Science Service System
You are here : Home : Proposals_Summary

Summary of Proposal HYD0776

TitleLevee screening using multipolar TerraSAR-X data to identify potential failure sites
Investigator Hasan, Khaled - Mississippi State University, Geosystems Research Institute
Team Member
Associate Aanstoos, James - Mississippi State University, Geosystems Research Institute
SummaryFailure of levee can produce disastrous consequences on life and property as evidenced in the aftermath of Hurricane Katrina, Hurricane Ike and flood events that occur regularly across the globe. This project will enhance efforts to develop tools and methods based on the use of SAR imagery to achieve rapid and efficient screening of levee reaches. SAR's all weather capability and ability to detect displacements in interferometric mode will be tasked to detect changes in levees that may signal impending failure. The pre-emptive screening using SAR images will identify levee sections that exhibit characteristics that make the reach vulnerable to failure under flood loading. This identification will lead to more detailed examination or repairs of these higher-priority sections. We are currently working with airborne NASA UAVSAR, a polarimetric L-band SAR and are proposing to investigate the usefulness of satellite based SAR imagery in carrying out the same task. The proposed work will examine the applicability of the RADARSAT-2 and the TerraSAR-X imagery, and will also demonstrate a new and efficient approach for taking in situ soil measurements. The primary objectives of this research are to: (1) Develop algorithms and software to utilize satellite-based SAR systems to screen levees for vulnerabilities, and (2) Test and demonstrate a new approach for taking in situ soil measurements for ground truth validation. We will analyze the polarimetric and interferometric SAR data and intend to test and refine the polarization feature detection algorithms we have developed for the UAVSAR data with the acquired satellite SAR data. Deformation maps from the interferometric SAR data will allow us to detect features indicative of weakness in the levee segment. DEM, optical data and ground truth will be used to assist information extraction from the images. The image analysis will investigate existing approaches and develop new algorithms and a study will be performed to determine which is preferred, based on required input, skill of the end-user, and quality of output. The ability of different sensors to detect different features will be utilized by combining the outputs from the analysis of TerraSAR-X and Radarsat-2 imagery in a decision fusion to determine the final classes. Knowledge -based expert system will be employed to incorporate soil type, elevation, volumetric water content and surface roughness as ancillary data for SAR image classification. The primary data for this research will be the high resolution polarimetric SAR data from TerraSAR-X and Radarsat-2. Image data at 1-3 m resolution will be required to detect slides, sand boils and high soil moisture content indicative of levee weakness. We plan to utilize the quad polarized images and determine the best-suited polarization mode to and also compare the efficiency of the ascending and descending pass images. To support the SAR analysis we will also use optical imagery. Soil moisture data will be collected from field measurements to match with image data acquisition. Similarly, data on climate, hydrology, soils, geology, topography and landuse-landcover will be integrated in the analysis. The primary deliverables are the prototyped algorithm for Levee Segment Classification. Results will be published in appropriate scientific journals and conferences. A technical report may be submitted to DLR if required. The current project is funded by US Department of Homeland Security under their SERRI program. Limited funds have been allocated in the budget for SAR image purchase and at the commercial price of Infoterra only a small number of TerraSAR-X images can be purchased, limiting the scope of the study. If approved for scientific pricing more images of various beam and polarization configuration can be acquired to make the research comprehensive which will allow a more vigorous evaluation/validation of TerraSAR-X imagery.
Final ReportOVERVIEW The overall purpose of this research is to develop methods for improving knowledge of the condition of earthen levees based on remote sensing technology, giving levee managers new tools to prioritize their efforts. This project focused on methods based on the use of space-based synthetic aperture radar (SAR). For this work we partnered with the US Army Corps of Engineers (USACE) which has responsibility for maintenance of many of the levees in the US. The study area for this project is most of the mainline levee system of the Mississippi River in the state of Mississippi (Figure 1). Data were collected to support the investigation of the use of remote sensing to analyze physical factors that could indicate problems in levee conditions, whether they arise from moisture content such as slope stability, hydraulic uplift, water seepage through levees, or underseepage resulting in sand boils on the land side. REMOTE SENSING DATA For this project we used TerraSAR-X and TanDEM-X, earth observation SAR instruments operated by the German Aerospace Center (DLR). Both instruments are satellite-borne, side-looking polarimetric X-band imaging systems acquiring images at 9.65 GHz corresponding to 3.12 cm wavelength. We chose to use High Resolution Spot (HS), SpotLight (SL), and StripMap (SM) modes to address different aspects of the study, with HS giving the highest detail and SM the broadest regional coverage. The SpotLight (SL) images provided the optimum balance between ground resolution and scene extent, and the majority of the images were collected in this mode. With the exception of four images for which timing was critical in order to cover extreme flood events, all were collected from the right-looking ascending path as that gave maximum parallelism between the satellite orbit and the course of the Mississippi River in our study area. Images of 2010 and early 2011 were collected with incidence angles ranging from 30° to 36°. FILTERING It is a common pre-processing approach to reduce backscatter speckle noise by applying filters that suppress the speckles and minimize the contrast within a SAR image. A study conducted by Hassan et al. (1999) found the Frost filters a good choice for mapping land cover in