3.1. Description of the Scenes and Field Samples
In both RADARSAT-2 scenes, co-polarizations (HH and VV) showed similar backscattering coefficients for most of the field samples belonging to different classes (Figure 3
), except for PFT E (tall grasslands, see Table 2
) for which backscattering was higher and more variable in HH than in VV. The analysis of field samples representative of each class suggests that using only one polarization is quite inappropriate for classifying the region of interest, especially with regard to bare soil discrimination (Figure 3
). However, HV (or VH) seems to be the preferred polarization for discriminating wetland classes with RADARSAT-2 scenes. Differences between information classes were higher in the cross-polarization band (HV) than in co-polarization bands (HH or VV). This issue has been pointed out in [11
]: the depolarizing character of the vegetation differs according to vegetation structure, biomass distribution and flooding state.
Characterized mainly by mirror reflection, water had low backscattering coefficients. However, water scattering was higher in the steep angle scene than in the shallow one, as expected for a slightly rough surface. The highest backscattering was observed for PFT A samples (bulrush marshes), probably due to the contribution of double-bounce scattering that was reported and modeled in [60
]. The lowest scattering among vegetation classes occurred for PFT D, grouping short plants with low cover and biomass, which probably allowed large signal penetration up to rough soil or water. Bare soil backscattering was always similar to the backscatter of at least one PFT, thus suggesting difficulty to discriminate bare soil in further classifications (especially if polarimetric information is not included). Although the backscattering of some classes showed differences between the scenes (Figure 3
), these differences were not evident in the Pauli representations (Figure 1
c,d) and high cross-polarization response (high |HV|) predominated in both scenes.
3.2. H/α Segmentations
The interaction between PFT targets and the polarimetric SAR signal was heavily influenced by the incidence angle (Figure 4
), assuming no substantial changes in vegetation, hydrometric or wind conditions during the two days between scene acquisitions (see Section 2.2
). In line with the predominance of volume scattering at C-Band from vigorous herbaceous vegetation, most of the pixels of both scenes had high H (H > 0.8) and intermediate α values (30° < α < 50°). Thus, segmenting the H/α plane in eight zones was not accurate for discriminating PFTs and bare soil: the non-water areas were grouped in the same class in both segmented images (Figure 4
a,b). The density of pixels in zones with low α and low or intermediate H, thus presenting simple interaction mechanisms, was higher in the steep angle scene than in the shallow angle scene (Figure 4
c vs. Figure 4
d). In the segmented image from the steep angle scene (Figure 4
b), those pixels were located in wetland areas with high H values in the shallow angle scene. These differential signal-target interactions according to the incidence angle may indicate the presence of flooded vegetation. Indeed, with a shallow incidence angle, volume scattering with one or more vegetation layers occurred, so that the signal was scattered before reaching water below vegetation. On the contrary, the steep signal was less attenuated by the vegetation and mainly interacted with water below it.
Few pixels with high α (indicating double-bounce scattering) occurred in both scenes. The ones with low H were located in the middle of the Paraná River or outside the region of interest in urban areas, pointing out dihedral reflectors (e.g., ships, buildings, metal roofs). Besides, the pixels with high α and intermediate or high H were probably related to double-bounce from vegetation targets. Their locations were coincident in both segmented images, but the covered area was slightly higher in the shallow incidence angle scene than in the steep one.
These analyses of the H/α space improved the previous knowledge about the response of herbaceous wetland vegetation to SAR signal. Most of the previous studies have described high H areas as forests and high α areas as cities (e.g., [23
]). However, high biomass vegetation assessed in this study featured very high H. According to the a priori
labeling criteria, the location of the PFT A pixels in the H/α plane was similar to the location of pixels with floodable trees and shrubs in the Amazonas floodplain [32
]. High α values in pixels dominated by the equisetoid herbs of PFT A (e.g., Schoenoplectus californicus
in bulrush marshes) presented high α values such as predicted due to its high double-bounce scattering in presence of a water film [60
3.3. Unsupervised Wishart H/α and H/A/α Classifications
The H/α Wishart classifiers converged to eight classes by shifting most of the centroids of the segmented plane to areas with high H and intermediate α (Figure 4
c,d). In both classifications, the centroid of one of the classes had very high α (α > 60°, dihedral reflectors, big ships). Thus, this class was excluded in further analyses. Next, the class with the lowest H and α was assigned to the open-water class. Among the remaining classes, the one with the lowest H (and low α) was assumed to belong to bare soil. The remaining classes were labeled as stated in Section 2.3.3
e,f). Both the bare soil and the PFT D (short grasslands, see Table 2
) classes had lower α in the steep angle scene than in the shallow angle one.
