Continental forests play a key role in the carbon cycle on Earth, particularly in tropical areas. Global and regular monitoring of these forests is necessary to better understand their impact and quantify the carbon stocks and exchanges with the atmosphere. The ESA-Biomass mission, to be launched in 2021-2022 time frame, has been selected to address this problematic [1]: satellite measurements with Synthetic Aperture Radar (SAR) system and a revisit time of 3-4 days will provide information related to vegetation characteristics and can be used to monitor the evolution of the forestry scenes (seasonality, degradation, reforestation …). The use of P-band for the SAR acquisitions has been chosen to ensure a good penetration through the canopy cover. A tomographic acquisition mode constituted by full-polarimetric repeat-pass interferometric orbits will permit the application of Tomography SAR techniques: from the slightly different orbits above an identical scene, Tomography SAR gives access to the 3D structure of the distributed media and has shown interesting capacities for the estimation of vegetation parameters such as ground elevation, canopy height and biomass [2, 3, 4].

To support the future Biomass mission, airborne campaigns are specifically designed to collect data in various contexts. It is the case of the ESA-sponsored AfriSAR campaign, which took place in Gabon (Africa) in 2015-2016, where ONERA conducted airborne acquisitions with its SETHI system in July 2015 [5]. Among the objectives for ESA, this campaign was constructed in order to test Biomass geophysical inversion model along with the assessment of P-band imaging for forests monitoring in this context of dense tropical forest where the biomass can reach 600t/ha. This ONERA campaign was also part of an on-going project with Total, in which one of the objectives was to analyze the potential of P-band SAR to retrieve undercanopy DEM and forest structure in order to understand the geomorphology of terrains and for traficability purpose.

Several acquisitions have been done by ONERA during this campaign, including a multi-baseline stack constituted by 10 parallel flights in P-band above the Lopé National Park in central Gabon. In this presentation, we focus on the processing of this data stack for the application of Tomography SAR techniques for the analysis of forest height and undercanopy topography. One important step is to compensate the phase disturbances that are present in the data stack due to the insufficient accuracy of the motion unit system on-board and are creating strong artefacts and defocalisation in tomographic profiles. In this work, this step is done with the Phase Center Double Localization method applied to the specifity of our data [6, 7]. Tomographic analysis is then applied on the corrected multi-baseline data stack, and simple and classic estimation techniques permit the extraction of the ground elevation and canopy height on the whole Lopé site from the HH and HV tomographic profiles. The quality of these estimations is assessed in comparison with the Lidar data available on the site, which generally lead to good results (even if non negligible biases can be observed in regions of strong slopes or at forest borders). However, if the results look promising in the context of airborne acquisitions, one essential question to address is the evolution of the performance of such tomographic analysis for the Biomass spaceborne configuration. A first study in terms of resolution and size of the analyzing window is conducted in this paper.

[1]: T. Le Toan, S. Quegan, M. W. J. Davidson, H. Balzter, Ph. Paillou, K. P. Papathanassiou, S. Plummer, F. Rocca, S. Saatchi, H. Shugart and L. Ulander, “The BIOMASS mission: Mapping global forest biomass to better understand the terrestrial carbon cycle,” Remote Sens. Environ., vol. 115, no. 11, pp. 2850–2860, Nov. 2011.

[2]: A. Reigber and A. Moreira, “First demonstration of airborne SAR tomography using multibaseline L-band data,” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 5, pp. 2142–2152, Sep. 2000.

[3]: Tebaldini, “Algebraic synthesis of forest scenarios from multibaseline PolInSAR data,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 12, pp. 4132–4142, Dec. 2009.

[4]: D. Ho Tong Minh, T. Le Toan, F. Rocca, S. Tebaldini, L. Villard, M. Réjou-Méchain, O. L. Phillips, T. R. Feldpausch, P. C. Dubois-Fernandez, K. Scipal and J. Chave, “SAR tomography for the retrieval of forest biomass and height: Cross-validation at two tropical forests sites in French Guiana,” Remote Sens. Environ., vol. 175, no. Supplement C, pp. 138–147, Mar. 2016.

[5]: P. C. Dubois-Fernandez, X. Dupuis, P. Capdessus and R. Baqué, “Preliminary results of the AfriSAR campaign,” in Proc. EUSAR 2016: 11th European Conference on Synthetic Aperture Radar, pp. 1–3, Jun. 2016.

[6]: S. Tebaldini, F. Rocca, M. Mariotti d’Alessandro and L. Ferro-Famil, “Phase Calibration of Airborne Tomographic SAR Data via Phase Center Double Localization,” IEEE Trans. Geosci. Remote Sens., vol. 54, no. 3, pp. 1775–1792, Mar. 2016.

[7]: V. Wasik, P. C. Dubois-Fernandez, C. Taillandier and S. S. Saatchi, "The AfriSAR Campaign: Tomographic Analysis With Phase-Screen Correction for P-Band Acquisitions," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 11, no. 10, pp. 3492-3504, Oct. 2018.

Acknowlegments: The ONERA AfriSAR campaign was supported by ESA, CNES and Total.

Area A - Board 72

P-band tomographic SAR (TomoSAR) data represent the combination of two trends in radar remote sensing of forest: Combining multiple acquisitions to obtain more information about forest structure and using longer wavelengths within the UHF-band to increase both the penetration into the canopy volume and the selective sensitivity to the woody elements comprising most of the forest above ground biomass (AGB). Evaluating AGB retrieval from P-band TomoSAR data is highly relevant with regard to ESA’s upcoming BIOMASS mission which will be the first spaceborne P-band SAR. BIOMASS has the specific goal of providing global AGB estimates and will have a repeat pass TomoSAR acquisition mode.

