Analysing PolSAR data from vegetation
by using the subaperture decomposition approach

The SAR imaging process consists of a high amount of low-resolution echoes from a scene collected by the moving sensor at azimuthal angles which are combined to create a full resolution radar image. As a result, each pixel in a SAR image represents an observation of an area across a specific range of angles constrained by the azimuth antenna pattern. The Doppler spectrum arising as a consequence of this time-varying azimuthal observation of the scene can be exploited to analyse the potential backscattering variations along the azimuth angle [2]. Figure 2 provides a simplified scheme of the subaperture decomposition concept by using three subapertures (indicated in black, red and blue colors).

Equation (3) provides the link among the Doppler frequency $f_{D}$, the sensor velocity $V_{radar}$, the azimuth look angle $\phi$ and the radar wavelength $\lambda$ [24]:

The relationship between azimuth resolution $\delta_{az}$ and the azimuth observation interval $\Delta\phi$ is as follows [24]:

This approach is performed by first applying the Fourier transform along azimuth dimension. Then, in the Doppler spectral domain a SAR signal for a particular range bin $r$ is expressed as follows [25]:

where $S_{t}(\cdot)$, $H(\cdot)$ and $W(\cdot)$ represent the Fourier transforms of the transmitted signal, the scene coherent reflectivity and the focusing reference function (both included in $H(\cdot)$) and the weighting function, respectively. An average sidelobe reduction window can be recovered to correct for the spectral imbalance $W(f_{D},r)$ in the full resolution image [2]. Once the weighting function has been compensated, the third step is the splitting of the original Doppler spectrum into several sublooks (with or without overlapping) represented by a reduced Doppler bandwidth around a particular Doppler frequency $f_{D0}$ by applying the windowing function $G(f_{D}-f_{D0},r)$:

It is noted that function $G(f_{D}-f_{D0},r)$ accounts also for the sidelobe reduction in each subaperture spectra before they are converted back to spatial domain by an inverse Fourier transform. This procedure leads to a number of lower resolution SAR images according to the selected amount of subapertures employed. It must noted, however, that results differ depending on the number of subapertures and the degree of overlapping among them in the spectral domain [26, 27] and this in turn is directly related to the resolution loss [7].

First, we have employed indoor polarimetric data from test vegetation samples (i.e. small fir trees and maize-see Figure 3) acquired under controlled conditions at the European Microwave Signature Laboratory (EMSL) [28]. These datasets have been widely employed for vegetation radar scattering studies [29, 30, 31, 22]. The antenna beamwidth uniformly illuminates the vegetation sample and the system is operated in a stepped-frequency mode in 10 MHz steps ranging from 1.6 to 9.4 GHz for the fir trees sample and from 2 to 9 GHz for the maize. Data acquisitions cover the whole azimuth interval at 5^∘ steps making available 72 measurement through the azimuth look angle. This set-up allows an analysis of the polarimetric signatures as a function of the azimuthal observation which is relevant for the purpose of the present study as the whole azimuthal span can be splitted into four quadrants (hereinafter also referred to as subapertures) thus providing a way to analyse vegetation scattering variations depending on the azimuth look angle.

The P-band polarimetric SAR datasets acquired by DLR’s E-SAR system during BioSAR 2008 campaign have been also employed [32]. This campaign was carried out in the Krycklan river catchment in northern Sweden, an area characterised by hilly terrain with elevation ranging from 100 to 400 m and surface slopes of up to 14^∘ within the monitored forest stands.

All P-band polarimetric features have been computed by employing a 9$\times$9 boxcar filter.

The selected product is an ALOS PALSAR Level 1.1 Single Look Complex (SLC) image. It was downloaded from the Alaska SAR Facility website (https://search.asf.alaska.edu/). The product ( ALPSRP176360090-L1.1) covers the tropical forest in the Paracou area, French Guiana, a well-known test site as it was monitored in the framework of the TropiSAR campaign conducted in 2009 [33]. Terrain exhibits gentle topographic variations of about 30 m. Figure 5 shows the Paracou test site.

In order to avoid the possible effect of slopes terrain, we have employed the Google Earth tool aimed at providing the elevation profile and the terrain slopes to focus our analysis in areas where the slopes are within the range $\pm$2% which corresponds to a $\pm$1.15^∘ interval. All L-band polarimetric features have been computed by employing a 9$\times$9 boxcar filter.

