PSI and DInSAR (3)

Abstract

The availability of long time series of interferometric data acquired all over the world from several synthetic aperture radar (SAR) satellite missions make possible to perform a worldwide assessment of the terrain and infrastructure stability by persistent scatterer (PS) SAR interferometry techniques. This technology is computationally demanding, in particular because it requires a 3D processing. When applied to large areas several problems have to be faced to handle huge amounts of data. In this work we present a significant example of PS big data processing performed at national scale (the whole Italian territory) with ERS, Envisat and COSMO-SkyMed data, and the main challenges/problems (and relative solutions) related to such objective, and to the possible worldwide extension with Sentinel data.

Index Terms— SAR, persistent scatterer interferometry, big data, big data, ground deformation

1. Introduction

Almost two decades ago the availability of long time series of interferometric data acquired all over the world from synthetic aperture radar (SAR) sensors (ERS 1-2, Envisat, Radarsat-1) called for the development of techniques able to exploit these series of data for measuring millimetric surface deformations. In particular, since the introduction of the fundamental ideas [1-2], persistent scatterer (PS) SAR interferometry has been successfully developed, validated and applied to detect and monitor slow ground displacements due to subsidence, landslides, earthquake and volcanic phenomena, and different PS interferometry techniques have been developed [3-5], wheras the more classical approach of working with small baseline interferometric pairs was developed and brought to maturation [6].

In the last few years, the new high-resolution SAR sensors like COSMO-SkyMed have opened the possibility of measuring the surface deformations with very high detail, and led us to developed the persistent scatterer pair (PSP) SAR interferometry technique to fully extract the available information and collect a huge number of PS measurements, both on natural terrains and structures [7-8].

In the next future, the availability of the Sentinel-1 mission recently launched will make possible to perform a worldwide assessment of the terrain and infrastructure stability.

PS interferometry is computationally demanding, in particular because it requires a 3D processing of a 3D set of data (a stack of 2D images). When applied to large areas several problems have to be faced to handle huge amounts of data. In this work we present a significant example of PS interferometry big data processing, performed with ERS, Envisat and COSMO-SkyMed data, to obtain a database of surface deformation measurements from interferometric SAR data over the whole Italian territory.

This ambitious task has required developing the most advanced PS interferometry processing algorithms and suitable procedures to manage hundreds of interferometric data-stacks, thousands of SAR images and billions of PS measurements, and represents a pioneering service for mapping and preventing geo-hazards.

The capability of the new C-Band Sentinel-1 sensor to collect interferometric SAR data globally with a short revisit time will open the possibility of monitoring the surface deformations worldwide and routinely, and will pose new challenges that can be afforded by exploiting cloud based models.

In the following section we will summarize the overall project, the main challenges/problems (and relative solutions) related to such objective, the results already obtained and the work still in progress. Then, we briefly discuss the possible worldwide extension with Sentinel data, and we draw some conclusions.

2. PS SAR interferometry analysis of
the whole italian territory

Starting from 2008, as part of a program called “Special Plan of Remote Sensing,” with the objective of mapping and preventing geo-hazards, the Italian Ministry of the Environment awarded to an industrial team lead by e-GEOS a series of contracts aimed at realizing a database of surface deformation measurements by PS SAR interferometry processing of all the available ERS and Envisat SAR data over Italy, and then at updating the PS measurement database based on COSMO-SkyMed data stacks. The project, referred as PST-A, is the first PS SAR interferometry project at national scale.

The PS processing of the whole ERS and ENVISAT of SAR images (over 10,000 images) acquired over Italy from 1992 up to 2010 was completed in 2013. The obtained products are made available through the national cartographic portal (http://www.pcn.minambiente.it
/GN/progetto_psi.php?lan=en) via WebGIS technology. A global view of the PS surface displacement measurements is reported in Fig. 1.

The PS  analysis of the ERS and ENVISAT interferometric SAR data till 2010 brought to the identification of about 14 million PS points with ERS data and 28 million PS points with ENVISAT data. For each PS point, a set of deformation measurements corresponding to every acquisition date is obtained, for a total of more than 1 billion deformation measurements in the analyzed period (1992–2010).

Currently, one hundred high-resolution interferometric data stacks acquired in 2010–2014 from the COSMO-SkyMed constellation, for a total of about 5,000 images are being processed, in order to update the Italian PS measurement database. The area covered in this phase of the project with high-resolution COSMO-SkyMed SAR data is reported in fig 2.