Other work has shown that P-band tomography can improve retrievals in tropical forest, with the backscatter at 30 m above ground being highly correlated with AGB. For boreal forest it has been shown that suppressing ground level contributions to the signal by removing the backscatter power originating from below 10 m above ground simplifies the dependence on AGB at L-band. Modeling the remaining volumetric TomoSAR intensity with a power law results in significantly improved retrievals compared to those obtained using SAR gamma nought values, independent of which polarization channel is used.

This work applies the same approach of removing ground level backscatter to P-band TomoSAR data acquired over boreal forest. An evaluation is done for airborne and simulated spaceborne data originating in the BioSAR 2 campaign data set from the Krycklan test site in northern Sweden, with reference values for model training taken from a lidar-based AGB map and in-situ measurements from 27 forest stands used for validation. It is found that for P-band TomoSAR volumetric intensity gives improved AGB retrievals compared to ground intensity in all cases where the vertical resolution is high enough to make a meaningful distinction. Volumetric TomoSAR HV intensity has smaller errors than SAR HV gamma nought in the best performing cases.

A comparison with results obtained from data based on the BioSAR 1 campaign in Remningstorp in the south of Sweden provides insight in the importance of ground topography. The terrain in Remningstorp is relatively flat, with maximum median ground slopes of 5° for the 58 forest stands providing reference AGB values compared to 15° for the 27 stands in Krycklan. The smaller vertical extent of the scene also enabled larger tomographic baselines, resulting in a vertical resolution spanning 16 m - 42 m for the in-situ stands compared to 26 m - 101 m for Krycklan.

AGB retrieval performance for Remningstorp P-band TomoSAR ground intensity and SAR gamma nought are similar, with worse results for volumetric intensity. In contrast to Krycklan the HH polarization is found to have smaller errors than HV. These observations suggest that the relatively flat terrain does not significantly affect the AGB dependence of the strong ground forest interactions seen in boreal forest at P-band. This leads to the conclusion that P-band TomoSAR with resolution on the order of forest height can help compensate for the adverse effects of large ground slopes such as found in Krycklan. However, there is little benefit with regard to AGB estimation for sites with more limited topography such as Remningstorp.

Area A - Board 73

This work addresses the retrieval of forest height in tropical areas based on multi-baseline data acquired at P-Band. Two different approaches are compared. In the first case, forest height retrieval is cast in term of a parametric estimation problem, under the assumption of the Random Volume over Ground (RVoG) model. This approach is conceptually similar to the one retained in PolInSAR literature, although the implementation in this paper is designed to work on single-polarization data as well, under the fundamental assumption that multi-baseline data are available. Interestingly, forest height can be inverted by this approach even without assuming the existence of a ground-free polarimetric channel, which is often not the case at P-Band, and wave extinction does not need to be set as a prior. The second approach here proposed is based on SAR Tomography. In this case, tomographic imaging is used to produce vertical backscattering profiles by jointly focusing data from multiple flights. Forest height is not directly produced by SAR Tomography, but it can be obtained by analyzing the shape of the vertical profiles. This approach is conceptually more similar to how forest height is retrieved from Lidar measurements, where it is computed by setting a threshold on Lidar returns according to some percentile. A drawback of this approach with respect to parametric inversion is that it requires the user to set a threshold for the extraction of forest height, whereas an advantage over parametric estimation is represented by its computational speed. These two approaches are here compared based on the data from the ESA campaigns TropiSAR, flown in French Guiana in 2009, and AfriSAR, flown in Gabon in 2015. Validation is carried out against reference forest height data from Lidar measurements. Results indicate that height retrieval by parametric estimation is in good agreement with Lidar height for various types of forests, resulting in no significant bias over the interval of forest height between 20 m and 50 m. For tomography, it was observed that the best results are obtained by adapting the user-defined threshold used for height retrieval to the different forest types. This approach was also observed to overestimate forest height in case of short vegetation, meaning that the optimal threshold should also depend on height. In contrast, forest height by tomography is observed to produce a lower dispersion at all forest sites investigated in this paper. Accordingly, our conclusion so far is that a strong complementarity exists between these two methods, suggesting the need for a combined estimator.

Area A - Board 74

Although tropical forests are largely recognized to play a major role in the global carbon cycle, our current knowledge about the distribution of forest above-ground biomass (AGB) at continental scales is not sufficient to reduce the major uncertainties in calculations of carbon stocks and fluxes associated with the terrestrial biosphere. To fill this gap, the future P-Band SAR Mission BIOMASS will be launched in 2021/22, with the primary objective to measure AGB worldwide to within 20% accuracy. After the commissioning phase, BIOMASS will be operated in tomographic mode for a period of 14 months. During this phase, commonly referred to as Tomographic phase, BIOMASS will acquire stacks of 7 acquisitions with a revisit time of 3 days, which will be processed through tomographic techniques to generate a 3D representation of the forest structure. A most relevant result in this context is the one produced by Ho Tong Minh et al., based on airborne data acquired over French Guiana, in which it was shown that the use of tomographic intensities brings a dramatic improvement of AGB retrieval accuracy as compared to traditional 2D SAR data. This finding immediately triggered the question whether this result could be replicated based on spaceborne acquisitions, which will be unavoidably affected by temporal decorrelation. This question was first addressed in a series of study based on ground-based data from the TropiSCAT campaign, which indicated a limited impact of temporal decorrelation on tomographic data. In this paper, we aim at adding new elements to the discussion of temporal decorrelation by analyzing the multi-temporal airborne data-set acquired by DLR at the forest site of Lopé, Gabon, during the AfriSAR campaign. The data-set consists of 16 SLC images, of which 10 were acquired on the second day, while each 3 were acquired on the other 2 days. The acquisition plan was not designed to check the temporal stability of tomographic products but rather interferometric ones, therefore it is not possible to do the same analysis as on TropiSCAT data. To tackle this problem, we use a recently proposed method referred to as Interferometric Ground Notching. The method takes as input a single pair of SAR images to obtain a ground-notched image by canceling out the backscattered power from the ground level. The methods operates coherently on the data, which enables a direct measurement of the temporal behavior of forest canopies by evaluating the interferometric coherence between ground notched images. Further analysis are carried out to study the impact of temporal decorrelation on ground-notched images, so as to evaluate the applicability of this technique in the context of BIOMASS.

