L(

1)I(p 1L

L

where I is the signal intensity and R is the groundreflectivity of a homogenous area. In this case,the mean value and the standard deviation of thesignal intensity are: µI = R, σI = R / L . As thestandard deviation varies with the mean value, wegenerally use the variation coefficient to expressthe signal variation: CI = σI / µI = 1 / L . Figure1 illustrates de multiplicative nature of the radarsignal. The variance increases with the meanvalue. The figure shows clearly the large varianceof the radar signal.

Intensity I

p(I)

P0

P0 = 1

P0 = 2

P0 = 4

Figure 1: The probability density function ofthe radar signal.

2.2- Filtering and segmentation

The spatial information is important infiltering and segmentation techniques. Classicallinear filtering techniques perform convolutionoperations using fixed window sizes. Thewindow is centered on the output pixel anddifferent weight values are assigned to eachwindow cell. Different filters result from theutilization of different weighting functions. Thisis a linear transformation with no decisioninvolved. High frequency attenuation filters areused to reduce the noise by smoothing the image.Gradient or pass-band filters are used into edgedetection processes. Adaptive image filteringtechniques want to smooth the image insidehomogeneous areas (regions) while preservinglines and region contrasts (edges). A decisionprocess is used to detect edges and avoid edgeblurring.

In SAR image filtering, the image is viewed asa stochastic process. Pixel density probabilities arecalculated. The image filtering is considered as anestimation problem: estimate the truth pixel valuesfrom the observed values. Generally, the bestestimate corresponds to the mean value over a setof homogenous pixels (from a same population).This results in calculating the mean value over awindow. Correct values are obtained if the windowcover homogenous areas. Adaptive techniques tryto detect if the window contains pixels fromdifferent fields and then adjust the averagingprocess. The goal is to preserve linear features andedges. The decision is based upon statistics (meanvalues) calculated over fixed and predefinedwindows (eventually, with varying shapes).

Image segmentation is the division of theimage plan into disjoint regions. All the differentapproaches to image segmentation involve adecision process to determine if 2 adjacent pixelsor sets of pixels are similar and should be puttogether (merged). The difference between the 2mean values is often used. A segment is a set ofconnected pixels. We will consider onlyhierarchical or region growing approaches wheresegments are sequentially merged to form largersegments. We could initially start with eachindividual pixel as segment. We then sequentiallytake 2 adjacent segments and merge them if theyare similar. The shapes and sizes of segment arenot predefined. The segments grow and expand asthe segmentation process progresses.

An important difference between filtering andsegmentation is the selection of the pixel set usedin calculation. In segmentation, the pixels areincluded into a segment as the result of decisions.When we calculate the mean value of a segment,we know that the used pixels have been foundsimilar. In filtering, the calculation is performedover predefined windows and we do not know apriori if the pixels in the windows are similar. SARadaptive filtering techniques try to select the mostappropriate window among a set of alternatives.Hopefully, the selected window will containsimilar pixels. Unfortunately, the decision ofteninvolves values calculated over windows that wedo not know yet if they are homogenous.

2.3- The Gamma filter

Many SAR adaptive filters [3] use a weightedaverage of the pixel value I and the mean value Iover a fixed size centered window:

)y,x(I)w1()y,x(Iw)y,x(R̂ −+=

where R̂ is the estimated ground reflectivityvalue and w is the weighting factor. Differentfilters use different formulas to calculate theweighting factor from the coefficient of variationestimated over the window. Near edges, theweighting factor should be close to 1 to preservethe initial value. Inside homogeneous fields, thevalue should be 0 to smooth out the noise. TheGamma filter [1] is similar but uses a morecomplex non-linear equation. The filter is used tosmooth the image while preserving the linearfeatures and the edges. It tries to detect imagediscontinuities and reduces smoothing to preserveedges. A 7x7 window is used. Figure 2 shows a 7-look airborne C/X SAR 580 image. There arelarge variations of pixel values between neighborpixels. In the Gamma filter result, Figure 3, wehave uniform surfaces inside fields. Near edges,the original pixel values are more preserved,evoiding the blurring of the edges. There are alsomany zones inside fields with large pixel valuevariations.