With the high-resolution COSMO-SkyMed SAR data, the PS measurements density increase by two orders of magnitude with respect to those obtained with ERS and Envisat data. In Fig. 3 and Fig. 4 two examples of comparison between the PS mean velocity surface displacements achievable with ENVISAT and Cosmo-SkyMed are reported. In particular Fig. 3 refers to an urban area, the city of Venice, Italy, and Fig. 4 refers to a rural area close to the city of Palermo, Italy. The density and the accuracy of the measurements obtained with the high-resolution Cosmo-SkyMed SAR data make possible the study of the stability of each single building and infrastructure and the analysis of landslides phenomena that were not observable with the old ERS and Envisat low-resolution SAR systems.

The extension to hundreds of thousands of square kilometers of this detailed analysis requires dedicated high performance computing (HPC) in order to manage, process and make accessible this big amounts of data.

Moreover, in the next future, with the capability of Sentinel-1 to cover regularly the globe with a short revisit time, it will be in principle possible to extend the area of interest to the entire world, providing routinely PS SAR interferometry displacement measurements. This scenario call for cloud based paradigms that will be necessary both to process such an amount of data and to make accessible to the users this product in a convenient way.

 
4. Conclusion

PS SAR interferometry proved to be an effective technology for measuring slow ground deformations due to subsidence, landslides, earthquakes and volcanic phenomena, and in general to monitor terrain and structure stability.

The project described in this work demonstrates that this cutting edge technology is an operative tool for mapping and preventing geo-hazards. On the counter side, this work also shows that the needs coming from large scale applications and big data availability further stimulate the research and the development of improvements to the PS technology.

The computational effort to process and manage this big amount of data is very demanding. Considering the high-resolution of the X-Band sensors, and the capability of the new C-Band Sentinel-1 sensor to collect interferometric SAR data globally with a short revisit time, cloud based paradigms must be considered in order to both produce and make conveniently accessible to the users PS SAR interferometry displacement measurements at worldwide scale.

10. References

[1] A. Ferretti, C. Prati, and F. Rocca, “Permanent scatterers in SAR interferometry,” IEEE Trans. Geosci. Remote Sensing, vol. 39, no.1, pp. 8-20, Jan. 2001.

[2] A. Ferretti, C. Prati, and F. Rocca, “Non-linear subsidence rate estimation using permanent scatterers in differential SAR interferometry,” IEEE Trans. Geosci. Remote Sensing, vol. 38, no. 5, pp. 2202-2212, Sept. 2000.

[3] A. Hooper, H. Zebker, P. Segall, and B. Kampes, “A new method for measuring deformation on volcanoes and other natural terrains using InSAR persistent scatterers.” Geophys. Res. Lett., vol. 31, no. 23, L23611, Dec. 2004.

[4] M. Costantini, S. Falco, F. Malvarosa, F. Minati, “A new method for identification and analysis of persistent scatterers in series of SAR images,” in Proc. Int. Geosci. Remote Sensing Symp. (IGARSS), Boston MA, USA, pp. 449-452, 7-11 July 2008.

 [5] A. Ferretti et al., “A new algorithm for processing interferometric datastacks: SqueeSAR,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 9, pp. 3460–3470, Sep. 2011.

[6] P. Berardino, G. Fornaro, R. Lanari, and E. Sansosti, “A new algorithm for surface deformation monitoring based on small baseline differential SAR interferograms,” IEEE Trans. Geosci. Remote Sens., vol. 40, no. 11, pp. 2375–2383, Nov. 2002.

[7] M. Costantini, F. Minati, F. Trillo, and F. Vecchioli, “Enhanced PSP SAR interferometry for analysis of weak scatterers and high definition monitoring of deformations over structures and natural terrains,” in Proc. Int. Geosci. Remote Sens. Symp. (IGARSS), Melbourne, Australia, Jul. 2013, pp. 876–879.

 [8] M. Costantini, S. Falco, F. Malvarosa, F. Minati, F. Trillo, F. Vecchioli, “Persistent scatterer pair interferometry: approach and application to COSMO-SkyMed SAR data,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 7, no. 7, July 2014.

