Terrain subsidence and landslides II

Structural health assessment is an important practice to guarantee the safety of infrastructure in general. In case of dam monitoring, it is necessary to control the structure itself and the water reservoir, to guarantee efficient operation and safety of surrounding areas. Ensuring the longevity of the structure requires the timely detection of any behaviour that could deteriorate the dam and potentially result in its shutdown or failure. Traditional structural dam monitoring requires the identification of soil movements, tilt, displacements, stress and strain behaviour.

The detection and monitoring of surface displacements is increasingly performed through the analysis of satellite Synthetic Aperture Radar (SAR) data, thanks to the non-invasiveness of their acquisition, the possibility to cover large areas in a short time and the new space missions equipped with high spatial resolution sensors. The availability of SAR satellite acquisitions from the early 1990s enables to reconstruct the historical evolution of dam behaviour, defining its key parameters, possibly from its construction to the present. Furthermore, the progress on SAR Interferometry (InSAR) techniques through the development of Differential InSAR (DInSAR) and Advanced stacking techniques (A-DInSAR) allows to obtain accurate velocity maps and displacement time-series.

The importance of these techniques emerges when environmental or logistic conditions do not allow to monitor dams applying the traditional geodetic techniques. In such cases, A-DInSAR constitutes a reliable diagnostic tool of dam structural health to avoid any extraordinary failure that may lead to loss of lives.

In this contest, an emblematic case will be analysed as test case: the Mosul Dam, the largest Iraqi dam, where monitoring and maintaining are impeded for political controversy, causing possible risks for the population security. In fact, it is considered one of the most dangerous dams in the world because of the erosion of the gypsum rock at the basement and the difficult interventions due to security problems. The dam consists of 113 m tall and 3.4 km long earth-fill embankment-type, with a clay core, and it was completed in 1984. It started generating power on 1986.

Specific objective consists in determining the degree of detail of dam surface strains that can be obtained from different satellite SAR datasets at different resolutions (microwaves X and C bands). Therefore, different datasets are analysed: the archive available SAR data (ERS and Envisat from ESA), the currently acquiring Sentinel data (EU Copernicus programme) and the high-resolution COSMO-SkyMed data (ASI program) over the study area (Mosul dam).

The different stacks of data are processed applying SBAS and PS A-DInSAR techniques; the deformation fields obtained from SAR data are evaluated to assess the temporal evolution of the strains affecting the structure. Obtained results represent the preliminary stage of a multidisciplinary project, finalized to assess possible damages affecting a dam through remote sensing and civil engineering surveys.

The Sacramento-San Joaquin delta (Delta) contains more than 1700 km of levees that protect various reclaimed lands from flooding. Most of the delta is experiencing subsidence at rates that can exceed 5 cm/yr locally, and which can affect the structural integrity of the levees. In-situ and airborne LIDAR monitoring of this extensive levee network is expensive, making Interferometric Synthetic Aperture Radar (InSAR) an attractive, cost-effective alternative that can provide uniform and consistent monitoring. InSAR has proven to be a powerful technique to study surface displacements at high accuracy (few mm/year), over large regions (up to 250 km wide swaths), and at a high spatial resolution (up to a meter). However widespread usage of InSAR, particularly within the application community, is challenged by several technical issues, the most significant of which are decorrelation noise introduced by a change of scattering properties (e.g., moisture and vegetation), and noise due to variation in atmospheric properties between different SAR acquisitions (i.e., tropospheric delay).  These effects are particularly limiting in the rural/agricultural setting of the Delta.  We demonstrate the usage of InSAR for spatially comprehensive subsidence monitoring both at the scale of the levees and at a scale that captures the intra-island variability.  The study uses data collected over a period of six years (2009-2015) with NASA’s Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) instrument, which is the prototype airborne instrument for the NISAR mission. We mitigate atmospheric noise by estimating a correction from state-of-the-art weather models, and reduce decorrelation noise by utilizing L-band SAR and using advanced time-series InSAR processing methods. Our analysis includes nine UAVSAR flight lines that cover altogether an area of approximately 8500 km², including the Delta and the surrounding areas. 

The Geohazards Exploitation Platform (GEP) is a European Space Agency (ESA) initiative within the ecosystem of Thematic Exploitation Platforms (TEP) focuses on the integration of Ground Segment capabilities and ICT technologies to maximize the exploitation of EO data from past and future missions. A TEP refers to a computing platform that deals with a set of user scenarios involving scientists, data providers and ICT developers, aggregated around an Earth Science thematic area. The Exploitation Platforms are targeted to cover different capacities and they define, implement and validate a platform for effective data exploitation of EO data sources in a given thematic area.

In this framework, the GEP aims at providing on-demand and systematic processing services to address the need of the geohazards community for common information layers and, finally, to integrate newly developed processors for scientists and other expert users.

One of the pilot services to be available in the first quarter of 2017, FASTVEL, is being developed by TRE-ALTAMIRA. The aim of FASTVEL is to provide a robust tool for generating terrain displacement mean velocity maps from a stack of SAR images, namely ERS, ENVISAT-SAR and Sentinel-1. The main developing effort concerns the automation of the unsupervised processing with the minimum number of initial parameters. In this work, the developed methodology in the context of a cloud-based workspace will be described.

FASTVEL will be an on-demand service. In order to test and validate the service a list of users among the geohazards community will be selected. In this testing phase, different scenarios (seismic, landslides, underground mining, urban,…) affected by ground motion will be analysed in order to provide a complete characterization of the service. In this work, representative study cases and the corresponding user experience will be described.

Two more related GEP services developed by TRE-ALTAMIRA, will be also described in this work. The first one is the already available GEP Diapason to whom it will be added the phase unwrapping step. The second one is a post-processing service for PSI results (“PSI Post-proc”) that will be also available in GEP. This service will allow projecting PSI displacement in the line-of-sight (LOS) direction onto a predefined direction of real ground displacement (east-west, up-down or down-slope) using one single orbit pass or both, filtering points by means of a phase quality indicator or geometrical distortion masks (foreshortening or layover), computing the acceleration field and changing the reference point.

Introduction and Rationale

In the current Remote Sensing scenario that is characterized by the huge availability of SAR data, the use of Cloud Computing platforms plays an ever-increasing role for performing DInSAR analyses at very large scale and maximizing the exploitation of such big data archives.

In this paper we focus on the latest advances in cloud computing solutions for the generation of Earth deformation time series through the Parallel Small Baseline Subset (P-SBAS) approach. Indeed, recently, the P-SBAS automatic processing chain has been implemented within the Amazon Web Services public cloud environment [1] and the attained scalable performances have been analyzed by also identifying the relevant major bottlenecks [2]. The P-SBAS chain is constituted of several processing steps – from the SAR raw data focusing up to the time series generation - very different from the algorithmic viewpoint and in terms of computational requirements (CPU usage, RAM occupation, Input/Output (I/O) throughput). They exploit both multi-core and multi-node programming techniques to parallelize the codes and cut down the processing times [3]. The achievement of a high scalability– which means decreasing the processing times when the number of computing nodes exploited for the analysis increases, by maintaining unvaried the computational efficiency – for a complex processing chain such as P-SBAS, within a cloud environment, is not a trivial issue. The major bottleneck is characterized by the very large data flow to be read and written during the P-SBAS processing, especially in the case of DInSAR analyses at very large scale, when the input datasets are of hundreds of GigaBytes and produce, in some intermediate steps of the processing chain, an I/O data flow of almost one order of magnitude greater [4]. Indeed, concerning the CPU and RAM issues, the adopted parallelization strategies and the large collection and typology of resources available within the AWS environment allow us to optimize the computation and to reduce the processing times by increasing the number of computing nodes that work in parallel. On the contrary, regarding the I/O workload, such a rationale is not straightforward applicable. Indeed, due to the nature of the implemented algorithm, there are some steps of the P-SBAS processing chain characterized by unsolvable data dependencies; therefore we cannot avoid a data sharing logic among different computing nodes. This means that, when the number of parallel processes of the P-SBAS processing increases, and therefore of exploited computing nodes, the simultaneous accesses to the shared data increases as well, thus generating a limitation for the scalability, even though the disk access bandwidth and the performances of the network linking the different computing nodes are high.

A solution to this issue is the implementation of a distributed storage computing architecture that is ad-hoc designed for the P-SBAS processing chain, aimed at splitting as much as possible the I/O workload among different nodes but maintaining the data sharing. This is realized attaching to each computing node its own storage disk with high I/O performances, and mutually connecting all the disks through a Network File System (NFS) protocol. In this way each node/disk can access/be accessed by all the other nodes/disks to read and write data. Moreover, to fully take advantage of such architecture, also the parallel jobs scheduling strategy has to be properly defined. In particular, each computing node has to work as much as possible on data that are physically located on its own local disk, minimizing the transfer of data among different nodes and therefore the network occupation. Consequently, for each step of the P-SBAS processing chain, depending on the specific operation that is carried out, the reading and writing of data has to be properly managed. In this way the overall architecture will have a I/O bandwidth equal to the one of a single disk multiplied by the number of exploited nodes, thus essentially eliminating the bottleneck.

The presented P-SBAS cloud computing solution allows carrying out extensive interferometric processing, moving the DInSAR analysis scenario from local to continental scale.

Results

To show the potential of the P-SBAS cloud computing solution presented in this paper, we carried out a large-scale DInSAR analysis regarding a South California area extending for about 90.000 km². In particular, we exploited the full SAR raw data (level-0 imagery) archive acquired over this region by the ENVISAT ASAR sensor, both from ascending and descending orbits. The considered dataset is composed of 35 ENVISAT frames (17 from ascending and 18 from descending orbits), each one having on average 40 SAR images, with a total size of about 400 GB.

The initial input dataset was stored into the AWS S3 storage and downloaded for processing.

Concerning the implemented computing architecture, we exploited in total 280 AWS instances that worked in parallel, 8 instances per each ENVISAT frame, with one instance corresponding to a single computing node (see Fig. 1). As for the storage volumes, we exploited 280 disks of AWS (provisioned IOPS SSD), which are suitable for I/O-intensive workloads, with a size of 120 GB and a disk access bandwidth of about 250 KiB/s each one. Therefore, the overall employed storage was of 33.6 TB.

The whole processing, including the overall computing architecture configuration phase (which has been automatically performed through bash scripts also exploiting AWS libraries) as well as the P-SBAS processing of the 35 ENVISAT slices, lasted approximately 8 hours and cost about 1900 USD. The average processing time and cost per each frame have been less than 6 hours and 53 USD, respectively.

In Figs 2 and 3 we show the overall LOS mean deformation velocity maps generated through the P-SBAS processing over the selected area, for the ascending and descending orbits, respectively. For the sake of uniformity, the LOS velocities have been computed by considering the same time period 2005-2010. Some discontinuities are present between the mean velocity deformation maps, which are due to the fact that these are spatially referred to different points, that adjacent tracks illuminate the same ground area with different look angles and that possible effects of residual orbital phase errors can be present.

Moreover, in Figs 4 and 5, we present some plots showing the comparisons between the displacement time series retrieved from the LOS measurements (black triangles) and the corresponding GPS ones (red stars), in some pixels affected by significant deformation patterns. As it is clear, a very good agreement is found between the SAR and LOS projected GPS measurements. It is worth noting that the regional trend was removed from the LOS mean deformation velocity maps presented.

[1] I. Zinno et al., "Cloud Computing for Earth Surface Deformation Analysis via Spaceborne Radar Imaging: A Case Study," in IEEE Transactions on Cloud Computing, vol. 4, no. 1, pp. 104-118, Jan.-March 1 2016.

[2] I. Zinno et al., "A First Assessment of the P-SBAS DInSAR Algorithm Performances Within a Cloud Computing Environment," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 8, no. 10, pp. 4675-4686, Oct. 2015.

[3] F. Casu et al., "SBAS-DInSAR Parallel Processing for Deformation Time-Series Computation," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 7, no. 8, pp. 3285-3296, Aug. 2014.

[4] I. Zinno; F. Casu; C. D. Luca; S. Elefante; R. Lanari; M. Manunta, "A Cloud Computing Solution for the Efficient Implementation of the P-SBAS DInSAR Approach," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol.PP, no.99, pp.1-16

Terrain subsidence and landslides II