High-resolution vertical polarimetric imaging of pine forests

Authors


Abstract

This paper describes a field campaign performed with a ground-based polarimetric tomographic imager designed to measure the vertical response of several forest plots in order to better understand the mechanisms contributing to the total radar response. The campaign took place in the Mende forest, an artificial forest of European black pines, in the south of France. The objective is to perform a sensitivity study on the radar observables based on biophysical parameters. This paper presents a description of the experiment hardware and procedure, the results obtained for all sites, and a discussion of the results in the light of the ground truth. The results show the vertical distribution of the backscattered intensity for several parameters like tree density, forest growth, presence of undergrowth, and ground slope. The evolution of the energy backscattered by the ground, the volume, and the whole forest structure is also displayed. These results have shown the interest of such a campaign to improve the understanding of the distribution of radar mechanisms along the vertical axis.

1 Introduction

The retrieval of the biophysical parameters of forest with remote sensing is a challenge nowadays. In particular, the biomass of the canopy, the moisture content of the ground, and the branches are the three parameters of interest. It is well known that the low-frequency radars may provide information about some features of forests, and in particular, P band is often proposed for biomass estimation [Le Toan et al., 1992; Cloude and Papathanassiou, 1998]. Also, the detection of under the canopy targets takes advantage of the properties of penetration of the low-frequency waves like P band. The assessment of the algorithms of retrieval and detection benefits from a good understanding of the mechanisms involved in the forest scattering.

A lot of work has been done with high-resolution imaging in P band [Sandberg et al., 2009; Dubois-Fernandez et al., 2012] as well as in L band [Shimada et al., 2008]. However, many interrogations remain about the vertical distribution of scatterers versus the polarization, even with the polarimetric-interferometric techniques [Garestier et al., 2005] or the polarization coherence tomography [Cloude, 2006, 2007]. Airborne radar tomography is a technique which provides access to the vertical resolution and shows some promising results [Huang et al., 2011; Tebaldini et al., 2011]; however, the observed resolution is usually too coarse to provide a detailed vertical profile or to determine finely which mechanisms generate the forest scattering which is observed. So many questions remain about the components of the scattering in the forest, on their geometrical spread, and their contribution in the radar response.

To go further, a previous experiment was performed with the TropiScat experiment [Albinet et al., 2012], which goal was to study the temporal evolution of P band coherence and intensity over a land plot of tropical forest with vertical imaging from a tower top.

In this paper, a new experiment is presented similar to the TropiScat one but with an improved vertical resolution and conducted over a temperate forest. It is a crane-based mobile experiment. Although similar work has been done in the past on agricultural plots [Brown et al., 2003; Lopez-Sanchez et al., 2006], it is the first time, to our best knowledge, that very high resolution tomography is performed over several forest plots in P band. The main objectives of this experiment are the detailed study of the vertical distribution of the scatterers and their integrated values on some forest plots presenting varying ground properties and the highlight of the respective roles of the canopy volume and the underlying ground on the scattering in horizontal transmit and receive (HH), vertical transmit and receive (VV), and vertical transmit and horizontal receive (HV) polarizations.

First, in section 2, the experiment is introduced, and the tomographic reconstruction algorithm is described. Then, section 3 presents the validation of the correct operation of the radar hardware for the parameters of acquisition. Finally, the results of the ground campaign are shown in section 4, and a detailed analysis of these data is performed in section 5.

2 Experiment Description

2.1 Description of the Site

The Mende forest is a public forest area created during the nineteenth century and composed of European black pine, as shown in Figure 1. The forest, which is situated on a mountain plateau, is managed by the public organism Office National des Forêts (ONF), which goal is to maintain the forest, including cuts, clearings, etc. The ONF has also defined and characterized some forest plots, i.e., areas of statistically homogeneous forest, with their characteristics in terms of species, biomass, height, density, and ground slope.

Figure 1.

Photograph of the Mende forest taken from the top of a bucket truck.