P-SBAS Service within ESA G-POD Platform for Unsupervised On-demand DInSAR Processing

De Luca C.^1,4, Cuccu R.^2,3, Elefante S.¹, Manunta M.¹ , Zinno I.¹, Rivolta G.^2,3, Casola V.⁴, Lanari R.¹, Casu F.¹

1) IREA – CNR, via Diocleziano 328, 80124 Napoli (Italy)

2) ESA Research and Service Support, via Galileo Galilei, 1, 00044 Frascati (Italy)

3) Progressive Systems Srl, Parco Scientifico di Tor Vergata, 00133 Roma (Italy)

4) Università degli studi di Napoli Federico II, DIETI, via Claudio 21, 80124 Napoli (Italy)

ABSTRACT

The current Remote Sensing scenario is characterized by the increasing number of applications related to natural (volcanoes, earthquakes, landslides) and man-made (urban and infrastructure monitoring) hazards that make a large use of satellite Synthetic Aperture Radar (SAR) data. Such a request is satisfied by the huge archives collected in the last 20 years by both ERS-1/2 and ENVISAT missions, operating at C-band, as well as by the X-band SAR sensors, including the Italian COSMO-SkyMed (CSK) and the German TerraSAR-X (TSX) constellations, which are still in operation. Moreover, a massive and ever increasing data flow will be further supplied by the Copernicus Sentinel-1 SAR mission, whose first satellite has been recently launched (April 2014), which is characterized by both a “free and open” access data policy and a global coverage acquisition strategy.

In this context, the availability to the scientific community of algorithms and tools suitable to effectively exploit huge SAR data archives for generating value added spaceborne products, is becoming more and more crucial.

Nowadays, Differential SAR Interferometry (DInSAR) is one of most used technique for the investigation of Earth's surface deformation phenomena from SAR data. Such a technique permits, indeed, to retrieve ground deformation maps with centimeter to millimeter accuracy [1] starting from SAR scenes relevant to the same area of interest acquired at different epochs. Originally, DInSAR was developed to analyze single deformation episodes [1] and afterwards evolved towards the study of the temporal evolution of the detected deformations, thanks to the availability of above mentioned SAR data archives and of the so-called multi-temporal DInSAR algorithms, able to generate deformation time series relevant to an observed area [2],[3].

Among several, a largely used multi-temporal DInSAR algorithm is the one referred to as Small BAseline Subset (SBAS) [3]. The SBAS approach relies on the use of a large number of SAR acquisitions and implements an easy combination of small spatial and temporal baseline interferograms generated from these data in order to mitigate the noise effects (referred to as decorrelation phenomena), thus maximizing the number of reliable measure pixels.

Recently, an advanced parallel computing solution of the SBAS algorithm, referred to as P-SBAS, able to effectively process big SAR data archives by making use of large computing resources, has been presented [4]. P-SBAS, which implements the whole SBAS processing chain, i.e., from raw data focusing up to displacement time series generation, makes use of different parallelization strategies and is able to exploit both multi-node and multi-thread computing architectures.

In the depicted scenario, the possibility to connect such kind of algorithms, as the P-SBAS one, and scalable distributed computing infrastructures (e.g. Grid and Cloud Computing (CC) platforms) together with the storage location of the SAR data archive, is fundamental for the generation of effective and high performance services for Earth Observation (EO) community.

To meet these needs, the ESA’s Grid Processing On Demand (G-POD) [5] provides fast access to large data volumes, as those of the ESA’s Virtual Archive 4, as well as an operative processing environment where specific applications that make use of EO data can be plugged into. Each application is encapsulated in a virtual environment within G-POD and can exploit distributed high-performance computing resources as well as a large volume of EO data to provide the scientific community with new EO-based services.

In this work, we present the implementation of P-SBAS algorithm within the G-POD platform together with the set-up of an operational service (see Figure 1), which is addressed also to users that are not expert in advanced interferometric processing, for the generation of SBAS displacement time series.

The challenge for the deployment of P-SBAS within G-POD has been to develop an algorithm able to work in a totally unsupervised and automatic way. In particular, the user's tasks are limited to the selection of the SAR images that have to be processed from ESA archives and to the set-up of some interferometric parameters, e.g., the maximum allowed spatial and temporal baselines and the reference stable area. After these few actions the user can i) submit and run the P-SBAS algorithm, ii) check the ongoing processing and iii) wait for the results that will be available for download from G-POD web portal, typically in a short time frame. The service is fully automatic and does not need any user interaction during the processing phase. To achieve such a result, the P-SBAS chain has been strongly improved in terms of efficiency and robustness. In particular, specific algorithms have been implemented to automatically retrieve the illuminated area that is common to all the input SAR acquisitions. The reference geometry (master image) is automatically selected to minimize the error sources during the co-registration step, which is performed by using orbital and topographic information and spatial coherence maximization criteria. The interferometric pairs are chosen in unsupervised way in order to guarantee as much as possible high coherence values in the corresponding interferograms. Finally, the Phase Unwrapping (PhU) step and the selection of the reliable measure points are carried out by automatically selecting spatial PhU network and the coherence thresholds on the basis of the characteristics of the processed DInSAR dataset.

