5.2. Results and Analysis
Figure 9 shows the photographs of all crops types at the date of the radar observation. In addition to the farmland area, two forest areas (one in near-range, the other in far-range) and one building area are also selected to test the decomposition performance. In order to assess the performance, we have also obtained the solutions from a number of traditional incoherent model-based decomposition methods, namely: Y4O (Yamaguchi four-component decomposition, Y4O [
5]), Y4R (Yamaguchi four-component decomposition with rotation, Y4R [
15]), S4R (Yamaguchi four-component decomposition with rotation plus dihedral volume scattering model, S4R [
18]), G4U (General four-component decomposition with unitary transformation, G4U [
21]). A spatial multilook with a 9 × 9 boxcar filter to reduce speckle noise was applied. The conventional RGB composites obtained from the six decomposition methods are shown in
Figure 10, whereas a comparative analysis of the results for each crop type is presented next.
In this paper, we will concentrate on the following fields (see location in
Figure 8):
- (1)
Sugar beet: Fields 102 and 460;
- (2)
Winter wheat: Fields 230 and 250;
- (3)
Maize: Field 222;
- (4)
Rape: Fields 101, 110, 130, and 140; and
- (5)
Winter barley: Fields 440 and 450.
Sugar beet: On the day of radar acquisition, Field 102 is a bare soil with few vegetation residual and Field 460 is a completely bare soil. The decomposition power contributions and best-fit volume scattering models for the two fields are shown in
Table 5 and
Table 6 respectively. It can be seen that, as expected, surface scattering is the dominant scattering mechanism in all the decomposition methods. Compared with Y4O method, the volume scattering components of the other five methods are smaller because of the orientation compensation. The double-bounce scattering components of Chen’s and proposed method increase and the volume scattering components decrease, respectively, when compared with other traditional methods. In addition, the proposed method maintains a high surface scattering component, similar to the traditional methods, whereas that of Chen’s method is the lowest one. For these two fields, the most used best-fit volume scattering models for Chen’s and the proposed method are the vertical dipoles model. However, because of the residuals, the results in the traditional methods show almost all vertical dipoles model in Field 102, whereas the mixture of vertical dipoles model and random dipoles model in Field 460. In general, it can be stated that for all methods there always exists some coupling of either double-bounce or volume components (or both) even in case of a bare rough surface. This suggests some limitations of the Bragg scattering model and its improvements for accounting for depolarization effects even at L-band. This is clearly seen in the volume component for all four traditional methods where it ranges between 17% and 24% of the total backscattered power. Otherwise the double-bounce is negligible in those cases. On the other hand, Chen’s and the proposed methods tend to balance the “residual” power not assigned to surface scattering between double-bounce and volume scattering. This negative effect is even more noticeable in the original method by Chen.
Winter Wheat: On that date, the winter wheat is in the early vegetative stages with an average height about 17–18 cm. It can be seen from
Table 7 and
Table 8 that a mixture of scattering components appears in the two fields from all decomposition methods. For both wheat fields, the best-fit volume scattering model for the traditional methods is the random dipoles model, the proposed method mostly chooses the vertical dipoles model, whereas Chen’s method mostly adopts the horizontal and vertical dipoles model. In addition, it must also be noted that the retrieved strongest scattering mechanism on Field 230 for the traditional methods is the direct surface whereas it happens to be the double-bounce for Chen’s and new methods. Otherwise, on Field 250, which is located at far range, the scattering mechanisms tend to be more similar among all them since the increase of the volume component leads to a redistribution of the power among mechanisms. This agrees with the expected behavior related to a higher incidence angle.
Maize: From
Figure 9, we can see that the maize field at that date is a bare soil with some vegetation residuals. In
Table 9, it is obvious that the dominant contribution in this field is surface scattering for all methods. The mostly used best-fit volume scattering model in traditional methods is the random dipoles model, whereas in Chen’s and the proposed methods is the vertical dipoles model. More interestingly, as happens for sugar beet fields (especially Field 102), the double-bounce is negligible (as expected) but, however, the volume scattering is 22.6% of the total power for Y4R, S4R, and G4U. While being a lower power than that provided by Y4O it becomes evident that none of the traditional methods is able to properly cope with the overestimation of volume component which seems to be an undesirable and essential attribute of these type of decompositions. In case of Chen’s and more clearly in the proposed method, it can be interpreted that they are able to diminish this drawback since it seems they tend to redistribute the residual power between double-bounce and volume scattering as it happened in sugar beet field. Notwithstanding this improvement, there is still 25% of the backscattered power in the best case (which is the proposed method) that is assigned to scattering mechanisms not expected on a bare surface at L-band. This effect clearly points out a recurrent flaw that must be further investigated. In this regard, the different scattering model choices in both traditional and Chen’s and proposed methods can be also indicative on the role that volume component could play in the decomposition as they could be acting as mere fitting components.