2.2 Hardware

The radar unit is based on a stepped frequency vector network analyzer (VNA). In this experiment, the intermediate frequency bandwidth is set to 10 kHz, and the output power is set to +3 dBm. From the hardware specifications, the dynamic range is 50 dB.

The VNA is remotely controlled by a personal computer which initiates the test, collects, and saves the acquired data on its hard disk, as displayed in Figure 2. The H and V antennas radiate from the bucket of a bucket truck with a depression angle of 30°. Thus, HH polarization is obtained with S11, HV and VH with S12 and S21, and VV with S22. The VNA measures the response for the emission of a frequency ramp between 400 and 600 MHz and stores the results in the frequency domain.

Figure 2.

(left) Instrument architecture and (right) crane deployment.

The antennas are wideband (400–1000 MHz) log-periodic antennas. They are characterized by relatively similar radiation pattern in both E and H planes, with a measured half-power beamwidth of 60° in the E plane and 80° in the H plane. In addition, the isolation between polarizations is better than 20 dB, with very low sidelobes and backward radiation less than −23 dB.

The main requirements for this experiment, which are tightly linked to the antenna constraints, are achieved.

  1. Illuminating a footprint which is large enough to be statistically representative of the scene under consideration, while eliminating the spurious remote echoes.
  2. Eliminating as strong as possible the echo of the metal structure of the arm, while illuminating the forest with the required incidence angles.
  3. Presenting similar radiation patterns in both E and H planes.

2.3 Method of Acquisition

The measurements are performed sequentially from positions that are aligned and spaced identically. To ensure the correct alignment of these points, the motion of the bucket is performed through the sole sliding of the highest part of the arm. To reduce the coupling between the antennas and the metallic structure of the arm, the latter is inclined by roughly 15° relatively to the vertical. For the determination of the successive position of the antennas, an inclinometer provides the inclination of the arm, and the height of the antennas is measured with an accuracy of 5 mm.

The proper functioning of the system was verified with some impulse responses measured on an adjustable dihedral-trihedral reflector with a side length of 2 m. Previously, in [Albinet et al., 2012], the feasibility of such vertical imaging was demonstrated with full tomograms on such a reflector. It was shown that the compression was achieved and that the phase was preserved for a target embedded in the forest and for an acquisition of equivalent duration. Note that the following tomograms do not present absolute or polarimetric calibration, as no calibration with reference targets was carried out. Consequently, the phase information between different polarimetric channels is not available with such a measurement system.

With Δf and df which are, respectively, the frequency bandwidth and the frequency step (Δf  =  (Nf − 1) × df), then the unambiguous range is math formula, and the range resolution is math formula, as defined in Figure 3.

Figure 3.

Illustration of the range resolution dr and the cross range resolution dcr compared to the measurement points M and the target T.

The cross range resolution dcr depends on the center frequency f0, the aperture Az of the antennas, the distance R between the antenna and the target, and the distance y to the ground:

display math

The choice was made to work with a synthetic array made of 33 measurement points and with 15 cm spacing. As a consequence, the synthetic array is 4.80 m long, and the cross-range resolution ranges from 1.95 m for the closest target at height 20 m and range 10 m, to 5.50 m for the farthest one on the ground at range 80 m.

2.4 Tomographic Reconstruction

The tomographic reconstruction, which is strictly vertical synthetic aperture radar (SAR) imaging, was performed with the matched filter algorithm [Carrara et al., 1995; Jakowatz et al., 1996]. The response I of a pixel at position (z, y) distant from the antenna of Rp(z, y), with measurements S made on K frequencies with P antennas, is computed to reconstruct the 2-D tomographic image:

display math(1)

This method is a good trade-off between computational complexity and image quality [Jakowatz et al., 2004; Jakowatz and Doren, 2006].

Several corrections are taken into account during the tomographic reconstruction process: the propagation is corrected considering the observed scene as a collection of point targets along the azimuth, the antenna gain for emission Ge and for reception Gr is compensated, apodization with Hamming windows H(fk) and H(p) are applied, and the error in the horizontal orientation of the bucket, and thus of the antennas which induces cross talk, is corrected with the method presented by Quegan [1994].

