InSAR with Sentinel-1 (2)

Sentinel-1A (S1-A) is the first of a family of satellites designed to provide a satellite data stream for the European environmental monitoring program Copernicus, previously known as GMES (Global Monitoring for Environment and Security) [1]. In particular, Sentinel-1 constellation has been specifically designed to make available, over land, advanced Differential Interferometric Synthetic Aperture Radar (DInSAR) products to analyze and investigate Earth’s surface displacements [2].

The satellite, which has been launched on April 3, 2014, provides scientific community with C-Band SAR data collected in continuity with the first generation ERS-1/2 and ENVISAT missions, guaranteeing further enhancements in terms of revisit time, coverage, timeliness and reliability of service. In particular, C-band S1-A SAR Instrument is designed to operate with four radar imaging modes, with dual polarization capability:

• Stripmap Mode (6 selectable swathes of 80 km each, 5 m x 5 m resolution)

• Interferometric Wide Swath Mode (250 km swath, 5 m x 20 m resolution)

• Extra-Wide Swath Mode (400 km swath, 20 m x 40 m resolution) and one single polarization ocean mode (HH or VV)

• Wave Mode (20 km x 20 km vignettes, 20 m x 5 m resolution).

Of particular interest is the new Terrain Observation with Progressive Scans (TOPS) acquisition mode [1] by means of which S1-A Interferometric Wide Swath (IWS) scenes, composed of three sub-swaths, are being collected. TOPS mode is quite similar to ScanSAR, since during the acquisition time the antenna beam is switched cyclically among the different sub-swaths. However, unlike ScanSAR, in the TOPS mode the sensor scans the image with very long bursts by rotating the antenna throughout the acquisition from backward to forward, resulting (opposite with respect to spotlight mode) in a worsening of the azimuth resolution. Such a new acquisition mode has led to the necessity to develop original, and innovative, processing chains for the extraction of Earth Observation information from SAR data (with particular emphasis on image co-registration and interferogram generation procedures).

This work is aimed at describing the development of an advanced and efficient interferometric processing chain, based on the well-known DInSAR algorithm referred to as Small BAseline Subset (SBAS) [3], for the generation of S1-A IWS deformation time-series. In this framework, the pursued strategy strongly takes into account the data acquisition characteristics of the TOPS mode. Indeed, IWS scenes contain one image per each sub-swath, consisting of a series of bursts that can be considered as separate acquisitions. This makes processing inherently parallel, approaching also problems connected with the use of SBAS-DInSAR chain in operational contexts, where dealing with large amounts of data represents a challenging task. Indeed, processing of large size datasets requires cutting-edge technological solution during all stages of SBAS-DInSAR processing chain for generation of deformation measurements.

The developed S1-A SBAS-DInSAR processing chain fully benefits from the S1-A burst partitioning data structure. Hence, the co-registration step is performed burst by burst through a coherent cross-correlation method followed by geometrical registration operation. Moreover, in order to further reduce misregistration errors on a sub-pixel basis, spectral diversity techniques are efficiently used in the overlapped intra-burst regions [4].

Once a sequence of differential SAR interferograms is obtained, they are then used to compute deformation time-series applying the SBAS algorithm. It basically consists in the cascade of the Phase Unwrapping (PhU) operation and the inversion of unwrapped interferograms through the Singular Value Decomposition (SVD) method. It is worth remarking that, since S1-A orbital tube is expected to be very narrow, small baseline interferograms will be mostly generated, thus making the application of SBAS approach particularly suitable. Moreover, being small baseline interferograms less affected by decorrelation noise effects, also phase unwrapping operation is also expected to be less critical. In particular, PhU step is performed by using Minimum Cost Flow (MCF) approach applied both in the time and space domains [5],[6]. The implemented processing chain also envisages an additional operational mode to be profitably used to “append” new interferograms to already processed DInSAR archives, as soon as new S1-A acquisitions are available. This processing strategy is particularly interesting for Sentinel-1 constellation that has recently started to image the Earth, for which the number of acquisitions on an area of interest is expected to grow up on a weekly basis. It means that there is no need for each new acquisition to run the processing chain from the very beginning; accordingly, updated deformation time-series can be available (using the “append” mode) in a reduced processing time.