Wide test and validation activities have been carried out to assess the performances of the P-SBAS processing chain within the G-POD environment as well as the quality of the provided service. To this aim, several datasets, acquired by ENVISAT satellite over different areas and spanning the 2003-2010 time interval, have been processed through the implemented P-SBAS G-POD service, whose main characteristics and elapsed processing time are shown in the attached Table 1. It is worth noting that the elapsed processing time strongly depends on the number of dataset images, the number of produced interferograms and the number of coherent pixels within the illuminated scene.

In particular, as shown in Table 1, the processing elapsed time for comparable datasets, in terms of both number of acquisitions and spatial coverage extension (that is of about 100x100 km, which corresponds to 25000x5000 full resolution SAR pixels) is, on average, approximately one day. This result demonstrates that the implemented P-SBAS service is able to generate in a very short time displacement maps and the corresponding time series from ENVISAT data sets, by properly exploiting the high-performance and sizeable computing resources provided by the G-POD platform.

The availability for the EO community of an on-demand DInSAR processing service based on the P-SBAS algorithm, which allows scientific users to generate in unsupervised way and in very short time ground displacement maps of large areas, will open new intriguing and unexpected perspectives to the comprehension of surface deformation dynamics at global scale.

REFERENCES

[1] D. Massonnet, M. Rossi, C. Carmona, F. Adragna, G. Peltzer, K. Feigl, and T. Rabaute, "The displacement field of the Landers earthquake mapped by radar interferometry," Nature, vol. 364, no.6433, pp. 138– 142, Jul. 1993.

[2] A. Ferretti, C. Prati, and F. Rocca, “Permanent scatterers in SAR interferometry,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 1, pp. 8–20, Jan. 2001.

[3] P. Berardino, G. Fornaro, R. Lanari, and E. Sansosti, "A new algorithm for surface deformation monitoring based on small baseline differential SAR interferograms," IEEE Trans. Geosci. Remote Sens., vol. 40, no. 11, pp. 2375–2383, Nov. 2002.

[4] F. Casu, S. Elefante, P. Imperatore, I. Zinno, M. Manunta, C. De Luca and R. Lanari, "SBAS-DInSAR Parallel Processing for Deformation Time-Series Computation," IEEE JSTARS, Vol.7, no. 8, Aug. 2014.

[5] P.G., Marchetti, et al., "A Model for the Scientific Exploitation of Earth Observation Missions: The ESA Research and Service Support," IEEE Geoscience newsletter, vol.162, pp. 10-18, 2012.

The current Earth Observation (EO) scenario is characterized by the availability of huge SAR data amounts such as those acquired during the course of the last 20 years by the long-term C-band ESA (ERS-1, ERS-2 and ENVISAT) and Canadian Space Agency (RADARSAT-1/2) satellites as well as those that are nowadays provided by the X-band generation SAR sensors, i.e. COSMO-SkyMed (CSK) and TerraSAR-X (TSX). Moreover, a massive and ever increasing data flow will be further supplied by the Copernicus (European Union) SENTINEL-1A SAR satellite, which has a repeat pass of 12 days that should be further reduced in 2016 with the advent of the SENTINEL-1B twin system [1]. Thanks to the free and open access Copernicus data policy and the global coverage acquisition strategy adopted by the SENTINEL program, new SAR data relevant to vast areas of the Earth surface will be soon available to the scientific community and they will be acquired on Land by using a SAR mode that has been specifically designed for Differential SAR Interferometry (DInSAR) applications [2].

DInSAR techniques, which are capable to generate Earth's surface deformation maps and displacement time series by using SAR data, are particularly suitable to exploit such long time data archives.

