Floods are the most frequent weather disasters in the world and the costliest. Numerous studies from the past demonstrated that SAR systems are suitable tools for flood mapping, thanks to high sensitivity to water and capability to provide data day and night, regardless of cloud cover. The processing of SAR data for flood rapid mapping generally relies on near real time (NRT) algorithms that are based on a very simple model of SAR return from floodwater, which assumes that the water surface is smooth and acts as a specular reflector of the radar signal, thus resulting in a very low backscatter. However, it is well known that this model is not able to represent situations, such as flooded vegetation, when plants emerge from the water surface, so that numerous omission errors may be committed. Mapping flooded vegetation through single channel SAR data is generally tackled in the literature taking advantage of change detection approaches aiming at detecting the Double Bounce (DB) effect (Hess et al., 2007; Pulvirenti et al., 2011).
The detection of this mechanism depends on sensor parameters (frequency, polarization and incidence angle), as well as vegetation characteristics (e.g., geometry, water content, density). The latter determines the cross section of the equivalent corner reflector and the attenuation of its radar echo. Consequently, it can be expected that, at X band, double bounce detection is possible only in particular conditions (e.g., stem dominated vegetation) and at horizontal (HH) polarization (i.e., less attenuated by vertical structures and more strongly reflected by the water surface). C and especially L band can be more sensitive to DB, thanks to the increased capability to penetrate the canopy.
In this paper we investigate the double bounce enhancement due to standing water in flooded agricultural fields to assess the capability of a radar to recognize the presence of floodwater under vegetation. The investigation was carried out by analysing a multitemporal, polarimetric and multifrequency SAR dataset (COSMO-SkyMed, Alos-2, Radarsat-2, Sentinel-1) collected over the Vercelli district in North Italy in Spring/Summer 2015, characterized by a widespread and intense cultivation of rice crop, were the fields were routinely artificially flooded and dried according to the agricultural practice.
The images from ALOS-2 (3 images) and from Radarsat-2 (4 images) were collected in fully polarimetric mode. The data set from COSMO SkyMed and Sentinel-1 took advantage from the better capability to monitor the site over time, although at a single polarization (COSMO SkyMed: 6 images HH) and dual pol (Sentinel-1A: 12 images VV, VH). The polarimetric data were processed by well-established methods, i.e. the Freeman scattering mechanism discrimination and the eigenvector (entropy/alpha) decomposition. The Sentinel-1A data were analysed by change detection methods thanks to the interferometric acquisition mode, e.g., extracting the complex coherence.
Both a general qualitative analysis of the satellite data, based on a visual inspection of the available images, and a more detailed quantitative data analysis, focused on some selected rice fields, were accomplished. It was found that a significant increase of the COSMO-SkyMed backscattering occurred when floodwater was present beneath rice plants at an advanced stage of growth. In fact, in July, when rice is well developed, high backscatter in X-band was observed in fields were the L-band polarimetric data recognized the double bounce return. The presence of a double bounce scattering enhancement at X-band was then established. At C-band the dihedral type of return was not clearly recognized by Radarsat-2 polarimetric data because of the smaller incidence angle of Radarsat-2 acquisitions and the high differential extinction between horizontally and vertically polarized backscattering. The high revisit time of Sentinel-1 in IW mode allowed to asses the potential of exploiting multitemporal techniques as opposed to polarimetric ones at C band.
Hess, L. L., J. M. Melack, D. Simonett, “Radar detection of flooding beneath the forest canopy: a review”, International
Journal of Remote Se sing, vol. 11, issue 7, pp. 1313-1325, 2007

Pulvirenti L., M. Chini, N. Pierdicca, L. Guerriero, and P. Ferrazzoli, "Flood monitoring using multi-temporal COSMOSkyMed
data: Image segmentation and signature interpretation," Remote Sens. Environ., vol. 115, no. 4, pp. 990-1002, Apr. 2011

Area A - Board 89

Keywords – Copernicus, open data, agriculture, solid land, vegetation, EM scattering

Abstract - The analysis of soil and land cover scattering properties, and their connection with the parameters of microwave scattering has been carried out for several years if not decades [1-3]. In recent years, however, the advent of modern space-borne microwave radiometers like SMAP [4] and SMOS [5] in addition to the trend towards open data for scientific use, created favourable conditions for refined models to be developed, and more accurate assessment of scattering parameters to be performed, on various types of land cover. In our work, we focussed our attention on agricultural land cover and the combination of multiple sources of passive and active microwave sensing data. The unfortunate failure of the L-band radar on SMAP contributed to boosting research on the use of alternate sources for active sensing data to be used jointly with SMAP radiometer data [6], and different models are being developed to incorporate data from Sentinel-1 or from other radar missions in order to reinstate the computation of scattering parameters. In this manuscript, however, we try and combine SMAP data with three different sources of Synthetic Aperture Radar (SAR) data at three different bands (C, L, X) on a vegetated land cover area in Italy, whose parameters were estimated from ancillary data derived from ground inspections or from proxies. Whereas preliminary findings are encouraging, the work is still in progress, and a more complete set of results will be presented with the full paper, together with our analysis, conclusions and directions of future work. This research has been carried out in the framework of the ESA Phi-Unet initiative [7] for geodata-based innovation.