2.5 Theoretical Validation of the Radar Configuration and the Tomographic Reconstruction With Simulated Data

In order to validate the choice of the parameters of the radar configuration, like the spacing between measurements or the number of measurements, and also in order to validate the tomographic reconstruction process, some raw data were simulated and used as input for tomographic reconstruction.

The simulated scene is composed of several point targets arbitrarily distributed in the scene. The radar configuration is as previously described. The complete frequency response is obtained for a distribution of diffracting points which represent the ground and the canopy. Then, the tomogram is reconstructed with equation (1) and is displayed in Figure 4.

Figure 4.

(left) Description of a 25 m high forest simulated with point targets uniformly distributed and (right) the tomogram obtained with the simulated raw data.

The tomogram shows that with the radar configuration that is proposed, it is possible to finely retrieve the vertical distribution of the scatterers and in particular, to isolate the backscattering from the canopy and from the ground in the case of a 25 m high forest. Note that the strong response in the top left part of the image is generated by the noise, which at this location is strongly amplified by the antenna radiation pattern correction. This artifact is not in the zone of interest and therefore can be ignored.

2.6 Influence of Internal Reflections in the Antennas

As previously explained, the HH and VV polarization data are acquired in reflection mode, i.e., with the same antenna for the emission and the reception. This induces spurious reflections inside the antenna, characterized by the standing wave ratio (SWR). The data corresponding to these spurious reflections (antenna in free space) may be approximated by orienting the antennas vertically toward the sky. To determine the consequences of this phenomenon on the quality of the tomograms, a tomographic reconstruction was computed with measurements made with the antennas facing the sky (Figure 5).

Figure 5.

Tomograms made from sky measurements to determine the effect of SWR in (left) HH and (right) VV polarizations. The black line is located at the ground and the gray ones at the highest possible height for the canopy.

One can see that in both HH and VV, the SWR induces artifacts in the tomograms, in the top part of the canopy, for a range larger than 50 m. This result shows that with an analysis performed from 20 m to 50 m in range, the SWR effects do not pollute the results. This is a side result of this study; with tomography, it is possible to rely on the S11 and S22 mode of acquisition, as the artifacts created by the SWR echoes are very local, and here they are located above the forest.

3 Field Campaigns

3.1 Imaged Forest Plots

The data were acquired during two campaigns. The first campaign took place on 24 and 25 November 2011 with two forest plots imaged and the second one from 9 to 11 May 2012 with eight forest plots imaged. The forest plots were selected to cover a large panel of forest configurations in terms of biomass, trees height, ground slope, homogeneity, and presence of undergrowth. Aerial photographs of their position in the forest are displayed in Figure 6.

Figure 6.

Location of the forest plots in (left) the northern part of the forest and on (right) the southern part.

3.2 Ground Truth and Biomass Estimation

The forest plots were characterized first in October 2011, during the airborne campaign [Angelliaume et al., 2011], which took place shortly before the first in situ radar campaign. They were characterized again in May 2012, during the second in situ radar campaign. For each forest plot, subplots are defined over which in situ data were collected. These measurements are used to compute biomass for all forest plots, based on equations given by Fung [1994] and Saleh et al. [2005]. Main ground measurement characteristics are summarized in Table 1. There is a clear correlation between the measured tree heights and the estimated biomasses, as displayed in Figure 7.

Table 1. Ground Truth and Biomass Estimation for All Forest Plots
ParameterP0P1NP1MP2P3P4P5P6P7P8
DateNovNovMayMayMayMayMayMayMayMay
Density (tree/ha)1800666499102563355067567518222857
Average height (m)625.525.5162426.5272767.7
Trunks mean diameters (cm)937.531.827.435.232.639391412.6
Ground slopeFlatFlatFlatFlatDownslopeUpslopeDownslopeUpslopeFlatFlat
Soil moisture contentWetWet33%≈30%31%28%32%≈30%27%25%
UndergrowthNONONONONONOYESYESNOYES
Biomass (ton/ha)132601401442061682992993349
Figure 7.

