Validity range of RVoG model

As discussed in Sec. 3.1.2, the ground-to-volume scattering ratioµ absorbs certain soil and vegetation properties into a single parameter. From the point of view of inverse modeling, this gives the opportunity to invert few coherence measurements to retrieve bio-physical parameters that are input to the model12, in particular forest height and ground topography. However, from the point of view of the forward modeling, a limited parameter space13 _{pose the question whether the model is able to predict reliably the} coherence in different conditions. For instance, a sloped terrain induces a variation of the ground-to-volume ratio which translates into a variation of the predicted coherence (according to the RVoG model). Is this variation of model predictions in accordance with observations?

We try to give an answer at this question for three environmental variables, i.e. the average tree height, the azimuth terrain slope and the range terrain slope. Each of those scene parameters affects the value ofµ and consequently impacts on the degree of co- herence. In other words, our objective here is to establish the range of validity of RVoG model against a variation of these three parameters. To compare model predictions with observations we should be able to assess the input parameters of the model from the data. Whilehv andϕg can be available (e.g. from LIDAR measurements), estimating µ may result a hard task. Indeed, if we were able to estimate µ at different polarization with proper techniques, such the blind separation of sources (Pham, 1996), the inversion strategies exposed in Sec. 3.2 would result much easier and accurate. Therefore in this study PSPSim numerical simulations are of fundamental importance: they generate the individual scattering mechanisms (direct-ground, direct-volume, volume-ground) from which the ground-to-volume ratio can be easily determined. In the following, the gen- eral methodology is first exposed and then the analysis over the three environmental variables is presented.

The approach for the range validity assessment is based on both the forward model- ing and the inverse modeling (Fig. 3.12). In the first case, we compare the coherence predicted by the RVoG model with the coherence estimated from PSPSim complex images. We indicate the former as the predicted coherence, and the latter as the ob- served coherence, since it requires an estimation procedure similar to real acquisitions. We do not enter into the details of the classical interferometric processing: it is known for long time (cf. Chapter 1) and several freeware and commercial packages exist for that purpose. We have implemented the interferometric processing chain for PSPSim imagery including interferogram flattening, vertical wavenumber computation, speckle

12. Obviously, the forest properties embedded into µ cannot be retrieved using the RVoG model and other approaches should be employed if we wish to retrieve them.

13. A parameter space is the manifold generated by the set of input parameters to the model (Taran- tola, 2005).

pixelwise using (3.3) b γpq = hspq1s ∗ pq₂i q hspq₁s∗pq₁ihspq₂s∗pq₂i , p, q = h, v. (3.63)

Note that coherence is evaluated here only for the three polarization states HH, HV and VV, although an optimization algorithm could be also employed. In order to calculate a predicted value of the coherence through RVoG model, the values of the mean extinctionκeand ground-to-volume ratioµ must be known. Given the selective backscattering matrices of the forest S(dv)_i and of the ground interactions S(dg)_i and S(gv)_i , the ground-to-volume ratio can estimated at different polarizations for each acquisition

b µpqi= σpq(dg)i + σ (gv) pqi σpq(dv)i = h|s (dg) pq_i |2i + h|s(dg)pq_i |2i h|s(df )pqi| 2_i , p, q = h, v, i = 1, 2. (3.64) and then averaged between the two acquisitions since the polarimetric stationarity condition holds (cf. Sec. 3.2.2)

b µpq= b

µpq1+µbpq2

2 p, q = h, v. (3.65)

The wave extinction in the canopy is also, in reality, polarization dependent. However, the randomly uniform nature of the volume layer in the RVoG model entails a constant extinction coefficientκe. Its estimated valuebκeis obtained by averaging the direct and double-bounce extinctions estimated at H and V polarizations, calculated during the numerical simulation as discussed in Chapter 1. Alternatively, κe can be estimated from the ratio between the attenuateds(dg)pq_i and the un-attenuated ˜s(dg)pq_i direct-ground

returns for a fixed height, e.g. hv= 15 m, b κe= 1 4 X i=1,2 X p=h,v cos θ 2hv ln |˜s (dg) pp_i |2 |s(dg)pp_i |2 (3.66)

wherein the double summation means averaging among polarizations and acquisitions. Using the ground topography phaseϕg and forest heighthvoutset for each simulation, and the estimated ground-to-volume ratio_bµpqand extinction bκe, the predicted RVoG coherence is, for each simulation,