References
[1] F. Ulaby, R. Moore, A. Fung, Microwave Remote Sensing Active and Passive Vol. III: From Theory to Applications, MA, Boston:Artech House, 1986.
[2] F. Ulaby, R. Moore, A. Fung, Microwave Remote Sensing Active and Passive Vol. II: Radar Remote Sensing and Surface Scattering and Emission Theory, MA, Boston: Artech House, 1986.
[3] J. C. Shi, J. Wang, A. Hsu, P. O'Neill, E. Engman, "Estimation of bare surface soil moisture and surface roughness parameter using L-band SAR image data", IEEE Trans. Geosci. Remote Sens., vol. 35, no. 5, pp. 1254-1266, Sep. 1997.
[4] D. Entekhabi et al., "The Soil Moisture Active Passive (SMAP) Mission," in Proceedings of the IEEE, vol. 98, no. 5, pp. 704-716, May 2010.
[5] Y. H. Kerr, P. Waldteufel, J. -. Wigneron, J. Martinuzzi, J. Font and M. Berger, "Soil moisture retrieval from space: the Soil Moisture and Ocean Salinity (SMOS) mission," in IEEE Transactions on Geoscience and Remote Sensing, vol. 39, no. 8, pp. 1729-1735, Aug. 2001.
[6] N. Das, D. Entekhabi, S. Kim, S. Yueh, P. O'Neill, "Combining SMAP and Sentinel data for high-resolution Soil Moisture product", Proc. of IGARSS, pp. 129-131, July 2016.
[7] Phi-Unet is an open-innovation network powered by ESA for geodata-based innovation – by leveraging Space data in particular in Universities 2.0. Web site: http://phi-unet.com/about/

Area A - Board 90

Remote Sensing offers the potential of mapping Earth surface, and in particular, city sizes and structure. With nearly one-quarter of the world population living in cities with more than 1 million inhabitants, this information is helpful in assisting urbanism design, in preventing the spread of contagious diseases, or in rescuing people after a natural disaster. Therefore, there is a need for further development of tools dedicated to the characterization of large urban settlements.

Considering modern advances in machine learning algorithms, especially deep learning, one of the first issue is to constitute a comprehensive database, primordial to design mapping or classification tools. In addition to the data acquired by the numerous satellite orbiting the Earth, simulated data can advantageously supplement a learning database, by adding also new scenarios that were not yet recorded before, such as very large flood events or new buildings construction.

In order to master the difference of textures between images with different resolution or frequency bands, in this paper we are interested in the Sentinel-1 (S1) constellation, constituted of two C-band sensors designed for continuous monitoring, and the TerraSAR-X (TSX) constellation, constituted of three X-band sensors designed for a high accuracy elevation measurement. Both constellations have complementary roles for urban monitoring: S1 images have a large coverage and a fixed temporal repeat cycle of 6 days while the 1m resolution of TSX images enables a better characterization of the monitored scene. Nevertheless, due to the difference in wavelength and large resolution gap, the aspects of the S1 and the TSX images are different.

In order to deeply analyze these differences, we use a dataset of S1 and TSX images acquired over the “plateau de Saclay” in the South of Paris in 2016 and 2017, for which an extensive ground truth has been performed.

We have selected both ascending images in order to minimize the differences in geometrical configurations. For the TSX image, only the HH polarization is available. The product is a spotlight 1.3m resolution image. For the Sentinel 1 image, VH and VV are available, and GRD 10m products are used.

Differences in the aspect of the SAR images are the result from both the change of wavelength and the change in resolution and also in some new buildings in TSX that have been constructed between both acquisition dates.

All images have been geocoded by applying a terrain correction. Then, they have been coregistered to visualize their differences on a 10mx10m common sampling grid. The resulting colored composition is given in the attached graphics. Qualitatively, S1-VV and TSX-HH polarizations are more similar than S1-VH and TSX-HH.

Then, a texture analysis has been based on features extracted from the grey-level co-occurrence matrix (GLCM): the energy, the contrast, the dissimilarity, the homogeneity, and the correlation. The GLCM are computed from speckle and urban patches co-localized in the S1 and the TSX images. Due to the differences in resolution, the patches are 30x30 pixels for the S1 image and 180x180 patches for the TSX image, as it can be seen in figure 1 a) and b). Each patch is converted in logarithmic scale before being separately discretized in 32 levels. The GLCM are constructed for distances between pixels from 2 (in order to avoid the correlation created by the zero-padding) to 14 pixels and angles from 0 to 180°. For both sensors, the GLCM features have a similar behavior toward distance: contrast and dissimilarity increase with the distance, the energy does not vary, while the correlation and homogeneity decrease with the distance. Nevertheless, the value of these features differs, particularly for urban patches: the value of contrast and dissimilarity are larger in the S1 patches while the value of energy, homogeneity, and correlation are larger in the TSX patches. The differences in aspect between the two images is more reflected in the differences in the angular pattern of the GLCM-features, especially for contrast and dissimilarity, as represented in figure 1 c) et d). On this example, one building appear brighter on the S1, creating a bright line while separate scatterers appears on the TSX image. The contrast is larger in the direction of the bright line, while there is no angular variability in the contrast for TSX image.

In synthesis, global textures are similar on bare soil speckle areas.
On the other hand, in urban areas, the textures are very different: even after decimation, the TerraSAR-X images have more underlying structuring patterns. It is likely that the density of permanent scatterer is greater. This point remains to be confirmed by analyzing time-series. Finally, despite the apparent differences in texture, it remains possible for a human operator to detect real changes (appearance of some buildings). Also, the simulation of multisensor textures would make it possible to envisage more robust detection algorithms between images coming from these different sensors.

Area A - Board 91

The use of multi-band SAR data is investigated on a specific archaeological area in Egypt where some structures are visible and some others are supposed to complete the complex, in accordance with epigraphic and historical sources.
The area is located south of the city of Cairo, and encompasses the fifth dynasty sites of Abusir and Abu Ghurab, which represent a very interesting and stimulating case of investigation of sacred landscape in Ancient Egypt, being characterized by a unique combination of royal monuments (pyramids and sun temples) settled in a geo-morphologically uniform and chronologically well-defined framework.
This paper aims at exploring the processes of construction and formation of this sacred landscape thanks to the combined, interdisciplinary analysis of the modern technology of remote sensing with the available historical cartography of the area, and the data coming from the "traditional" topographical and archaeological investigation.
Considering the remote sensing approach, the area is characterized by a dry desert landscape with an expected high penetration of microwaves in terrain. Data from different SAR missions were exploited for this analysis with the main objective of providing an as much complete as possible representation of the full area, paying special attention to the identification of the four missing sun temples.
Data from X, C and L band missions (Alos-2, Sentinel-1 and COSMO-SkyMed) were considered, taking into account, in addition to the different wavelength, the different technical specifications, in terms of spatial resolution, polarization and others.
The joint analysis seem to suggest that two of the four missing sun temples might be located in Abu Ghurab, in between Userkaf's and Niuserra's sun temples, thus confirming the usefulness of SAR data in mapping subsurface architectural objects in a desert environment.

Area A - Board 92

Remote sensing technology is an essential component of the global monitoring of the ocean surface. In particular, SAR imagers are efficient sensors to detect maritime pollution, to guide cleanup operation or to identify polluters for prosecution. As radar imagery can be used any time of day or year and in almost any weather conditions, the use and processing of such remote sensing data is usually preferred to optical imagery. Nevertheless, images collected in the optical domain provide access to key information which is not yet available with SAR instruments, such as the layer thickness or the volume of pollutant. To fill this gap a methodology based on the joint use of a scattering model and SAR images is presented and applied to a set of remote sensing data collected by a low frequency imaging radar over controlled release of oil.
During an oil spill cleanup exercise organized by NOFO in 2015, mineral oil was released at sea in a control manner, with the main objective of testing mechanical recovering system in a realistic marine environment. The oil-on-water exercise was carried out in the North Sea between the Norway and the United Kingdom within 10 Nautical Miles of position 59° 59’N, 02° 27’E, from the 8th to the 14th of June. In this paper, we focus on remote sensing data collected on June 9, during the so-called MOS Sweeper exercise. The released product is an emulsion of mineral oil in water and consists of a mix of water, Oseberg crude oil and a small addition of IFO 380 (Intermediate Fuel Oil or marine diesel oil, with viscosity of 380 mm2.s-1). For the trial, 45 m3 of mineral oil emulsion was discharged at sea and recovered by the MOS Sweeper system.
Polarimetric SAR (POLSAR) images were recorded during the oil-in-water exercise by SETHI, the airborne remote sensing system developed by ONERA. POLSAR data were collected at L-band (EM frequency of 1.325 GHz) with a transmitted bandwidth equal to 150 MHz. The imaged area is 9.5 km in azimuth and 1.5 km in range, with incidence angles ranging from 34° to 52°. SETHI is a high resolution imaging radar characterized by a low instrument noise floor allowing a sufficiently high SNR over marine slick for efficient surface properties estimation.
With a remote sensing imaging radar collecting data at L-band, the Polarization Ratio (PR=HH/VV) is only slightly impacted by changes in surface roughness and depends mainly on the incidence angle and the dielectric properties of the surface. The proposed methodology for estimating the proportion of oil in an oil-in-water mixture is based on the comparison between the experimental values of the PR with the theoretical values predicted by an analytical scattering model, namely the Universal Weighted Curvature Approximation.
The methodology has been applied to L-band SAR data collected by SETHI at 09:37, 10:02, 10:27 and 13:30 UTC. The L-band SAR images and the estimated oil concentration maps have been projected onto the ground geometry and the corresponding normalized histograms are displayed below. The oil concentration evolved only slightly between 09:37 and 10:27 UTC (only 50 minutes), with no visible significant variation between the three histograms. Compared to the first three histograms, the last histogram (acquisition at 13:30 UTC, 3-hours later) is shifted to the lower values of oil content. This decrease of oil concentration is due to weathering process (wind, waves, evaporation …) whose impact is clearly highlighted on the oil concentration histograms.
It is reported in this paper that a combined utilization of a rigorous scattering model and L-band SAR images collected with a sufficiently high SNR allows to efficiently estimating the concentration of oil within a marine oil slick.