3- A region-growing "filter"

We propose a region-growing approach todefine the set of pixels used in mean value or otherparameter calculation. The definition of the pixelset or segment is performed independently andsequentially for each pixel. For each pixel, we firstcreate a segment containing this only pixel. Foreach 4-neighbour pixel, we calculate its absolutedifference with the initial pixel. The segment growsby including the neighbor pixel with the smallestdifference. The list of neighbor pixels is updatedafter each merge. The merging continues until thesegment size reaches a predefined stopping value.The mean value of the segment is calculated andassigned to the output pixel (segment startingpoint).

Figure 4 shows the result obtained when theregion growing is applied to the SAR image ofFigure 2. The region growing is stopped when thesegment contains 30 pixels. This is not what weexpect from a filtering process. There is nosmooth image surface as in Figure 3. The processhas replaced the speckle texture by a new type oftexture where the range of the gray level valuesinside region is smaller. The formation ofunconstrained shape segments results in thepreservation of edges while reducing thesmoothing aspect (because of the preservation ofcontrast on segment boundaries).

Figure 2: A 7-look 300x300 SAR image. Figure 3: Gamma filter result (7x7 window).

Figure 4: Region growing "filter" result.

We do not consider that this region growingprocess could be used as a filter by itself. Itshould be used as a part of a more completeprocess. We are also interested by a modifiedversion for extremum value reduction.

4- Reduction of extremum values

SAR images are characterized by speckle noisewith large variance. They can be described ascomposed of alternating bright and dark spots.There is some similarity with salt-and-pepper noise(or impulse noise). The impulse noise could beviewed as outlier values that should be removed.The mean value of a set of pixels minimizes themean-squared-error and will be greatly affected byoutlier values. Therefore, median filter is moregenerally used with impulse noise [4]. Often, themedian value could be used instead of the meanvalue. To calculate the median, we need to orderthe pixel values, witch is computer timeconsuming, and select the middle one. It isexpected that the values in the middle of the listwill be rather similar and that outliers will belocated at both ends. A fair amount of outlierscould be accepted without producing an incorrectmedian value. Median filter has been used for SARimages. When the processing window cover ahomogeneous area, the mean value is a betterestimate than the median.

The bright and dark spots of SAR images arenot really outliers. They are values coming fromboth tails of the signal distribution. The large valuerange complicates the processing of the images. Itis the purpose of image filtering to reduce the valuerange. However, this is a difficult task. We areconsidering the fact that image filtering should beeasier if we could reduce the value rangebeforehand. To do so, we will consider only thevalue coming from both ends of the distribution. Inthe case of the median filter, those extreme valuesare replaced by the middle value of the ordered list.We are looking for a more limited modification. Ifthe value is at the high end of the list, we couldselect the Mth lower value in the list. If the value isat the low end, we select a higher value.

The bright and dark spots correspond to localextrema. The reduction process will start only fromthese extrema. We also take advantage of spatialinformation by applying a region growing processfrom each of these extrema. If we start from amaximum (minimum) point, then all its neighborswill have a lower (higher) value. The region willgrow by including the neighbor pixel with thehighest (lowest) value. The merging process willstop when the segment (region) has reached therequired size M. All the pixels inside the segmentare replaced with the lowest (highest) value of thesegment. We should use small value for theparameter M (2 to 5).

5- SAR image processing

The extremum reduction technique could beused as a pre-processing step for SAR imagefiltering and segmentation. This takes into accountthe probability information about the radar signal(rank statistics) and the spatial information (imagefield structures). The advantage of the process isillustrated by the filtering of a SAR image with andwithout pre-processing. For the extremum valuereduction pre-processing, we use 3 pixel segments(M=3). Figure 2 shows the original SAR imagewith 7 looks and 300x300 pixels. Figure 3 showsthe Gamma filter result. Figure 5 shows the resultof the extremum reduction pre-processing step. Itlooks rather the same as the original image. Figure6 shows the result of the Gamma filter applied toFigure 5. The pre-processing step contributes toimprove the inside field smoothing.

Figure 5: Extremum reduction result.

Figure 6: Extremum reduction and Gammafilter result.

Figure 7: Region growing and Gamma filter.

Figure 8: 1-look 400x400 Raradsat image.

