SeaWinds Resolution Enhancement: Initial Results
David G. Long

Oct. 21, 1997

Abstract

This is an experiemental web page to report the initial results of the resolution enhancement of Seawinds scatterometer measurements. Basic idea: given one rev of simulated measurements (Level 1.5), multiple revs of noisy sigma-0 measurements of a small test region were simulated. Both full `egg' and slice measurements were simulated. Non-enhanced and SIRF-enhanced images were then generated from the simulated measurements. The results clearly show that the slices result in higher resolution images than the eggs, even if resolution enhancement is used. In addition, though the individual Seawinds sigma-0 measurements are noisier than NSCAT measurements, they better cover the ocean's surface, requiring less data to make images. Adequate quality images can be produced with only 1 day of data (versus 3 for NSCAT). This will be useful in tracking rapid temporal variations in sea ice.

Introduction

NSCAT sigma-0 measurements have proved to be very useful in land and ice studies. So much so, that we desire to continue making global Ku-band measurements of sigma-0 from scatterometer measurements in the future. Unfortunately, the original design for Seawinds produced measurements of much lower resolution than NSCAT, which would limit the utility of the Seawinds measurements compared to NSCAT measurements when used in land/ice science studies. The measurement resolution was determined by the area of the pencil-beam footprint on the earth (known as the `egg'). This discrepancy between NSCAT and Seawinds resolution motivated the motification of the Seawinds design to enable measurement of sigma-0 over small regions (known as `slices') of the pencil-beam footprint. The improved resolution of the measurements is useful for both wind retrieval and land and ice studies. The improved wind measurement resolution is particularly useful near coasts.

The purpose of this report is to briefly consider the effects of the change from eggs to slices on land/ice imaging. The effects of applying the SIRF resolution enhancement algorithm to both cases is studied. The results show that the slices result in significantly better resolution than the egg measurements and that although the measurement noise in the slice measurements is higher than for the eggs, the SIRF algorithm works well with the slices to produce high resolution sigma-0 images. In the following sections I describe the simulation technique used, present sample results, and finally offer conclusions.

Methodology

As a pencil beam scatterometer, Seawinds has a much different measurement geometry than NSCAT. The scanning geometry produces a denser sampling of surface for Seawinds with higher signal-to-noise ratios than NSCAT. However, the dwell time is much smaller for Seawinds. As a result the individual sigma-0 measurements tend to be noiser for Seawinds than for NSCAT.

NSCAT makes measurements of sigma-0 over a range of incidence angles. The dependence of sigma-0 on incidence angle, theta, is expressed as:

One of the big (pun intended) differences between NSCAT and Seawinds is the size of the data files. While one rev of NSCAT L1.5 data occupies 20 MB, the equivalent sample Seawinds file (L1B_kpm0.7.DAT) kindly supplied by Vincent Hsiao is 1.2 GB in size! (The sample file has not been optimized in any way so should end up more like 200 MB when in final form). Unlike wind retrieval, land/ice imaging requires multiple passes (revs) to generate useful images. The large file sizes can make multiple rev simulations difficult.

We have selected a very small region to evaluate the imaging resolution because of the large file sizes involved. Even so, the intermediate `setup' file for the very small test region is over 150MB for Seawinds compared to 4.2 MB for NSCAT. For reference, an Antarctic NSCAT image requires a setup file of >350MB. Making Seawinds images will require A LOT more CPU/disk resources than does NSCAT!

Measurement Locations

For convenience, I am using a small area in Antarctica to extract the sampling geometry and cells to create the simulated images. The region is approximately 1000 km x 800 km and is centered at -74.5 deg latitude and 128.5 deg longitude. This area is observed approximately 5 times per day by Seawinds, though some passes only cover part of the region. (Note that a similar size region near the equator would be observed somewhat less often.)

Instead of having Vincent generate multiple rev files and since this is a simulation, I've taken the single rev of data and generated synthetic revs by merely adjusting the longitudes of the measurements at a rate of 360*101/1440 deg/rev. Assuming that the rotation rate is not synchronized to the ascending node, I also added a small amount of latitude `jitter' to the measurement locations.

Vicent's file contains the corner location of the slices and their center. To compute the eggs, I used the outline of all the slices. This is not completely accurate but is close. Figure 1 shows the outline and relative locations of 3 succeeding eggs from one record. The upper left eggs are the outer scan while the lower right are the inner scan. Figure 2 shows the outlines of the center 10 slices of the first two inner beam measurements. Figure 3 shows the egg locations for multiple scans over a small region.

Fig. 1: Egg locations from the simulation file consisting of six consecutive measurements in the file. The outline of the 12 slices in the file are shown.

Fig. 2: Slice locations from the simulation file for the first two inner scan measurements. Only the center 12 slices are shown.

Fig. 3: Egg locations for part of one pass of a small test region.

Simulation Procedure

Simulated sigma-0 measurements are generated by laying the sigma-0 slices onto a synthetic image of the surface (see Fig. 4 below). The effective sigma-0 is the weighted average of the pixels of the synthetic image. Then, given sigma-0, the coefficients of Kpc and the information in the file, Monte Carlo noise is added to generate a simulated sigma-0 measurement. The simulated measurements are used in the imaging below.