The developed SBAS-DInSAR chain has been tested both on single interferometric SAR data pairs acquired by S1-A satellite during these first months of operation (see S1-A interferogram shown in Figure 1) and on TOPS RadarSAT-2 interferometric dataset acquired over Mexico City (Mexico) from April to November 2013. The achieved results clearly demonstrate the capability of the developed processing chain to effectively provide deformation maps relevant to wide areas in reasonable time frames by exploiting SAR data acquired with TOPS mode.

References

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

[2]  Rucci, A. Ferretti, A. M. Guarnieri, and F. Rocca, "Sentinel 1 SAR interferometry applications: The outlook for sub millimeter measurements", Remote Sens. Environ., vol. 120, pp. 156-163, May 2012.

[3]  Berardino, P., Fornaro, G., Lanari, R., Sansosti, E., “A new Algorithm for Surface Deformation Monitoring based on Small Baseline Differential SAR Interferograms”, IEEE Trans. Geo. Rem. Sens., 40, 11, 2375-2383, 2002.

[4]  R. Scheiber and A. Moreira, “Coregistration of Interferometric SAR Images Using Spectral Diversity” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 5, September 2000.

[5]  M. Costantini, “A novel phase unwrapping method based on network programming”, IEEE Trans. Geosci. Remote Sens., vol 36, no 3, pp 813-821, May 1998

[6]  A. Pepe and R. Lanari, “On the extension of the minimum cost flow algorithm for phase unwrapping of multitemporal differential SAR interferograms”, IEEE Trans. Geosci. Remote Sens., vol. 44, no. 9, pp. 2374–2383, 2006.

The launch of Sentinel-1A in April 2014 was a pivotal moment in satellite radar interferometry. Previous missions were not specifically designed for measuring surface displacements, only acquiring images monthly, at best, and we have been fortunate to make any progress with data from them. Sentinel-1 is a 20-year program that will see, from 2016 onwards, a constellation of 2 satellites in orbit at all times. Together, these satellites will acquire data over all the tectonic areas of the planet every 3 days on average, including ascending and descending passes, a tenfold improvement over previous missions that will transform our capacity to monitor our hazardous planet.

The systematic acquisitions and free and open data policy for Sentinel-1 provide new opportunities for routine analysis and operational applications of InSAR. We have been funded by the UK Natural Environment Research Council through COMET (*) and the LiCS (**) project to provide processed results to the whole community for all the tectonic and volcanic regions of the planet. We will provide products including simple interferograms, time series, and regional velocity/strain maps, and use these to systematically derive products including earthquake source models, seismic hazard maps, and volcanic deformation alerts.

The data volumes and continuous long-term observations offered by the Sentinel-1 constellation demand a new approach to InSAR data processing. We have started to develop a new time series processing approach, with the aim of efficiently and continuously processing the data to deliver continuously-updated ground deformation products with the highest possible accuracy. Our target is to measure a relative line-of-sight velocity of 1 mm/yr between two points separated by 100 km (10 nanostrain/yr), which is a comparable accuracy to existing sparse regional GNSS networks. To be able to achieve this on a regional scale, our proposed approach requires the development of an automatic system that integrates time-series estimation methods with routine atmospheric correction products and refined orbits.

In this presentation we describe the main stages of our processing system: 1) the precise coregistration of TOPS (Terrain Observation with Progressive Scans) SAR images; 2) the formation of interferograms with short temporal baselines; 3) the identification of coherent pixels; 4) adaptive filtering and phase unwrapping; 5) calculation of time series; 6) incorporation of atmospheric models, with their uncertainties; 7) evaluation of uncertainties and the detection of blunders. Ultimately, our system will use a data assimilation approach to ingest new data into the time series. Each new acquisition will only influence the most recent part of the displacement history for a point, and hence reprocessing of the entire time series will only be required if there are major improvements in the processing chain.

We will present the current status of the COMET/LiCS processing system. We will show preliminary results obtained for the creeping section of the North Anatolian Fault at Ismetpasa and for Mexico City, where there should be sufficient Sentinel-1 data to detect ground movement by the time of the meeting.

*COMET is the UK NERC's Centre for the Observation and Modelling of Earthquakes, Volcanoes and Tectonics. http://comet.nerc.ac.uk

**LiCS is a NERC large grant, funding much of the development work for this processing system: http://comet.nerc.ac.uk/current_research_look_inside.html

Interferometry is a technique well known for its complexity and usability. Since it's beginnings, and especially with introduction of time-series concepts for deformation monitoring, an always pending questions is how accurately InSAR can measure and capture a deformation phenomena of interest. Inherently, the quality control of InSAR is also difficult. The lack of a well-defined scattering center, absence of redundant measurements, and the phase ambiguity resolution uncertainties, are just a number of issues restricting a straightforward quality assessment. Only using a qualitative heuristic approach the quality standards of InSAR technique can be estimated.

