Methodology and Techniques - DInSAR

During the last decade several InSAR time-series approaches have been developed in response to a non-ideal acquisition strategy of SAR satellites with large spatial and temporal baselines and with non-regular acquisitions. The small baseline tubes and regular acquisitions of new SAR satellites such as Sentinel-1, allows forming connected networks of interferograms and simplifies the InSAR time-series analysis to a weighted least squares inversion of an over-determined system of equations. Such robust inversion allows to better understand the different components of the InSAR time series and to evaluate the uncertainties.

We present an open source python-based package for InSAR time series analysis with unique functionalities for obtaining unbiased ground displacement time-series, geometrical and atmospheric correction of InSAR data and quantifying the InSAR uncertainty. Our implemented strategy for InSAR time-series analysis in PySAR (as shown in the Fig. 1 in attached pdf file) contains several features including: 1) improved spatial coverage using a coherence-based network of interferograms; 2) unwrapping error correction using phase closure or bridging; 3) tropospheric delay correction using atmospheric models and empirical approaches, 4) geometrical correction (e.g.; DEM error); 5) automatic outlier detection and optimal selection of the reference date, 6) quantifying InSAR uncertainty due to the residual tropospheric delay after corrections, 7) variance-covariance matrix generation of final products for geodetic inversion.

We demonstrate the performance and efficiency of PySAR using SAR datasets acquired by ALOS/ALOS-2, Envisat, TerraSAR-X, Cosmo Skymed and Sentinel-1. We show applications of PySAR for studying the spatio-temporal evolution of ground deformation caused by volcanic activities in Japan and Ecuador, interseismic tectonic deformation in Pakistan and Afghanistan, and deglaciation in Greenland. Our results show: 1) wide-spread volcanic deformation in Kyushu, SW Japan (Fig. 2), including the volcanic cycle of 2011 Shinmoe-dake, Kirishima volcano eruption event with both pre- and co-eruptive ground displacement up to 4 cm in LOS direction; and precursory inflation up to 3.4 cm in vertical direction in Cotopaxi volcano prior to its 2015 eruption (Fig. 3); 2) a 340 km long, shallow creeping segment along Chaman Fault with maximum surface creep rate of 8.1+/- 2 mm/yr in LOS direction, accommodating 30% of the relative plate motion between India and Eurasia (Fig. 4); 3) deglaciation-induced uplift and seasonal melting in Petermann glacier ice margin, Greenland (Fig. 5).

(Please refer to the attached file for the full abstract)

1. Introduction

Very high resolution SAR images in dense urban area are not trivial to interpret due to the inevitable layover caused by the side-looking imaging geometry. With the growing attention on very high resolution SAR data, the fusion of optical and SAR images in dense urban area has become an emerging and timely topic, because the complementary of these two data types can lead us to unprecedented insights and findings, such as the unique scattering mechanisms of different urban infrastructures. Lying at the basis of such fusion topic is the challenging task of the co-registration of SAR and optical images. Such two images are acquired with intrinsically different imaging geometries, and thus are nearly impossible to be co-registered without a precise 3-D model of the imaged scene. Only until recently, the “SARptical” [1], [2] system proposed a promising solution to tackle this challenging task. SARptical can trace individual SAR scatterers in corresponding high resolution optical images where we can analyze the geometry, material, and other properties of the imaged object. Vice versa, the similar study can also be done in the SAR image coordinate. This paper demonstrate the capability of the SARptical system, and its potential in various different applications including lamp poles deteciton for geodetic InSAR, object-based multibaseline InSAR, and the optical and SAR image matching.

2. The SARptical System

The general framework applies to a stack of SAR images and a pair of (or more) optical images. The focus of SARptical is put on linking the attributes from optical image to the SAR image by 3-D matching and projection. The basic idea is to match the 3-D models derived from SAR and optical images respectively. As a result, the 2-D SAR and optical images will also be matched. Based on the matched images, subsequent tasks such as semantic label texturing and joint deformation analysis can be conducted. The detailed procedures of SARptical are as follows.