Fig. 4: Synthetic `truth' A image. Grayscale range is from -20 to -5. Note that `A' is sigma-0 normalized to 40 deg incidence angle. This synthetic truth image is one we've used extensively. It simulates an Amazon-basin like image.

NSCAT Simulation

Just for reference, Fig. 5 shows the results of applying the SIRF resolution enhancement imaging algorithm to simulated NSCAT measurements of this same region and synthetic image. Note, however, that 6 days of v-pol NSCAT data have been used. NSCAT can achieve higher resolution than Seawinds.

Note that while the original synthetc image is rectangular, the projected study region does not fill the full output image area due to the equal area Lambert map projection. Comparing the NSCAT and truth image shows that NSCAT does a pretty good job of recovering the details of the truth image with some smearing along the edges.

Fig. 5: SIRF'd enhanced resolution (~4.5 km pixel resolution) A image generated from 6 days of simulated NSCAT measurements. Grayscale range is from -20 to -5.

Results

A number of cases will be compared. Using the simulated sigma-0 measurements, images were computed using (1) the traditional gridding approach and (2) the SIRF resolution enhancement technique for both eggs and slices. For the examples presented below, all the non-enhanced grid images have a resolution of approximately 25 km while the SIRF image have a pixel resolution of approximately 4.5 km.

Because the Seawinds measurements densely overlap (and we get lots of measurements per pass, particularly for the slices), reasonable images can be made from only one day of data in this polar region. However, as will be illustrated, the noise level in the images can be reduced if multiple days are combined. NSCAT can make reasonable v-pol images in 3 days, though for the polar region six days are used since this is the period required to make good quality h-pol images.

Gridded (non-enhanced) images

Figure 6 shows a gridded image created from Seawinds eggs. Note that while hints of key features are visible these are poorly represented and may not be fully visible. For comparison, Fig. 7 shows a gridded image produced from the slice measurements. We note the improved resolution of the image figures. (Fig. 8 is a repeat of the truth image inserted to make the comparison to the truth image easier.)

Fig. 6: Gridded (25 km resolution) A image generated from 1 day of simulated `egg' measurements. Grayscale range is from -20 to -5.

Fig.7: Gridded (25 km resolution) A image generated from two days of simulated `slice' measurements. Grayscale range is from -20 to -5.

Fig. 8: Synthetic `truth' A image. Grayscale range is from -20 to -5.

Resolution enhanced images

For comparison with previous images, I have produced images using a modified form of the NSCAT SIRF algorithm. I note that I have not `tuned' the SIRF algorithm for the Seawinds case. In particular, I have not incorporated antenna pattern weighting into the algorithm but have used a uniform weighting. As a result, it should be possible to generate better quality images than the ones presented here.

Figure 9 shows a SIRF'ed image generated from egg measurements while Fig. 10 shows a SIRF'ed image generated from slice measurements. Note that SIRF'ing the data improves the effective resolution and visibility of key features. The enhancement is most profound for the egg measurements, though the enhanced slice images have better resolution. While noisy, the SIRF'ed slice measurement is excellent. I extended the imaging period to use two days of data to produce Fig. 11 which is a SIRF'ed slice image. This image has similar resolution to the one day image but lower noise due to the additional measurements in the longer time period. Note that longer periods further reduce the noise. Also note that the two day SIRF'ed slice image is nearly as good as the NSCAT image, suggesting that Seawinds can be used for many of the same studies of NSCAT. Longer time periods can improve the quality the SIRF'ed egg images (compare Fig. 13 with Fig. 9).

Fig. 9: SIRF'd enhanced resolution (~4.5 km pixel resolution) A image generated from 1 day of simulated `egg' measurements. Grayscale range is from -20 to -5.

Fig. 10: SIRF'd enhanced resolution (~4.5 km pixel resolution) A image generated from 1 day of simulated `slice' measurements. Grayscale range is from -20 to -5.

Fig. 11: SIRF'd enhanced resolution (~4.5 km pixel resolution) A image generated from 2 days of simulated `slice' measurements. Grayscale range is from -20 to -5.

Fig. 12: Synthetic `truth' A image. Grayscale range is from -20 to -5.

Fig. 13: SIRF'd enhanced resolution (~4.5 km pixel resolution) A image generated from 2 days of simulated `egg' measurements. Grayscale range is from -20 to -5.

Conclusions

As expected, using slices rather than eggs improves the effective resolution of land/ice image produced from Seawinds. SIRF further improves the images. When slice sigma-0 measurements are used with SIRF, land/ice images comparable to NSCAT images can be produced. However, while NSCAT measured sigma-0 over a wide range of incidence at dual polarization, Seawinds will measure sigma-0 at only one incidence angle per polarization, and these angles are different. This will adversely impact the application of Seawinds data for some applications. For example, the NSCAT ice edge algorithm we have developed is based on the pol ratio and the incidence angle dependence of sigma-0 and so can not be used with Seawinds. We are currently evaluating alternative algorithms.