However, the effective processing of the above mentioned massive SAR data flow, within short time frames, requires not only large and powerful computing resources, but also efficient algorithms able to properly exploit such computing facilities. To provide a contribution towards this direction, a parallel version of the widely diffused SBAS-DInSAR algorithm, namely P-SBAS, has been recently proposed [3]. P-SBAS permits to generate Earth’s surface displacement maps and related time series by taking full benefit from parallel computing architectures, also including Cloud Computing (CC) environments. In particular, CC is progressively gaining consensus within the EO scenario [4], [5], [6], since it provides highly scalable and flexible architectures that, at a relative low cost, gives users the opportunity to process very large dataset by appropriately customizing the huge computing resources at their disposal. Moreover, the increasing availability of public Cloud infrastructures (such those provided by Amazon, Google, Microsoft, Helix Nebula), and their easy-to-use nature, thanks to advanced Application Programming Interfaces (API) and web-based tools, is further pushing towards the use of such a technology in scientific applications [7], [8].

Within such a framework, in this paper we present and discuss the integration of the P-SBAS DInSAR processing chain within a Cloud Computing environment, focusing on both the scalable performance assessment and the major advantages and drawbacks analysis.

The integration has been carried out within the Amazon Web Services (AWS) cloud platform by building up a proper virtual cluster based on customized Virtual Machines connected in virtual private cloud mode, thus resulting in a logically isolated section of AWS.

An experimental analysis has been undertaken by exploiting a real interferometric data set acquired over the Naples' bay area and composed by 64 ENVISAT SAR data frames spanning the 2003–2010 time interval. Such an analysis is aimed at evaluating the P-SBAS scalable performance achieved on cloud.

The preliminary results show that the P-SBAS algorithm is well suited to process ENVISAT datasets within Cloud environments in relatively short times and at reduced costs. In particular, the times elapsed to run the P-SBAS processing chain on AWS cloud by exploiting an increasing number of nodes (up to 16 nodes) have been computed, together with the associated Cloud costs. The P-SBAS processing times passed from 41 to less than 7 hours when using 1 and 16 nodes, respectively, with a correspondent cost ranging approximately from 110 to 245 USD. Further details can be found in the provided Table 1.

Moreover, in order to quantitatively evaluate the scalable performances of the P-SBAS processing chain, the behaviour of the achieved speed up [9] as a function of the number of employed nodes, together with the corresponding Amdahl's law, which represents the theoretical maximum possible speed up, have also been computed. Even in this case, the achieved performances are definitely satisfactory since the speed up curve is very close to the Amdahl's law, as clearly shown in Fig. 1. In particular, the maximum deviation between the Amdahl's law and the actual speed up behaviour occurs in correspondence of the 16 nodes test case and is of about 17%.

Finally, the performed analysis provided insights on the issues that can be critical in the perspective of processing big amounts of SAR data, such as those coming from the existing COSMO-SkyMed or upcoming SENTINEL-1 constellations, when a large number of acquisitions, and therefore of processing nodes, has to be considered. In particular, a detailed analysis on the several aspects that can represent significant bottlenecks in this case, such as the Input/Output workload, the concurrent reading/writing operations and the network characteristics, will be provided.

References

[1]       S. Salvi, S. Stramondo, G. J. Funning, A. Ferretti, F. Sarti, and A. Mouratidis, "The Sentinel-1 mission for the improvement of the scientific understanding and the operational monitoring of the seismic cycle," Remote Sens. Environ., vol. 120, May 2012.

[2]       A. 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]       F. Casu, S. Elefante, P. Imperatore, I. Zinno, M. Manunta, C. D. Luca, and R. Lanari, "SBAS-DInSAR Parallel Processing for Deformation Time-Series Computation," Selected Topics in Applied Earth Observations and Remote Sensing, IEEE Journal of 2014.

[4]       P. Rosen, K. Shams, E. Gurrola, B. George, and D. Knight, "InSAR Scientific Computing Environment on the Cloud," in AGU Conference, San Francisco, 2012.

[5]       S. Elefante, P. Imperatore, I. Zinno, M. Manunta, E. Mathot, F. Brito, J. Farres, W. Lengert, R. Lanari, and F. Casu, "SBAS-DInSAR time series generation on cloud computing platforms," IEEE IGARSS, 2013, Melbourne, Australia, 2013.

[6]       Google. (2014). Google Earth Engine. Available: http://www.google.it/intl/it/earth/outreach/tools/earthengine.html

[7]       K. Yelick, S. Coghlan, B. Draney, and R. S. Canon, "The Magellan Report on Cloud Computing for Science," U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research (ASCR)December 2011.