• 3-D reconstructions

a) Retrieve the 3-D positions and deformation parameters of the scatterers from the SAR image stacks. Since urban area is of our main interest, tomographic SAR inversion (TomoSAR), including SAR tomography and differential SAR tomography, is employed in order to resolve a substantial amount of layovered scatterers.

TomoSAR is the most computationally expensive step in the framework. In addition, TomoSAR and other multipass SAR interferometry (InSAR) algorithms typically requires a fairly large SAR image stack (>20 images). The computational and image resource are the main limitation for this step.

b) Retrieve the 3-D positions of points from the optical images using stereo matching with structure from motion (SfM) if necessary. For covering large urban area, aerial or spaceborne images are preferred. This step also calibrates the camera parameters.

Stereo matching and SfM are well studied topics. Many matured algorithms and software are readily available.

• 3-D matching: Co-register the TomoSAR point cloud and the optical point cloud.

The main challenges present in this step are the different modalities of optical and TomoSAR point clouds, i.e. nadir-looking and side-looking, as well as the relatively large anisotropic noise in the TomoSAR point cloud. However, considering the large amount of points compared to the few co-registration parameters to be estimated, the co-registration accuracy is expected to be high enough for the following steps.

• Optical image classification: applying semantic classification to the optical images.

This part is not the focus of SARptical. Depending on the application, different classification algorithms can be applied.

• Semantic texturing: Texture the InSAR point cloud with the attributes derived from optical images, e.g. RGB color, semantic classification label, object bounding box, etc.

The main challenge of this step is to project the optical image to TomoSAR point cloud without explicit 3-D surface reconstruction in the TomoSAR point cloud. Therefore, we choose point-based rendering technique.

The main limitation of this step is the relatively poor positioning accuracy (1 to 10m) of spaceborne TomoSAR point cloud. This error will directly translate to the projection accuracy of the TomoSAR points in optical image.

3. Demonstration of Applications

3.1Automatic lamp poles detection for geodetic InSAR

Only until recently, it has been demonstrated that absolute localization with centimeter accuracy can be achieved for manually matched persistent scatterer (PS)s from TerraSAR-X images acquired from cross-heading geometries [3]. For automatically localizing a large network of such PSs for geodetic applications like the “Geodetic SAR Tomography” [4], we found that cylindrically vertical structures like lamp poles on the street are, most probably, the only natural ones visible in SAR images acquired from both ascending and descending orbits [5]. Detecting these PSs in SAR images can be extremely challenging, while it is much promising to achieve in optical images.

Thus, the demonstrated methodology includes the identification of lamp posts from high resolution optical data and project them into the cross-heading SAR images using the SARptical system. The precise absolute 3-D coordinates of the points are retrieved from the corrected TerraSAR-X timing measurements using the stereo SAR method [3]. Results for a test site in the city of Berlin acquired from TerraSAR-X high resolution spotlight mode will be demonstrated in the full paper.

3.2 Object-based multibaseline InSAR

Deformation monitoring by multi-baseline repeat-pass synthetic aperture radar (SAR) interferometry is so far the only imaging-based method to assess millimeter-level deformation over large areas from space. Past research mostly focused on the optimal deformation parameters retrieval on a pixel-basis. Only until recently, the first demonstration of object-based urban infrastructures monitoring by fusing InSAR and the semantic classification labels derived from optical images was presented in the SARptical system [1], [2], [6].

In this paper, we proposed a general framework, given such classification label in the SAR image, for object-based InSAR parameters retrieval where the estimation of the parameters is achieved in an object-level instead of pixel-wisely. Another new development presented in this paper is to introduce a robust phase recovery step in prior to the parameters inversion, in order to handle outliers in real data. The demonstrated method outperforms the current pixel-wised estimators, e.g. periodogram, by a factor of as much as dozens in the accuracy of the linear deformation estimates, at various situations such as signal-to-noise ratio, and outlier percentage. For practical demonstration, we presented a full workflow of long-term bridge monitoring using the proposed approach in the final paper.