γ_pq = γg,v(ϕg, hv,µbpq,bκe) , p, q = h, v (3.67) where γg,v is the RVoG model function showed in (3.17). Although not shown, in (3.67) we utilize some system parameters of the interferometer, namely the wavelength λ, the look angle θ and the perpendicular baseline B⊥. The temporal baseline T is meaningless here since PSPSim simulates only the spatial decorrelation. Looking at Fig. 3.12, the first qualitative assessment of the reliability of RVoG model is the comparison between_bγpq andγpq,p, q = h, v. Any deviation in the predicted coherence propagates through the inversion procedure and leads to errors on the height and ground topography estimates. The quantitative range of validity is established with a threshold on this error, related with the requested accuracy of the height retrieval, usually10% of the total height (Papathanassiou et al., 2005a). In order to estimate the RMS error, the observed coherence is inverted using the hybrid magnitude/phase inversion strategy outlined in Sec. 3.2.4, disregarding the temporal correlation factor. Then the estimated forest height bhv and topography phaseϕbg are compared with the respective true value given as input to the numerical simulations and the validity range is assessed. In the case of slope analysis, the procedure is the same, with the exception thathvis fixed and b

µpq varies as consequence of the terrain tilt. The forest scenario is described hereafter along with the qualitative and quantitative results.

Varying the forest height

The imaged scenario has been described in detail in Sec. 2.3 and comprises a Scots pine forest, an underlying ground surface and a layer of short vegetation above the ground. In the height parametric analysis, sixteen different realizations of this scenario are generated by varying trees height from 6 m up to 25 m on a mean flat terrain. The ground surface is located at zero reference (ϕg= 0) and is characterized by a small scale roughness with 0.034 m correlation length, and a large scale roughness with 5.425 m correlation length. Dielectric value of soil is fixed at 9.717-j1.316. The short vegetation is 0.30 m tall and is a uniform random layer of stems and leaves with volume fraction

approaches the critical baseline. This aspect is well illustrated by the SAR frequency system model described in Chapter 1. The radar observes the targets from two different look angles, and two slices of the target spectrum are transferred into the images. The frequency shift of these two slices corresponds to the spatial decorrelation. A way to reduce this effect is to apply a process named spectral shift and filtering (Gatelli et al., 1994) which selects the constructive common bandwidth of the two acquisitions. As a result of the spectral filtering, the correlation is unitary (for pure surface scatterers) and spatial resolution of the interferograms reduces. In our case, this process would increase the coherence of aboutB⊥/B⊥,cr' 0.05 and hence it is disregarded.

Fig. 3.13 shows the set of backscatter outputs at different forest height, including the total and the individual returns, i.e. direct-volume, direct-ground and ground-volume return. For each scattering return, the backscatter in the three polarization channels HH, HV, VV is shown. As general behavior, in the total and direct-volume backscatter images the grey intensity increases as trees become taller. On the contrary, direct- ground component decreases if the forest height increases, as it is evident from the darker circular area in the image. Interesting is the case of the ground-volume mech- anism. The co-polar channels HH and VV backscatter increases from 6 m to 15 m, and then decreases again above 15 m. This is due to two concurrent effects: the larger canopy depth absorbs more ground-volume co-polar power and, at the same time, in- creases the ground-trunk interactions (VV) and the ground-branches interactions (HH). The ground-volume cross-polar return, on the contrary, originates mainly from the in- teractions between ground and canopy (or curved branches), which tend to augment with the forest height. A more accurate qualitative analysis is conducted by plotting the value ofµ averaged in the central vegetated area versus hv. The plot in Fig. 3.14a

14. The resulting height of ambiguity is ha ' 31.4 m. Although this value might be small for real

forests, our simulations are obtained with hv< 25 m, which avoids the folding of the phase along the

show that the ground-to-volume ratio decrease for taller trees in the co-polar channels and is almost constant in the HV channel. The value ofµ in the HV channel is about 10-20 dB lower than HH and VV channels.

The interferometric coherence has been estimated over11 × 11 pixels and the HH, HV, VV, HH+VV and HH_{−VV coherence magnitude and phase are shown in Fig. 3.15. It} appears evident in all polarization channels that the scattering phase center lifts off the ground with the increment of forest height. This is in accordance with the plots of the RVoG model in Fig. 3.3b and Fig. 3.4b where the normalized phase height has a monotonic trend versus canopy height and ground-to-volume ratio15. On the contrary, it is not true in general that the coherence magnitude decreases as trees become taller. This is a peculiar aspect of volumetric media and arises from the combined contribu- tion of ground and canopy, as confirmed by the model analysis of Fig. 3.4a. In our observations, these behaviors appears stronger in the VV channel of Fig. 3.15, wherein the coherence magnitude decreases until about 18 m and then increases again.