In the shallow angle classification product, four classes had H > 0.90 and were provisionally assigned to PFT E, A, B and C (in increasing order of H) (Figure 4
e). For these four classes with H > 0.90, a progressive classification procedure was followed by analyzing H/A/α subclasses, i.e.
, those classes obtained with the H/A/α classifier within each H/α class (see scheme in Figure 2
). Within PFT B, five H/A/α subclasses occurred. The geographic location of one out of these H/A/α subclasses was coincident with field records of sites dominated by PFT A, whereas other H/A/α subclass was coincident with field records of sites dominated by PFT C. The remaining H/A/α classes were retained in PFT B class. Thus, PFT discrimination was slightly improved by reassigning these two subclasses. The number of reassigned pixels was a minority compared to the total pixels in PFT B class. None of the pixels of PFT A, E and C were reassigned according to the H/A/α subclasses. The second labeling procedure (i.e.
, maximizing the Kappa index when assigning classes obtained to information classes) was coincident with the labeling carried out by means of our a priori
In the steep angle classification product, only one class had a mean H higher than 0.90 (PFT C) (Figure 4
f), so that the progressive classification on H/A/α planes was omitted. The two labeling procedures led to different results. Class assignment to information classes following the a priori
criteria is shown in Figure 4
d,f. Besides, Kappa index was maximized for this scene when pixels first assigned to PFT B were re-assigned to PFT E, and vice versa
. This result suggests that our physical hypotheses on vegetation were contrary to the observed pattern for these two PFTs, at least for the steep incidence angle scene. Our a priori
criteria could be misleading due to the difficulty to infer mean H and α values from biomass and non-quantitative vegetation structure data (distribution of biomass, degree of randomness in the canopy). In addition, the fact that our working hypothesis seems to be more accurate for the shallow incidence angle scene than for the steep one, is consistent with the general good performance of shallow incidence angle scenes for discriminating herbaceous wetland vegetation in the literature [11
3.4. Comparison between Incidence Angles and Accuracy Assessment
The spatial pattern of PFT classes differed according to incidence angles (Figure 5
a vs. Figure 5
b). Both maps were quite homogeneous without much noise (i.e.
, zones of vegetation were identifiable). The zonation of vegetation is typical in wetland ecosystems, where abiotic constraints are important for plant community development, and has been described for the study area [18
]. The classification products obtained with the shallow and the steep incidence angle scenes were coincident in 30.0% (a concordance of 17.8% in comparison with the randomly expected concordance, as estimated by the Kappa index). Water was almost equally classified in both scenes. Most of the differences between the resulting products arose for PFT E (tall grasslands), for which we identified a total area of 31 km2
in the shallow angle product and an area of 149 km2
in the steep angle one. Besides, 70.4% of the pixels assigned to PFT D in the shallow angle product were assigned to PFT B in the steep angle one, confounding short grasslands and short broadleaf marshes. In addition, much of the areas identified as PFT C in the shallow angle product were identified as PFT E in the steep angle one, and vice versa
; confounding tall broadleaf marshes and tall grasslands. This mismatch between the maps suggests that broadleaf and graminoid herbs interact differently with the SAR signal depending on the incidence angle.