Figure 6 shows the variation of $1-m_{FP}$ as a function of frequency for the four quadrants of the whole azimuthal span and the corresponding $1-m_{FP}$ values obtained by averaging all samples acquired by covering the whole azimuth interval (shown before in Figure 1 and repeated here for convenience). Most of $1-m_{FP}$ values, either for partial or the whole azimuth span, exhibit levels very close to or higher than 0.5 (except at L-band and around 4 and 5 GHz for two particular subapertures). In terms of backscattered power this means that the depolarising effects dominate the radar response. However, the $1-m_{FP}$ plots for all four azimuth quadrants clearly show different behaviours for different frequency bands which translates into different contributions of the diffuse backscattering power to the total power of the particular subaperture. At L-band there is a 9% difference between the first and both the second and fourth subapertures. The highest differences have been found to be 17% at about 3 GHz between the second and fourth subapertures and also at 7.5 GHz between the first and the fourth azimuth intervals. It must be noted here that speckle noise has been properly accounted for as the number of samples averaged is 162 (i.e. 9 samples within a bandwidth of 1 GHz around the central frequency and 18 azimuthal samples per frequency) and, hence, the observed variations must be attributed to different scattering behaviours.

Together with the depolarising effects substantiated in $1-m_{FP}$ the contributions of the total backscattered power of each individual subaperture to the total span of the whole image have been also computed. It is the ratio $Span_{i}/Span$, being $Span_{i}$ the span of the subaperture $i$ and $Span$ the total backscattered power for the whole azimuth illumination. They are shown in Figure 7. As the full azimuth interval has been equally split into four quadrants, all four ratios would be ideally 0.25 whenever the assumption of an isotropic distributed target is observed. As shown, the highest differences appear at S-band between 1st-2nd and 3rd-4th quadrants but the ratio values are within the 22-28% interval which is pretty close to the expected 25% ratio.

The same analysis has been carried out for the maize sample and the corresponding plots are shown in Figures 8 and 9. This case shows a qualitative difference with respect the fir trees cluster. The $1-m_{FP}$ values are consistently below 0.5, being most of them below 0.4. Indeed, from 4 GHz onwards values follow a decreasing trend up to about 0.25 at higher frequencies. This behaviour indicates that polarised components become dominant in the radar signature which means that anisotropic effects occur in the wave propagation, despite the power ratios are also equally distributed among all subapertures (except for the second subaperture at higher frequencies which decreases to about 20%). This observation is in agreement with the conclusions drawn in [22] by employing the coherence tomography approach. It is noted that the differences among $1-m_{FP}$ for all four subapertures reach values about 10% (up to 12.5% at 8 GHz). These values do not reach the ones for the fir trees but certainly they represent noticeable differences regarding the depolarising effects as a function of the azimuth look angle.

The slope map provided in the BioSAR2008 campaign is displayed in Figure 10 where the monitored stands are overlaped. The pixels with slopes outside the interval $\pm$5^∘ have been saturated and are shown in yellow and dark blue colors on the map. For our analysis we have focused on zones where terrain slopes effects are negligible so we have selected transects whose slopes range between -2.5^∘ and 1.25^∘. These transects are indicated on the slope map with red color lines and letters.

Figure 11 displays the $1-m_{FP}$ factor for individual subapertures as a function of terrain slopes for the transects selected and indicated with letters A-F in Figure 10. In general, the observed values are mostly limited within the $[0,0.3]$ range, therefore the diffuse scattering contribution for individual subapertures can be at most 30% of the total backscattered power for a particular azimuth look angle. This is not a dominant volume scattering which agrees well with SAR tomography studies [8] reporting a dominant polarised ground signature at P-band. Also it is observed that there is no clear pattern when comparing the $1-m_{FP}$ values among subapertures, thus suggesting that anisotropic effects arise depending upon the azimuth look angle and that they can be present in more than one subaperture (see transect A for slopes around -1^∘ or 0) or transect B for -0.5^∘ slopes. Transects A, B, C and D exhibit also some parts where depolarising effects in some subapertures are more noticeable reaching values higher than 0.3.

In addition to the analysis of the $1-m_{FP}$, we have also applied the stationarity test proposed in [2] where a Maximum Likelihood (ML) ratio test is designed to discriminate between anisotropic and isotropic scattering signatures as follows:

In Eq.(7) $Nsub$ represent the number of subapertures, $\mathbf{T_{i}}$ is the coherency matrix for subaperture $i$, and $N$ is the number of samples for averaging. It is noted that this is an expression different to that provided in [2] but totally equivalent.

