Conference Agenda

Modern SAR technology offers various approaches for processing stacks of interferometric SAR data. For surface motion estimations in urban areas, normally short wavelength data, like X- or C-band, is preferred. For applications in urban areas and infrastructure monitoring. Long wavelength data offers a certain amount of penetration capability and they are less sensitive to temporal decorrelation.

PS-InSAR is a widely used method for surface motion estimation from interferometric SAR data stacks. It is used in commercial applications and also in many projects in the Dragon program, starting from Dragon-1 until today. We consider it a stable technique, proven to successfully and reliably offer surface motion estimations in numerous projects. We used PS-InSAR in Shanghai and Wuhan for estimating urban subsidence and infrastructure stability. With the availability of Sentinel-1 and the large global SAR archive, it is nowadays possible to process PS-InSAR and estimate subsidence in regions of interest all over the world, opening this field up to the public even further.

SAR tomography with long wavelength SAR data, preferably with P-band data, allows foliage penetration and the true 3D reconstruction of the SAR signal under the foliage. This can be used for various applications, e.g. for the estimation of above ground forest biomass. SAR tomography here allows to measure the biomass, instead of estimating it based on tree canopy heights, on a global level. ESA will use this with the upcoming BIOMASS mission.

There are still several problems to be solved though. On problem is the temporal decorrelation. Although P-band is less sensitive to temporal decorrelation, it is still not immune to it. Especially changes in rainfall and canopy water content / water content layers, can cause problems in the 3D reconstruction. With one of the main areas of interest along the tropical rainforest, rain and changes in the rainfall patterns are to be expected though. Minimizing the amount of data necessary for a tomographic inversion is therefore important to allow a good biomass estimation with few acquisitions.

In Dragon-4 we are working closely together towards these goals, continuing our research on PS-InSAR and related techniques, but also extending towards the 3D reconstruction using SAR tomography with long wavelength SAR data.

In recent years, advanced techniques such as polarimetric SAR interferometry (Pol-InSAR) and SAR tomography (TomoSAR) have been widely used to retrieve forest parameters by means of SAR measurements. Pol-InSAR was developed based on the Random Volume over Ground (RVoG) model, which assumes a penetrable volume layer consisting of randomly oriented particles over an underlying rough surface. On this basis, Cloude and Papathanassiou proposed a parametric inversion scheme to retrieve forest height, which has been successfully applied for a variety of forest sites at different frequency bands.

SAR tomography is instead an imaging technique based on the collection of multiple flight lines. It allows focusing the received signal not only in the range/azimuth plane, as in conventional 2-D SAR imaging, but also in elevation, hence providing 3-D resolution capabilities. The retrieval of canopy height using SAR tomography has been considered since the early experiments. Indeed, wave scattering from forested areas is bound to occur between the terrain and the top of the canopy. Hence, canopy height can be retrieved, at least in principle, by tracing the upper envelope in tomographic sections.

In this paper, we aim at presenting an experimental assessment of vegetation height retrieval in tropical forests based on P-band SAR acquisitions. Two approaches are considered: i) parametric height estimation under the assumption of the Random Volume over Ground (RVoG) model, and ii) thresholding the vertical backscattering profiles that are focused by SAR tomography. The data-set under analysis is from the ESA AfriSAR campaign that was flown over Gabon in 2016. Results show that both of the two approaches are able to retrieve forest height to within an accuracy of about 3 m or better over the interval of forest height between 30 m to 50 m at a resolution of 25 m × 25 m

Several upcoming SAR satellite constellations, are going to be operated in bi-static mode, like TanDEM-L or TwinSAR-L, or may have a bi-static companion, like Sentinel-CS. Until now, bi-static data is mainly used for DSM generation, as in the TanDEM mission. In the future, the goal is to use such data also for surface motion estimation. However, current multi-baseline D-InSAR approaches are not well suited for processing this data and need to be adjusted.

The main advantage of a bi-static operation is the minimization of the temporal decorrelation and the atmospheric influence. But, a temporal difference close to zero between the acquisitions also means that ground deformations cannot be measured. Motion related phase components will only appear with a significant time difference between the acquisitions. By acquiring several image pairs over the same area, bi-static missions can deliver such repeat-pass acquisitions with a required temporal baseline, but these interferograms will again suffer from temporal decorrelation and atmospheric effects like the standard acquisitions.

In terms of PSInSAR or related processing methods, that is to say that we would expect an improved estimation of PS point height, but not necessary a better estimation of the deformation phase components, as the most severe problems still occur.

Even more, standard processing chains for PSInSAR will not work well, or at all, with such stacks. In our experiments, we used pursuit mono-static data from the TanDEM-X science phase. The along track baseline is extended to 10 seconds between the satellites, allowing both satellites to transmit and receive data undisturbed from each other. The data is therefore not bi-static and generally suitable for standard InSAR and PSInSAR processing. However, the very small temporal baseline of 10 seconds compared to the 11 days repeat-pass baseline can cause numerical problems in the estimation of the deformation phase.