Biomass estimated as a function of measured height during the field campaign.

Note that the ground was saturated during the first campaign, and the moisture contents were estimated to be around 30% during the second campaign.

3.3 Data Processing

The tomograms were computed for all forest plots and polarizations with a pixel size of 1 m, and the images were rotated so as to have a flat ground. Figure 8 shows the example of an inclined parcel for all polarizations and the corresponding horizontal image with apodization. The tomogram or its 2-D representation provides detailed information in terms of distribution of scatterers as a function of height and range and thereby incidence angle.

Figure 8.

Vertical tomogram for (left) HH, (middle) HV, and (right) VV reconstructed with the measurements performed on P1M. (top) Initial images and (bottom) images with slope correction and apodization. The black line corresponds to the soil location, and the gray lines correspond to the measured minimum and maximum height of the canopy.

An average vertical profile can be computed from these tomograms, by summing the intensity along all horizontal lines of pixels, between 20 and 50 m of range for all polarizations. The result obtained on one forest plot is displayed in Figure 9, and the corresponding analysis is done in section 5.

Figure 9.

Polarimetric backscattering profiles for P1M. The black solid line corresponds to the soil location, and the gray solid lines correspond to the minimum and maximum measured height of the canopy. The dotted lines show the vertical limits of the integration for the computation of backscattered energy by the volume and by the soil, respectively, in gray and black.

The energy backscattered by the ground is computed as the integral of the vertical backscattering profiles on a horizontal layer situated between −3 m and 3 m high for all forest plots. The volume contribution is also the integral of the vertical backscattering profiles but computed between the bottom local minimum of the volume response and its top local minimum. These high and low limits vary for each scene. For forest plots with low biomass, the volume response is overlapping with the ground response and cannot be separated, due to the limited cross-range resolution in the vertical plane. This is the case for P0, P7, and P8.

4 Data Analysis

The following analysis is performed on the 10 tomograms acquired during the experiment, each one corresponding to a different forest plot, including a forest plot before and after clearing.

4.1 Vertical Profiles Analysis

The goal of high-resolution tomography is to separate the scatterer response along the vertical axis. In Figure 9, we can see that the volume backscattering can be separated from the ground one. The ground response is localized at the ground reference level for all polarizations. Thus, a first conclusion is that the underground is not visible. The cross polarization backscattered by the ground is higher than what is expected from the models [Villard and Borderies, 2007]. Its localization at the ground level is also an important result. For the volume, the response is characterized by a high backscattering of the top of the canopy and by a decrease which is linked to a clear attenuation of the electromagnetic wave in the canopy. A follow-on study will explore the estimation of the attenuation as a function of height, based on the approach introduced by Lopez-Sanchez et al. [2006]. The resulting vertical attenuation profiles could then be used for the forest height retrieval with the polarimetric interferometric SAR method [Treuhaft and Siqueira, 2000].

In Figure 10, the vertical profile associated with a short forest, here 6 m high, has no local minimum, and as a result, the volume and the ground contributions cannot be separated. However, a local maximum is visible at a 5 m height for HV only.

Figure 10.

Polarimetric backscattering profiles for P7. The black solid line corresponds to the soil location, and the gray solid lines correspond to the measured height of the canopy.

It is possible to identify the peak corresponding to the ground response and the strong decrease of the canopy backscatter from top to bottom in all three polarizations for the high forests. For the short forests, this can also be observed. In this way, an estimated height may be derived on most forest plots. The corresponding results are displayed in Figure 11.

Figure 11.

Forest heights measured and forest heights estimated with tomography.

There is a very good agreement between the forest heights from the in situ measurements and those estimated by the polarimetric tomography. This establishes that the tomographic technique can be applied to retrieve the forest height with good accuracy if the vertical resolution is fine enough, for all polarizations. For the biomass retrieval, as previously shown in Figure 7, plots of forest with 25 m high trees can be characterized by biomasses from roughly 150 to 300 ton/ha. Thus, the information of forest height is not sufficient to retrieve the biomass, as discussed for example by Caicoya et al. [2010]. Other observables like the multipolarization intensity should be taken into account and embedded in a more global inversion process.