According to [2], the $\Lambda$ parameter is expected to increase in natural covers due to the stationary spectral behaviour from distributed targets. Contrarily, $\Lambda$ will decrease for anisotropic pixels as the stationarity assumption is not fulfilled.

In order to better understand the correspondence between the $\Lambda$ parameter and the anisotropic-isotropic behaviour, we have computed the ML ratio for the zone located at near range where several anisotropic and man-made structures can be seen which are surrounded by natural areas (see the area close to the coast in Figure 5). Figure 12 shows the results of $\Lambda$ in logarithmic scale (i.e. $\log\Lambda$) together with and RGB image in the lexicographic basis (R=HH, G=HV, B=VV).

The $\log\Lambda$ is displayed in a grey scale image where black level corresponds to values from 0 to -1, gray levels represent a transition interval between -1 and -2, and white level ranges from -2 to the lowest $\log\Lambda$ value found in the image. All structures showing a preferred orientation are clearly visible in white and light grey shades whereas the isotropic behaviour is displayed in black.

According to the previous results for this particular dataset, we set the threshold for discriminating between anisotropic and isotropic pixels in $\log\Lambda$=-1. By doing so, it seems possible to detect heterogeneities in forest covers. We have computed the ML ratio for a long transect (named G in Figure 10) where slopes are within the $[-1.9^{\circ},1.25^{\circ}]$ range. Figure 13 shows this result in solid line. Circles indicate pixels where a difference in $1-m_{FP}$ between any two subapertures is higher than 0.1 and where $\log\Lambda>$-1. Stars indicate pixels where a difference in $1-m_{FP}$ between any two subapertures is higher than 0.15 and where $\log\Lambda>$-1. The former case represents 28.39% of pixels for that transect whereas the latter case accounts for the 14.84% of pixels. Therefore, despite the stationarity test yields levels associated to an isotropic medium, the differences among depolarising factors suggest that vegetation heterogeneities could be affecting the scattering signature in a significant amount of pixels.

To support the previous interpretation based on the $1-m_{FP}$ differences and the $\log\Lambda$ indicator, we have computed a $\log\Lambda$ map corresponding to the highest forest stand (the one which contains transect B in Figure 10 whose average height is 21.408 m [32]). Figure 14.a) displays the terrain slope of the selected area where the black line represents the limits of the forest stand. Figure 14.b) shows a grey scale map where scattered darker patches show the pixels that simultaneously exhibit a $1-m_{FP}$ difference between any two subapertures higher than 0.2 and where the stationarity hypothesis is fulfilled (in our analysis, $\log\Lambda>$-1). Light grey shades inside the dotted blue line region are pixels where the slope terrain is within the $\pm$2^∘ interval. We focus on this particular zone to avoid that polarimetric signatures get modified by slope terrain. The number of pixels fulfilling both conditions represents 25.05% which means that, for this P-band dataset, one fourth of those pixels exhibit polarimetric signatures not compatible with the expected isotropic signature from forest covers. It is noted that if a 0.1 difference among subapertures $1-m_{FP}$ was considered, the number of pixels where potential anisotropic effects would appear increases to 58.71%.

For this dataset, this means that the relative contribution of the diffuse scattering to the total backscattering of an individual subaperture can noticeably vary. This leads to different balances between polarised and unpolarised components depending on the considered subaperture even though the ML ratio test characterise those pixels as isotropic targets. Therefore, this fact suggests that the assessment of anisotropic pixels in forests can be carried out on the basis of the information provided by $1-m_{FP}$. Additionally, the total backscattering power gathered by an individual subaperture, i.e. $Span_{i}/Span$, can also provide additional information, as seen next.

A zoomed area from pixel 5000 to pixel 6000 in transect G is displayed in Figure 15 where, for each subaperture, the $1-m_{FP}$ factor and the relative contribution of total backscattering are displayed. For comparison purposes, the full resolution $1-m_{FP}$ has been also plotted (black solid line) on the top image in Figure 15.

