Speckle reduction in low-contrast areas by dedicated SAR processing

Abstract

The standard ERS SAR PRI product has 25 meter resolution and 3 looks. For some applications, the noise level of this product is too high; this is notably the case for the low contrasts associated with sea bottom topography. By processing RAW data using special weighting functions, an image can be produced with the same spatial resolution but a lower speckle noise level. By using combinations of overlapping weighting functions of various form, the synthetic beam shape and the speckle level can be adjusted to the situation at hand. Both theoretical results and examples with ERS data are shown. The resulting images are used for the extraction of bathymetry, where they are expected to lead to more accurate results. On the basis of the same principle, RAW data may be processed to reduce image artifacts due to strong scatterer side lobes; results can be applied in the determination of wind fields close to the coast.

Keywords: SAR processing, speckle reduction, bathymetry.

Introduction

One of the larger disadvantages of SAR imagery is its noisy nature caused by the speckle. The ERS PRI product, for example, has, at 27 m resolution and 3 looks, a speckle level of -2.4 dB (1). When weaker features need to be discerned in an image, this is troublesome. For many applications, the speckle level is actually the limiting factor; contrasts related to the sea bottom topography, for example, are often only a fraction of a dB, and therefore well below the speckle. It is possible to lower the speckle level by spatial smoothing, but for many applications it is not desirable to degrade the resolution. The speckle problem has been the focus of much attention, and much effort has been spent in the development of filters. While some of these filters can produce quite good results, the more powerful ones are only applicable for certain types of data or applications (e.g., in assuming the image to be constructed of point, line or edge features). Instead of using filters, which operate on the image, this paper concentrates on techniques that can be applied in the production of the image, i.e. in the SAR processing. Special attention will be paid to the low-contrast signatures of sea bottom topography.

Method

There is indeed some room for improvement in the ERS data. The standard multi-look image product, PRI, was designed for general use (for all types of scenes), for limited processing load, and for limited complexity to ensure product consistency between the different PAFs. When creating a multi-look image from raw SAR data, the azimuth bandwidth is divided into a number of sub-apertures (or looks), which are separately processed and incoherently added. The four main parameters that define the way the azimuth bandwidth is divided into the multiple looks are: the shape of the azimuth tapering (or weighting) functions; the width of each look; the number of looks, N; and their overlap. The shape of the tapering function determines the shape of the synthetic beam, i.e. the image point spread function (PSF). The width of each look determines the width of the PSF, and therefore the spatial resolution. The number of looks determines the speckle level, as the SNR (/) of the multi-look image is proportional to the square root of the number of looks. The PRI product uses three adjacent Hamming tapered looks with 320 Hz bandwidth each. It is, of course, possible to increase the number of looks, and thereby decrease the speckle level, by decreasing the bandwidth of the looks; however, this results in a lower spatial resolution and is therefore not desirable. Alternatively, it is possible to use another tapering function. A uniform (or rectangular) tapering function yields the same resolution at a narrower bandwidth; therefore, more looks can be accommodated within the total available azimuth bandwidth. Of course, the uniform tapering has as disadvantage its higher side lobes. In addition, it is possible to increase the overlap between the looks. The overlapping looks, being not fully independent anymore, still result in a reduction of the speckle level, although the dependence on N is weaker than N. The equivalent number of looks (ENL), defined as (/)², can be computed as a function of the overlap on the basis of the correlation coefficient between the different looks. The use of overlapping looks also obviates the need to divide the total available azimuth bandwidth into an integer number of adjacent sub-apertures, making it possible to use the total bandwidth to the fullest extent, given a target spatial resolution and tapering function. For the ERS, this makes it possible to use much more of the azimuth bandwidth than the 3x320 Hz of the PRI product (the SLC product uses approximately 1350 Hz, and Hamming tapering).

Instead of using overlapping looks of the same type, it is also possible to try to make optimum use of the available bandwidth by combining overlapping looks of different types. One technique that has been proposed for this is the "Improved Multi-Look" (IML) technique [1]; here, the azimuth bandwidth is divided into a set of wide Hamming windows with 50 % overlap, and again into a similar set of narrow Hamming windows, the latter of half the width. The relative weighting of the broad and narrow windows is a free parameter, which determines the PSF and the ENL.

Considering that making the windows overlap makes optimum use of the raw data, except at the extremes of the total band, another alternative for combining looks of different tapering functions is by adding a narrow window at the band extremes. Taking 50 %-overlapping Hamming windows again and making the width of the two extreme windows half that of the main windows, their relative weighting is again a free parameter determining the PSF and ENL. Alternatively, it is possible to use rectangular extreme windows; in that case, their width can be chosen so as to yield the same spatial resolution as that of the main Hamming windows.

Results

Two sample results of computing the ENL as a function of the fractional overlap are plotted in Figure 1, for the cases of uniform and Hamming tapering. It can be seen that the increase of the ENL begins to saturate at overlaps of more than 60 %. With Hamming tapering and 4 independent looks, a 65 % gain in the ENL can be achieved; with fewer independent looks to start with, this gain is less. Even with uniform tapering, overlap can yield an increase in the ENL, though not as large (25 % for 4 independent looks). (Note that the ENL values of Figures 1a and b cannot be directly compared, as they pertain to different resolutions.) In the case of ERS data, it is possible to process to standard PRI resolution (27 m) with 8 looks, 40 % overlap and uniform tapering, yielding a theoretical ENL of 6.3.

It is possible to compare the results of the various multi-look techniques by plotting the ENL as a function of the spatial resolution. This is done in Figure 2. It can be seen that, whereas the use of one type of tapering function leads to a linear increase of the ENL with spatial resolution, the IML technique at first gives a faster increase, which levels off at larger PSF. The technique that uses two narrow Hamming windows at the extremes shows an even faster increase at small PSF, but quickly drops off thereafter. The last technique mentioned in the previous paragraph has, by definition, a fixed beam width and is not plotted in Figure 2; its increase in ENL is maximally 4.1 %.