3.3 SAR and optical images matching

The identification of similar image patches certainly is a frequently demanded task in remote sensing-related image analysis, especially in the framework of stereo applications. While many established feature-based approaches, specifically designed for the matching of optical images e.g. SIFT [7], already exist. To this date, the matching of images acquired by different sensors still remains an open challenge. This particularly holds for a joint exploitation of SAR and optical imagery. The challenge is caused by two completely different sensing modalities: optical imagery is acquired passively in a perspective projection at visible to inferred band (hundreds of THz), whereas SAR imagery is acquired actively in a cylindrical projection at microwave frequency (several GHz). Thus, particularly structures elevated above the ground level, such as buildings in urban areas, show strongly different appearances in both image types.

Thanks to the SARptical system, we have collected tens of thousands of matched SAR and optical images patches which can be used for exploitation the SAR and optical patch similarity. We demonstrate a convolutional neural network (CNN)-based approach, which allows to identify similar patches of very high resolution (VHR) optical and SAR imagery of complex urban scenes. The underlying similarity function is learnt directly from automatically generated training data and does not resort to any hand-crafted features. First evaluations show that the network provides an overall accuracy of more than 93% with a false alarm rate of 0%, thus indicating great potential for further development to a generalized multi-sensor matching procedure. We will show the example using real data in the final paper.

4. Summary

SARptical is a novel concept that allows a pixel-level matching between high resolution SAR and optical images of dense urban areas. This is probably the first time that we see an optical image in dense urban area in a SAR geometry, and vice versa. We demonstrated the potentials of SARptical, including difficult target detection, object-based InSAR, and SAR optical image matching which were otherwise very challenging without the aid of SARptical.

5. Reference

[1] Y. Wang and X. X. Zhu, “Fusing Meter-Resolution 4-D InSAR Point Clouds and Optical Images for Semantic Urban Infrastructure Monitoring,” IEEE Trans. Geosci. Remote Sens., 2016.

[2] Y. Wang and X. X. Zhu, “InSAR Forensics: Tracing InSAR Scatterers in High Resolution Optical Image,” presented at the Fringe 2015, 2015.

[3] C. Gisinger et al., “Precise Three-Dimensional Stereo Localization of Corner Reflectors and Persistent Scatterers With TerraSAR-X,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 4, pp. 1782–1802, Apr. 2015.

[4] X. X. Zhu, S. Montazeri, C. Gisinger, R. F. Hanssen, and R. Bamler, “Geodetic SAR Tomography,” IEEE Trans. Geosci. Remote Sens., vol. 54, no. 1, pp. 18–35, 2015.

[5]        S. Montazeri et al., “SAR Ground Control Point Identification with the Aid of High Resolution Optical Data,” in IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2016, Beijing, China, 2016.

[6] Y. Wang and X. X. Zhu, “Semantic Fusion of SAR Interferometry and Optical Image with Application to Urban Infrastructure Monitoring,” presented at the La Grande Motte, France, La Grande Motte, France, 2015.

[7] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, 2004.

DInSAR techniques for discriminating between surface and buried targets using Sentinel-1 images

The aim of this research is to develop a scheme for buried target detection under a soil using the Differential Interferometric Synthetic Aperture Radar (DInSAR) measurements from satellite SAR data. Here, we report on an investigation into the use of DInSAR for discrimination between a buried and surface point targets using Sentinel-1 images.