Area A - Board 75

This paper describes a geometrically rigorous approach generating a spacebore SAR tomographic data-stack starting from an airborne data-stack. This simulation approach is explicitly designed to handle the differences in acquisition geometry between spaceborne and airborne surveys. Indeed, range and incidence angle exhibit large variations within the imaged swath in the case of airborne surveys. As a result, the vertical resolution of tomographic airborne data usually varies dramatically from near to far range. This is especially true in the case where tomographic acquisitions are gathered by flying horizontal baselines, where vertical resolution can vary by a factor 3 or 4. Moreover, aircraft trajectories can deviate from ideal straight lines by several meters (depending on aircraft control), which causes vertical resolution to vary in azimuth as well. In contrast, in spaceborne surveys range and incidence angle are nearly constant within the imaged swath, and flight trajectories are bound by orbital mechanisms, resulting in the possibility to perform tomographic imaging over large areas with a negligible variation in vertical resolution. This is indeed an advantage of spaceborne tomography over airborne tomography, especially concerning the radiometric accuracy of tomographic products. Another relevant difference is linked to the orientation of the tomographic resolution cell in the (ground range, height) plane, which varies in the airborne case and stays constant in the spaceborne case. For a small bandwidth system like BIOMASS, this causes vertical resolution at steep incidence angles to be limited not only by baseline aperture, but also by pulse bandwidth. Again, this condition cannot be directly reproduced in airborne data, due to incidence angle variation from near to far range. The simulation approach we followed to simulate spaceborne data consists essentially in “adapting” the spectral content of airborne data to fit that of spaceborne data, which is carried out though a peculiar technique we refer to as “wavenumber warping”. The proposed approach is here validated against synthetic airborne and spaceborne data. As an example of application, the simulator is used to assess the sensitivity to BIOMASS tomographic data to Above Ground Biomass (AGB) in tropical forests. The assessment is carried out by simulating BIOMASS data based on the P-Band data acquired by ONERA at the forest site of Paracou, French Guiana, in the context of the TropiSAR campaign, and by validating the results against reference AGB values available coincident from in-situ surveys.

Area A - Board 76

The aim of this talk is to discuss the technical challenges and potentials associated with three-dimensional tomographic imaging of natural media by spaceborne Synthetic Aperture Radar (SAR) systems, with a special attention to P- and L-Band. A tomographic acquisition is obtained by collecting the Radar echoes backscattered from the illuminated targets as the Radar system is flown along multiple trajectories. This particular configuration allows the formation of a two-dimensional synthetic aperture, which results in the capability to resolve the targets in three dimensions. Spatial resolution in the direction of azimuth and range is the same as in the case of conventional SAR systems, whereas resolution in the direction of elevation depends on the number of trajectories and their relative positions. There are two fundamental requirements for this approach to work. The first one is that the signal backscattered from the illuminated scene remains coherent from one pass to another. While this assumption can be retained for urban scenario, the case of natural media needs to be treated carefully. Radar backscatter from forested areas at P-Band is mostly determined by trunks and larger branches, which can be at least in principle supposed to remain stable over time scales of days or even weeks. However, this is generally not the case at L-Band band or higher frequencies over natural scenes. A possible way out of this blocking circumstance is provided by assuming a bistatic SAR system, as in the case of Tandem-X. Indeed, bistatic systems allow for an instantaneous measurement of the interferometric correlation between the mono- and the bistatic acquisitions, which corresponds to a particular vertical wavenumber of the imaged scene depending on the interferometric baseline. By collecting multiple pairs with varying across-track baseline it is then possible to get multiple wavenumbers, which allows to resolve in targets in elevation virtually in the absence of temporal decorrelation. The other requirement for the feasibility of tomographic imaging is that the knowledge of sensor position and propagation conditions is accurate enough to predict variations of the distance travelled by the wave to within an accuracy much better than the system wavelength. This condition is never met in the context of a spaceborne mission, due to errors about orbital knowledge, tropospheric and ionospheric propagation, and clock synchronization errors in bistatic systems. For this reason the implementation of a spaceborne tomographic system is intrinsically related to the development of an accurate phase calibration scheme.
Throughout this talk, we will present and discuss relevant results concerning the themes outlined above based on studies related to the upcoming P-Band Mission BIOMASS and the proposed L-Band bistatic missions SAOCOM-CS and PARSIFAL. The overall processing schemes developed for both frequency bands will be outlined, and final performance discussed based on theoretical developments and end-to-end simulations of spaceborne tomographic acquisitions derived from campaign data.