Figure 1. Equivalent number of looks (ENL) as a function of fractional look overlap, for uniform (above, a) and Hamming tapering (below, b). Drawn line is for 2 independent looks (i.e., the width of one look is half the full aperture), dashed line for 4 independent looks.

Figure 2. ENL as a function of 3-dB beam width, for overlapping identical looks (drawn line), the IML-technique (dashed), and the technique with two narrow extreme Hamming windows (dash-dotted).

The spatial resolution used in Figure 2 is the -3 dB beam width. Although this is of course a well-defined measure, this manner of comparison is not conclusive, since the various techniques have different PSFs. While the PSF of the Hamming taper is well behaved, the PSF of the IML technique is characterized by broad "wings". When using two narrow Hamming tapered windows at the band extremes, the resulting PSF also has wings, but at a lower level. As can be expected, the same technique but using two narrow rectangular windows has relatively high side lobes, though of course lower than those from a uniform tapering.

These techniques have been implemented in the Generic SAR Processor (GSP) of TNO. The GSP is a software package that runs on a SUN workstation and is able to process both airborne and satellite SAR data. One result, concerning sea bottom topography signatures, is shown in Figure 4. Processing and resulting image parameters are listed in Table 1.

Table 1. Parameters of the images of Figure 4. H: Hamming; U: uniform; (1): PRI product; (2); measured.

When inspected on a monitor or high-quality print, it can be seen that the speckle level is decreased for the cases processed by the GSP with the non-standard settings, as compared to the PRI product. In the examples using a uniform tapering, no effects of the higher side lobes can be distinguished by eye; this is mainly due to the low contrast of the scene and the lack of strong scatterers nearby. The resolution as listed in Table 1 was measured from an assumed point source (not within the printed part of the image), and the speckle level from an assumed homogeneous area. These values are therefore not very accurate, and for the ENL certainly lower limits. Nevertheless, it can be concluded that the ENL can be raised by a factor of 2 with respect to the PRI value.

Image 4b has been used in the Bathymetry Assessment System [2], a data assimilation scheme designed to extract bathymetric information from SAR images and in-situ data. It was found that the use of the speckle-reduced image indeed improved the bathymetry estimation, by of the order of centimeters. This is illustrated in Figure 3. In general, the exact amount of improvement very strongly depends on the situation at hand; in particular, it depends on the weighting the SAR data receives with respect to the in-situ data in the assimilation process. In the case of bathymetry applications, an image taken under favorable hydrometeorological conditions is quite valuable, since these conditions may not be very frequent; the fact that improved SAR processing can increase the ENL by a factor of two (which is equivalent to having a second, statistically independent image) in that sense therefore doubles the value of such an image.

Figure 3. Depth profile in the area of Figure 4. Horizontal scale is in km, vertical scale in cm. Plotted are in-situ measured bathymetry (drawn) and results from the Bathymetry Assessment System of ARGOSS (dashed); the latter with a PRI image as input (short dash) and the low-speckle image of Figure 4b as input (long dash).

Discussion

The use of heavily overlapping Hamming windows yields high ENL values without producing adverse side lobes; disadvantage is that it is computationally of the order of twice as intensive as a simple no-overlap image. The use of uniform tapered, non- or slightly overlapping windows yields high ENL values while remaining computationally favorable; however, the high side lobes degrade the image contrast somewhat, and the presence of strong nearby scatterers may severely disturb the image; this method is therefore suited for low-contrast areas. More complex techniques that lead to high ENL values which use combinations of overlapping looks are in general computationally intensive, and may lead to high side lobes (though not as high as for uniform tapering) and to PSFs with less favorable shapes (wings). It is to be expected that the use of images with these kinds of PSFs for bathymetry assessment will result in a slight underestimation of very small-scale detail in the bathymetry; whether this is offset by the increased bathymetric accuracy expected to result from the lower speckle level is not easy to assess.

When images produced by uniform tapering are used in a data assimilation scheme, it is possible to alleviate the effects of the strong side lobes to some extent. When data are used in an assimilation scheme, a weighting factor can be assigned on a pixel-by-pixel basis, to reflect the quality of each pixel. It is possible to define such a quality measure based on side lobe contamination, so that pixels with a high estimated side lobe contamination are used with a lower weighting in the assimilation. Such a quality measure can be generated from the observed image and the PSF. The use of this approach is currently being explored, also in the context of bathymetry assessment.

The freedom to choose the PSF of the SAR image by adjusting the processing parameters can also be exploited for such an application as extracting high-resolution wind fields from SAR images in inland waters or in the coastal zone. Such an application is often hindered by side lobes from strong scatterers located on the shore; by applying even smoother tapering functions than Hamming, these side lobes may be much more suppressed.

Conclusions

It is possible to produce SAR images from raw ERS data with the same resolution as PRI images, but with an ENL twice as high. In this way, the value of the data can be increased twofold for certain applications. In low-contrast areas, and when using the data as input in an assimilation scheme, this need not have computational consequences, through the use of uniform tapering and a side lobe-based image quality measure. By using overlapping Hamming windows, the same gain in ENL can be achieved even without the above restrictions on contrast and assimilation, however, at the cost of a much increased processing time. These facilities are implemented in TNO's Generic SAR Processor, and work is currently being carried out to bring this processor up to commercial operation.

Acknowledgments

Gerard Hesselmans (ARGOSS) is acknowledged for providing Figure 3. This work was partly funded by the EC MAST program, contracts MAS2-CT94-0104 and MAS3-CT95-0035.

References

Moreira, A., 1990: