VALIDATION OF PERSISTENT SCATTERERS INTERFEROMETRY OVER A MINING TEST SITE: RESULTS OF THE PSIC4 PROJECT

Transcription

1 VALIDATION OF PERSISTENT SCATTERERS INTERFEROMETRY OVER A MINING TEST SITE: RESULTS OF THE PSIC4 PROJECT Crosetto M. (1), Agudo M. (1), Raucoules D. (2), Bourgine B. (2), de Michele M. (2), Le Cozannet G. (2), Bremmer C. (3), Veldkamp J.G. (3), Tragheim D. (4), Bateson L. (4), Engdahl M. (5) (1) Institute of Geomatics, Parc Mediterrani de la Tecnologia, Av. del Canal Olímpic, s/n E-08860,Castelldefels (Spain), michele.crosetto@ideg.es, marta.agudo@ideg.es (2) BRGM, Orleans (France), d.raucoules@brgm.fr; b.bourgine@brgm.fr; G.LeCozannet@brgm.fr (3) TNO, Geological Survey of the Netherlands, Utrecht (The Netherlands), chris.bremmer@tno.nl; hans.veldkamp@tno.nl (4) BGS, Nottingham,(United Kingdom), dgt@bgs.ac.uk; lbateson@bgs.ac.uk (5) ESA-ESRIN, Frascati (Italy), Marcus.Engdahl@esa.int ABSTRACT This paper describes the results of a Persistent Scatterers Interferometry (PSI) validation project. The project, which is named PSIC4 (Persistent Scatterer Interferometry Codes Cross-Comparison and Certification for Long-Term Differential Interferometry), has been funded by the European Space Agency. Within PSIC4, eight teams performed, under blind conditions, PSI processing over a test site. They had no a priori information about the type and location of deformation in the chosen test area. A consortium of independent entities analysed these results, performing validation and inter-comparison tasks. The paper describes the project and the general characteristics of the test site. Then it illustrates some of the most important validation and inter-comparison results. Finally it includes the main conclusions of the PSIC4 project. 1. INTRODUCTION The Persistent Scatterers Interferometry (PSI) techniques represent a group of advanced Differential Interferometric SAR (DInSAR) methods to monitor land deformation from long series of SAR images. During the 2003 FRINGE workshop, ESA decided to initiate a project devoted to the assessment of the performances of the PSI techniques for deformation monitoring. The main focus of the project, named PSIC4 (Persistent Scatterer Interferometry Codes Cross-Comparison and Certification for Long-Term Differential Interferometry), is the assessment of the quality of the products coming from the PSI techniques, rather than the algorithms used by the different techniques. The key questions addressed in the PSIC4 project are: - How well PSI describes land deformation, spatially and temporally? - How accurately, and how precisely? - How consistent are the results obtained by different PSI techniques? In order to answer to the above two questions the project has involved some validation activities, where the PSI results achieved over a test site have been compared with ground truth data, coming from repeated campaigns of spirit levelling surveys. In addition, in order to answer to the third question, some intercomparison activities have been performed. In the PSIC4 project eight independent PSI teams have processed a set of SAR data acquired over a test site, which covers a coal mining area located in Provence (France). They have been asked to process the SAR data according to their own PSI methodology, without having a priori information about the location and the type of deformation occurring in the test area. A consortium of independent entities has analysed the eight PSI results, performing validation (comparison against ground truth levelling data) and intercomparison activities. The validation and intercomparison have been performed on an anonymous basis. This paper describes the most important components of the PSIC4 project. Firstly, the general context of the project is described. Then the characteristics of the PSIC4 test site are discussed, describing the type of available ground truth data. Some of the most important validation and inter-comparison results achieved in the project are illustrated in Section 4. The key validation results include the analysis of the deformation velocity maps and the analysis of the deformation time series. The inter-comparison results include the analysis of spatial distribution and density of the PSI data and the inter-comparison of velocity maps. The last part of the paper includes the main conclusions of the PSIC4 project. Proc. Envisat Symposium 2007, Montreux, Switzerland April 2007 (ESA SP-636, July 2007)

2 2. CONTEXT OF THE PSIC4 PROJECT In the PSIC4 project, six of the eight PSI teams used different implementations of the Persistent Scatterers technique, while the other two teams used slightly different concepts (basically two coherence-based approaches). In this document the eight teams are named PSI teams. The eight PSI teams are: - Altamira Information, Spain [1], - Atlantis Scientific Inc. (now is Vexcel Canada), Canada [2], - Delft University of Technology, Holland [3], - Gamma Remote Sensing, Switzerland [4], - German Aerospace Center (DLR), Germany [5], - IREA-CNR, Italy [6], - Technical University of Catalonia, Spain [7], - TeleRilevamento Europa (TRE), Italy [8]. These teams have been asked to process the SAR data according to their own PSI methodology, without having been given a priori information about the type and location of deformation in the chosen test area. The test area was selected by ESA among those proposed by the validation group considering the characteristics of the deformation and the quality of the available ground truth data. The key PSIC4 activities involved the following steps: - Eight PSI processings over the test area were run in parallel by the different PSI teams. - Few pre-processing tasks were performed on the PSI results, which included coordinate transformations, the correction for global geocoding errors, the reference of all data to the same reference area, and some spatial and temporal interpolation. - The validation and inter-comparison of the PSI results were performed by an independent consortium consisting of geological survey organizations (BRGM, BGS, TNO) and the Institute of Geomatics. The validation and inter-comparison were performed on an anonymous basis: the consortium did not know which team produced which PSI results. - Starting from the results achieved in the validation and inter-comparison analyses, the validation team prepared a set of open questions, which were discussed during the PSIC4 final presentation and the PSI validation workshop, held in ESA-ESRIN (Italy) the 18 th and 19 th September Most of the PSI teams involved in PSIC4 participated in these events. Some of the most relevant outcomes of the discussion on the PSIC4 results, and the associated open questions, are included in the conclusions of this paper. It is worth mentioning that within the GMES Terrafirma project [9], an additional inter-comparison study has been performed on the same dataset of PSIC4, which only concerns four of the eight teams of PSIC4. The main results of this study, named Provence inter-comparison, are described in [10]. When going through the results of the PSIC4 exercise, one should keep in mind the following key issue. In order to assess the applicability of the PSI techniques, the PSIC4 exercise was conceived as a blind test. The eight teams did not have knowledge of the type of deformation occurring on the test site, i.e. information such as the linearity/non-linearity of the deformation, the driving mechanism, the location of the deformation, start and end dates of the deformation, and expected deformation magnitude. Moreover, the eight teams were not aware of which spatial subset of their data will be validated and compared against the PSI processing (i.e. whether the validation would be performed on a relatively stable area or on a deforming area). More in general, the teams received no information on what exactly was the deformation signal of interest, i.e. it was not clearly specified the goal of the PSI analysis. By contrast, the validation was focused on a specific deformation phenomenon, i.e. the deformation associated with the mining area of Gardanne. This point is important because the PSI data processing has different parameters that can be tuned and tailored to a specific application goal. For instance, the processing can be modified to get a suitable density of PS in a given area. The advantage of a tuneable PSI processing was not exploited within the PSIC4 project. As a consequence, each team was left free to deliver PS data points where they judged them the most reliable. 3. PROJECT TEST SITE The areas covered by the PSI results vary from team to team. In different cases they cover more than 1000 km 2. However, the results of the project only refer to a mining area located in Gardanne (Provence, southern France), which is about 100 km 2, where Charbonnages de France collected a rich spirit levelling dataset. The location of the test area, between Aix-en-Provence and Marseille, is shown in Fig. 1. The levelling data consists of 12 survey lines along railway lines, canals, and roads, see the levelling network in Fig. 2. The levelling measurements have an inter-distance of m, and were acquired once or twice per year. The area covered by the levelling data is about 8 by 6 km. During the period of interest ( ), the observed ground subsidence is mainly associated with the coal mining exploitation, which is based on the long wall mining technique. For the concerned period the exploited panels are located at a depth between 600 and 1100 m, with a width of about 250 m. The thickness of

3 of the exploited coal layers ranges between 2 and 3 m. The cumulated deformation in the considered period ranges from zero (stable areas) up to some decimetres, see Fig. 3. This figure shows the total amount of deformation (cumulative deformation) that occurred across the test site over the whole considered period, between 1992 and This cumulative deformation, calculated by levelling values resampled at SAR acquisition dates was used as reference to validate the deformation measured by PSI. The Gardanne mining area represents a difficult case for the C-band PSI techniques. The above deformations typically occur over a few months, a rather short period for the (at most) monthly SAR acquisitions. Therefore, we can state that the above deformations are strong deformations from the viewpoint of C-band PSI with the current acquisition capabilities of ERS and Envisat, which are used in the context of this project. However this cannot, in principle, be extended to other types of configurations, e.g. the SAR systems based on L-band sensors, or those that can offer more frequent SAR acquisitions. In PSIC4 the performance of C-band PSI techniques was tested at the very edge of its current capability. Additional information on the test site is shown in Fig. 4, which displays the wrapped and unwrapped interferometric phases of a 105 day interferogram. In this interferogram, which covers a relatively short period, one may notice at least four fringe patterns, which correspond to four major deformation areas. The strongest deformation phenomenon has about two fringes, i.e. 6 cm line-of-sight deformation. This confirms the strong deformation (and the fast deformation rates) of the considered study area. Figure 1. Location of the PSIC4 project test area: the Gardanne mining area (southern France). 4. VALIDATION AND INTER-COMPARISON RESULTS This section summarizes some of the most relevant validation and inter-comparison results of the PSIC4 project. The first one is related to the deformation velocity maps of the teams, which represent a key PSI product. In Fig. 5 the deformation maps of two of the eight teams, teams 6 and 8 are shown. The two maps have rather different characteristics, concerning in particular the density of Persistent Scatterers and the deformation pattern. Team 8 has a relatively high PS density and identify most of the deformation signal, while team 6 has clearly chosen another measurement strategy, concentrating on only providing good PS, which mainly fall in the stable and gently deforming areas. As it is discussed in the conclusions, this is related to the general context of the PSIC4 project, and in particular the fact that the project was conceived as a blind test. Figure 2. The levelling network overlaid on a cartographic map of the Gardanne area. Figure 3. Cumulative deformation measured by levelling, in space and time.

4 Figure 4. Wrapped phase (left) and unwrapped phase (right) corresponding to an interferogram covering the period (105 days). Fig. 6 shows additional useful information related to the PS density. It is a density map generated from team 3 s data, which covers an urbanized area. One may notice that most of the covered area has a good PS density (this is indicated by the reddish colours), while there is a part of the image, which is highlighted in grey, where the PS density is very low. The important point is that at least a part of the area characterized by low PS density is urbanized. Over such an area a high PS density should in principle be expected. The main reason for the lack of PS in such an area is that it is a strongly deforming area, where most of the PS are lost because the deformation models used in the PSI processing techniques do not fit well with the actual deformation. That is to say, the area surely includes several good permanent scatterers (i.e. points that provide good interferometric phases) but they are discounted because the associated observations do not fit well with the deformation model used. Starting from the above results the following open questions arose: Why are there large differences in the PS densities? Why there are no PS in most of the deformation areas? And why are most of the deformations missed? The second set of results concerns the validation of the deformation velocity. The velocity validation was based on the temporally resampled levelling data (at each SAR acquisition date) and the spatially resampled PS data (at each levelling point location). Fig. 7 shows the map of vertical velocities retrieved by levelling data and by 2 PSI teams. [mm/yr] Figure 5. Velocity maps for team 8 (top) and team 6 (bottom).

5 Points per 100m 2 Figure 6. Zoom over the map of PS density, computed for team 3. A visual assessment shows that the spatial sampling of team 2 outlines the main pattern shown by the levelling velocity map. Nevertheless, the main subsidence areas are detected with fewer PS, which underestimate the deformation rates of the order of 3-4 mm/year. The sampling density provided by team 7 is slightly lower, and the estimated deformation rates are smaller than those of team 2. Apart from these two examples, a more global assessment of the velocity validation, for the eight teams, can be derived from the global statistics shown in Tab. 1. The mean differences range from 2.6 to 4.7 mm/yr. Note that all the teams have the same mean sign. This tells us that all the teams systematically underestimate the deformation velocities. The standard deviation of the differences range from 5.1 to 7.2 mm/yr. The inter-comparison of the deformation velocities is shown in Tab. 2. The lower triangle in this table shows the mean velocity differences between pairs of teams, which range from to 0.46 mm/yr. In this case, team 8 shows the biggest discrepancies. The upper triangle indicates the standard deviations of the velocity differences which range from 0.63 to 1.86 mm/yr. It is worth underlining that these statistics are remarkably better then those of Tab. 1. These values are much better because they mainly reflect the behaviour of PS outside the major areas of deformation, where most of the teams have velocity estimates. Two examples of time series generated in PSIC4 are shown in Fig. 8. The first plot shows the time series estimated by six teams and the reference from levelling data. This is an example of good agreement between the time series, which corresponds to a case of moderate deformation (the cumulated deformation is about 30 mm). Figure 7. Vertical velocity map derived from the levelling data (above). Vertical velocities of PS resampled at levelling points location, for team 2 (in the middle) and team 7 (below).