[8]       O. Terzo, L. Mossucca, A. Acquaviva, F. Abate, and R. Provenzano, "A Cloud Infrastructure for Optimization of a Massive Parallel Sequencing Workflow," presented at the The 12th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid 2012), Ottawa, Canada, 2012.

[9]       H. El-Rewini and M. Abd-El-Barr, Advanced Computer Architecture and Parallel Processing, 2005.

InSAR time series techniques, such as persistent and distributed scatterer (PS/DS) time series in all their variations, have in common that they need to disentangle two highly correlated parameters: the quality of the time series (coherence) and the parameters of interest (topographic, kinematic and atmospheric signal and phase ambiguities). Quality estimates are required to select which of the billions of reflections in SAR images can be used to successfully estimate the parameters of interest. Especially the presence of the unknown integer phase ambiguities make this problem cumbersome. 

Since quality estimates and parameter estimates are not independent, all methods use heuristic solutions for this problem. Practically, the methods use harsh assumptions, either on the expected behavior of the signal in time (a certain ‘smoothness’ of the kinematic signal), or in space (a ‘smoothness’ between nearby points). This is mostly parameterized using temporal or spatial estimators of coherence.

Although there are several classes of problems where some of these assumptions seem to be fair, as the many successful InSAR applications in literature show, in its origin the dependency on such assumptions make the techniques extremely vulnerable. This has been the outcome of several (ESA) studies, where different expert groups processed the same data,  leading to completely different  results. More specifically, even a single expert can nowadays use different processing tools, and obtain significant differences in the results using the same input data. Nowadays, this is one of the main bottlenecks for the operational acceptance of InSAR time series by end users who are not radar experts.

The other bottleneck in the operational use of InSAR time series is the inability to analyze the extremely large datasets showing the kinematic behavior of points. Visualizations in terms of ‘velocity maps’ (essentially an aggregate variable based on the time series) are poor, as they do not  come close to convey the rich information  content in the data.

To address these problem, we start with postulating some guiding principles. First, there is no single assumption on signal behavior that satisfies all millions of InSAR measurement points, the world is too complex for this. Consequently, processed results of millions of locations will have inhomogeneous quality. Second, the InSAR data are obtained opportunistically, at times dictated by satellite orbits, and at positions provided by mere chance based on the size, shape and orientation of the targets. Third, due to the phase ambiguity problem, the estimated parameters are extremely correlated, and there are more solutions given an observed phase time history, with different degrees of likelihood.

Given these postulates, we propose a new probabilistic approach to tackle this problem. The approach is currently limited to post-processing of estimated InSAR time series results, but could be applied at different stages in the processing chain. The approach consists of two parts. In the first part, the estimated kinematic time series are  scrutinized as a whole. Most importantly, in this stage we  estimate the noise of the reference point, using the per epoch correlation between all arcs relative to the reference point. Then, the second phase consists of multiple model hypothesis testing. We established a library of 100+ potential temporal models that may be applicable to the time series of a specific point. We then start by defining a null-hypothesis, which is based on steady-state behavior, and compute the proper test statistic to decide whether the null hypothesis needs to be rejected. If this is the case, we  test the whole library of potential alternative models with different physical realistic parameters against the null hypothesis. The first  main innovation in this approach is that we account for the different degrees of freedom (or dimensions) of the various tests, depending on the amount of parameters that need to be estimated. This is based on Baarda’s B-method of testing. Secondly, for computing the test statistics for the 100+ alternative hypotheses, we present a computationally very efficient method, using only the residuals under the null-hypothesis, and an a priori defined matrix per alternative hypothesis. This way, it is not necessary to compute parameter results  for each individual alternative hypothesis. Finally, using test-ratio’s it is possible to select the most likely model for each point. We can evaluate hundreds of thousands of points within a few minutes on a standard computer.

In the figure we show results for real (persistent scatterer) data, identifying temperature dependent signals, breakpoint signals, unwrapping errors, exponential decays.

Finally, we are able to give likelihoods to the estimation results, on a point-by-point basis.