In this study, a network of seven corner reflectors were deployed in the test area, which is a farm located in Nottingham, UK. This farm grows different crops in different seasons with no man-made feature inside the farm and trees are only located on the borders, making this farm an ideal place to conduct a corner reflectors (CR) experiment. The reflectors were installed rigidly and permanently for over one years’ worth of observations, commencing in September 2015. The interferometric phase of these CRs were analysed for their stability. Each CR in the network acted as a point target in the SAR image. Five of the seven CRs were covered from their base with a thin layer of sand, 1cm thickness, to simulate buried targets. For each satellite pass, the moisture content of the covering sand layer was precisely measured.

A stack of 38 Sentinel-1 images were collected for the period from September 2015 to September 2016. Both the amplitude and the interferometric phase for the CRs were analysed and compared with the moisture content of the sand layer for the covered (buried) targets. It can be reported that the backscatter of the covered trihedral reflectors showed a clear reduction with increased moisture contents. It is noticed that the targets became invisible with moisture content of 10%. Thus, the analysis implemented in this study was limited to moisture content from 0% (dry sand) to 10%. In terms of phase, there is a clear difference in the phase returns from covered and exposed reflectors. The differential interferometric phase of the exposed targets shows stable phase with an irregular 3^o oscillation, on the contrary to the covered reflectors phase which show a strong positive linear relationship with the moisture content of the covering sand layer. We can conclude that the signal from the covered corner reflectors can be easily distinguished from the un-covered reflectors.

This technique can be applicable for arid regions to study/detect pipelines networks which are buried at a shallow depth, in contrast to the current underground feature detection techniques which relay on in-situ measurements.

Keywords

Sentinel-1, DInSAR, Corner Reflectors 

DInSAR technologies have already demonstrated over the past decades their capabilities to effectively study and follow deformation phenomena related to natural and anthropic hazards, with centimeter to millimeter accuracy. In particular, the advanced DInSAR technique referred to as Small BAseline Subset (SBAS) algorithm [1],[2],[3] have proven to be an effective tool able to carry out multi-scale and multi-sensor analyses of surface deformation, providing more insights on the spatial and temporal pattern of the investigated displacements at the regional (low resolution analysis) and local (full resolution analysis) spatial scales [4]; [5]. This reveals to be particularly suitable to detect, map and monitor displacements affecting urban areas [6] or archaeological and historical sites [7], [8]; moreover, the full resolution SBAS-DInSAR analysis allows capturing also very localized deformation signals associated with both large man-made features and portions of a single historical monument or building. Figure 1 shows an example of the good performances retrieved by applying the full resolution SBAS-DInSAR method (in terms of number of coherent points and detected structures) to a consistent COSMO-SkyMed SAR dataset relevant to the well-known archaeological site of Pompeii (Southern Italy). It is clear how very localized displacements related to extended man-made features or portions of buildings can be captured with great spatial and temporal details, thanks to the large number of coherent targets detected over the investigated structures.

The widespread use of advanced DInSAR approaches throughout the scientific communities and the increasing application of such techniques in the Solid Earth science field are going together with the consequent technological progress, oriented on the one hand towards the effective exploitation of the DInSAR method performances (i.e. implementation of efficient algorithms), on the other hand toward the development of new SAR sensors and satellite missions, characterized by different frequency bands, spatial resolution, revisit times and ground coverage. The DInSAR scenario is nowadays characterized by a steady increase in the availability of satellite SAR systems since 1992, starting from the "first-generation" C-band SAR missions (ERS-1/2 and ENVISAT of the European Space Agency and RADARSAT-1 of the Canadian Space Agency), moving to the "second-generation" SAR constellations, specifically the X-band COSMO-SkyMed (CSK) and TerraSAR-X (TSX) systems, which are particularly appropriate to follow the space-time characteristics of the detected deformation phenomena at the scale of single buildings, allowing to map nearly all the man-made structures of an investigated area, also revealing possible intra-building differential movements, as well as to monitor the temporal evolution of the displacements also in presence of small rates, fast-varying and non-linear deformation phenomena. However, such an improved observation characteristics (reduced revisit time and higher spatial resolution) have led to the creation of very large SAR data archives to handle.