To avoid this, we separated the estimation of the topographic phase from the estimation of the deformation phase component and use different image pairs in both cases. We estimate the topographic phase only from the 10s pairs. Based on the estimated heights from this first step, we process the deformation phase using the repeat-pass images. Having two images per time can reduce the noise, however we found no significant difference in the performance from this. In areas with high skyscrapers, like our testing area in Guangzhou, China, the deformation estimation can vastly benefit from the much better height estimation of this approach. However, unfortunately, the amount of data available is currently very limited, so that we can only present preliminary results for deformation estimation, showing only slight improvements in this regard.

In synthetic aperture radar (SAR) remote sensing, Differential SAR Tomography (Diff-Tomo) is developing as a powerful crossing of the mature Differential SAR Interferometry and the emerged 3D SAR Tomography, producing advanced 4D (3D+Time) SAR imaging capabilities extensively applied to urban deformation monitoring.

More recently, it has been shown that through Diff-Tomo, identifying temporal spectra of multiple height-distributed decorrelating (forest) scatterers, the important decorrelation-robust forest Tomography functionality is obtained.

To loosen application constraints of the related main experimented full model-based processing, and develop other functionalities, this work presents an advanced adaptive, just semi-parametric, generalized-Capon Diff-Tomo method conceived and developed at University of Pisa (UniPi) for extraction of height and dynamical information of natural distributed (volumetric) scatterers. In addition to robust Tomography, particular reference is to separation of decorrelation mechanisms in forest layers.

Simulated and P-band results are shown. A review of other advanced Diff-Tomo tools developed at UniPi for information extraction in decorrelating forest scenarios is also presented.

Ack.: the Authors thanks Dr. Francesco Cai, formerly at UniPi and now with Leonardo Company, for his support in the SW development.

Reigber, A., Moreira, A.: ‘First demonstration of airborne SAR tomography using multibaseline L-band data,’ IEEE Trans. Geosci. Remote Sens., 2000, 38, (5), pp. 2142-2152

Pardini, M., Papathanassiou, K.: ‘On the estimation of ground and volume polarimetric covariances in forest scenarios with SAR tomography,’ IEEE Geosci. Remote Sens. Lett., 2017, 14, (10), pp. 1860-1864

Huang, Y., Ferro-Famil, L., Reigber, A.: ‘Under-foliage object imaging using SAR tomography and polarimetric spectral estimators,’ IEEE Trans. Geosci. Remote Sens., 2012, 50, (6), pp. 2213-2225

Azcueta, M., Tebaldini, S.: ‘Non-cooperative bistatic SAR clock drift compensation for tomographic acquisitions,’ Remote Sensing, 2017, 9, (11), pp. 1-11

Lombardini, F.: ‘Differential tomography: a new framework for SAR interferometry,’ IEEE Trans. Geosci. Remote Sens., 2005, 43, (1), pp. 37-44

Lombardini, F., Cai, F.: ‘Temporal decorrelation-robust SAR tomography,’ IEEE Trans. Geosci. Remote Sens., 2014, 52,(9), pp.5412-5421

In this paper, the GPUs are used to accelerate the processing efficiencies in time domain (TD) SAR simulation and time domain back-projection (TDBP) focusing. The raw data simulation and back-projection reconstruction are both implemented in the time domain for handling the scenarios of highly non-linear trajectories. The processing inefficiencies, however prevent extensive applications of TD SAR simulation and TDBP focusing. Thus, we utilize the massive parallelism of GPUs to enhance the processing efficiencies. In this contribution, we develop an optimized time-domain SAR simulation algorithm with complexity O(n³). We also discuss the drawback of the optimized simulation method and our contributions to mitigate this problem. Furthermore, both parallel simulation and back-projection focusing algorithms are fully optimized under the NVIDIA’s Compute Unified Device Architecture (CUDA) framework to guarantee a relevant acceleration compared with CPU counterparts. As a result, the GPU-based TD SAR simulation gains a 78x speed-up factor over the CPU serial version. The GPU based TDBP implementation achieve an over 100x speed-up factor compared with the CPU counterpart. To demonstrating the validity of our methods, we apply our GPU based TDBP focusing methods in simulated SAR raw data from highly deviated trajectory and circular trajectory.

Forest above ground biomass (AGB) retrieval by P-band Synthetic Aperture Radar (SAR) tomography has been extensively studied in recent years in the context of the forthcoming spaceborne mission BIOMASS. Most studies made use of airborne data collected in a single day, for which temporal decorrelation could be neglected. This fortunate situation will clearly not be repeatable in the case of BIOMASS, for which the revisit time will be of 3 days. The impact of temporal decorrelation on tomographic observables was analyzed in previous studies using data from the ground-based experiment TropiSCAT, which provided continuous tomographic observations at the expense of covering a small area and providing no azimuth resolution. This paper is meant to complement those studies by investigating the effect of temporal decorrelation on forest canopies over large areas, based on the airborne data-set acquired by DLR during the AfriSAR campaign. The analysis is carried out based on the recently proposed ground-notching technique, which is used to single out volume scattering based on single-baseline acquisitions gathered at a time lag of 4, 5, and 9 days. Results show that volume temporal coherence is consistently between 0.6 and 0.85 when forest height is larger than about 25 m, whereas low vegetation areas appear to be significantly more affected by temporal decorrelation. As a result, the intensity of volume-only scattering is observed to vary to vary by less than 1 dB when ground notching is performed using acquisitions from different dates.