Area A - Board 77

Synthetic Aperture Radar (SAR) represents an attractive way to rapidly map wide areas and, when installed on spaceborne platforms, in a systematic way. SAR systems operated at larger wavelength (typically 30cm at P band) may be used to characterize dense forests, due to the penetration of electromagnetic signals into the forest cover, and to their interactions with various kinds of scattering bodies within the cover.
Nevertheless, characterizing tropical forest using classical 2-D reflectivity images may reveal, in practice, limited by several factors, such as a saturation, i.e. bad conditioning, of the relationship between SAR reflectivity and forest biomass, a lack of knowledge of the local geometry, i.e. topography and three height, which may heavily modulate the intensity of SAR responses, varying ground humidity conditions… Interferometric SAR data processing, and its extension to the case of polarimetric signals, may help to overcome some of the aforementioned limitations, but its use to characterize forest covers still relies on several assumptions, mainly related numerous and varied contributions originating for different parts of the observed volumetric medium, and falling within a resolution cell. SAR tomography uses several interferometric acquisitions to produce a 3-D image of the observed, which discriminates scattering contribution in the elevation direction too, with a resolution steered by the acquisition geometry.
A series of recent studies demonstrated that P band Tomographic SAR (TomSAR) processing could be used to estimate the Above Ground Biomass (AGB) of tropical forests in a reliable way, from a measure of their 3-D electromagnetic reflectivity sampled at a given height.
This paper presents a refined analysis of a tropical forest, aiming to estimate its AGB, tree height and underlying ground topography, using polarimetric tomographic SAR data acquired at P band. The performance of the proposed techniques is compared to those of existing approaches, and an evaluation is led using airborne SAR data acquired by the ONERA SETHI sensor in French Guyana. Finally, the behavior of the characterization techniques proposed here is evaluated over data sets with much coarser spatial resolutions, simulating data that will be acquired in the frame of the future ESA P-band SAR mission BIOMASS.

Area A - Board 78

This paper presents a light-weight and low-cost Wide-Band (WB) MIMO radar system, entirely developed at the IETR laboratory, University of Rennes, France, that may be easily and rapidly deployed over remote natural environments, in order to produce high-resolution 3-D images. This device coherently processes WB signals acquired along a large virtual antenna array in order to produce 3-D electromagnetic images, with a spatial resolution of a few centimeters in range, and of a few tens of centimeters in azimuth and elevation at short range.

The system architecture consists of a low-noise oscillator coupled with a DDS which deliver high-quality signals, over low PRF frequency sweeps, that are fed into a WB PLL in order to synthesize C-band FM Continuous Waveforms (FMCW). The output RF signal spectral domain lies between 4.6 GHz and 6.4 GHz, corresponding to a range resolution of 8.15 cm. The antenna system consists of 6 Tx and 6 Rx individual antennas, arranged as linear non uniform Tx and Rx arrays. Each antenna was realized at the IETR, and consists of a folded dipole, with 24 deg and 52 deg -3 dB apertures, in azimuth and elevation, respectively, with a gain of 13 dBi, and frequency band 4.5 GHz - 8.2 GHz. It is important for the targeted application to note that each antenna only weights 100 g and its realization is extremely cheap.
The radar behaves as a MIMO system, with all Tx and Rx antennas operating at the same time, i.e. without classical TDMA multiplexing that would seriously slow down the acquisition process. The orthogonality between waveforms is preserved through slight spectral shifts between Tx signals, implemented using low-cost hardware. The decoding of the multiple waveforms is realized through simple software processing, after the simultaneous acquisition of the multiple Rx channels.

This system covers a wide range of applications and may be used as a tool able to capture the very-local 3-D electromagnetic “ground truth” of Sentinel 1 acquisitions.
In particular, it has been used to measure complex seasonal Alpine snow packs, containing large grain and partially wet layers, to image in 3-D agricultural vegetation layers, and to monitor forest canopies in order to quantitatively characterize scattering mechanisms.

Area A - Board 79

3D SAR Tomography (Tomo-SAR) is an emerged experimental multibaseline mode, stemming from the mature phase-only Interferometry and framed among techniques of coherent combination of complex (amplitude and phase) SAR data; it allows full 3D imaging through elevation beam-forming, i.e. spatial (baseline) spectral estimation. This is producing applications, some going to be operational, of information extraction in complex volumetric or layover scenes, like derivation of forest height/biomass, sub-canopy topography, and heights/reflectivities in dense urban areas.
Another coherent data combination mode termed Differential SAR Tomography (Diff-Tomo) has been originated at University of Pisa (UniPi), integrating advanced D-InSAR and Tomo-SAR concepts, deeply exploiting multipass multibaseline data, to allow “opening” the SAR cell in complex dynamic scenes (UniPi patent). Based on proper (sparse) space (baseline)-time spectral analysis, it allows joint resolution of multiple elevation and deformation motion components of scatterers mapped in a same SAR cell, i.e. 4D (3D+Time) or higher order imaging, with successfull application e.g. to subsidence monitoring of dense layover urban areas and critical infrastructures. Even more generally, Diff-Tomo can identify scattering components distributions in the domain of spatial (height) and temporal frequency of harmonics in which the signal from scatterers with temporal decorrelation can be decomposed; this both allows avoiding their misinterpretation in Tomo-SAR processing, and enables new 3D characterizations of such dynamical processes.
In this work, an overview of UniPi research and results and recent advances are presented of these Diff-Tomo 4D/5D SAR imaging techniques, with particular reference to information extraction in complex distributed (volumetric) dynamic (decorrelating or deformating) scenarios.
In particular, a portfolio of three different 4D/5D Diff-Tomo algorithms is illustrated, spanning from more flexible semi-parametric to matched fully model-based, tailored to various enabled information extraction functionalities, ranging from 3D Tomography robust to temporal decorrelation, through elevation separation of different temporal decorrelation mechanisms i.e. decorrelation profiling, to deformation monitoring of volume-buried scatterers (or extraction of temporal trends of collective phase shifts).
Specifically regarding the algorithmic possibilities , after recalling the generalized-MUSIC Diff-Tomo, and an efficient subspace-based algorithm, the semi-parametric adaptive generalized-Capon Diff-Tomo method, originally conceived by Unipi in 2013, is formalized here and reported together with a series of considerations and insights. Generalized-Capon Diff-Tomo is sidelobe suppressing and superresolution yet non-parametric in height, to get flexibility for distributed scatterers applications, still exploiting the higher-order processing incorporation of a spectral model of the (long-term) decorrelation in the temporal frequency domain, as the other two algorithms do. A simulated comparison testing pros and cons of the different methods, and their degrees of suitability for the different advanced dynamic information extraction functionalities, is also presented.
Several extensive experimental results from airborne P-band E-SAR boreal forest data will be also shown and commented, for the three Diff-Tomo algorithms and the different functionalities (decorrelation-robust Tomography, decorrelation profiling, volumetric monitoring), to prove and illustrate the capabilities of these advanced Diff-Tomo methods, and the related potential for new SAR information products.
The described enabling Diff-Tomo tools may be fruitfully further developed and tailored in particular for possible application in the framework of the incoming or possible missions Biomass, TanDEM-L, NISAR; perspectives for the illustrated and possible other future applications are hinted, also within specific context of these missions.