6 Table 1. Overall differences between PS and levelling velocities (in mm/yr). Differences are expressed as mean and standard deviation of the teams minus levelling, for each team. mm/yr T1 T2 T3 T4 T5 T6 T7 T8 T T T T T T T T Table 2. Global inter-comparison of the deformation velocities. The lower triangle, in brown, shows the mean differences between pairs of teams [in mm/yr], while the upper triangle, in blue, shows the standard deviation of differences [in mm/yr]. The second example concerns three teams and the reference levelling time series. In this case the discrepancies are considerably larger: although one time series follows the reference profile rather well, the other two clearly underestimate the deformation. A global validation analysis of the time series is summarized in Fig. 9, which plots the RMSE (RMS of the differences between the PSI time series and the reference one) for five classes of deformation velocities. As it can be seen, the lowest RMSE values are found for those time series which fall within the class -5 mm/yr to >0mm/yr. About two-thirds of the levelling time series of these velocity classes have a maximum displacement inbetween SAR acquisition dates of less than 1.4cm. The average RMSE is about 10 to 15mm, apart from processing chain 8 which has an RMSE more than 20mm. The RMSE increases considerably for those time series which fall in the class <-10 mm/yr. An interesting comparison of the reference levelling data and the PSI deformation estimates is shown in Fig. 10, which displays nine spatio-temporal deformation profiles. Here the PS time series can spatially and temporally be compared with the levelling data along one levelling traverse. For the levelling data, the temporally interpolated data have been used, so that at every SAR acquisition date we have an estimated levelling displacement. Figure 8. Two examples of time series compared with the levelling data (in black). RMSE (mm) > <-15 velocity class (mm/yr) Figure 9. RMS of the differences between the PSI time series and the reference one. For the PS data the spatially interpolated data have been used. so that at every levelling location we have a time series. The results concern the AXE levelling traverse. The plots show the levelling transect as discrete levelling points on the horizontal axis and time as acquisition date on the vertical axis

7 Figure 10. Spatio-temporal profile (X: distance along levelling line, Y: time) of the levelling data along the AXE levelling traverse (top) and profiles of the eight PSIC4 teams.