In agreement with previous observations, the $1-m_{FP}$ noticeably varies among subapertures. For some parts the full resolution depolarising factor is dominated by one of the subapertures (first and third strong peaks of the transect and around pixels 5200 and 5760, in case of the second subaperture). In other regions there appears a more similar behaviour among all three azimuth look angles or even there could be cases where two out of three subapertures resemble the full resolution pattern whereas the third one is either above or below it. In addition, there also appear some cases where the full resolution $1-m_{FP}$ takes values higher than any of the ones for individual subapertures. On the other hand, the relative contributions of individual subaperture backscattering exhibit a highly variable pattern with high dynamic range. In order to jointly explore the patterns exhibit by both features, the scatter plots of $1-m_{FP}$ vs. $Span_{i}/Span$ have been represented in Figure 16 for the whole transect G.

In general, these scatter plots exhibit a higher variance of $1-m_{FP}$ at low values of $Span_{i}/Span$ up to 0.3 or 0.35, i.e. when the relative contribution of the subaperture backscattering to the whole image span is not higher than 35% . On the contrary, when an individual subaperture collects the highest amount of backscattered power, then the $1-m_{FP}$ factor greatly decreases to values below 0.2 which is compatible with strong polarised radar returns. These scatter plots reveal an average inverse relationship between $1-m_{FP}$ and $Span_{i}/Span$ which is consistent with potential target structural variations (ignoring terrain slope effects as it is the case) causing anisotropic signatures in vegetation radar scattering.

A particular case to exemplify this behaviour is displayed in Figure 17 where pixel 5760 is considered. Subaperture #3 gathers 65% of total span but it is a totally polarised signal, as inferred by the $1-m_{FP}$ value which is less than 0.01. Indeed, the full resolution MF4C decomposition for that pixel yields percentages of 40.6%, 28.37%, 26.5%, and 4.51% for Pd, Ps, Pv, and Pc, respectively. Hence, the polarised components made up by double-bounce and surface scattering represent around 69% of the total backscattering. In addition, for the same pixel, it is also pointed out that subaperture #2 only covers 15.3% of total span but it senses depolarising effects that mostly determine the depolarising factor $1-m_{FP}$ of the full resolution image, which is 0.265, despite that for subaperture #1 is as low as 0.16 and even less than 0.01 for subaperture #3, as mentioned above. This case clearly differs from others also shown in Figure 17, as for the pixel intervals 5700-5715 and 5730-5740 where the full resolution $1-m_{FP}$ gets closer to the average of all three $1-m_{FP}$ subaperture values.

Additionally to the mentioned two situations (i.e. the full resolution $1-m_{FP}$ either dominated by only one individual subaperture or, alternatively, with a value closer to the average of the $1-m_{FP}$ subapertures), there could be also a third case for which the full resolution $1-m_{FP}$ becomes much higher than any of the values for individual subapertures. Figure 18 displays a portion of transect G where these different behaviours are illustrated. It is noted that letters A, B and C represent, respectively, pixel intervals for cases where $1-m_{FP}$ is dominated by only one individual subaperture, it is closer to the average, or it becomes higher than any individual value.

The conclusion from this particular P-band dataset is twofold. On the one hand, all these observations provide an additional evidence of the anisotropic effects taking place in forest covers at P-band, where the random volume assumption is likely to break down due to local vegetation heterogeneities [7, 9]. On the other hand, the reported variation of factor $1-m_{FP}$ for every subaperture in comparison with its value for the full resolution allows us to hypothesise that depolarisation effects even when not being dominant and being only sensed through a small portion of the synthetic aperture could lead to overestimated retrievals of the volume scattering for the full resolution image.

Three different transects have been selected for the analysis with this L-band dataset. Figure 19 shows the first one indicated with a red line overlapped on the SAR composition and a Google Earth image. At the bottom of the figure the elevation profile is also plotted. Maximum and minimum values for slopes are 3.2% and -4.8%, i.e. 1.83^∘ and -2.75^∘, respectively. Hence, any dominant slope effect potentially coupled in the radar returns can be ruled out.

The $1-m_{FP}$ factor for every subaperture has been computed and plotted in Figure 20. In general, it exhibits an increase of $1-m_{FP}$ values due to the stronger vegetation radar signature at L-band in comparison to the previously analysed P-band data, as expected. Some areas exhibiting a high variability among subapertures have been indicated together with the corresponding slopes (in %). Numerical differences of $1-m_{FP}$ can reach values up to 0.3 which entails a remarkable difference in depolarising effects induced as a function of the azimuth look angle.