From the polarimetric point of view, we can draw two important observations. The first is that, in general, we cannot state that certain polarimetric channels have higher coherence level than other channels. Indeed, depending on the ground-to-volume ratio, the coherence observed in HV channel, for instance, can be higher or lower than the HH coherence, as appears for shorter trees in Fig. 3.15. Note that the coherence magnitude affects only partially the position of the PolInSAR coherence points along the line in the complex plane, since it is more the coherence phase that determines their relative position. The second observation is indeed on the coherence phase. Fig. 3.15 confirms the physical link among polarization states (or scattering mechanism) and scattering phase center height. As example, we notice how the HH_{−VV scattering phase center} lies always below the HV phase center. The comparison between observed and pre- dicted coherence is plotted in Fig. 3.16. The observed coherence is averaged in the central forested area of each simulation. Since we have used realistic values of ground- to-volume ratio combined with forest height and mean extinction, we can assess the scattering phase center location with high fidelity. For instance, in the case of forest 20 m tall with moderate density, the HH phase center is located around 5 m above the ground and the HV phase center about 5 m below the top of the canopy.

The predicted coherence is also plotted with dashed lines. The main conclusion from these plots is that the RVoG model is generally very reliably for different forest height, with some exceptions of coherence magnitude for higher canopy. However, these errors are negligible since the inversion procedure is robust enough and the forest height re- sults retrieved with accuracy better than10% over all the range of tested forest height (cf Fig. 3.14b). We remark that this is true for pine trees and in presence of only spatial

15. Note that in the height parametric analysis the ground-to-volume ratio is expected to decrease with the forest height in all polarimetric channels.

zero-slopes case in the HV channel, and remains almost constant in the HH and VV channels. The same trend can be observed on both the predicted and the coherence estimated from PSPSim simulations. In Fig. 3.19, the total coherence HH−VV reveals the ground tilt with a linear change of its phase. In that channel, the phase shift induced by the forest is almost negligible. In order to enhance the slope analysis and to highlight the contribution of the tilted ground, we can single out its contribution in the coherence. In other words, we can define a sort of residual coherence factor due to topography from the ratio between the estimated coherence and the coherence estimated forαa= 0 b γs(αa) = b γ (αa) b γ (0) (3.68)

wherein the dependence of the polarization has been omitted. As a complex number, b

γs(αa) gives the variation of the magnitude and phase of the complex coherence as consequence of the tilted terrain only. Its magnitude can be greater than one as it is not a decorrelation source but rather a residual factor. The curves associated with b

γs(αa) versus the azimuth slope αa are plotted in Fig. 3.20. Notably, the HV returns decorrelate in presence of 7% azimuth tilt of about 0.08, and their scattering phase center moves down about 2 m. Fortunately, this variation is well followed by the RVoG predictions as the dashed lines in the same plot confirm. Therefore, we conclude that predicted and observed coherence in presence of azimuth terrain slope show a very good agreement for both the magnitude and the phase. As expected, the RVoG inversion retrieves the canopy height and the underlying topography with accuracy better than 10% over the full range of tested azimuth slope (cf. Fig. 3.18b).

Varying the range terrain slope

A similar analysis has been conducted for range terrain slope, i.e. by varying the ground tilt between_{−7% and 7%, which corresponds to α}r' ±7 deg, and fixing hv= 15 m.

Positive values correspond to a surface faced to the radar antenna. The backscatter output is illustrated in Fig. 3.21 wherein the most evident effect is the decreasing of co-polar ground-volume return in presence of sloped terrain as consequence of the de- formation of the double-bounce geometry at base of trees. In general, the interpretation of range-sloped results is more critical due to the inherent imaging characteristics of the radar. As received samples are arranged on the basis of the range time delay, topogra- phy causes geometrical (and radiometric) distortions, usually classified as forshortening, layovering and shadowing phenomena in a single SAR image. The interferometer is also affected by the topography: as the terrain is tilted, the target is actually observed from look angles different that the nominal ones and this impact on the spectral shift between the two received signals. Hence to calculate the vertical wavenumber and the critical baselineB⊥,cr, the look angleθ must be replaced by the local incidence angle θ− αr, yielding (Bamler and Hartl, 1998)