The recent launch of the C-band Sentinel-1 (S1) constellation, within the framework of the Copernicus (formerly GMES) Programme of the European Union, is pushing toward the present Earth Observation scenario to up-to-date research and monitoring frontiers, opening new possibilities to the investigation of surface deformation phenomena at a continental scale, thanks to the innovative acquisition mode referred to as Terrain Observation with Progressive Scans (TOPS) [Torres et al., 2012], specifically devoted to advanced DInSAR applications, which allows collecting S1 Interferometric Wide Swath (IWS) scenes. In particular, such a C-band system allows generating SAR images with a spatial resolution comparable to that of the ERS and ENVISAT satellites, but with a remarkable increase in the range coverage (about 250 km). Moreover, the reduced revisit time (6 days) ensured by the fully operative Sentinel-1A (S1A) and Sentinel-1B (S1B) twin satellites, together with a “free and open access” data distribution policy, permits to systematically generate highly coherent interferometric products over very wide areas.

All these systems have enabled us to collect, over the past two decades, huge SAR data archives that have permitted us to continuously investigate surface displacements over a wide temporal and spatial extent, with different spatial resolutions and revisit times. In this context, a massive data volume will be supplied in the next few years, and petabytes of DInSAR measurements (raw data, interferograms, displacement maps, deformation time series) have to be processed, archived and handled, so that the DInSAR scenario is moving toward a Big Data challenge, with a strong impact on the data storage and the computational requirements needed to generate the advanced DInSAR products. To profitably exploit the performances of the current SAR sensors, it is crucial to develop innovative and appropriate solutions, aimed at automatically and efficiently handling these huge SAR data archives, more and more increasing in terms of both temporal and spatial resolutions, as well as of ground coverages. These solutions are based on the one hand on the exploitation of advanced methodologies and algorithms acquired from the new Information and Communication Technologies (ICT), which guarantee high efficiency in terms of portability, scalability and computing performances; on the other hand, it is also crucial to develop much more advanced DInSAR methodologies (and codes) able to effectively squeeze the information associated with these huge amounts of SAR data. Accordingly, both large storage and high performance computing capabilities are needed, as well as efficient algorithms have to be developed to tackle this huge data flow. 

In this paper, we present an innovative parallel computing solution for the full resolution SBAS-DInSAR processing chain [2], which is particularly appropriate to exploit the current available parallel hierarchal platforms. In particular, two parallelization levels are considered. The first one is based on a coarse/medium granularity-based approach (mainly applied to the whole processing chain); the second one relies on a fine-grained parallelization and is implemented only for the heaviest computational steps, in terms of computing time and allocated memory.

The coarse/medium-grained parallelization strategy addresses to the exploitation of multiprocessor systems with distributed memory, computations that can be parallelized by requiring a minimal effort to partition the application into independent parallel parts. This kind of processing essentially exhibits minimal dependencies in terms of data, synchronization, or ordering.

The fine-grained parallelization approach is mainly based on the use of Graphical Processing Unit (GPU), which allows significantly increasing the computing performances, in terms of optimization of the available memory on the GPU, reduction of the Input/Output operations on the graphic unit, and consequent reduction of the processing time, by efficiently exploiting parallel processing architectures as CUDA. However, this strategy requires a strong effort to re-design some key steps of the overall full resolution SBAS-DInSAR processing chain, in order to strongly benefit from the efficiency achieved through the use of the GPU; at the same time, this guarantees to reach the best performances in terms of processing time and scalability. The GPU parallelization strategy is mainly applied to the processing blocks that work on a pixel-by-pixel basis (for example, the step dedicated to the computation of the velocity and topographic phase components of the used model within the full resolution SBAS-DInSAR processing chain, through the maximization of the temporal coherence). In this case, by simultaneously exploiting the very large number of the processors within the GPU, it is possible to compute in parallel the same operations involving single pixels, thus reducing the whole computational time related to some processing blocks (e. g. the maximization of the temporal coherence) up to two orders of magnitude (Figure 2) with respect to the corresponding sequential processing implementation, particularly critical when dealing with very huge DInSAR datasets.  A detailed analysis of the performances of the proposed implementation through widely used metrics (such as speedup, efficiency, and load balance) is still in progress, and the preliminary results achieved over a dataset of 40 COSMO-SkyMed SAR images acquired from ascending orbits over the city of Roma (Italy) are very promising, demonstrating that the proposed solution can be particularly relevant for the continuous monitoring of complex deformation phenomena over large urban areas, as well as for the development of preservation strategies for the archaeological sites and historical buildings all over the world.