Area A - Board 80

New technologies are urgently needed to observe forest biomass stocks globally, providing data for efforts to mitigate greenhouse gas emissions by reducing deforestation and forest degradation. To meet this challenge, the next European Space Agency Earth Explorer BIOMASS will collect synthetic aperture radar (SAR) data by employing a multiple-baseline orbit to do tomography. Recent studies, conducted in the Paracou and Nouragues sites, French Guiana (South America), have already shown that SAR tomography (TomoSAR) significantly improves the aboveground biomass (AGB) retrieval in dense forest areas, with an error of only 15% at 1-ha resolution, for biomass ranging from 200 to 500 Mg/ha However, the degree to which this TomoSAR approach can be transferred from South America to other continental sites (e.g., Africa) remains unknown. Since BIOMASS is a global mission, careful assessment of this unknown was required. In order to address this issue, a dedicated campaign was flown over several sites in Gabon (Africa). Data from this mission (AfriSAR) were acquired for TomoSAR processing at many tropical forest sites, including the sites Lopé, Rabi, and Mondah.

Here we provide the first cross-continental assessment of the performance of TomoSAR by conducting a comparison at five tropical-forest sites in South America and Africa.

By exploiting the ensemble of all flight lines in TomoSAR processing, the multiple SAR data can be transformed into a new multi-layer composite SAR image where each layer is associated with a certain height above ground level. We refer to each image within the multi-layer data stack simply by the associated height (i.e.: 15-m layer, 30-m layer ...), or as ground layer for the image focused at 0 m. We define a specified height (above ground level) which corresponds to the most concentrated intensity location in the vertical direction as a phase center height.

We consider a tomography-derived biomass model that includes not only intensity of the 30-m layer but also information on the phase center height. Results of the calibration and validation with field data, presented in Figure a (See figure attached), showed a model RMSE of 12.2%. This indicates the very good performance of tomographic data in estimating forest biomass over the full range of biomass values. Using training samples from Africa and validation samples from South America, and vice versa, the biomass retrieval resulted in a RMSE of 14 - 18% using 1-ha plots. The strong correlation between biomass retrieval and in-situ data (using training plots from South America, available only in the range 200-500 Mg/ha and validation plots from Africa (from 50-500 Mg/ha)) shows that our model prediction would succeed in capturing variation in low-biomass range sites as well. Thus, our TomoSAR biomass retrieval method can be generalizable to other study sites, even those in different continents, and can be used to map variation in forest biomass at both regional and global scales, fulfilling an important requirement for the BIOMASS mission.

To conclude, our present results considerably reinforce the scientific basis for the BIOMASS mission, increasing our confidence that it can provide accurate mapping of global forest biomass, and thereby enabling progress on global carbon fluxes, REDD initiatives, and related topics.

Area A - Board 81

Strategies for topographic mapping have been constantly improving for about three centuries witnessing the importance of this topic. As a matter of fact, digital terrain models (DTMs) are employed is several different fields: military, scientific and commercial. Among others, DTMs are used for setting defensive emplacements, planning infrastructures and resource exploitation. Concerning the scientific side, DTMs are not only necessary to characterize the environment to be explored but also are subject of investigation as their temporal and spatial dynamics are often connected to underlying processes. Several techniques are available to retrieve the elevation map of a region, they differ for the kind of signal they use (acoustic or electromagnetic) and for the retrieval strategy. Each one is suitable for a specific environment: underwater surveys are feasible by using sonars whereas a wide range of the electromagnetic spectrum can be exploited to retrieve elevation over the ground. Image triangulation allows the retrieval of the tree-dimensional structure of the target from optical images whereas laser scanners directly exploit the travel time of highly localized light pulses. Moving toward the less energetic part of the EM spectrum SAR systems are found. The exploitation of a pair of SAR images provides sensitivity to the 3D structure of the region resulting in powerful InSAR techniques. Satellite-borne InSAR systems were able to deliver high quality maps of the whole world as SRTM or TanDEM-X DEMs. Forested environments pose specific challenges to each of these techniques, tropical forests in particular represent the most difficult environments to deal with due to the thick vegetation covering the object under investigation. LiDAR systems must rely on the few light pulses that are not intercepted by any leaf of branch, reach the ground and come back to the on-board receiver. Dense forests and understory strongly reduce the number of photons that interacts exclusively with the ground before being recorded thus making it difficult to get a clear localization of the ground level. InSAR systems experience a similar effect: the penetration through the vegetation attenuates the radar pulse so that the echo reaching the ground level is possibly below the noise floor when it comes back to the receiver. The forest extinction can be minimized by using longer wavelengths: P-band signals (around 450MHz) generally provide a good visibility of the ground level. However, the exploitation of longer wavelengths radars leads to another difficulty: each elementary element of these semi-transparent media backscatters a fraction of the EM energy and let the rest pass. It follows that the information associated with each elementary scatterer reaches the receiver mixed up with many others, this means that the elevation returned by the InSAR processing results from a weighted average of the elevations associated with each scatterer where the weight is given by the backscatter intensity. This difficulty can be overcome by jointly exploiting a whole stack of coherent SAR images instead of a pair only. SAR tomography (TomoSAR) jointly processes a multi-baseline stack to retrieve the full 3D structure of the target under investigation; it allows to move from a stack of images associated with slightly different look angles to another stack of radar images associated with different depths inside the vegetation layer. Results obtained in the recent years proved the effectiveness of TomoSAR in estimating the DTM under boreal forest, however the extension to the tropical case has not been assessed yet. The evaluation of the performances of TomoSAR in retrieving DTMs under tropical forests is particularly important in view of the forthcoming ESA BIOMASS mission. Main driver of this mission is the estimation of the biomass amount in the tropical belt whereas DTMs are part of its secondary objectives. Still, many AGB retrieval algorithms rely on a clear identification of the ground level, in order to set a reference elevation or to reject its contribution. In this perspective, the quality of the DTMs directly impacts on the quality of the AGB estimates.
This work presents the DTMs estimated by TomoSAR in four different tropical forests, three in Gabon (western Africa) and one in French Guiana (South America). This test sites were explored in the framework of the AfriSAR (DLR) and TropiSAR (ONERA) campaigns respectively. The need for tomographic acquisition is assessed by exploring the quality of the InSAR estimates varying incidence angle and ground slope. The quality of TomoSAR estimates is evaluated by comparing the elevation maps with LiDAR estimates varying ground slope and tree top height. The overall agreement between TomoSAR and LiDAR DTMs is very good; furthermore, a larger discrepancy between the two results if large footprint LiDAR is selected, thus suggesting that TomoSAR might be the most accurate estimate in this case.