Area A - Board 94

Since the beginning of the 1990's, the number of satellites having on board SAR instruments has significantly increased. In addition, the spatial resolution of the currently provided products is improved reaching the submetric dimension for the TerraSAR-X (TSX) and the COSMO-SkyMed (CSK) missions. At the same time, the Sentinel-1 (S-1) Copernicus SAR data, generated by a constellation of two polar-orbiting satellites, will be free distributed for many years and with a few days revisit time. Moreover, new SAR configurations are planned to be launched in the very next future. Italian Space Agency CSK mission, thanks to its IEM (Interoperable, Expandable, Multimission) characteristics, will be expanded in the next months in order to form, jointly with the SAOCOM Mission (an L-Band SAR Argentinian mission), the SIASGE (Sistema Italo-Argentino di Satelliti per la Gestione delle Emergenze) System of Systems. The four CSK satellites already produce more than 1000 images per day and so, together with future SAOCOM data, a huge amount of different sensors data will be available to the SIASGE users. This new situation opens new challenges and new fields of research in Earth Observation, and, in particular, in SAR image processing. Indeed, one of the main issues is that such a wealth of information might be difficult to be managed by final users who therefore might not take fully advantage of the operations carried out by the space segments. If users, as usually happens, process all the available data separately, the interesting capability of effectively combining different sources of information may remain unexploited. In this context, the importance of data fusion algorithms is steadily increasing. Various techniques have been presented in literature regarding data fusion for remotely sensed data. A review on data fusion techniques between multispectral and hyperspectral data has been recently published [1] while in [2] fusion methods between SAR and optical images or including lidar and vector data are considered. However, even if multi-band SAR analysis of specific targets has been often performed [3], the design of specific pixel-based data fusion algorithms for SAR images acquired at different bands has been scarcely addressed so far. Neural networks (NNs) have been already used in the past as an effective approach to pixel based data fusion. Their competiveness for SAR and optical fusion was shown in [4] while in [5] they were exploited for dimensionality reduction of hyperspectral data, in [6] for the fusion of hyperspectral and panchromatic images, in [7] for the generation of products based on hyperspectral-multiangular-multitemporal acquisitions.
In this paper, an Auto-Associative Neural Network approach for the generation of SAR X-band CSK and C-band S-1 fused products is described. Indeed, with their capability of positively handling nonlinear relationships among data, Neural Networks seem a suitable technique for the management of the different scattering mechanisms stemming from multi-band acquisitions. In the study, the effectiveness of the data fusion technique is analysed in the case of the generation of land cover maps. In fact, the validation of the proposed approach is twofold. On one side the fused products are compared with the original ones by means of the results obtained on the corresponding classification maps. On the other side, a Principal Component Analysis (PCA) technique is considered as a benchmark algorithm for data fusion [8]. It will be shown that the fused products have various advantages. First, they retain all relevant information contained in the original data. In fact they slightly increased the overall accuracies with respect to the one obtained using the original data. Moreover, data dimensionality can be effectively and significantly reduced (halved in our case), which, on its turn, involves simpler processing modules and faster training times. In the applicative context, these results suggest that, in future, the distribution of “higher level” products, such as the NLPCs, might be an option to easy the data exploitation by final users. It will be also shown that, given the complex nature of the scattering mechanisms to be taken into account, a simpler data fusion approach based on a standard PCA is not as effective as the neural one. Next objectives of this research activity will be the inclusion in the merging process of either other SAR data (e.g. L-Band SAR SAOCOM) or also of optical data, in particular of Sentinel-2 data which have coverage and revisit time similar to S-1.
[1] N. Yokoya, C. Grohnfeldt, J. Chanussot, “Hyperspectral and Multispectral Data Fusion: A comparative review of the recent literature,” IEEE Geoscience and Remote Sensing Magazine, Vol. 5, No. 2, 2017
[2] J. Zhang, “Multi-source remote sensing data fusion: status and trends,” International Journal of Image and Data Fusion, 1:1, 5-24, 2010
[3] D. Latini, F. Del Frate and C. E. Jones, “Multi-frequency and polarimetric quantitative analysis of the Gulf of Mexico oil spill by means of different SAR systems,” Remote Sensing of Environment, v. 183, pp. 26-42, 2016
[4] F. Pacifici, F. Del Frate, W. J. Emery, P. Gamba, J. Chanussot, “Urban mapping using coarse SAR and optical data: outcome of the 2007 GRS-S data fusion contest,” IEEE Geoscience and Remote Sensing Letters, vol 5, n. 3, pp. 331-335, July 2008
[5] G. Licciardi and F. Del Frate, “Pixel unmixing in hyperspectral data by means of neural networks,” IEEE Transactions on Geoscience and Remote Sensing, vol. 49, n. 11, pp. 4163-4172, November 2011
[6] G. A. Licciardi, M. M. Khan, J. Chanussot, “Fusion of hyperspectral and panchromatic images using multiresolution analysis and nonlinear PCA band reduction”, EURASIP Journal on Advances in Signal Processing, December 2012, 2012:207
[7] R. Duca and F. Del Frate, “Hyperspectral and Multi-Angle CHRIS Proba Images for the generation of land cover maps,” IEEE Transactions on Geoscience and Remote Sensing, vol. 46, n. 10, pp. 2857-2866, October 2008
[8] C. Pohl & J. L. Van Genderen, Review article Multisensor image fusion in remote sensing: Concepts, methods and applications, International Journal of Remote Sensing, 19:5, 823-854

Area A - Board 95

Several SAR sensors are today orbiting the Earth and acquiring data with various wavelength and resolution. This offers new opportunities in many Earth Sciences domains but it also calls for new tools to process the massive amount of data and to benefit from the diversity of sensors.

We present here an automated mass processing InSAR tool based on
- an InSAR processor developed at the Centre Spatial de Liège (CSL), namely the CSL InSAR Suite (CIS; see Derauw et al. this session),
- a collection of shell scripts,
- the Multidimensional Small Baseline Subset (MSBAS) software (see Samsonov et al., this session)

This fully automated mass processing tool aims at performing optimized and self-evaluating mass processing of a large amount of radar interferometric pairs, from any kind of SAR sensor (ERS, ENVISAT, RADARSAT 1 & 2, TanDEM-X, TerraSAR-X, ALOS, CosmoSkyMed, Kompsat-5 and Sentinel-1 A&B) and with any orbital geometry, polarization, wavelength and ground based resolution.

The obtained geocoded deformation maps (which can be optionally interpolated to fill small gaps and/or detrended to remove residual phase plane) are automatically fed into the MSBAS software to obtain vertical and horizontal time series of ground deformation.

The shell scripts perform optimized sequences of processing for large amount of data, e.g. by sharing intermediate results between compatible processes etc. It does not parallelize the computations but split the number of InSAR processes in a number of parallel sessions chosen by the user. These sessions can be run on separate hard disks in order to minimize the performance loss due to large amount of simultaneous read and write operations on a given hard drive.

The processing also self-evaluates the quality of some steps such as the coregistration: if it is not satisfying, it will automatically change the coherence threshold and re-run the coregistration of that specific pair.
For Sentinel-1 data, when the precise orbit of an image already used in some interferometric pairs is made available, the processing tool will update all the required products and results.

The processing is also incremental, that is, each time (a) new SAR image(s) is (are) received, only the new InSAR compatible pairs satisfying chosen temporal and spatial baselines criteria are automatically processed. These new deformation maps are then added to the database ingested by the MSBAS processing, which only takes a few minutes to hours to be re-run depending on the amount of data.

Moreover, additional scripts tools are designed:
- to assist the quality check, the result integrity check, to quick look the data at several steps of the processing if required etc.
- to plot any kind of maps (amplitude, deformation, coherence etc…), or time series of deformation at single pixels or for double difference between selected pixels,
- to create sequence of amplitude images (incl. gif animation),
- to create coherence and/or water bodies masks
- to launch multiple run of MSBAS to assess the best regularization order etc.

In this poster, we illustrate the capabilities of the processing tools with the measure of the ground surface deformations affecting the city of Bukavu, DR Congo. Bukavu is a city of 800.000 inhabitants located south of Lake Kivu in the East African Rift, at the border of Rwanda and Congo. The region is characterized by steep topography affected by heavy precipitations which makes it prone to landslides. Urbanization is very dense and rapidly growing, often without appropriate planning.

We simultaneously use 109 Sentinel-1 and 234 Cosmo-Skymed SAR images acquired in ascending and descending orbits between March 2014 and October 2017 to analyze vertical and horizontal ground deformations in the city of Bukavu. The 343 SAR images allow to compute 1063 interferograms with temporal and perpendicular baselines of less than 70 days and 200-350 meters respectively.

The analysis of the time series allows the monitoring of a 1.5 km2 landslide which is identified and studied by field observations in Funu (a slum neighborhood in the city of Bukavu). The results are consistent with ground-based measurements performed during dGPS campaigns and with results from classical InSAR time series analysis (Nobile et al, 2018) performed with StaMPS (Hooper, 2008), that is a method which computes time series of ground deformation, though only expressed in the satellite line of sight and only for a given satellite in a given geometry at a time.

MSBAS also allows to identify several other landslides in the region. The method can also be used to detect and analyze any kind of ground deformation, from natural or anthropogenic origin (volcano monitoring, post mining or fluid related ground deformation such as from CO2 storage, water pumping, sediment compaction, monitoring of manmade structures such as bridges, warehouses, dams, highways, railways… ).

All the component are running on Linux or Mac. They can be obtained from the authors:
- CSL InSAR Suite can be obtained on request from the CSL Signal Processing Laboratory (contact : aorban@uliege.be)
- Shell scripts suite can be obtained on request from Nicolas d’Oreye (ndo@ecgs.lu)
- msbas can be downloaded from http://insar.ca/.

Area A - Board 96

Being part of the ESA Fringe group since its very first meeting in October 1992, the Centre Spatial de Liege (CSL) has a long-standing experience in the frame of SAR processing and SAR interferometry. Through numerous projects mainly funded by the Belgian Science Policy (BelSPo) and ESA, the CSL Signal Processing Laboratory has developed numerous tools now known as the CSL InSAR Suite (CIS).

CIS, as all tools of its kind, allows performing all processing steps for InSAR, DInSAR, PolSAR, PolInSAR, … including geoprojection of results in raw or ENVI format.

Additionally, it is a command line processor (coded in C language) having its own peculiarities:
- It allows handling most of the former and currently available sensors (ERS, ENVISAT, RADARSAT 1 & 2, TanDEM-X, TerraSAR-X, ALOS, CosmoSKyMed, Komsat-5 and Sentinel-1 A&B),
- In addition to Snaphu, it includes two homemade phase unwrapping processors,
- It allows performing Split Band interferometry (e.g. for absolute phase unwrapping or ionospheric phase screen estimation) and spectral coherence measurements,
- It allows masking the interferometric products before unwrapping based on masks provided in geographic coordinates to avoid perturbations due to low coherence areas,
- It incorporates tools for interpolation of small gaps and residual phase plane removal,
- It allows pixel coherence tracking.

With respect to Sentinel,
- It allows reading and processing only selected bursts encompassing the area of interest, or even cropping a region of interest within the burst(s) based on a kml file, to limit processing time
- For zone of interest that would span more than one image, it performs automatically the stitching of all the required images to cover in a single scene the desired footprint (provided by a kml file)
- Its high precision orbitographic computation allows performing Sentinel1 data coregistration without requesting Extended Spectral Diversity processing.

In many aspects, each modular InSAR processing function was fine-tuned and optimized to ease its full automation through shell scripts (see d’Oreye et al., this session) allowing to process all modalities (ascending and descending orbits, vertical and horizontal polarizations, whatever the satellite, the wavelength and the ground based resolution) from the reading of raw images up to the InSAR pair selection and the generation of differential phase, displacement, amplitude, coherence… layers in geo-projected coordinate system.

New modalities may be added as they come to extend the time series.

We will present some CIS specificities thit a powerful tool, specifically for – but not limited to - (M)SBAS processing (see Samsonov et al., this session). In particular, it allows referencing all SAR data of a given modality to a single super master image and performing (D)InSAR processing of all pair in the geometry of this super master.
Differential phase layers are then sent to the MSBAS processor to generate or update displacement time series, allowing to perform near real time monitoring of natural or anthropogenic ground deformation.

In this presentation, we will review the above-mentioned specificities and show some examples issued from full automation like:
- Vertical and horizontal MSBAS ground deformation results,
- time series analysis of shadow measurements for lava-lake leveling
- time series coherence analysis.



(Illustration: https://sentinels.copernicus.eu/documents/247904/3483128/Nyiragongo-Shadow-animation-Sentinel-CSK-small.mp4)

Area A - Board 97

Nowadays, several satellite missions are available providing interferometric SAR data at different wavelengths, spatial resolutions, and revisit time. A new interesting opportunity is provided by Sentinel-1, which has a spatial resolution comparable to previous ESA C-band sensors, and revisit times improved up to 6 days. According to these different SAR space-borne missions, the present work discusses current and future opportunities of Multi-temporal InSAR (MTI) applications to ground instability monitoring.
The mission parameters, which impact on the quality of the MTI products, are the wavelength of the SAR signal (decreasing from L, ~23 cm, to C, ~5.5 cm, to X band, ~3.1 cm), the revisit time, the spatial resolution, and the orbital tube. The figures that allow estimating the quality of the displacement maps are: the number of measurable coherent targets on the grounds (e.g. the PS density), the maximum detectable displacement velocity, the minimum number of processed images or the observation time span needed to derive reliable products, and the noise affecting both the final displacement and residual height estimation.
Issues related to coherent target detection, mean velocity precision and product geo-location are addressed through a simple theoretical model assuming backscattering mechanisms related to point scatterers. In particular, a comparative analysis is carried out, aimed at addressing specific advantages and limitations of different satellite missions.
This performance analysis may provide support for the SAR data selection, or for the definition of the geometrical and radiometrical configurations of future missions.
For instance, high resolution data are found to increase the density of coherent targets, thus improving the monitoring of local scale events. Short X-band wavelengths improve the sensitivity to displacements. Short revisit times allow collecting large data stacks in short times, and improve the temporal sampling, thus increasing the chances to catch pre-failure signals (high-rate, nonlinear signals). The precision of the displacement rate detection depends on the number of images and on the phase noise, while the precision of the residual height depends also on the orbital tube size. These relations have been formalized in a conceptual model, by assuming typical values for target characteristics such as persistent scatterer shape and signal to clutter ratio. The model allows to derive quantitative, although indicative predictions for the performance of several missions (present and future), based on the above mentioned mission parameters. In particular, for example, Sentinel-1 (S1), which will provide SAR data for the next years with short revisit times, results able to provide reliable displacement estimations at large scale, and in quite limited observation time spans (accuracy of about 1 mm/y for time series covering 1 year). However, the narrow orbital tube size, may limit the height precision, and consequently the geo-location quality (e.g. vertical precision above 20 cm with observation time span below 12 months, assuming S1-A/B 12 days repetitivity). X-band missions as COSMO-SkyMed (CSK) or TerraSAR-X, show the best performance in terms of precision on the estimated mean velocity and height, especially for limited observation time spans (e.g. precision of 2 mm/y for velocity and 3 cm for height, with 12 months of observation time and 1 acquisition per month). By considering the forthcoming new SAR missions (e.g. SAOCOM, Tandem-L, NISAR, RADARSAT Constellation Mission, PAZ, COSMO-SkyMED-SG, TerraSAR-X-NG), the resolutions and revisit times will be able to further improve the interferometric performances.
The paper also presents examples of multi-sensor ground instability investigation. More than 20 years of MTI SAR data have been processed, coming from both legacy ERS and ENVISAT missions, and last-generation RADARSAT-2, COSMO-SkyMed, and Sentinel-1 sensors. Moreover, a comparison has been performed between the outcomes of the model developed for assessing the MTI product precision, and the results obtained by processing the real datasets from ERS, ENVISAT, RADARSAT-2, COSMO-SkyMed and Sentinel-1 missions available for a real test case.

Area A - Board 98

INTRODUCTION
In the framework of the management of environmental risks, current space missions for Earth observation, associated with synthetic aperture radar (SAR) payloads, convey remarkable potential. In particular, the opportunity of joint availability of COSMO-SkyMed and RADARSAT-2 imagery offers very high resolution (VHR), all weather, day/night, short revisit time, and multifrequency (X- and C-bands for COSMO-SkyMed and RADARSAT-2, respectively) acquisition capabilities. The present abstract summarizes the scientific outcome of the SOAR-ASI 5245 project (2014-2016), which was conducted within the joint announcement of opportunity for COSMO-SkyMed and RADARSAT-2 data issued by the Italian Space Agency (ASI) and the Canadian Space Agency (CSA). The project involved three research laboratories based at the Universities of Genoa (principal investigator), Siena, and Pavia (Italy).
It focused on the development and experimental validation of multifrequency, multitemporal, and multisensor image-analysis methods for applications to risk monitoring. The integration of COSMO-SkyMed and RADARSAT-2 data with VHR optical imagery through data fusion was also addressed in collaboration with a research laboratory at INRIA (Sophia Antipolis, France). The following sections summarize the main approaches developed and tested within the project.