8 The cumulative displacement is colour coded. The levelling data show a distinctive subsidence process, starting in the first quarter of 1993 and ending in July The subsidence bowl starts to form at the benchmark AXE , and travels southward along the levelling section in conjunction with the progression of the mining works. Instantaneous subsidence velocities amount here to a maximum of about 12 cm/yr. Apart from processing chain 8, none of the processing chains shows a full coverage of the section where the subsidence has taken place. Two of the processing chains (3 and 6) fail to capture any signal of subsidence along this levelling section. Outside the stretch where subsidence takes place, all processing chains show a fairly good to good coverage. All processing chains underestimate the displacement in time along the subsiding stretch. The underestimation of the subsidence is largest for that part of the subsidence starting before July Probably this has to do with the fact that there is a large gap of SAR acquisitions between 1993 and 1995, thus posing problems with phase unwrapping for points showing large displacements. However, although the onset of subsidence is visible in the part of the section where subsidence starts well after July 1995, an underestimation of the displacement in time can be seen. 5. DISCUSSION AND CONCLUSIONS The most relevant aspects of the PSIC4 project have been described in this paper. The PSIC4 project represents an unique experiment for the number of PSI teams involved and for the quality of the available ground truth data, which involves more than 1000 levelling benchmarks measured once or twice per year. The project involved an important validation effort, which included the preparation of a detailed validation plan, different pre-processing steps, and several validation and inter-comparison activities. Furthermore, a special effort was devoted to a comprehensive crosscheck of the key activities and tasks of the project. This cross-check showed that the outcomes of the validation and inter-comparison results are consistent and reliable: they represent a technically sound base to be used to assess the PSI performances over the considered test site. The main outcomes of the validation and intercomparison activities over the Gardanne test site have been described in this work. The main conclusions of this work are summarized in the following points. One of the most important conclusions of the project concerns the characteristics of the mining test site of PSIC4, which include abrupt non-linear movements with magnitudes that range from few centimetres up to some decimetres. These are severe characteristics from the viewpoint of C-band PSI with the temporal acquisition capabilities of ERS and Envisat. Why are the deformation characteristics so important? Because in principle PSI can measure surface displacements with millimetric precision, but this can only be achieved under the following conditions: The right model to describe the deformation is adopted. This is difficult to accomplish with abrupt nonlinear movements. The aliasing due to low PS density and/or low temporal sampling with respect to deformation, which may cause phase unwrapping errors, is controlled. This is difficult or impossible with strong deformation magnitudes. The assumptions to separate the atmospheric contribution from deformation are correct. This typically fails in presence of non-linear motion. Most of the results of PSIC4 can be understood in the context of the above conditions: none of them is fully accomplished in the mining area of Gardanne. It is worth underlining that the above consideration on strong deformations holds for C-band PSI with the current temporal acquisition capabilities of ERS and Envisat. They cannot in principle extended to other types of SAR missions based on different band frequencies and more frequent SAR acquisition capabilities. The PSIC4 project was conceived in a specific framework, where the teams worked under blind conditions, with no a priori information on deformation type, driving mechanism, deformation magnitude, etc. Furthermore, they had no information on what exactly was the deformation signal of interest, i.e. the goal of the PSI analysis. By contrast, the validation was focused on a specific deformation phenomenon, i.e. the deformation associated with the mining area of Gardanne. This point is important because it plays a key role in the PSI processing. In fact, instead of tailoring the processing to a specific objective of the analysis, the teams used a standard approach and a processing which is feasible with reasonable resources. It is worth emphasising that the area covered by most of the teams is considerably larger than the 100 km 2 area used for the validation. None of the PSI teams have performed an advanced or refined PSI analysis, because neither the area of interest nor the goal of the refinement were defined. This again explains most of the PSIC4 results, e.g. the lack of PS in the mining area of interest. It is worth analysing a specific consequence of the above point, which explains the different densities achieved by the teams. The PS densities are different because there is no definition of what exactly is a PS. The teams used their standard PSI approach (instead of an advanced or tailored one) and delivered the PS only

9 where both velocity and time series could be extracted with reasonable reliability. Unfortunately the validation area represents a difficult area, where phase unwrapping errors represent the main problem. Due to the high probability of this type of error many teams did not deliver unreliable information. Note that this did not occur outside the mining area. Considering the above points and the results achieved in the Gardanne mining area we can say that PSIC4 has clearly demonstrated that the PSI limitations are real, i.e. that PSI is not applicable everywhere. Though this was already clear to many PSI specialists, now this evidence has been widely documented. The PSIC4 results only concerned the difficult mining area of Gardanne, while most of the PSI teams covered considerable larger areas, which include gentle deformation and stable areas. For this reason, within the ESA-founded project Terrafirma it was decided to run a additional inter-comparison study over the same dataset of PSIC4. This work, named Provence Intercomparison was restricted to four of the eight teams of PSIC4. The main results of Provence Intercomparison can be found in [10]. This paper offers an interesting complementary source of results with respect to PSIC4. When comparing the results of the two projects, one may notice an increase of the performances in the case of the Provence intercomparison for both the deformation velocity maps and the time series. It is worth underlining that the results of the two projects (Provence inter-comparison and PSIC4) are not contradictory. In some cases they simply show different complementary aspects. For instance, the Provence inter-comparison is largely based on data outside the Gardanne mining area, which were simply not analysed in PSIC4. In other cases the results are rather similar. Finally, the limitations of PSI over deformation areas with similar characteristics to Gardanne open some new important issues for the future. We briefly mention two of them. The first one is the importance of a feasibility study before running a PSI analysis. This may help in avoiding false expectations and disappointing results. Note that a feasibility study is now implemented within the Terrafirma project. A second issue concerns the appropriate ways to inform the PSI users of the limitations of the technique, especially in those cases where PSI is employed under non-ideal conditions. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] Crosetto, M., Agudo, M., Capes, R. and Marsh, S., (2007). GMES Terrafirma: Validation of PSI for users: Results of the Provence inter-comparison. Proceedings of the Envisat Symposium, April 2007, Montreux, Switzerland, ESA Special Publication SP-636, CD. ACKNOWLEDGMENTS The authors thank Charbonnages de France for kindly providing the levelling dataset used as reference data in this work. This work has been funded by the European Space Agency.