The region growing process could also be used as apre-processing step. Figure 7 shows the result ofthe Gamma filter applied to Figure 4. The 2 lastfigures show that the 2 segmentation-basedprocesses could be useful to improve thesmoothing of homogeneous areas. However, thereis some deterioration of linear features and edges.Further studies are needed to assess the importanceof this deterioration while taking into account theimprovement of smoothing. As proposed in [3], thepre-processed image could be used to calculate thecoefficient of variation only. The initial image is

used for the calculation of the output mean value ofthe adaptive filter.

The following figures show the results for a 1-look SAR image where the noise variance is larger.Figure 8 is the original 1-look 400x400 Radarsatfine mode image. Figure 9 shows the Gamma filterresult (5x5 window). Figure 10 shows the Gammafilter result after extremum reduction (M=3) andFigure 12 after region growing (M=10). The pre-processing step contributes to improve the insidefield smoothing.

Figure 9: Gamma filter result.

Figure 10: Extremum reduction and Gammafilter result.

Figure 11: Region growing result (M=30).

Figure 12: Region growing and Gamma filterresult.

Evaluation of SAR image filters is a complextopic [5]. We have used synthetic images toanalyze the effect of filters on edges (fieldboundaries). 4-look images with 2 regions areproduced. The left region has a mean value of 1and the right region a mean value of 4 or 8. The toprow of Figure 13 shows the value distributionsinside each region. A 5x5 window is used. The 2nd

row shows the distributions of the filtered values at3 column from the edge where the window coversonly one of the 2 regions. The next row, at 2

column from the edge where the window includesone column (among 5) from the other region, andthe last row, next to the edge where 2 columns ofthe window are inside the opposite region. Thesmoothing corresponds to the reduction of thewidth of the distribution. The edge is blurred if themean value (the center) of the distribution isshifted. The adaptive nature of the Gamma filtermeans that the smoothing is reduced as we getcloser to the edge. There is a shifting of the valueinside the right region only (mean of 4 or 8). The

utilization of an extremum reduction pre-processing step results into more smoothing(thinner distributions). The mean value of the rightregion is not displaced while there is a smalldisplacement for the left region.

6- Conclusion

Image filtering and segmentation are basicallytwo different approaches to image processing. Wehave presented 2 segmentation–based techniquesthat could be useful as a pre-processing step inSAR image filtering, a region growing "filter" andan extremum reduction process. Appropriatelyadapted filtering techniques could also be useful aspre-processing step to SAR image segmentation.

Acknowledgements

The author would like to thank Dr RidhaTouzi and the Canada Centre for Remote Sensingfor their supports.

References

[1] A. Lopez, E. Nezry, R. Touzi and H. Laur,"Structure detection and statistical adaptivespeckle filtering in SAR images," Int. J.Remote Sensing, 14(9), 1735-1758, 1993.

[2] Chris Olivier and Shaun Quegan,UNDERSTANDING SYNTHETIC APERTURE

RADAR IMAGES, Artech House, Norwood,MA, 1998.

[3] Armand Lopes, Ridha Touzi and E. Nezry,"Adaptive speckle filters and sceneheterogeneity," IEEE Trans. Geosc. &Remote Sensing, 28(6), 992-1000, 1990.

[4] Sanjit K. Mitra and Giovanni L. Sicuranza,NONLINEAR IMAGE PROCESSING, AcademicPress, 2001.

[5] Hua Xie, Fawwaz T. Ulaby, Leland E.Pierce and M. Craig Dobson,"Performance metrics for SAR speckle-suppression filters," in Proc. 1999 Int.Geosc. Remote Sensing Symp.,IGARSS’99, Hamburg, Germany, 1999.

Figure 13: Signal probability distributions before and after filtering.

0 1 2 4 6 8 0 1 4 8 1 2 1 6

0 1 2 4 6 8

0 1 2 4 6 8

0 1 2 4 6 8

0 1 4 8 1 2 1 6

0 1 4 8 1 2 1 6

0 1 4 8 1 2 1 6

0 1 2 4 6 8 0 1 4 8 1 2 1 6

0 1 2 4 6 8

0 1 2 4 6 8

0 1 2 4 6 8

0 1 4 8 1 2 1 6

0 1 4 8 1 2 1 6

0 1 4 8 1 2 1 6

3

2

1

contrast: 1 to 4 contrast: 1 to 4contrast: 1 to 8 contrast: 1 to 8GAMMA FILTER PRE-PROCESSING & GAMMA FILTER