Figures 21 and 22 show a second transect where slope values range between 2.7% and -5%, i.e. 1.54^∘ and -2.86^∘, respectively. Again, there are some parts where the depolarising effects greatly vary depending on the subaperture with differences over 0.3, in agreement with the outcomes from transect 1.

As in the case of P-band data, we have also computed the $\log\Lambda$ indicator which is shown in Figure 23 together with a HV channel image for reference. Black level and dark grey shades represent those pixels where the stationarity hypothesis is fulfilled. Expectedly [2], this is the case for most of pixels in this image which mostly includes natural covers. The values of $\log\Lambda$ are within an interval ranging from -1.15 to 0, being mostly of them within the [-0.3,0] interval, as shown by the histogram in Figure 24. Both the absolute values and dynamic margin of $\log\Lambda$ are noticeably reduced with respect the P-band case. There are some isolated white pixels located inside the yellow rectangle which belong to areas with man-made oriented structures, thus indicating that anisotropic effects dominate there. The main conclusion to be drawn from the $\log\Lambda$ map for this L-band dataset points towards an unquestionable validity of the stationarity assumption and, hence, the absence of orientation effects in the forest cover with regard L-band data. Nevertheless, this conclusion seems to contradict the high variation found at different azimuth observation angles in the depolarising factor $1-m_{FP}$, which can be explained by structural variations of vegetation, in line with our previous analysis at P-band.

A third transect has been also selected along range covering a lower slope terrain area within the interval $\pm$2.6%, i.e. $\pm$1.5^∘. This transect has been indicated with a red rectangle in Figure 23. In this case we have also included in the analysis the $Span_{i}/Span$ signatures together with the $1-m_{FP}$ variation for all subapertures as well as for the full resolution data. These are displayed in Figure 25.

On the one hand, the $Span_{i}/Span$ ratios take values quite close to $Span/3$ which represents a proportional distribution of the total span among all three subapertures. This is in agreement with our previous observations on indoor data from vegetation samples (Section 4.1) where from L- and S-band up to X-band the $Span_{i}/Span$ ratios were almost equal. This fact, however, is in contrast with the clear variation of $Span_{i}/Span$ along the azimuth look angle shown at P-band (Section 4.2).

On the other hand, the qualitative behaviour exhibited by $1-m_{FP}$ differences among subapertures in Figure 25 matches those observed in all cases previously presented not only in this section at L-band but also for the two previously analysed datasets. The differences reach values higher than 0.3 which suggests important depolarising effects as a function of the azimuth look angle. In addition, the full resolution $1-m_{FP}$ factor shows the three different behaviours reported above at P-band (see Figure 18), i.e. 1) values dominated by only one individual subaperture, 2) values closer to the average, or 3) values higher than any individual value. The overall conclusion from this dataset is not only consistent with findings reported in previous studies [13] regarding boreal forest heterogeneities at L-band, but it also gives support to the hypothesis that non-dominant depolarisation effects could lead to overestimated volume backscattering power in the full resolution image.

The previous analysis based on the subaperture decomposition provides an additional insight for the understanding of the volume backscattering overestimation in PolSAR decompositions. This issue has been object of a intense research (see for example [34, 35, 36, 19, 37]) since the original work by Freeman and Durden [38]. Despite the great progress achieved on the topic, no full consensus has yet been reached as to whether there is an optimal way to avoid or minimise this inconsistency, despite the progress made in model-free decompositions [19] or approaches which merge both model-free and model-based schemes [39, 40]. Therefore, there is still a need for analysing the ranges of applicability of the proposed models [37].

Here, we propose a heuristic method for volume component correction based on the assumption that the vegetation anisotropy sensed by any of the subapertures should lead to a decrease of the depolarising effects present in the full resolution image. However, our previous analysis on the $1-m_{FP}$ factor seems to suggest that depolarisation effects only sensed through a small portion of the synthetic aperture and without being a dominant mechanism could lead to a overestimation of the full resolution volume scattering instead. Therefore, we propose to weight every subaperture $1-m_{FP}$ value with the ratio $Span_{i}/Span$ which corresponds to the proportion of the total backscattered power gathered by any individual subaperture $i$. Hence, the corrected value of the full resolution $1-m_{FP}$ factor is defined as a function of $Span_{i}/Span$ and the individual $1-m_{FP}^{i}$ factor as follows:

Once the full resolution $1-m_{FP}$ is recalculated according to Eq.(8), then the corrected diffuse backscattering power is obtained as in Eq.(2).