The method presented at Fringe is based on a submitted paper to IEEE TGARS

InSAR can be used to estimate double-difference scatterer displacements (in space and in time) based on complex-valued radar reflections stemming from these scatterers. When used in a time-series approach, the double-difference displacements can be generalized to, e.g., single-difference (space-only) velocity estimates, or displacement behavior parameterized via polynomial coefficients. Whatever the parameterization is, a key property is that the estimates are inherently relative. If we focus on changes in elevation (or in the radar line of sight), the estimates are always relative to a reference point with a particular location (or, equivalently, the average of scatterers within a reference area), and a reference epoch. The value assigned to the InSAR reference point can be arbitrary; often, for convenient interpretation of the displacement estimates, a 'conventional' value is chosen, e.g. zero displacement or velocity for a reference point situated in an area assumed stable. In these cases the value is deterministic and has zero standard deviation. When InSAR studies use a point or area 'assumed to be stable' as reference, the results are subject to this assumption. If the assumption is incorrect, i.e. the reference point is unstable in reality, the other estimates are essentially incorrect as well. 

Although effectively all geodetic positioning techniques are relative, there is sometimes the need for InSAR displacement estimates which can be considered to be 'absolute', i.e., estimates which can be expressed in a well-defined terrestrial reference frame (TRF) and related to the results of other techniques. In other words, whenever two or more sets of estimates stem from specific reference frames or datums, there is a need for datum connection. For this, the value for the InSAR reference point displacement can be connected to other measurements, e.g. Global Navigation Satellite Systems (GNSS). High-precision geodetic GNSS measurements can be used to derive the 'absolute' position (as a function of time) of the InSAR reference point in a global TRF. The displacement of the InSAR reference point in this TRF is then stochastic, and the errors can be propagated. Here we propose such a method for datum connection, to merge the results of InSAR surveys and studies referring to a global TRF with associated noise variance-covariance (VC) matrices, using the existing international or national networks of GNSS stations. 

An opportunistic approach for datum connection is to utilize the coherent InSAR scatterers that happen to occur in the vicinity of these GNSS stations as reference points, assuming they show the same displacement behavior as that measured by GNSS. However, this harsh assumption does not always hold in practice, because of local variations in displacements. Also, coherent InSAR scatterers can often result from multiple-bounce reflections, and could contain a component from ambient effects (e.g. swelling/compaction in the surrounding soil) which would not be measured by a well-founded GNSS receiver. Additionally, it is not guaranteed to have coherent InSAR scatterers in the vicinity of a GNSS station. 

In the proposed method, we mechanically attach phase-stable radar transponders to GNSS stations. This way, no assumptions are required regarding the relative motion between the transponder and the GNSS receiver, since their sturdy connection ensures that they both experience the same movement. We thus obtain a network of radar beacons with independently observed position time series in a TRF. As these reference GNSS stations are usually located spatially dense enough to have one or more stations in a typical radar image (170x250 km for Sentinel-1), there is an ’absolute’ InSAR reference point in each scene, connected to a TRF such as the International/European Terrestrial Reference Frame (ITRF/ETRF). The available VC matrices of the GNSS stations can be used when transforming relative InSAR displacement estimates to displacements with respect to the reference ellipsoid. Subsequently, the (time series of) vertical positions can then be referred to an international or national height system using the local state-of-the-art geoid. Existing VC matrices for these geoids can then be used for proper error propagation. 

An application of the proposed approach is to tie overlapping or non-overlapping radar datasets together, yielding datasets comparable over large distances, even across oceans and continents. Additionally, region-wide deformation can be linked to sea-level changes via collocated InSAR-GNSS measurements, thereby contributing valuable information, e.g. towards flood risk assessment. Another benefit of this approach can also be the correction for residual orbit errors in different radar datasets. Phase ambiguities can moreover be estimated in an absolute sense, as the connection via the GNSS station yields exact geometric ranges. The collocated transponder-GNSS approach can perform datum connection between InSAR survey results stemming from different satellites and looking directions, simply by programming the transponder accordingly. The regular, standardized and frequent acquisitions of the Sentinel-1 mission will also make the connection of all future radar observations to a TRF possible. 

Here we present (i) the mathematical approach for connecting InSAR to a TRF, (ii) experimental results with collocated InSAR-GNSS measurements, and (iii) a feasibility study for the implementation on a national scale, extensible to a European and a global scale. We have performed this study at selected permanent GNSS stations in the Netherlands using Radarsat-2 data. We focus in particular on the case of IJmuiden, with collocated InSAR and GNSS measurements at a tide gauge (Fig. 1). We show an 'absolute' displacement map of the area, and discuss in depth the various factors that should be considered for precise datum connection. 

Round Table