4.2 Intensity Analysis

4.2.1 Global Results

As previously explained, the relative intensities backscattered by the volume and by the ground were derived separately from the vertical profiles. These results are displayed for all forest plots in Figure 12. Again, for the first time, to our best knowledge, volume and ground contributions can be separated accurately.

Figure 12.

(left and right) Polarimetric relative energy backscattered by the volume and the soil, respectively, for all sites.

The behavior of the volume backscattering is similar for all forest plots: HH is the highest, HV is on the order of 6 dB lower, and VV is about the average of HH and HV. This peculiar polarization distribution can be explained by the fact that the volume is an oriented medium. Similar results showing clear differences between polarizations were obtained with indoor experiments [Cloude et al., 1999]. In fact, for European black pines, the branches are mainly horizontal, and the only vertical parts which are visible are the high part of the trunks. For modeling purposes, at P band, the canopy is generally considered as a collection of dielectric cylinders of finite length. The effect of such a geometric distribution on the relative levels of the polarization intensities is evident, since the results are far from those of a uniformly random distribution of dielectric cylinders, for which HH and VV would be of the same order and HV on the order of 6 dB lower. One may think about such a property, demonstrated here experimentally, to derive some volume structure features [Borderies and Villard, 2010].

For the ground, the intensity backscattered in all the polarizations varies greatly, more so than what is observed for the volume. Actually, the energy backscattered at the ground level comes from direct backscattering and double bounce and can be affected by slopes, presence of undergrowth, branches and trunks on the ground, and by the propagation in the volume, which makes the interpretation of the results rather complex. Nevertheless, one can note that HH and VV signals are relatively close in most cases and that there is no dominance of one over the other. The results confirm that a significant backscattering in HV is generated at the ground level, even if it can be attributed to various reasons as will be seen in the next section.

In addition, the relative total energy backscattered by the whole scene and the ratio of the volume contribution over the total contribution were computed and are shown in Figure 13.

Figure 13.

(left) Polarimetric energy backscattered by the total scene and (right) the share in percentage of the volume energy upon the total response.

The previous components are added so that in general, the total backscattering follows the trend of the volume, even if the share of the contribution originated by the volume varies in the total response. The volume represents from 20 to 80% of the full response, which means that the ground contribution is always significant.

4.2.2 Detailed Analysis

It is interesting to compare the forest plot P1, in November 2011 (P1N) and in May 2012 (P1M), after the clearing. For most forests, at P band, the variation of intensity as a function of biomass presents saturation for values higher than 200 ton/ha [Rosenqvist et al., 1999]. When comparing the spot before and after clearing, the values for the volume remain of the same order of magnitude, which makes sense, since both scenes are in the saturation domain.

The ground contribution of P1M behaves as a random volume made of cylindrical scatterers, with HH and VV with similar values and HV which is roughly 6 dB lower. This may be explained by the fact that, after the clearing, a large number of branches were abandoned on the ground. The ground before clearing behaves quite differently: HH and HV present as expected, whereas VV is much lower. Indeed, double bounce is strongly attenuated due to a strong attenuation in VV originated by the large and dense vertical trunks [Borderies and Villard, 2010]. Finally, it is to note that it is difficult to quantify which part of these effects is related to seasonal variations.

P1M and P2 are two different plots characterized by a similar biomass level, with very different structure. The first one is composed of high trees (25.5 m), while the second one has an average tree height of 16 m, with a tree density twice as big as for the P1M plot—500 trees/ha for P1M and 1000 trees/ha for P2. The volume contribution is larger for P2 than for P1M; this can be observed with all polarizations, while the ground contribution decreases, also with all polarizations. One can see that here, with a constant biomass, the increase of the trees density has a higher influence than the decrease of the height on the variation of both the volume and the ground contributions. This is potentially due to a different distribution of the branches with the age of the forest.