Rape: Results for four rape fields are shown in
Table 10,
Table 11,
Table 12 and
Table 13, respectively. As an overall comment beyond the numerical differences among methods, these results epitomize the noticeable disagreement between both types of decomposition methodologies. As shown, in rape Fields 110, 130, and 140, the traditional methods show similar levels of both surface scattering and volume with higher percentages than double-bounce scattering which represents 11% of total power at most. The least frequently used volume scattering model in these three fields is the vertical dipoles model, whereas the most frequently chosen volume model alternates between horizontal dipoles and randomly oriented ones. On the other hand, Chen’s and proposed methods retrieve in Fields 110, 130, and 140 a mixture of three scattering mechanisms without an obvious dominant scattering mechanism. According to these results and even though we have some knowledge of the status of rape fields on that date (see
Figure 9) we cannot provide any conclusive statement on which decomposition methodology characterizes better the scattering processes on rape Fields 110, 130, and 140. Considering now the rape Field 101 located at the furthest point from nadir, a different decomposition behavior is observed. In all methods, volume scattering is the dominant scattering component, which is consistent with a shallower incidence since the sensor becomes more sensitive to the volume. However, the volume power for the traditional methods is about 72%, which is clearly higher than 56% for Chen’s method or 62% for the proposed one. Regarding the selection of the type of volume model, there is again an obvious disagreement among the set of traditional methods, Chen’s and the proposed method. Among all them, the vertical dipoles model is hardly ever chosen.
Winter barley: From
Table 14, it can be seen that in Field 440 the surface scattering and volume scattering are the two dominant scattering mechanisms in traditional methods, whereas Chen’s and proposed methods exhibit double-bounce and volume scattering mechanisms as the two strongest mechanisms but the surface component is also noticeable. On the other hand, for Field 450 whose results are shown in
Table 15, all methods retrieve surface and double-bounce scattering as the two main scattering components. In this regard, both sets of techniques yield the same qualitative description for this field. In addition, for both fields, the mostly used volume scattering model in traditional methods is the random dipoles model, whereas in Chen’s and proposed methods are the horizontal and vertical dipoles model, respectively. It is important to emphasize that the incidence angle has hardly changed between Fields 440 and 450. This allows stating that: (1) all different decomposition methods adapt to the structural changes of soil and plants; and (2) whether these disagreements are due to design flaws on either one group of methods or the other or both cannot be ascertained according to the present study.
Forest: We tested two forest areas in the image, one in the near-range region and the other in the far-range region. From the results in
Table 16 and
Table 17, it is clear that volume scattering is the dominant scattering mechanism in these two forest areas in all methods, as expected. It is noted that in the traditional decomposition processing, the power of the surface and double-bounce scattering components are forced to be zero when the sum of the volume and helix scattering components is over the total span. Then the decomposition algorithm jumps to a two-component decomposition and the volume scattering will be directly calculated by just using the difference of total span and the helix power. For the forest area in near-range, after computation, there are 85.29%, 39.78%, 39.78%, and 39.78% of pixels in Y4O, Y4R, S4R, and G4U methods, respectively, in which this happens. For the forest area in the far-range, the percentages are 81.70%, 34.90%, 34.90%, 34.90%, respectively. Note that as the results of Chen’s and proposed methods are the optimized and adapted solutions by the algorithm, they do not suffer from this limitation. In addition, it must be highlighted that the random dipoles model is the mostly used volume scattering model in traditional decomposition methods, whereas the entropy model is the mostly selected volume scattering model in both Chen’s and proposed methods. It is also pointed out that some contribution of double-bounce mechanism is also expected from forest areas at L-band according to a number of previous works in the literature. As shown, Chen’s and the proposed methods fulfill this expectation better than all four traditional methods. However, at this moment we cannot provide any conclusive statement on this issue.
Building: According to the results, as expected, it is evident from
Table 18 that the double-bounce scattering is the dominant component in all methods. The volume scattering components in S4R and G4U are reduced because of the effect of the selection of the volume model caused by the oriented dihedral structures. The mostly used volume model in Chen’s and proposed methods is the entropy model, and the volume scattering component is further reduced compared with the traditional method. Finally, the proposed method shows the highest double-bounce scattering contribution, which demonstrates its consistency also for build-up area.
If we analyze the whole scene within the defined regions of interest (i.e., crops, forests and built-up areas), the mean values of the normalized minimum residuals are 0.0020 and 0.0062 for Chen’s and the proposed method, respectively. The values are both low enough, even though the proposed method shows a larger residual because of the tighter boundary conditions.
Regarding the retrieved values of
, the feasible range according to the model (
Figure 3) is from −0.5695 to −0.0516 for our tested E-SAR data.
Figure 11 shows the histograms of the values of
in all fields from all different methods. Since the traditional methods impose no limitation on the final retrievals and we found the values of some pixels are very high or low, in order to show the histograms, we limit them to the interval [−1,1]. It is evident that due to the assumption made in traditional decomposition methods the values of
are zero in many cases. On the other side, the result from Chen’s method is very likely to fall outside the physical range, whereas the proposed method yields solutions that are distributed in the physical range.
For , because of adopting the magnitude and angle of as the unknown parameters in the equations solving system and adding the constraint of the magnitude in the proposed method, there are no cases where the absolute value of is higher than one. However, 61.01% of pixels in Chen’s method exhibit retrieved higher than one. This is a consequence of using real and imagery parts of as unknowns in the nonlinear optimization processing and without adding the constraint of magnitude in the system.