When talking about InSAR validation, usually the corner reflectors (CR) are being resorted to. Knowledge of their phase center, and usually monitoring of their vertical movement with additional in-situ techniques (e.g., leveling, GPS) makes them extremely useful, and perhaps the only option, for an independent technique validation. A number of validation efforts are reported in the scientific literature. Although valuable, considering a TOPS mode and bursted data, and spatial extent of Sentinel-1 (S-1), a straightforward repetition of previous experiments has to be handled with great care since it might lead to over/under-estimation interferometric phase quality.

In this contribution we build on top of all the previous experiences and we extend them to S-1. Specifically we are running a number of campaigns in parallel that are optimized for S-1. The first campaign is conducted in Poland, on the three test-sites (Babiak, Lewino, Berejow). The other one is performed in Norway, on the Nordnes site, and it is an extension of an already on-going validation campaign to S-1. The both campaigns are long term efforts.

In Poland, the campaign is designed having in mind a validation activities of relevance for the shale gas exploration. Specifically, the sites are equipped with 20 corner reflectors (in total 60). Their CR control networks cover the area of approx ~5x5 km, and are being routinely monitored by high-precision leveling and GPS techniques. In addition an atmospheric and permanent GPS stations are deployed in proximity of the sites, as well as an information related to the ground water and overall deformation activity in the area is collected. In designing the experiment we fully utilized the spatial coverage of S-1 by having multiple sites in the same data take, with approx distance of 150 km between them, allowing to run a different type of integration and validation experiments within a single S-1 scene.

The Nordnes site in Norway, is designed for monitoring and validation of InSAR for rockslides. It is much smaller in terms of the number of CRs and its spatial extent wrt Polish sites. But it is equally well monitored by in-situ. However, importantly, in the contrast to the Polish sites, that are located mainly on the moderately flat terrain, in Nordnes the height difference between CRs is a couple of hundred of meters - CRs are distributed along the slopes of fjords. This completely different terrain configuration, and type of deformation phenomena gives an additional dimension to our validation efforts, since impact of the vertically stratified atmosphere on interferometric phase can be studied and quantified.

This contribution will discuss in detail on methodologies used in the validation campaigns. We will report in detail on the initial validation results, and achievable quality standards of Sl-1 for the deformation monitoring applications. We will also elaborate on the stability of the annotated system parameters of S-1 (e.g., stability of the local oscillator). By utilizing multiple CR sites in the same data takes we will assess the quality of orbits, and discuss on the integration strategies in the context of bursted data.

In addition, considering that all sites are also routinely monitored by TerraSAR-X and RADARSAT-2, we will report on the shoot-out between the different sensors and ground truth data for deformation monitoring applications.

The standard imaging mode of Sentinel-1 is the Interferometric Wide Swath (IWS) mode. This wide swath mode is based on principles of Terrain Observation by Progressive Scans (TOPS), and has similar coverage and resolution properties as the ScanSAR mode, but operates without any scalloping effects. Due to its coverage, the TOPS mode is appealing for the interferometric monitoring applications. However, the steering of the antenna along the track during the acquisition of a burst, results in signal properties that require the careful review and optimization of the established InSAR processing algorithms. Due to this antenna steering the signal has an azimuth varying Doppler centroid. This variability is much larger than the Pulse Repetition Frequency (PRF) on the burst edges, and as a consequence it significantly increases the requirements for the coregistration accuracy, typically 1/100th of pixel or even better. If the coregistration is not accurate, phase jumps are introduced between bursts.

The coregistration of TOPS data can be viewed as an extension and better implementation of the conventional approaches used for the stripmap and spotlight modes. Approaching the problem more carefully, we can reduce the problem of coregistration to the problem of estimating the baseline error as a function of time along the entire dataset. With availability of precise orbits, this error is typically constant or very slowly varying, such that only a very small set of parameters needs to be estimated.  An additional major consideration is that if bursts are coregistered individually using a conventional approaches, the residual orbital error will in general be discontinuous across each single burst boundary. This inherently leads to a single set of error parameters to be estimated for each single burst overlap due to the sensitivity to azimuth coregistation errors.