References

[1]     Berardino, P., Fornaro, G., Lanari, R., and Sansosti, E. (2002). A new Algorithm for Surface Deformation Monitoring based on Small Baseline Differential SAR Interferograms. IEEE Trans. Geosci. Remote Sens, 40, pp.2375-2383.

[2]     Bonano, M., Manunta, M., Marsella, M., and Lanari, R. (2012). Long-term ERS/ENVISAT deformation time-series generation at full spatial resolution via the extended SBAS technique. Int J Remote Sens 33:4756–4783

[3]     Lanari, R., Mora, O., Manunta, M., Mallorquí, J.J., Berardino, P., and Sansosti, E. (2004a). A small baseline approach for investigating deformations on full resolution differential SAR interferograms. IEEE Trans. Geosci. Remote Sens., 42, 1377-1386.

[4]     Bonano, M., Manunta, M., Pepe, A., Paglia, L., and Lanari, R. (2013) From previous C-Band to New X-Band SAR systems: assessment of the DInSAR mapping improvement for deformation time-series retrieval in urban areas. IEEE Trans Geosci Remote Sens 51 (4):1973–1984

[5]     Sansosti, E., Berardino, P., Bonano, M., Calo`, F., Castaldo, R., Casu, F., Manunta, M., Manzo, M., Pepe, A., Pepe, S., Solaro, G., Tizzani, P., Zeni, G., and Lanari, R. (2014). How second generation SAR systems are impacting the analysis of ground deformation. Int J Appl Earth Obs Geoinf 28:1–11. doi:10.1016/j.jag.2013.10.007.

[6]     Arangio, S., Calo`, F., Di Mauro, M., Bonano, M., Marsella, M., and Manunta, M. (2013). An application of the SBAS-DInSAR technique for the assessment of structural damage in the city of Rome, Struct Infrastruct Eng Maint, Manag Life-Cycle Des Perform: 1–15

[7]     Scifoni S., Bonano M., Marsella M., Sonnessa A., Tagliafierro V., Manunta M., Lanari R., Ojha C., and Sciotti, M., (2016). On the joint exploitation of long-term DInSAR time series and geological information for the investigation of ground settlements in the town of Roma (Italy). Rem. Sens. Env, 182, 113–127, doi:10.1016/j.rse.2016.04.017

[8]     Zeni, G., Bonano, M., Casu, F., Manunta, M., Manzo, M., Marsella, M., Pepe, A., and Lanari, R. (2011). Long term deformation analysis of historical buildings through the advanced SBAS-DInSAR technique: the case study of the city of Roma Italy. Journal of Geophysics and Engineering, 8, S1 doi:101088/1742-2132/8/3/S01.

[9]     Torres, R., Snoeij, P., Geudtner, D., Bibby, D., Davidson, M., Attema, E. , Potin, P., Rommen, B., Floury, N., Brown, M., Navas Traver, I., Deghaye, P., Duesmann, B., Rosich, B., Miranda, N., Bruno, C., L'Abbate, M., Croci, R., Pietropaolo, A., Huchler, M., Rostan, F. (2012), GMES Sentinel-1 mission, Remote Sens. Environ., 120, 9-24.

3.04.b   Methodology and Techniques - DInSAR