Area A - Board 82

The European Space Agency is currently preparing a multi-frequency SAR campaign in support of forest applications.

The test site is the Kermeter site in the Eifel National Park in North Rhine-Westphalia in Germany. The Kermeter is an upland region, up to 527.8 m above sea level, which is part of the North Eifel in the southwestern part of the state of North Rhine-Westphalia in Germany. The Kermeter ridge is covered by a forest, that is one of the largest contiguous, deciduous forests in that region. It covers an area of 3,592 hectares, of which about 3,300 hectares is a single forested area (the so-called Kermeter-Hochwald or Kermeter High Forest). Beech woods dominate the shaded, damp northern slopes (24%), in places with trees that are over 200 years old. Oak woods hold sway on the drier, southern slopes (26%), interrupted by rocky outcrops (Felsheide). Almost half the forest area still consisted of spruce trees, a consequence of reforestation measures after the Second World War. However the spruce stock is continuously being reduced in favour of deciduous woods.
In spring 2019, part of the site will be covered by single pass bistatic C- and L-band SAR and by P-band tomographic acquisitions. To support the SAR data analysis, in summer 2018 the site has been covered with airborne acquisitions of the hyperspectral sensors HyPlant and AVIRIS-NG and with a small footprint lidar. In addition during the campaign terrestrial lidar scans for selected sites will be taken and forest inventory data will be made available for 0.1 ha plot on a regular 250 m grid.
All together this data will make one of the most complete airborne SAR datasets ever collected over a forest site. The data will allow unique investigations of advanced SAR measurement techniques like SAR tomography and bistatic SAR at C-, L- and P-band frequencies.
In this paper we will present the campaign site, the collected data and first analysis results.

Area A - Board 83

Today synthetic aperture radar (SAR) tomography with resolution capability in the so-called slant height direction for remote sensing of complex scenario is in the operational level. Since about two decades, SAR tomography (TomoSAR) with peculiarity of layover solution has been the cynosure of remote sensing urban mapping [1-3]. Beside this, the 3D imaging of vegetation scenario to aim of vertical structure retrieval was the other most investigated topic by SAR tomography [4-8]; and by today two main spaceborne missions (NISAR and Tandem-L) have been under development to the goal of forest structure monitoring and biomass mapping using SAR tomography.
Typically, TomoSAR focusing algorithm employs multi-baseline (MB) complex data obtained by multi-pass SAR sensor under the interferometric conditions. However, MB data often are affected by temporal coherence loss of the natural scatterers in the interferometric time interval, which impairing TomoSAR focusing by some blurring and defocusing effects. Some theoretical models and robust focusing techniques may mitigate the degradation effect of temporal decorrelation [9, 10], while this issue may become a critical problem in some multi-temporal baseline configuration. Beyond some practical reports on the coherence loss, practically no effort has been dedicated on the assessment of temporal decorrelation with respect to the polarization employed in TomoSAR, while a couple of studies have been dedicated to analyze effects of temporal decorrelation with respect to the employed frequency of the data [11, 12].
Yet, to the aim of optimal data selection for the feasible TomoSAR application in natural scenario, an a priori assessment and comprehension of the expected level of temporal decorrelation effect on each employed polarization are required. In this framework and in this study, practical and statistical analysis on the dispersion of the reconstructed vertical structure in presence of temporal decorrelation with reference to employed polarization are presented. The analysis is based on the synthesizing of any polarimetric response of a given scatterer from its measurement in a linear orthonormal basis. The most important consequence of polarization synthesis process is fully characterization of the scatterer under temporal decorrelation, allowing a better identification and study on the optimal polarization. This pave a comprehension of the expected effect of temporal decorrelation in TomoSAR focusing with respect to the employed polarization and complementing the frequency oriented analysis.
Experiments using some realistic simulated and real data on the assessment of vertical profile dispersion in any polarization for varying temporal decorrelation degrees are reported. In particular, simulation on both typical short- and long-term decorrelation models are presented and evaluated. Moreover, points and hints on the importance and application of this practical and statistical analysis of vertical profile’s behaviour are finally given.