kz= 4πB⊥ r0λ sin(θ− αr) , B⊥,cr= Wrr0tan(θ − αr) c (3.69)

whereWris the system range bandwidth (cf. Chapter 1). Whenαrincreases from zero, the vertical wavenumber increases and the critical baseline reduces, yielding greater sen- sitivity but lower correlation; for values ofαrclose toθ, i.e. in the region of blind angles, the sensitivity is completely loosed since the spectral shift exceed the range bandwidth; finally, whenαris greater than the blind angle, the radar works in the layover region. On the contrary, whenαris negative, the sensitivity is lower and the spatial correlation is higher (the spectral shift reduces); for higher negative angle, the surface is not illu- minated by the radar and the shadowing occurs. We are far away from such extreme cases in our analysis: the look angle is 45deg and the surface slope is αr ' ±7 deg. The associated spatial (i.e. baseline) decorrelation in the worst case isB⊥/B⊥,cr' 0.07 and hence it is still disregarded.

However, there is a second point that we consider. When the coherence is evaluated, the interferometric phase is proportional to the difference between the slant-range travel paths from the antenna to the target. This difference depends on the horizontal and vertical relative target location. When the interferogram is flattened including the hori- zontal shift introduced by the viewing geometry, the remaining coherence phase depends only on the vertical location of the target, which is a desiderated effect in our context. This is clearly shown in Fig. 3.15, where flattened interferogram exhibit a variation only due to the forest layer.

In the case ofα6= 0, the relative horizontal location of the target is modified by the terrain slope, i.e. the local fringe frequency of the terrain increases or decreases with respect to the flat Earth case. It follows that the flattened interferogram, obtained removing only flat Earth fringes, contains still a residual fringe pattern induced by

b

γs(αr) = b b

γ (0) (3.70)

have been plotted in Fig. 3.24.

Conversely to the case of azimuth slope, µ varies only in the HH and VV channels, decreasing of about 8 dB for sloped terrain (cf. Fig. 3.22a). The cross-polar scattering ratio is almost constant, with a loss of about 1 dB in the tilted terrain case. As expected, the coherence varies mainly for the HH and VV channels: while the VV coherence level seems to benefit of the range slope, the HH coherence level exhibits slightly lower values with respect to the zero-sloped case; the coherence phase, on the contrary, reveals that the HH and VV scattering phase centers rises further 5 m with as consequence of the topography. The comparison with the RVoG predictions is generally in good agreement with the observations, though some errors can be observed in the HH coherence magnitude and phase. These errors superimposes to the ones discussed in the height analysis and lead to less accurate estimates of the ground topography and vegetation height in presence of topography. We have found, using Scots pine simulations, that a good estimation of forest height within10% error is possible only if the terrain slope is comprised between_±2%.

Figure 3.13: Total and selective backscatter of a Pine forest obtained through numerical simulations using PSPSim. The near-range is the bottom of the images.

Figure 3.14: Ground-to-volume ratio of a Pine forest versus forest height obtained through numerical backscatter simulation (a). Performance of RVoG model inversion (b). The grey intensity in vertical bars is proportional to RMS error.

Figure 3.15: Coherence of a Pine forest obtained through PolInSAR processing of backscat- tering numerical simulations.

Figure 3.16: Coherence of a Pine forest obtained through PolInSAR processing of numerical simulations of complex backscatter (PSPSim).

Figure 3.17: Total and selective backscatter of a Pine forest obtained through numerical simulations using PSPSim. The near-range is the bottom of the images.

Figure 3.18: Ground-to-volume ratio of a Pine forest versus azimuth terrain slope obtained through numerical simulation (a). Performance of RVoG model inversion (b). The grey intensity in vertical bars is proportional to RMS error.

Figure 3.19: Total coherence of a Pine forest obtained through PolInSAR processing of backscattering numerical simulations. The near-range is the bottom of the images.

Figure 3.20: Residual coherence of a Pine forest obtained through PolInSAR processing of numerical simulations of complex backscatter (PSPSim).

Figure 3.21: Total and selective backscatter of a Pine forest obtained through numerical simulations using PSPSim. The near-range is the bottom of the images.

Figure 3.22: Ground-to-volume ratio of a Pine forest versus azimuth terrain slope obtained through numerical simulation (a). Performance of RVoG model inversion (b). The grey intensity in vertical bars is proportional to RMS error.

Figure 3.23: Total coherence of a Pine forest obtained through PolInSAR processing of backscattering numerical simulations. The near-range is the bottom of the images.