Coherence Exploitation

In the 1990’s ERS-1 C-band data acquired with short 3-day repeat intervals and after the launch of ERS-2 also ERS-1/2 Tandem acquisitions with only one day interval permitted developing a range of applications based on the interferometric coherence, such as land use characterization [1] and the retrieval of vegetation parameters [2,3]. While a significant potential was found for the C-band data acquired with 1 and 3 day intervals, the potential was much lower with the 35-day repeat interval of the ENVISAT mission. Now with the launch of Sentinel-1A, there are again C-band data with shorter intervals of 12 days, and later on in combination with Sentinel-1B even just 6 days, available. It is therefore of interest to consider again more closely the interferometric coherence as a valuable parameter for a range of thematic mapping applications.

InSAR processing of Sentinel-1 data is challenging because of the special characteristics of the interferometric wide-swath (IWS) data acquired in the TOPS mode. The steep azimuth spectrum ramp in each burst results in very stringent co-registration requirements. Our existing co-registration method that considers also the scene topography was updated, so that it is applicable to the TOPS data. The main geometric transformation function is derived based on the orbit and surface geometry and only a small correction is applied to this which is either determined using intensity cross-correlation or based on the InSAR phase differences in the along-track burst overlap regions. The co-registered TOPS SLC pairs and stacks are then used to calculate a differential interferogram and estimate the interferometric coherence. Compared to ERS data, the range resolution of the Sentinel-1 TOPS data is significantly better in range direction and significantly lower in azimuth direction. Applying a stronger multi-looking in range instead of in azimuth direction results in Sentinel-1 interferograms with a comparable spatial resolution and number of looks as for ERS or ENVISAT. Major advantages of Sentinel-1 compared to these sensors are the much larger spatial coverage (IWS and EWS modes) and the availability of coherence also at cross-polarization.

First results indicated that the C-band coherence is significantly lower over vegetation after 12 days (Sentinel-1 ) as compared to after 1 or 3 days (ERS). Consequently, we observe for Sentinel-1 data that the land use characterization with a coherence product, which is an RGB composite of the coherence, the backscatter intensity and the backscatter change between two observations [1] is fuzzy between forest and other vegetation classes. We also find that this longer interval clearly reduces the potential to retrieve forest parameters from the coherence as compared to shorter-interval C-band InSAR coherence [2]. Figure 1 shows a section of a Sentinel-1 coherence product derived from a summer pair over Sicily Italy, and Figure 2 the corresponding optical imagery found in Google Earth. For dry bare fields and urban areas high coherence values are observed. The coherence is significantly reduced though even over some of the agricultural fields and clearly reduced for orchards with small fruit trees. On the other hand, we expect some parameter retrieval potential at lower vegetation levels.

Additional results will be presented at the conference with more Sentinel-1 data becoming available to the user community.

References:

[1] Wegmüller U. and C. L. Werner, SAR interferometric signatures of forest, IEEE Trans. Geosci. Remote Sensing, Vol. 33, No. 5, pp. 1153-1161, 1995.

[2] Askne J. and M. Santoro, "Automatic model-based estimation of boreal forest stem volume from repeat pass C-band InSAR coherence," , IEEE Trans. Geosci. Remote Sensing, Vol. 47, 2, pp. 513-516, 2009.

[3] Wegmüller U. and C. L. Werner, Retrieval of vegetation parameters with SAR interferometry, IEEE Trans. Geosci. Remote Sensing, Vol. 35, No. 1, pp. 18-24, 1997.

Due to their crucial role in Earth’s carbon cycle, it is important to monitor forests for detection of deforestation and for a better understanding of changing processes within them. A possible solution for this purpose is the TanDEM-X/TerraSAR-X constellation, which allows the acquisition of two spaceborne synthetic aperture radar images simultaneously. This single-pass interferometry offers the opportunity to generate global digital surface models from the interferometric phase even for areas covered with dense vegetation. However, the main drawback of radar interferometry is the phase ambiguity that has to be resolved by phase unwrapping. Sudden jumps between tree crowns and the ground terrain challenge conventional phase unwrapping techniques.

Therefore, in this study, a novel approach for estimating tree heights based on robust coherence and residual phase calculation is introduced. The approach is able to deliver smooth residual phases that can be unwrapped with conventional phase unwrapping algorithms. The separation of ground and tree height allows, for instance, to obtain both a Digital Terrain Model and a Digital Surface Model.

A widely used method for coherence computation between two SAR images is spatial averaging. For meaningful results, a sufficient number of samples and therefore a large enough estimation window needs to be applied. Hence, the resulting coherence image suffers from an unavoidable loss of spatial resolution. Another major drawback is the bias introduced by the interferometric phase itself due to topography, deformations, or atmospheric change. This bias can be considered as a result of topography alone, since single-pass TanDEM data are considered. In case of variations of the topographic phase within the estimation window an underestimation of coherence is caused. A simple way to remove this bias is to subtract the topographic signal with knowledge of a DSM first. Unfortunately, this information is rarely given if volume height estimation is the pursued purpose.

Another solution is to lower the image resolution, so topographic variations are mitigated. However, in addition to the evident disadvantage of lower resolution, this approach is limited by violation of a fundamental assumption underlying spatial coherence estimation: neighboring resolution cells are assumed to show similar scattering behavior. This assumption only holds if an object is observed at an appropriate resolution. For lower and higher resolutions, homogeneous areas become heterogeneous.

Therefore, in this study, a novel approach for estimating coherence on a pixel by pixel basis in the frequency domain is proposed. In detail, our approach exploits the increased system bandwidth of the recent high resolution SAR satellite generation. Based on a Fourier transform, the complex SAR image can be split into multiple subbands with lower - but due to the high system bandwidth still reasonable - resolution. Thus, resulting subbands can then be stacked and evaluated for coherence estimation. The advantage of this approach is that the resulting lower resolution does not depend on anything but the required signal samples needed for coherence estimation. Further, the spatial resolution is not decreased by the coherence estimation window and issues at borders and heterogeneous areas are resolved in contrast to a spatial estimation. However, the frequency domain coherence is affected by a linear residual topographic phase due to a different center frequency for each subband. In contrast to the spatial topographic variation that depends on the local terrain trend, the introduced phase ramp shows linear behavior and can easily be removed. We could observe that the processing chain implemented for frequency domain coherence yields better values than spatial coherence estimation. The frequency domain coherence exhibits more stable results for different resolutions and blurring at the edges can be avoided.

Furthermore, we will show that exploitation of the frequency domain coherence values for tree height estimation delivers promising results if multiple images are used. In detail, five SAR images with similar baseline are stacked to estimate tree heights using the RVoG model and by subtracting an approximate ground phase. For this, it is assumed that the ground-to-volume ratio changes from spring to summer because of phenological changes. However, this approach is restricted to the use of seasonal time slots of SAR data and to tree species with changing tree crowns. Still, the estimation of homogeneous areas leads to a high success rate for the given test site.

Following a plethora of validations and demonstrations SAR Interferometry (InSAR) has been established as a mature space geodetic technique for providing valuable insights for various phenomena related to geohazards. One of the main advantages of space borne SAR systems with respect to GNSS is the continuous spatial coverage. However, the impact of temporal decorrelation especially in repeat-pass interferometry has been observed during the historical development of InSAR applications. Interferometric coherence can be considered as expression of temporal decorrelation. It is understood that interferometric coherence decreases with time between SAR acquisitions because of changes in surface reflectivity, reducing the quality of SAR phase measurements. This is an intrinsic characteristic of the design of SAR systems that has a significant contribution at longer time scales. Since the majority of geohazards rely on long term observation scenarios, the effect of temporal decorrelation is evident as coherence becomes dominated by temporal changes. Although in the past there was not sufficient amount of SAR data to extract robust statistical metrics, in the present study it is demonstrated that tailored analysis of interferometric coherence by exploiting the large SAR archive as well as data from operational missions such as the Sentinels, available by the European Space Agency (ESA), enables the accurate quantification of temporal decorrelation. A methodology to translate the observed rate of coherence loss into decorrelation times over two volcanic landscape, namely the Santorini volcanic complex and Mt. Etna is the subject treated in this study. Specifically, a sensitivity analysis was performed on a large data stack of interferometric pairs to quantify at a pixel level the time beyond which the interferometric phase becomes practically unusable due to decorrelation. The estimation and mapping within a geostatistical context of the spatial distribution of the temporal decorrelation times in an area without a necessary a priori knowledge of its surface characteristics is a fundamental parameter for the design and establishment of local GNSS networks as well as the definition of optimal monitoring strategy for various geohazards. The dependence of decorrelation on various land cover/use types is also analysed by post-processing in a GIS environment. Though the dependence of decorrelation on various land cover/use types is already documented the provision of additional information regarding the expected time of decorrelation is of practical use especially when EO data are utilized in operational activities. The performed analysis is viewed within the improved capacity of current and future SAR systems, while underlining the necessity for exploitation of archive data.

Differential SAR Interferometry (DInSAR) is a geodetic technique that allows investigating terrain-deformation-related geophysical processes. It takes profit from the phase difference between two SAR scenes in order to estimate surface motion achieving sub-centimeter to sub-millimeter precision (Gabriel et al, 1989). A great deal of phenomena can be studied by using this technique, including active tectonics, earthquakes, urban subsidence, ground water, oil and gas exploitation, volcanism, landslides and anthropogenic structures like tunnels and mining facilities (Massonet & Feigl, 1998)

            ESA SAR Missions ERS-1/2 and ENVISAT provided over two decades of useful data for DInSAR. Availability of a great number of scenes repeatedly acquired over AOIs promoted the development of the so called time-series techniques (MT-InSAR) which allow investigating the deformation history of every coherent pixel. Most popular between them are Small Baseline Subsets (SBAS) (Berardino et al, 2002) and Persistent Scatterers (Ferretti et al, 2001). SBAS makes profit of distributed scatterers by using multilooked interferograms in order to enhance the SNR ratio. Spatial and temporal decorrelation is minimized by selecting interferometric pairs characterized by small spatial and temporal baseline. By the contrary, PS solves for point scatterers. It generates full resolution interferograms between the available scenes and a common master one. As a consequence, high density of results is achieved in favorable areas, like urbanized ones.

Key parameter that gives an idea of the phase quality in an interferogram is the interferometric coherence. It is approximated from the data by means of a maximum likelihood estimator (Seymour and Cumming 1994), and takes values between 0 and 1. Temporal evolution of this parameter at a generic resolution cell is an indicator of the time this cell will allow computing reliable results. Furthermore, coherence is a measurement of phase error used for improving phase unwrapping (Pepe and Lanari, 2006). It depends not only on the scatterers properties but also on the acquisition geometry and other SAR system characteristics (Zebker and Villasenor, 1992).

Second generation SAR systems has been launched during the last years. X-band missions COSMO-SkyMed (ASI) and the TerraSAR-X (DLR), orbiting the Earth since 2007, are characterized by reduced revisit time and high spatial resolution. Increased availability of SAR data makes interesting investigating decorrelation with X-band data and time-series processing. In particular, COSMO-SkyMed constellation allow acquiring data with HH, VV, HV and VH polarizations (COSMO-SkyMed, 2010). Questions are: Which acquisition mode is convenient for long term interferometric exploitation? Is decorrelation different at different polarization? Are coherent pixels in HH scenes the same as in VV ones?

Trying to answer those questions, we analyzed the behavior of the temporal coherence in a COSMO SkyMed dataset acquired by interleaving HH and VV acquisitions covering a common AOI. Temporal coherence depends on the changes in the individual scatterers inside the resolution cell. Those changes are due to scatterers re-distribution and/or re-orientation and change in dielectric properties of the observed surface. Scatterers re-distribution and reorientation in a rural environment can be due to fields plowing and irrigation. Natural phenomena like vegetation growing, wind, rainfall, snowfall and temperature change introduce modification both in the relative location of scatterers and in the soil moisture content, leading to temporal decorrelation. The diversity of causes makes temporal coherence a magnitude not easily modelled. Zebker and Villasenor (1992) proposed a theoretical exponential model taking into account horizontal and vertical scatterers random movement. However, that model is not easily extrapolable to a real environment because individual scatterers displacement is not available in a general case.

In this work we present the results of an empirical study regarding temporal decorrelation occurred in HH and VV COSMO SkyMed scenes time series. Area of Interest is Mendoza city in western Argentina. It covers the Mendoza metropolitan area and surroundings including different coverage types: urban, bare soil, agricultural and mountainous (rock outcrops). We used 62 scenes of which 32 are HH and 30 are VV, conforming two datasets interleaved in time. Time span is one and a half year between June 2010 and March 2012.