[1] G. Fornaro, F. Lombardini, A. Pauciullo, D. Reale, and F. Viviani, "Tomographic Processing of Interferometric SAR Data: Developments, applications, and future research perspectives," IEEE Signal Processing Magazine, vol. 31, no. 4, pp. 41-50, 2014.
[2] D. Reale, G. Fornaro, A. Pauciullo, X. Zhu, and R. Bamler, "Tomographic Imaging and Monitoring of Buildings With Very High Resolution SAR Data," IEEE Geoscience and Remote Sensing Letters, vol. 8, no. 4, pp. 661-665, 2011.
[3] X. X. Zhu and R. Bamler, "Demonstration of Super-Resolution for Tomographic SAR Imaging in Urban Environment," IEEE Transactions on Geoscience and Remote Sensing, vol. 50, no. 8, pp. 3150-3157, 2012.
[4] H. Aghababaee, G. Ferraioli, G. Schirinzi, and M. R. Sahebi, "The Role of Nonlocal Estimation in SAR Tomographic Imaging of Volumetric Media," IEEE Geoscience and Remote Sensing Letters, vol. 15, no. 5, pp. 729-733, 2018.
[5] Y. Huang, L. Ferro-Famil, and A. Reigber, "Under-Foliage Object Imaging Using SAR Tomography and Polarimetric Spectral Estimators," IEEE Transactions on Geoscience and Remote Sensing, vol. 50, no. 6, pp. 2213-2225, 2012.
[6] S. Tebaldini, "Single and Multipolarimetric SAR Tomography of Forested Areas: A Parametric Approach," IEEE Transactions on Geoscience and Remote Sensing, vol. 48, no. 5, pp. 2375-2387, 2010.
[7] S. Tebaldini and F. Rocca, "Multibaseline Polarimetric SAR Tomography of a Boreal Forest at P- and L-Bands," IEEE Transactions on Geoscience and Remote Sensing, vol. 50, no. 1, pp. 232-246, 2012.
[8] M. Tello, V. Cazcarra-Bes, M. Pardini, and K. Papathanassiou, "Forest Structure Characterization From SAR Tomography at L-Band," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 11, no. 10, pp. 3402-3414, 2018.
[9] H. Aghababaee, A. Budillon, G. Ferraioli, V. Pascazio, and G. Schirinzi, "Assessment of Temporal Decorrelation in Differential SAR Tomography for forestry applications," in EUSAR 2018; 12th European Conference on Synthetic Aperture Radar, 2018, pp. 1-5: VDE.
[10] F. Lombardini and F. Cai, "Temporal Decorrelation-Robust SAR Tomography," IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 9, pp. 5412-5421, 2014.
[11] A. Hamadi, L. Villard, P. Borderies, C. Albinet, T. Koleck, and T. L. Toan, "Comparative Analysis of Temporal Decorrelation at P-Band and Low L-Band Frequencies Using a Tower-Based Scatterometer Over a Tropical Forest," IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 11, pp. 1918-1922, 2017.
[12] Y. Morishita and R. F. Hanssen, "Temporal Decorrelation in L-, C-, and X-band Satellite Radar Interferometry for Pasture on Drained Peat Soils," IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 2, pp. 1096-1104, 2015.

Area A - Board 84

In this work, we study the use tomographic SAR data for the task of supervised land-cover classification.

Land-cover classification from SAR images has received a strong interest since, contrary to optical images, the backscattered echoes in the microwave domain are directly related to geometrical and intrinsic physical properties of the media. Moreover, SAR allows all weather and day/night independent imaging which makes it a powerful tool for temporal monitoring. Polarimetric SAR (PolSAR) is the most common type of SAR data used for classification due to the existence of target decomposition theorems allowing to retrieve different types of canonical scattering mechanisms. Unsupervised classification based on such theorems has reached a tremendous success in the past decades. However, this type of classification is limited to describe simple physical characteristics of the media. For many applications, a semantic labelling of complex classes such as city, field, or forest is required that cannot be modelled by simple physical models. Supervised systems allow to learn the statistical properties of such classes from training samples that have been labelled manually. PolSAR has also been successfully exploited for supervised learning.

SAR Tomography (TomoSAR) has been of interest for more than a decade but has been only very recently considered as being potentially useful for classification. TomoSAR uses several images acquired with different view angles to reconstruct the reflectivity in the elevation direction. TomoSAR has been exploited for the 3-D reconstruction of buildings, the imaging of semi-transparent media such as forest canopy as well as ice sheet sounding. The so-called tomograms are the reflectivity profile in the elevation direction and are obtained by spectral focusing methods such as Fourier, Capon beamforming, MUSIC or even more sophisticated compressive sensing methods.

Since tomograms contain information about the number of scatterers present in the resolution cell as well as their respective position, power and extent, they should provide good features for a classifier provided the considered classes exhibit different vertical structure and the baseline configuration allows to resolve this structure. In this work we extend the work of our recent publication [1] by considering the fusion of TomoSAR and PolSAR features, studying the influence of the number of training samples, and applying our framework to a recent dataset.

Due to sparse irregular sampling, decorrelation and positioning errors, tomograms are often noisy and exhibit artifacts. Moreover, for some classes such as bare soil, only a few coefficients are relevant. Therefore, using the raw tomogram as a classifier input has been shown to lead to poor performance. Our contribution is the design of relevant low dimensional features to improve classification accuracy. In this work, we consider two types of features: Pixel-based and spatial features. Pixel based features are statistical descriptors such as statistical moments, value and position of the different peaks, number of non-zero elements, coefficient of variation and Shannon entropy. The spatial descriptors encompass intensity based features and tomogram based features. We combine features describing the statistical properties of regions obtained by a prior over-segmentation of the image into segments of homogeneous intensity with features describing the local texture of the image.

We use the well-established Random Forest (RF) classifier for our experiments. This classifier combines the outputs of simple binary decision trees and exploits randomness to increase the robustness of the classification. At each node, a tree performs a splitting of the data along a randomly chosen dimension. The best split is obtained by minimizing an impurity measure. The classification map is obtained by computing the average of posteriors over all the trees for each label. The advantage of using a multitude of single trees is that overfitting cannot happen if the number of trees is increased. Moreover, performance is not affected by data scaling since each feature dimension is considered independently.