Accuracy was higher for the shallow incidence angle scene than for the steep one. On the one hand, the product obtained with the shallow incidence angle scene had an overall accuracy of 61.5% or of 52.4% excluding the water class (Table 3
). The class with lowest commission error was PFT B, whereas the one with lowest omission error was PFT C (Table S1
). PFT D and PFT E (short and tall grasslands, respectively) were confounded. Sixty percent of the pixels belonging to PFT E were omitted and assigned to other PFTs and to bare soil. The accuracy differed from the chance-adjusted expected accuracy and the Kappa index was 54.8% or 42.5% (including or not water class, respectively) (Table 3
). In comparison with the Isodata classification on HH, HV and VV, the Wishart classifier improved the accuracy by 20.7% (including the water class) or by 34.2% (without the water class), in line with previous comparisons of the performance of Wishart classifications derived from quad-pol imagery and of classifiers based on standard linear dual-pol imagery [66
]. Regarding the omission errors (Table S1
), some of the pixels classified as open water in marsh areas might actually correspond to flooded vegetation. In areas of the Amazonas floodplain dominated by macrophytes, mirror reflection has been reported for emergent biomass (above water level) up to 200 g/m2
, from which volume scattering starts increasing [13
]. Thus, some the flooded areas of PFT D and E could be underestimated and assigned to the water class.
On the other hand, the product obtained with the steep incident angle scene had an overall accuracy of 46.2% or of 35.0% excluding the water class (Table 3
). Kappa indexes were low and, when excluding the water class, the 95% confidence interval included the zero. High omission errors were observed for all the PFTs and were the highest for PFT E (Table S1
). Optimum class labeling in terms of Kappa maximization improved the results and led to an overall accuracy of 53.3% and a Kappa index of 42.9%. Even with the Kappa maximization criterion, the accuracy was higher for the shallow incidence angle scene than for the steep one, being the former the more reliable product.
The results suggest that our hypotheses between PFT and SAR signal interactions were accurate only for the case of shallow incidence angles. In addition, our results were consistent with the reported preference of shallow incidence angle SAR signals for discriminating vegetation types and of steep incidence angle signals for detecting water below marsh vegetation and for flood monitoring [11
]. In our study, at least two observations support this fact: in comparison to the shallow incidence angle scene, the steep one showed a higher density of pixels with low H and low α (Figure 4
d vs. Figure 4
c); and the centroids of classes PFT B and C featured lower H and α. The observed pattern can be explained by considering that steep SAR signal is less attenuated by vegetation layers and can interact to soil or water below the vegetation.
Regarding double-bounce scattering, the interaction mechanism seems to be more common for targets observed with high incidence angle scenes: pixels with medium H and high α in the shallow incidence angle scene showed low H and low α in the steep one. In the case of flooded herbaceous vegetation with low emergent biomass, a vertical incidence angle may diminish the probability of a first bounce in vegetation and a second one in the water (or vice versa
), increasing the probability of reaching water. This observation is in marked contrast to what happens in forest environments, where for a given polarization and frequency, double-bounce scattering is more important in steep than in shallow incidence angles scenes [12
In the Lower Paraná River floodplain, previous vegetation maps performed with optic data have overestimated the forest coverage at the expense of herbaceous wetlands [68
]. Thus, the map obtained with the shallow incidence angle scene is a better start point than optical-derived maps for monitoring wetland vegetation in the floodplain. In the future, the study of SAR/optical fusion schemes will be considered. Our results suggest that most of the area was dominated by a ruderal strategy or an intermediate ruderal-competitor strategy (PFTs B, C and E; 58.7% of the total area or 63.8% excluding water). Ruderal plants are characterized by fast growth rate, short life span and high seed production, thus are favored by chronically disturbed but potentially productive environments [52
]. The hydrological regime periodically disturbs the marsh areas, along with cattle activity. Ruderal plants can quickly regenerate between two ordinary flooding events, thus exploiting and nutrients brought by the river and taking advantage of resource opportunities. In addition, plant species with intermediate ruderalness-competitiveness may play a crucial role in vegetation adaptation to wetland states or in disturbance resistance. The fact that most of the floodplain seems to be covered by ruderal or intermediate ruderal-competitor plants gives to the environment a high ability to recover after disturbances and ordinary floods [69