SUPERVISED JOINT CLASSIFICATION OF MULTIRESOLUTION COSMO-SKYMED, RADARSAT-2, AND PLEIADES IMAGERY
First, a novel method was developed for the supervised classification of multisensor SAR and optical images associated with multifrequency acquisitions at different spatial resolutions. The focus was on the fusion of X-band COSMO-SkyMed and C-band RADARSAT-2 SAR data along with optical visible and near-infrared (VNIR) Pléiades data. The proposed technique built on the theory of hierarchical MRFs [1] both to address the challenging task of the supervised classification of data that are multiresolution, multisensor, and multifrequency, and to focus on the synergy among three topical satellite data sources.
The rationale was to benefit from the fusion capabilities of hierarchical MRFs and avoid the, usually difficult, estimation of joint statistics of multisensor data. Inference was accomplished by generalizing an approach that was previously introduced in [2] for multitemporal fusion and was based on the marginal posterior mode rule. Probability density estimation was addressed using mixtures of Gaussian and generalized Gamma distributions and the stochastic expectation-maximization technique [3]. A case-specific topology composed of two distinct quad-trees [1] was introduced. Methodological details can be found in [4].
A complex VHR data set, collected over Port-au-Prince, Haiti, and made of a Pléiades pansharpened image at 0.5-m resolution, an HH-polarized COSMO-SkyMed spotlight image at 1-m resolution, and an HH-polarized RADARSAT-2 ultrafine image with 1.56-m pixel spacing, was used for experiments (Figure 1). The proposed method obtained rather high overall accuracy (OA) on the test set and remarkable spatial regularity (Figure 1). To explore its capability to benefit from X-band, C-band, and VNIR data, comparisons with the results obtained by using Pléiades data by themselves or together with either SAR image were performed. The results obtained using only SAR data led to poor accuracy. Quantitative assessment on the test set confirmed that the joint use of all three sensors led to substantially higher OA as compared with the use of a subset of the input sources. The use of only Pléiades data led to poor detection of the “urban” class (accuracy of only 61.66%). The joint use of this VNIR image and of X-band and C-band SAR allowed improving the discrimination of “urban” to 91.45%. These results suggest the potential of the proposed method in exploiting the complementarity of multisensor, multifrequency, and multiresolution data from current VHR satellite missions.
Further details and the discussion of other experimental comparisons with previous techniques can be found in [4].

ROBUST CROSS-MISSION CHANGE DETECTION WITH COSMO-SKYMED AND RADARSAT-2 IMAGES
This second proposed method addressed the problem of exploiting VHR multifrequency SAR data collected by COSMO-SkyMed and RADARSAT-2 to support risk monitoring and assessment in urban and suburban areas. An unsupervised change detection technique was developed, in which feature-level fusion was applied to a satellite image time series made of both COSMO-SkyMed and RADARSAT-2 acquisitions to improve the detection results as compared to those obtained through single-frequency (and single-mission) processing.
On one hand, the use of a time series of SAR multifrequency images may generally favor improved detection as compared to single-frequency analysis, because of the possibility to exploit the complementary properties of the two SAR sources to minimize the false alarm rate. On the other hand, a major difficulty in the application of conventional change detection methods is the heterogeneity of the images in the series, in which pixel intensities correspond to different frequencies. Indeed, the same type of event exhibits different signatures, depending on the characteristics of the sensor, thus making a direct, pixel-based, comparison between pre-event and post-event data almost unfeasible.
However, approaches based on global sample statistics and information theory [5] may address this problem and make it possible to perform robust change detection from multi-frequency SAR. Here, first, multitemporal features aimed at emphasizing ground changes [6] were separately computed from single-frequency and multi-frequency image pairs. Then, the resulting features were combined to detect changes. The former stage was based on the change feature in [6], which was based on the integration of information-theoretic distances with the mean-shift clustering algorithm. The latter stage was accomplished by combining the individual features resulting from the previous stage through case-specific normalization and geometric mean procedures. The change map could be automatically computed by applying an automatic thresholding to the resulting cross-mission feature, as in [5-7]. Further details can be found in [8].
The case study was the earthquake occurred on 2010 in Haiti. The 2.5 km × 2.5 km area under observation was centered on Port-au-Prince. Spotlight COSMO-Skymed acquisitions (1m-resolution) were used (pre-event on December 12, 2009, and post-event on December 23, 2011). A post-event 3m-resolution RADARSAT-2 image, acquired on December 8, 2011, was used (Figure 2). All such images were HH-polarized.
The change maps obtained from the single-mission and the cross-mission features are shown in Figure 2. An improved detection result was provided by the cross-mission feature, both in terms of geometric accuracy and false alarm rejection. Through a manually generated test map, a quantitative performance assessment was performed. It indicated Cohen’s kappa coefficients of 0.550 and 0.651 for the change detection results obtained by using only the COSMO-SkyMed images and the proposed cross-mission approach, respectively. On one hand, these kappa values were relatively low, a result due to undesired changes in the considered scene and to inaccuracies in the manually prepared test map. On the other hand, when considering the same detection problem, an 18% improvement in kappa was obtained owing to the joint use of COSMO-SkyMed and RADARSAT-2 image data for change detection and mapping purposes.

REFERENCES
[1] J.-M. Laferté, P. Pérez, F. Heitz, “Discrete Markov image modeling and inference on the quadtree,” IEEE Trans. Image Process., 9:390-404, 2000.
[2] I. Hedhli, G. Moser, J. Zerubia, S. B. Serpico, “A new cascade model for the hierarchical joint classification of multitemporal and multiresolution remote sensing data,” IEEE Trans. Geosci. Remote Sens., 54-6333-6348, 2016.
[3] G. Celeux, D. Chauveau, J. Diebolt, “Stochastic versions of the EM algorithm: An experimental study in the mixture case,” J. Stat. Comput. Simul., 55:287-314, 1996.
[4] I. Hedhli, G. Moser, J. Zerubia, S. B. Serpico, “Classification of multisensor and multiresolution remote sensing images through hierarchical Markov random fields,” IEEE Geosci. Remote Sens. Lett., 14:2448-2452, 2017.
[5] B. Aiazzi, F. Bovolo, L. Bruzzone, A. Garzelli, D. Pirrone, C. Zoppetti, “Change detection in multitemporal images through single- and multi-scale approaches,” in Mathematical Models for Remote Sensing Image Processing, G. Moser and J. Zerubia, Eds., Springer, 2017.
[6] B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, C. Zoppetti, “Nonparametric change detection in multitemporal SAR images based on mean-shift clustering”, IEEE Trans. Geosci. Remote Sens., 51:2022-2031, 2013.
[7] A. Garzelli, C. Zoppetti, “Optimizing SAR change detection based on log-ratio features”, Proc. Multitemp 2017, Bruges, Belgium, pp. 1-4.
[8] A. Garzelli, G. Moser, S. B. Serpico, “Robust change detection from COSMO-SkyMed and RADARSAT-2 multitemporal images,” Proc. IGARSS 2017, Fort Worth, TX, pp. 5252-5255.