If we now compare forest plots P3 and P4, which are forest plots with, respectively, downslope and upslope and P5 and P6 with a similar downslope and upslope configuration but with significant undergrowth, we can see that their volume responses are similar. Concerning the ground response, the influence of the slope can be clearly observed between P3 and P4, for which the ground response is much lower in HH and VV. This is certainly due to the effect of the slope of the ground on the double bounce which is located at the ground level. The presence of undergrowth in P5 and P6 tends to reduce the ground response and its variation with slope, making it similar to that of a medium of dielectric cylinders with a uniformly random orientation like for P1M. This phenomenon, that is also visible on the total response, is due here to the undergrowth composed of small trees that have many branches and thin trunks. So it is interesting to observe that the strong effects of slopes are cancelled by the presence of undergrowth on the ground.

P0, P7, and P8 are forest plots composed of small trees. These trees are similar to those found in the undergrowth of P5 and P6. For these forest plots, the total polarimetric backscattering behaves again similarly to the response of a medium with a uniform random distribution of cylinders. P0 is farther away because it is very heterogeneous, with holes in it.

5 Conclusion and Future Prospects

A high-resolution polarimetric tomographic imager was designed; the processing chain was set up, and the whole system was validated. Tomograms with a metric resolution were reconstructed, and the corresponding vertical profiles and backscattering contributions were deduced. Thereby, the feasibility of such a high-resolution vertical imaging has been demonstrated, and the interest of the resulting vertical images was clearly exhibited. The findings provide new insights in the understanding of the scattering mechanisms in the forests.

The results which have been obtained exhibit interesting features, keeping in mind that to make them more general, a greater number of measurements would be necessary. The most significant result is that the tomographic process provides a profile of the forest from the ground to the forest top, allowing a precise estimation of the forest height, and separating clearly the volume scattering and the ground one (direct ground added to double bounce). In addition, relevant observations could be done with the corresponding physical interpretation. The ground backscattering is localized at the ground level. This observation discards any significant underground contribution, low branches contribution or volume multiple scattering in it, because their backscattering would appear above or below the ground level. The ground contribution presents high values, even for HV polarization, which was generally attributed to the presence of cut branches on the ground or of undergrowth. On the spots treated here, the slope of the ground was observed to have a strong influence which becomes masked when undergrowth is present. It was also shown that the analysis of the backscattered intensity in HH, VV, and HV over volume and ground gives some insight on the structure of the forest under consideration.

To conclude, it was shown that this technique allows a detailed measurement of the ground and volume backscattering, of the forest height, of the volume/ground ratio, and possibly the presence of undergrowth. Even if it was not done in this study, the canopy extinction could be retrieved with the inversion of the profiles of the volume backscattering. The obtained vertical resolution allows an easy analysis of the results.

In terms of relation with models, this separation of the different contributions leads to a strong improvement, since it allows a comparison of the relative contributions more insightful than a simple analysis of the total backscattering. It opens the path to improving the models by modeling some of the various effects which have been observed and which up to now are not always accounted for, like for example those related to the slope, the presence of the undergrowth and a cross-polarization component on the ground.

In the future, the radar will be upgraded with a fully automatic and faster instrument; it will be then possible to acquire several tomograms per forest plot and to image a significant number of forest plots and possibly explore 3-D imaging. In addition, the future radar will be able to provide fully calibrated polarimetric data and to consider not only the vertical profiles of the backscatter power for the three polarimetric channels but also for linear decompositions (Pauli channels), polarimetric decompositions (Cloude-Pottier, Freeman-Durden, or Yamagushi), and other observables (phase difference between HH and VV). These results may complement the analysis already carried out in this paper with new interpretations and comfort existing ones.

Acknowledgments

The authors would like to thank the Office National d'Etudes et de Recherches Aérospatiales (ONERA) for having founded this study with the Federating Research Project ENVIRO, an internal project at ONERA which two major objectives are to demonstrate the ability to provide maps for estimating ground moisture and biomass estimation from the fusion of radar and optical data at metric resolution. They particularly want to thank Emmanuel Rosencher, from ONERA, for his support to the project and the ONF for having provided some precious field measurements.

Ancillary