Horizontal Versus Vertical Polarization

While SASS generally operated in vertical polarization modes, it also made a limited number of horizontally-polarized measurements. Over the study region, these were made primarily during two one-week periods in the first month of the mission. Coupled with the criss-cross coverage pattern (refer to [7]), horizontally-polarized measurements were not made at all over several diamond-shaped regions in our study area. In addition, because there were fewer horizontally-polarized measurements than vertically-polarized measurements, the intrinsic resolution of any reconstructed and images from the horizontally-polarized measurements would much coarser than the vertically-polarized and images.

While the paucity of the horizontally-polarized data limits our ability to draw firm conclusions, we have found that addition of the horizontally-polarized measurements increases our ability to distinguish between some varieties of tropical vegetation. Our results are presented below. In the following discussion subscripts are used to distinguish and values derived from vertically-polarized and horizontally-polarized measurements.

In Section 4.1, we presented the results of a supervised discrimination experiment using only vertically-polarized () image data. In this section we consider how these discrimination results are affected by including the vertically-polarized () image and the horizontally polarized and images ( and ). Three of the eighteen vegetation classes used in Section 4.1 were excluded from this experiment because of the lack of coverage by horizontally-polarized SASS measurements. The statistics for fifteen remaining vegetation classes are shown in Table 4.

As evident in Table 4, the coefficients were, as a rule, lower than the corresponding coefficients. It is also apparent the smallest differences between the average and values exist between members of the forest group (), and the difference increases slightly in the woodland group () and dramatically in the grass-shrubland group (). Formations having the greatest proportion of grassland, such as campos sujos/limpos (), pantanal (), and grassland with palms () display the greatest differences between and values. The standard deviations shown in Table 5 also indicate the variability in the backscatter coefficients was smallest in the forest group, somewhat larger in the woodland group, and greatest within the grass-shrubland group. With the exception of the woodland standard deviation values, the vertical coefficients displayed less variation than the horizontal coefficients.

When compared on a pixel by pixel basis, there is a significant linear correlation between the and coefficients for several of the classes. An interesting pattern is evident for the forest, woodland, and grass-shrubland groups. There is only slight correlation () among corresponding and coefficients for the forest group, with some members of the group (e.g., very moist forest) showing almost no correlation. For these members, this indicates that the difference () between the and values for individual pixels is primarily due to randomness or noise rather than the measurement polarity. The linear correlation between pixels increases through the woodland () and grass-shrubland () vegetation groups. For these groups, this indicates that the polarity of the measurements is a factor in the value of the coefficients. It is interesting to note that the greatest polarization dependence (as shown by ) is within the agriculture class ().

In order to explore the effects of polarization on actual discrimination results, the supervised discrimination experiment described in Section 4.1 (but with only the fifteen vegetation classes) was conducted using all possible combinations (there are fifteen) of the , , , and image data. As in the previous experiment, the fifteen vegetation categories were subdivided into forest, woodland, and grass-shrubland before conducting the quadratic discriminant analyses.

The results for the fifteen classification experiments are summarized in Table 6. This table shows the coefficient variables used in each classification experiment tabled against the simple percentage and Cohen's kappa accuracy figures resulting from the classification. Kappa is an accuracy metric which, unlike a simple percentage, accounts for the classification accuracy which could be expected by the operation of random chance alone. (The reader unfamiliar with Cohen's kappa may wish to review [11] for a discussion of its calculation and merits.) In comparing classifications, kappa is a more useful measure of discriminating power than is percent, particularly in experiments producing low accuracies and/or involving very few classes. A kappa of 0 indicates that there is no discrimination between the classes while a kappa of 1.0 indicates that errorless discrimination may be achieved.

From the table, a number of conclusions can be drawn. First, discrimination using coefficients derived from the vertically-polarized data is always superior to discrimination predicated on their horizontally-polarized counterparts. Second, experiments utilizing a single theme (i.e., , or ) exhibit significantly enhanced discrimination capability when the corresponding coefficient is added. The enhancement, however, is more noticeable in the vertical polarization case. Finally, if and data are utilized, a small but significant amount of discrimination ability is gained in adding horizontally-polarized coefficient data.

Based on the results summarized in Tables 4, 5, and 6, it appears there is significant polarization dependence in 14.6 Ghz backscatter for tropical woodlands, grasslands, and some tropical forest formations. It also appears the difference between vertical and horizontal is generally small in tropical forest formations but increases in tropical woodland and grass-shrubland formations. This implies improved vegetation discrimination between tropical forest and other vegetation groups can be achieved when both vertically-polarized and horizontally-polarized measurements are available. Finally, despite the variance in the coefficient data, the coefficient data does support a small amount of discrimination (particularly, ).

Despite the results from the classification comparisons, the reader should not assume that vertically-polarized 14.6 Ghz measurements are superior to horizontally-polarized measurements for discriminating between subtropical vegetation. It must be remembered that the reconstructed and images used in these experiments are based on much fewer measurements than the reconstructed and images. Since all four images were reconstructed to the same resolution, there is more noise in the and images than in their vertically-polarized counterparts (see Table 5 and the resolution versus noise tradeoff in the companion paper [7]). Hence, an alternative interpretation of the decreased classification accuracy of the horizontally-polarized measurements may simply be that there is less noise in the and estimates and, by extension, better class discrimination.