ACKNOWLEDGMENT
The authors thank the French Space Agency (CNES), ASI, and CSA for providing Pléiades, COSMO-SkyMed, and RADARSAT-2 images, respectively. The SAR data were procured within the SOAR-ASI 5245 project. The authors also thank Prof. J. Zerubia (INRIA) and Dr. Ihsen Hedhli (Laval University, Canada) who collaborated to optical-SAR fusion.

Area A - Board 100

Low-land permafrost areas, with ice/water-rich active layers (the top layer of permafrost), are subject to intense vertical surface deformation processes, due to phase changes from ice to liquid water. Annually, downward movement of the land surface (subsidence) associated with seasonal thaw in summer is compensated by upward movement associated with frost heave in fall. Seasonal changes in elevation can reach decimetres every year. If seasonal thawing in summer dominates in the long term average over frost heave in fall, multiannual ice in an ice-rich permafrost body underlying the active layer is melted and an effective multiannual subsidence of the surface can occur. The precise vertical position of the Earth’s surface over low-land permafrost over multi-annual time scales can thus be a direct measure of climate change impacts on permafrost.

Satellite SAR interferometry (InSAR) has been applied in the past to measure surface deformation over low-land permafrost during thawing seasons and to derive remotely-sensed seasonal changes in active-layer thickness. Seasonal as well as year to year developments in the freeze-thaw cycle and subsequent subsidence have been identified using SAR data from various satellite missions. Thereby, interferometric phase coherence over these typically moist and vegetated areas is a critical factor. Over the Yamal Peninsula, located within a region of continuous permafrost in Russia where tundra lakes and river flood plains are the most prominent landscape features, we analyzed a series of satellite SAR data at L-band from ALOS-2 PALSAR-2 between 2015 and 2018, C-band from Sentinel-1 between 2015 and 2018, and X-band from Cosmo-SkyMed between 2013 and 2017. In our contribution we will discuss the effect of phase coherence for the processing of short-time (i.e. a few days) and long-time (i.e. one year) interferograms, present and compare summer and multi-annual subsidence maps derived from the different sensors and wavelengths, and give an outlook over possible future synergistic contribution of L, C, and X-band data for the monitoring of surface deformation in low-land permafrost areas.

Area A - Board 101

The Campine Coal Basin was extensively mined in the 20th century. After closure of the last mines in the early 1990s, pumping of mine water stopped, and water levels started to rise. Increasing pore pressures produced by the rising groundwater has caused significant uplift in the former mine concession areas. Cultural heritage and infrastructure may suffer from structural instabilities caused by differential ground movements.

In this study we analysed timeseries of ERS, Envisat, Sentinel-1 and Cosmo-SkyMed data through PS-InSAR processing. Combined, these timeseries provide 25 years of deformation history of the area. The observed displacement patterns are consistent with the extent of concession areas, timing of mine closures, and the location of faults in the subsurface. The results clearly demonstrate that satellite-based radar measurements provide a spatio-temporally dense network of observations and offer important advantages for the analysis of complex deformation mechanisms in (former) mining areas.

Further work is focussing on the integration of the deformation maps with other geospatial data layers such as geology, land-use and the location of the built heritage; feature-based data fusion techniques and decision rules based on geomechanical expertise will be combined to create ground movement risk maps. In addition, selected case-studies will be investigated at high resolution by means of on-site monitoring techniques as well as stability analysis to evaluate the applied approaches.

The presented work is part of the GEPATAR project (GEotechnical and Patrimonial Archives Toolbox for ARchitectural conservation in Belgium). Aim of this project is to create an online interactive geo-information tool that integrates information about Belgian heritage buildings and the occurrence of ground movements. The toolbox will allow the user to view and be informed about buildings potentially at risk due to differential ground movements and thus help improving the management of the built patrimony.

Area A - Board 102

Izmit Bay, tectonically, under control of the strike slip North Anatolian Fault System (NAFS) is a depression basin bounded by Kocaeli Peneplain in the north and Samanlı Mountains in the south. After the big earthquake, Mw=7.4 in 1999, many geodetic and geophysical research projects have focused on the monitoring of the post-seismic and inter-seismic crustal movements along NAFS, towards the Marmara Sea. Apart from the GPS derived kinematics, horizontal and vertical deformation patterns in Izmit Basin are investigated using multi-band InSAR techniques based on image archives for the time period of the last 5 years. Three sets of images, two of them from the Sentinel 1A and the other one from COSMO-SkyMed (CSK), have been used to map spatio-temporal, horizontal and vertical displacements induced by NAFS, soil instability, liquefaction, etc. An automatic processing strategy of a stacked SAR data was utilized to generate the time series of interferograms. The active deformations are spreading into the basin depending on the tectono-morphology of the Cenozoic era. The estimated time series of the small baseline subset algorithm reveal significant subsidence occurrences over the low topographical reliefs in the range of 0-50 m. The tendency of subsidence following mainly river beds increases its intensity as getting closer to the shore of sea or lake which are potentially very hazardous areas in the moment of a seismic activity. In addition to vertical displacement movements up to 8 cm per year in the basin plains, the fault parallel horizontal motion of the Anatolian plate of 1.5-2.0 cm/yr has been identified with respect to the Eurasian Plate.

This study aims the evaluation of coherency of CSK with Sentinel 1A results. The coherency analysis is performed in terms of horizontal and vertical velocity components. A further research will be proceeded by using GNSS CORS data for evaluation and verification of the InSAR velocity fields within the study area.

Area A - Board 103