            Using those scenes, we computed 496 coherence maps in HH-HH polarization and 435 ones in VV-VV polarization. For estimating temporal coherence we used the methodology proposed by Wang et al. (2010). Briefly, they compute the range spectral shift (Gatelli et al, 1994) that is a function of the perpendicular baseline, incidence angle, terrain slope and system parameters. Azimuth spectral shift is directly computed from Doppler centroid difference between scenes. Linear relationship between geometric coherence and spectral shifts is exploited for estimating the former. Then, temporal coherence is computed by dividing the observed coherence by the estimated geometric coherence. 

Temporal coherence maps were studied by defining Regions of Interest (ROIs) relevant for different land coverage: 1) urban, 2) bare soil, 3) agricultural and 4) mountainous. In particular, we analyzed percentage of coherent points and seasonal behavior as a function of polarization. A pixel is considered coherent when its temporal coherence is higher than a defined threshold in a significant number of interferograms. For this study, we set the temporal coherence threshold at 0.25 and the percentage of interferograms to 30%. This criteria is similar to the one employed in the SBAS implementations for deciding which pixels remain in the final solution (Pepe and Lanari, 2006).

Table I shows the pixel status, in terms of coherence, for data at different polarization inside the ROIs. As can be appreciated, vast majority of pixels show similar behavior in HH and VV data: 91% in urban environment, 84% in bare soil, 93% in mountainous terrain and 98% in agricultural fields. As expected, most pixels in agricultural ROI are incoherent in both polarizations, whereas coherent pixels are many more in urban and bare soil terrain. Mountainous terrain behaves differently because, despite it is composed of stable structures like rock outcrops, it receives heavy snowfall in winter, and so only 18% of pixels remain coherent according to the used criteria. HH polarization shows a marginally higher ratio of coherent pixels when compared with VV polarization. This effect is observed in urban, bare soil and mountainous environment.

In order to analyze the behavior in time of the temporal coherence, we categorized the coherence maps according to the time span between master and slave scenes. Then, for each ROI and time span we computed the temporal coherence mean, median, standard deviation and minimum and maximum value. Figure 1 show the results for bare soil, urban and agricultural ROIs considering 16- and 128-days time spans as a function of master scene date.

Seasonal behavior in agricultural soil is noticeable when time span is 16 days (Figure 1 a-b). Temporal coherence reach peaks of ~0.60 in winter and drop to less than 0.20 in summer. This effect is not surprising because after the crop season (January to march) fields remain mostly undisturbed until the spring (September). Coherence rapidly increase from March to August and then drop towards the summer when the growing season peaks. Two coherence maps in July 2010 show anomalous behavior with temporal coherence of ~0.10. We suspect that they can represent short term coherence loss like the one provoked by snowfall. It remains as future work testing that hypothesis. When time span is set to 128 days (agricultural ROI) the seasonal behavior is less noticeable. Winter peaks are still observed (0.25), but temporal coherence is around 0.15 most of the time.

            Bare soil areas (Figure 1 c-d) behave similarly to the agricultural ones. Coherence is relatively high reaching peaks of ~0.80 during the winter whereas summer valleys are about 0.30. This effects can be due to growing of sparse natural vegetation located within these areas. Same behavior is observed with 128 days’ time span, although coherence is always higher than in the agricultural case.

            The urban case (Figure 1 e-f) is somewhat different. Temporal coherence at 16 days’ time span is high and stable independently of the master scene date. However, we observe also higher dispersion of values than in the other ROIs. This effect could be due to the fact that in Mendoza city the urban vegetation (trees) is really significant. Thus, temporal coherence in urban ROIs are a mixture of man-made structures coherence and deciduous trees coherence. When 128 days’ time span is analyzed, a slight drop of coherence is observed, but it remains stable all year long.

            One important observation relevant for all the analyzed ROIs is that the temporal coherence show no significant difference when HH and VV polarized scenes are studied. Note that in the figure, red bars are VV and blue ones are HH.

References

Berardino, P., Fornaro, G., Lanari, R., Sansosti, E., 2002. A new algorithm for surface deformation monitoring based on small baseline differential SAR interferograms. IEEE Trans. Geosci. Remote Sens. 40, 2375–2383.