We have run experiments on a 6 image L-Band quad-pol dataset acquired by the F-SAR sensor from the DLR (German Aerospace Center) over a mainly forested area close to the town of Traunstein, Germany. For training and evaluation, we manually labelled part of the images into three classes: Urban, Field, and Forest. The tomographic processing has been using the Capon method for focusing on a single polarimetric channel (VV). RF has been applied with 60 trees.

For comparison, we have applied the same classifier to PolSAR data where the features we extracted from the coherency matrix as described in [1]. For PolSAR and TomoSAR we have augmented the pixel based features by the spatial features described in [1]. Since the information contained in PolSAR and TomoSAR are complementary, it is interesting to fuse polarimetric and tomographic features. Therefore, we have also run RF classification with the combination of both feature sets. Since for many remote sensing application labelled data is based on in situ acquired ground truth, it is common that a restricted number of samples are available. Therefore, we have also run an experiment with a number of training samples limited to 30 per class.

In the case of a large number of training samples, PolSAR and TomoSAR were showing similar accuracy although slightly better for TomoSAR. In the case of the smaller number of samples, however, TomoSAR classification outperformed PolSAR by several percent. Our study also shows that adding PolSAR features to tomogram ones leads to a slight increase in accuracy compared to TomoSAR alone.

[1] D’Hondt, O.; Hänsch, R.; Wagener, N.; Hellwich, O. Exploiting SAR Tomography for Supervised Land-Cover Classification. Remote Sens. 2018, 10, 1742.

Area A - Board 85

Spaceborne SAR sensors operating in long wavelengths are prone to combined effects of Faraday rotation and system distortions (cross-talk and channel imbalance). The latter degrade the SAR signal, which can have a substantial impact on the estimation accuracy of bio and geophysical parameters such as above-ground-biomass (AGB). Hence, strict requirements need to be set on these system parameters, which need to be carefully traded off with the complete end-to-end system. In this study, we present the results of a closed form analysis on the impact of cross-talk, channel imbalance and Faraday rotation on the AGB estimation accuracy for the Biomass mission. The impact is analysed for different biomes and forest types to assess the adequacy of the specified requirements.
To build on representative test cases, covariance matrices are extracted from full-polarized airborne P-Band campaign datasets of forest stands with known AGB values. The analysed campaign sites are Paracou (tropical rainforest, French Guiana), La Lopé (savannah and tropical rainforest, Gabon), Mabounie (disturbed tropical rainforest, Gabon) and Remningstorp (hemiboreal forest, Sweden).
A simulator is used to study the impact of the system distortion. The measured covariance matrices are being corrupted with a given system distortion and Faraday rotation. The resulting distorted matrices are used to estimate AGB using a simplified, locally adapted inversion model based on the power-law relationship between AGB and the HV cross-polarized backscattering coefficient gamma nought. Each combination of cross-talk and channel imbalance magnitudes yields a worst-case AGB estimation error under the assumption of the phase of the distortion terms being uniformly distributed. Different scenarios are analysed in order to identify maximum allowable cross-talk and channel imbalance limits for each test site. It is shown that the driving parameter for the impact of these system terms is the sensitivity of the AGB inversion model. Preliminary results of this analysis suggest that the current requirements used by industry could be relaxed without violating the goal of estimating AGB with an accuracy of 20%.

Area A - Board 86

SAR tomographic methods have proven extremely adept at measuring vegetation vertical structure at a variety of wavelengths including L and P-bands and resolving vertical urban environments. Measuring the three dimensional structure of vegetation and its changes resulting from either natural or anthropogenic causes are key parameters in monitoring ecosystems. Accurate 3-dimensional representations or urban structure and change are needed for many applications monitoring the expansion or urban environments.

UAVSAR is a NASA/JPL L-band fully polarimetric synthetic aperture radar employing an electronically scanned array whose primary design goal was to enable robust repeat pass radar interferometric measurements of deforming surfaces either from natural or anthropogenic causes. The radar is housed in a pod mounted to the fuselage of a Gulfstream III jet. Electronic steering of the L-Band antenna is tied to the inertial navigation unit so that consistent pointing is achieved regardless of the platform yaw. The platform was modified to include a precision autopilot (PPA) that allows the aircraft to fly a specified trajectory within a 5 m tube. This enables UAVSAR platform to fly a series of flight lines with well-prescribed interferometric baselines. The radar can also be configured to as a P-band polarimetric radar or a Ka-band single pass interferometer. The talk will discuss the utility of the UAVSAR radar for tomographic measurements in urban and forest settings.

For tomographic mapping of vegetation we will show data collected over boreal forests in L-band and P-band over Delta Junction in Alaska and in L-band over the Berms site in Canada. During the Tomosar acquisition at Delta Junction, the birch trees that dominate the area were mostly leaf-off. Present within the site are small ponds and stands of lower stature trees and bushes, as well as manmade features such as paved and dirt roads. The underlying terrain topography was mostly gently undulating, with a small amount of mostly grassy ground foliage. Selection of the UAVSAR TomoSAR boreal forest site was primarily on the availability of an airborne LIDAR forest height and hyperspectral image products from the NASA G-LiHT system. The Berms site in Canada was selected based on the availability LVIS lidar data sets and in situ soil moisture measurements during the data acquisitions. UAVSAR acquisitions were done in coordination with L-band and S-band tomographic observations acquired by the DLR F-SAR radar.

Munich, being close to DLR Oberpfaffenhofen facility was a convenient urban site for collecting UAVSAR tomographic lines for an urban setting during Spring 2015 deployment to Europe. Tomographic lines were collected with the smallest ambiguity heights only a couple of meters allowing for very fine vertical resolution. We will discuss the first experimental results of tomographic processing of UAVSAR data in an urban environment.


This research was conducted at the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.

Area A - Board 87