Obviously, all aformentioned leads to the geometric approach, where the available precise orbits are first used to resample a number of patches, and then to subsequently estimate two out of the three dimensions (slant range and azimuth) of a constant baseline error by patch cross-correlation. The key point is that according to the geometric model, all patches at a constant slant range have offsets with very slow variation in azimuth. Thus, patch correlation functions of each slant range might be averaged/smoothed in azimuth, leading to a very precise estimate.

The major steps of the proposed solution are:

Assume master orbit is correct (even tens of meters absolute error does not change relative phase much, so this is only important for absolute geolocalization).

Correct each slave orbit by geometric approach, assuming that all of the baseline error is due to slave.

Importantly, due to the high sensitivity to azimuth coregistration errors at the burst edges, the quality of the performed coregistration can be quantified by utilizing the phase difference in coherent overlap areas. Any bias detected by this quality measure, might be used to further refine the coregistration in azimuth.

We will elaborate on the proposed coregistration algorithm in detail, demonstrate its performance on a number of Sentinel-1 interferometric pairs, and assess its reliability on a number of interferometric data stacks.

In this contribution, we discuss issues related to time series interferometry that stem from the bursted nature of TOPS data. The coregistration problem for single Sentinel-1 bursts is discussed elsewhere, and not addressed here. After successful coregistration, there is still a number of open questions and algorithmic options for the subsequent interferometric time series analysis.

Independently from the type of time series analysis to be performed (e.g., SBAS, PSI, integrated SBAS/PSI, etc), two burst stacking strategies, in space and time respectively, can be envisaged:

First process, then stitch -- “stitching in the deformation measurement domain”.

First stitch, then process -- “stitching in the phase domain”.

While the choice obviously depends on the application, and the extent of the area of interest, we outline general considerations from the processing and algorithmic development perspective. Note that for both strategies, we are considering only the azimuth direction. Because of less scattering diversity and larger overlap, it is likely to be straightforward to stitch in the range direction by direct phase comparisons, in particular if the azimuth stitching is already successfully carried out.

Two strategies can be in short summarized as:

When stitching in the deformation measurement domain (the first strategy), each stack of bursts is treated independently. The results per burst stack results are then merged in the final processing step, similar to an existing Wide Area Processing (WAP). Processing of single burst stacks is straightforward and frameworks for spotlight InSAR can be directly reused. Nevertheless, dependency on the reliable unwrapping for successful stitching might be a major issue in scenes with many disjoint coherent subareas. For example, due to the short length of a burst in the azimuth direction (about 20 km), a completely decorrelated zone extending over the entire azimuth direction of a burst is more likely to happen (e.g. a big forest or a fjord). This inherently leads to a difficult and error prone situation for both unwrapping and consequently estimation of the Atmospheric Phase Screen (APS).

When stitching in the phase domain (the second strategy),  the number of coherent areas that are divided by a burst boundary is minimized. This strategy benefits unwrapping and APS estimation, i.e., stitching is done before either. For multilooked interferograms, the procedure is rather straightforward. In fact, this step might have already been done in the process of coregistration in order to reach the azimuth coregistration accuracy necessary to avoid phase jumps between bursts. However, for the full resolution applications (PSI, etc), pixel characterization is an issue. The point target clouds from each of the overlapping bursts might be different due to the different view angle, thus a statistical approach might be necessary rather than direct phase comparison.

Obviously, either strategy has its strong and weak points. We will discuss and elaborate this in more detail, and propose a general processing strategy that is building on strengths of both approaches. For example, a “too big interferogram” will be challenging to handle as a full unit, so there is still room to stitch to a set of partial interferograms in phase domain, e.g. each swath independently, and then exploit an existing WAP algorithms to merge the swaths in deformation measurement domain.

Finally, a relevant question is whether stitching of complete IW scenes is needed at all. For many applications, analysis of the full 250 km x 200 km is not required. We argue that no matter how small your area of interest is, it might be located on a burst boundary. Thus, even in this case the challenges addressed in this contribution apply, though on a smaller scale.

In summary, it all boils down to how to identify and connect regions small enough to be easily handled, and large enough to avoid unnecessary problems with unwrapping and APS estimation. Often there is a thin line between “big enough” and “too big”. We will try to indicate where to draw this line.

During the round table, seed questions proposed by the chairs will be discussed with the audience.