COSMO SkyMed System, 2010. Description and Userguide: http://www.e-geos.it/products/pdf/csk-user_guide.pdf

Ferretti, A., Prati, C., Rocca, F., 2001. Permanent scatterers in SAR interferometry. Geosci. Rem. Sens. IEEE. On 39, 8–20.

Gabriel, A.K., Goldstein, R.M., Zebker, H.A., 1989. Mapping small elevation changes over large areas: differential radar interferometry. J. Geophys. Res. Solid Earth 1978–2012 94, 9183–9191.

Gatelli, F., Monti Guarnieri, A., Parizzi, F., Pasquali, P., Prati, C. and Rocca, F. (1994), The wavenumber shift in SAR Interferometry, IEEE Transactions on Geoscience and Remote Sensing, 32(4):855–865.

Massonnet, D. y K. Feigl (1998). "Radar Interferometry and its Application to Changes in the Earth's Surface." Reviews of Geophysics. Vol 36, Nº 4, pp. 441-500

Pepe, A., Lanari, R., 2006. On the extension of the minimum cost flow algorithm for phase unwrapping of multitemporal differential SAR interferograms. Geosci. Remote Sens. IEEE Trans. On 44, 2374–2383.

Seymour, M. S. and Cumming, I. G. (1994), Maximum Likelyhood Estimation For SAR Interferometry, in: International Geoscience and Remote Sensing Symposium, Pasadena, CA, USA, 8–12 August 1994, pp. 2272–2275.

Wang, T., Liao, M., Perissin, D., 2010. InSAR Coherence-Decomposition Analysis. IEEE Geosci. Rem. Sens. Let. 7, 156–160.

Zebker, H.A., Villasenor, J., 1992. Decorrelation in interferometric radar echoes. Geosci. Remote Sens. IEEE Trans. On 30, 950–959.

ARIA (Advanced Rapid Imaging and Analysis) is a joint JPL/Caltech coordinated project to automate InSAR and GPS imaging capabilities for scientific understanding, hazard response, and societal benefit. Under ARIA project we have developed a prototype damage detection algorithm using Interferometric Synthetic Aperture Radar (InSAR) coherence, and the algorithm is being incorporated into ARIA data system that will automatically respond to future major disaster events. We present results that show capabilities of the algorithm and the data system in terms of quality of products, latency, and feasibility of use for analysis of large SAR data sets and for disaster response activities.

ARIA team has responded to a number of natural disaster events including Christchurch Earthquake, Tohoku Earthquake and Tsunami, Hurricane Sandy, and Super Typhoon Haiyan in the Philippines, and South Napa Earthquake, producing change detection maps using X-band, C-band, and L-band radar data from COSMO-SkyMed, Envisat, and ALOS satellites, respectively. The figure shows one of recent ARIA products – a damage proxy map (DPM) for Super Typhoon Haiyan. The 40-by-50 kilometer DPM covers a region near Tacloban City, where the massive storm made landfall. The map was derived from COSMO-SkyMed radar data. The areas in red color reflect the heaviest damage to cities and towns in the storm's path. Each pixel in the damage proxy map is about 30 meters across.

The latency for data discovery, access, and processing combined was 6 days for Hurricane Sandy response. We were able to reduce this latency down to 11 hours for Super Typhoon Haiyan response. This difference highlights the importance of a inter-agency collaboration and system’s efficiency on data handling. Future automated ARIA system should achieve further reduction in latency. The distribution of detected pixels was compared with the European Commission’s Copernicus products and showed high correlation. The total latency since the landfall of Haiyan in Tacloban City until the production of the damage proxy map was three days.

We also explore the sensitivity of InSAR coherence to different types of disaster events in both spatial and temporal senses. For M6.3 February 2011 Christchurch Earthquake, we demonstrate a way to improve the quality of DPM with temporal characterization of coherence for each pixel and use of high-resolution LiDAR data, for detecting building damage and liquefaction damage. For 2011 Kirishima volcano eruption, we estimate the lower bound on the detectable ash layer depth. We also examine the coherence’s detectable damage grade of residential structures from 2013 Moore tornado damage. Our coherence analyses reveal potentials and limitations of coherence-based change detection algorithms. 

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