Timothy Warner
Assistant Professor of Geology and Geography
Department of Geology and Geography
P O Box 6300
West Virginia University
Morgantown, WV 26506-6300

twarner2@wvu.edu

This report describes the data fusion of Synthetic Aperture Radar (SAR) and optical data for mapping roads at Fort Benning, GA. Directional filters were applied to the individual data sets, and then combined. This fusion of data from disparate wavelengths was found to provide a good method of ruling out some of the confusion in the individual data sets. In addition, because directional attributes are retained, this approach would facilitate the application of direction-following algorithms to isolate roads further.

Classification of Daedalus imagery has shown that roads are poorly classified on 3 meter optical imagery with a limited spectral range. A spatial model was able to improve the road identification somewhat by identifying linear arrangements of certain classes, such as shadow. This paper describes an alternative, the fusion of SAR imagery with optical data. Such methods hold great promise because they exploit fundamentally different physical properties of the targets.

The SAR data was originally collected at approximately 1 meter spatial resolution. This data was then aggregated to a 3 meter resolution (Figure 1), to match that of the Daedalus data. In addition, this reduced the speckle. The SAR data were acquired in a number of E-W strips. The mosaicking of these strips, and the lack of calibration of the raw data, means that it is not possible to infer quantitative backscatter properties from the raw DN values. In addition, it is not possible to calculate local incidence angles because the boundaries of each strip were not known, and are not necessarily discrete.

Figure 1. SAR backscatter image.

With such non-calibrated data, conventional classification is unlikely to succeed. Indeed, the SAR data actually degraded the classification if simply added as another band in a De Daedalus land cover classification. The inherently noisy nature of SAR makes alternative strategies more attractive. Texture is an attribute likely to be more robust in the presence of shifts in offset, though to a lesser extent in the case of shifts in gain (Barber and LeDrew, 1991). There are a bewildering variety of texture algorithms (Haralick, et al., 1973), but they are all attempts to quantify the scale-dependant spatial arrangements of the backscatter or reflectance of objects in the scene. The scale of the texture evaluation is normally determined by the size of a moving window or kernel. Thus for example, a common texture measure is the variance within a 3 by 3 kernel.

Texture is one of the major clues used in human vision. It therefore is counter-intuitive that it has been only moderately successful in applied image analysis. The major problem is that often the scale of the texture in the class is large (i.e. the texture is coarse), and therefore can only be measured with a large kernel. This unfortunately has the effect of blurring the image (Figure 2), producing artifacts on class boundaries. In the case of the Fort Benning Data, this can be seen most clearly on the edges of the zones of no data. One strategy that has been applied in an attempt to overcome this problem is an adaptive filter in which the new pixel value is the lowest value all the kernels that include that pixel (Ryherd and Woodcock, 1996). The disadvantage with this approach is that it inevitably produces a rather blocky texture image.

Figure 2. SAR texture image: 9 by 9 moving window of local variance.

In this research, an alternative approach was developed that has broad application to suppressing this edge-artifact problem. In addition, it exploits the expected linear attribute of roads. The method is to calculate many directional texture attributes for each pixel. For each texture measure, the kernel consists of zeros for all positions, except those along the angle of interest. The texture measure uses the non-zero pixels to calculate the kernel variance, which is returned to the central pixel. The simplest kernels are the horizontal (90 degrees) and vertical (0 degrees) directions. The following are two directional kernels for 5 by 5 data.

0° Directional kernel.

45° Directional kernel.

Larger kernels allow more directions to be evaluated. In this work a 9 by 9 kernel was used, which can be used for a maximum of 16 different directions, each approximately 11° apart. For each direction a variance was calculated, and the lowest was assigned to the final texture output (Figure 3). In addition, a second layer was generated, which was coded according to the direction of the minimum variance (Figure 4).

Figure 3. SAR minimum directional texture image.

Figure 4. SAR direction of minimum texture

As expected, the SAR directional data is highly noisy. Therefore the Daedalus data were also processed in a similar method (Figure 5) and combined with the SAR data. Prior to calculating the Daedalus directions, however, the original data were rotated through a principal component transformation, and only the second principal component (which is similar to a vegetation index) entered into the directional filtering program. The Daedalus and SAR direction of minimum texture data were combined by the rule:

Figure 5

Thus only directions that are nearly the same on both images are kept, and the rest of the image is set to zero. The image was cleaned further by thresholding the image such that the PC 2 was less than 23 (i.e. to be road, vegetation must be low), and also excluding the Water and Swamp spectral classes from the nPDF classification.

The results (Figure 6) are very promising, especially in the forested area. There are several very narrow roads, which are hard to see in the original data sets (especially the SAR), but are effectively isolated in this way. The fact that the direction of each pixel is also obtained offers the potential of applying a road following algorithm to this data.

Figure 6

The directional filters identify any linear feature, and thus the railway line is also well-discriminated. Although it's sinuosity suggests it might be a railway, there is no direct way to determine this from the directional texture image. In the regeneration areas, alignments along the regularly spaced conifers produce a network of spurious directions, which degrade the utility of this technique, unless further filtering can be applied to isolate the roads in these areas. One simple technique is to sieve out small clumps. Simply removing clumps smaller than 2 pixels does reduce the false positives somewhat (Figure 7), though at the expense of losing some of the isolated pixels that help define the narrow roads in the forest areas.

Figure 7

Directional minimum texture images are able to identify the spatial attributes in the data that are less affected by the mosaicking process. The SAR directional data are highly noisy, but when combined with directional texture derived from optical data, roads are well-isolated. Unfortunately, the technique needs further refinement to filter out false positives in the regeneration area.

Barber, D. G. and E. F. LeDrew, 1991. SAR sea ice discrimination using texture statistics: a multivariate approach. Photogrammetric Engineering and Remote Sensing 57 (4): 385-395.

Haralick, R. M., K. Shaunmmugam, and I. Dinstein, 1973. Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics SMC-3 (6): 610-621.

Ryherd, S. and C. Woodcock, 1996. Combining spectral and texture data in the segmentation of remotely sensed images. Photogrammetric Engineering and Remote Sensing 62 (2): 181-194.

1. SAR backscatter image.

2. SAR texture image: 9 by 9 moving window of local variance.

3. SAR minimum directional texture image.

4. SAR direction of minimum texture.

5. Daedalus direction of minimum texture.

6. Combined Daedalus and SAR direction of minimum texture.

7. Combined Daedalus and SAR direction of minimum texture, sieved to remove small clumps.

Timothy Warner

August 4, 1997

Link to Tim Warner's home page