We have tested this idea for some transects at P- and L-band previously considered above. Results are shown in Figure 26. In both cases, all corrected $1-m_{FP}$ values are lower than the original ones (but for some isolated cases around pixel 5150 or 5700 at P-band, or pixel 460 at L-band). It is observed that differences can reach values up 0.28 at P-band and 0.16 at L-band but, however, there are also some regions in both transects where the corrected $1-m_{FP}$ values do not differ too much from the original ones. This is specially evident at L-band.

According to these new $1-m_{FP}$ values and following the model-free decomposition framework proposed in [19], this correction leads to the modification of the rest of backscattering powers as they all depend on $m_{FP}$. In particular, it can be seen that one direct consequence of this correction on $1-m_{FP}$ is that the helix backscattering power $P_{c}$ is inversely modified with respect to the volume backscattering $P_{v}$. The $P_{c}$ expression is [19]:

where $\tau_{FP}$ is the scattering asymmetry parameter defined in [19].

As seen in Eq.(9), whenever $1-m_{FP}$ gets lower then $m_{FP}$ is increased and so does the helix component. An expression for the new value of $P_{c}$, i.e. $P_{c}^{{}^{\prime}}$, in terms of the variation of volume scattering power, $\Delta P_{v}$, can be obtained as follows. The new volume backscattering $P_{v}^{{}^{\prime}}$ (due to the new depolarising factor) can be defined as:

Then, considering the new degree of polarisation $m_{FP}^{{}^{\prime}}$ and Eq.(2), the new volume component is:

From Eq.(11), the new degree of polarisation is given as:

Therefore, by using Eq.(9) and substituting the new expression for the degree of polarisation $m_{FP}^{{}^{\prime}}$ therein, and considering that

and

then, the new value for the helix component as a function of the variation of volume scattering power, $\Delta P_{v}$, is written as follows:

Eq.(15) yields a higher value for the new helix component whenever the new volume power gets reduced (i.e. a negative $\Delta P_{v}$) or, contrarily, it would be decreased in case that the volume scattering increases after the correction (i.e. a positive $\Delta P_{v}$).

It is important to bear in mind that this constitutes a preliminary attempt to derive a method to assess the volume overestimation issue and further research on this topic is beyond the scope of this study. Results provided in Figure 26 only offer an exploratory perspective of the problem and they must be scrutinized in a more systematic way by undertaking additional analysis on larger vegetated areas.

The present study has several limitations that need to be addressed in further work. Firstly, we have employed the subaperture decomposition by assuming three different subapertures with no overlap between adjacent spectra. Previous research [2, 41, 26, 27, 4] have extensively analysed these issues and found differences depending on these parameters. As we have only examined one possible situation, caution must be applied when interpreting these results. In addition, the degradation of azimuth resolution and the number of samples employed for averaging (which influences both azimuth and range resolution after speckle filtering) have also an impact in the retrieval of scattering mechanisms [7] and should be taken into account. Note that, for example, the upcoming ESA’s BIOMASS mission is designed to provide 12.5 m and 25 m in azimuth and slant-range coordinates, respectively. Hence, these issues entail reasonable uncertainty regarding the application and performance of any approach based on subaperture decomposition.

Secondly, despite the new insights provided on the overestimation of the vegetation canopy scattering, this study does not demonstrate whether the proposed method for the correction of the depolarising factor $1-m_{FP}$ yields realistic decomposed backscattering powers. This is a drawback also shared by other existing and widely employed methods based on either mathematical or physical constraints and, certainly, it still constitutes a fundamental open issue in this field.

Thirdly, it must taken into account that the width of the azimuth observation interval is the key element in the detection of anisotropies. As is known, different microwave propagation effects are more clearly identified whenever this interval becomes greater. Both the multi-frequency indoor data and the P-band airborne data are inherently acquired under wide azimuth intervals whereas the L-band dataset was acquired by a satellite-borne sensor, thus reducing the azimuthal angle diversity. Despite that in the present results remarkable quantitative differences in the depolarising effects along the azimuth dimension have been observed in all cases, the interpretation provided for P-band data and its practical implication for P-band spaceborne acquisitions must be still confirmed as the anisotropic effects may be diminished under a narrower azimuth interval.