floodplains, reducing contrast while retaining feature boundaries. For images with different ground resolution and extent, different window sizes and coefficients of variation were used. HS images were filtered with a 19×19 window size and 0.45 coefficient of variation; SL, with 11×11 windows and 0.4 coefficient of variation; and SM, with 11×11 windows and 0.45 coefficient of variation. (Figure 2) INCIDENCE ANGLE CORRECTION Despite having relatively consistent slope within a section of levee, there can be significant variations in the orientation of the levee with respect to the radar look direction. Using a digital elevation model (DEM) obtained from high resolution LiDAR data and the orientation and local incidence angle computed from satellite orbit parameters we corrected the TerraSAR-X image (Figure 3) by applying the calculated local incidence angles to the calibrated backscatter coefficients. GROUND DATA To provide detailed information on soils on the levees in our study area, “ground truth” data were collected by both USACE and our own team. This data included the exact location and timing of slump slides, photographs of the vicinity, and notes on grass height. In addition, we took samples of soil moisture and measured soil electrical conductivity (EC). (Figure 4) SLIDE CLASSIFICATION ALGORITHMS AND RESULTS The flow chart in Figure 5 illustrates the approach used to test and assess slide classification algorithms. Due to the relatively small number of ground truth pixels available a “leave-one-out” cross validation technique was used to estimate overall accuracy. In this approach, all samples but one are used to “train” the classifier, and the sequestered sample is used for testing. This is repeated in a round-robin manner until all samples have been tested. The features used as inputs to the classifier included both the magnitudes of the polarimetric channels HH and VV as well as a type of texture feature – that is, one that is calculated from a neighborhood of pixels around the pixel under evaluation. The texture feature we used is based on the wavelet transform, a kind of spatial frequency analysis (Burrus 1997). To obtain the wavelet coefficients, a transform using a Daubechies mother wavelet was applied to each pixel of the HH and VV TerraSAR images using a 7 × 7 sliding window. The window was decomposed into seven wavelet coefficient sub-bands (horizontal, vertical, and diagonal detail coefficient sub-bands for levels 1 and 2 along with an approximation coefficient sub-band). By calculating the mean and standard deviation of the coefficients’ energy for each of the seven sub-bands, 14 corresponding values are obtained for each pixel of the HH and VV image, which are the wavelet features we used. Some of the algorithms tested were supervised – requiring training—and others were unsupervised. The best of each class were the support vector machine (SVM) algorithm and the RX Anomaly Detector (RXD). SVM attempts to find a separating hyperplane in a given feature space. The input data are first transformed into a feature based on a kernel function. Next, a hyperplane that separates the classes is computed by applying an optimization method. Figure 6 shows the results of applying this method over the area of study. Pixels classified as healthy (non-slide) are green, and those classified as slide pixels are blue. It can be seen that the actual slide area was well-detected, but there are a number of false positives. The density of these can be seen to be much lower than the true positives. Figure 7 shows how the accuracy of this SVM classifier is affected by the quantity of training samples used. The RXD algorithm detects the signatures that are distinct from the surroundings with no prior knowledge (Chang 2002). The algorithm uses the covariance matrix which calculates the Mahalanobis distance from the test pixels to the mean of the background pixels. The RXD algorithm was applied to the September 15 TerraSAR-X HS data in the area of an active slide. Using a window size of 4x4, the features input to RXD were the magnitudes of the HH and VV polarimetric backscattering coefficients and the wavelet features computed from them. The classification map obtained from the RX detector unsupervised classifier is displayed in Fig. 8. The output produced by the RX algorithm is a grayscale image, and anomaly detection is performed by visual interpretation. A threshold can be applied to distinguish targets from the background. In the figure, the magnitude of the output is color-coded. The range of values coded as blue appears predominantly in the slide area. If a threshold is chosen to highlight those values, then the slide would be well detected in such an image. In practice, the threshold appropriate for a given dataset may differ, so a tool that allows the user to vary a threshold while observing the result would facilitate the interpretation. CLASSIFICATION OF ANCILLARY FEATURES In addition to testing algorithms for direct classification of levee slides, we studied the feasibility of using radar data to estimate other features related to levee vulnerability such stressed vegetation and water extent Stressed vegetation is often found around slump slides on levees. Classification of water extent could support the detection of areas of through-seepage and possible sand boil flows. We used TerraSAR-X imagery to classify water extent for five selected dates in 2011 and 2012. The Frost filtered images were used in this simple threshold-based classifier. Within the different polarizations, the threshold was exactly the same for water class and varied slightly for the other three classes. After the four-class classification, comparisons were made between different dates. The effect of polarization was investigated by comparing results of HH and VV outputs for the same date. The changes between them were insignificant, and in every case VV images classified about 1% more areas as water compared to HH. In general the water extents for all dates were around 11% of the tested area. The May 2011 flooding was a 73 year record event, which doubled the water extent in this area, as detected by this method to 22% of the area covered. Although the 2011 floodwaters on agricultural land on the land side of the levee were detectable, none of these water-covered areas could be associated with reported sand boils (Figure 9). We investigated potential for using multi-temporal TerraSAR-X data to classify levee vegetation characteristics that may allow us to detect vulnerable levee segments by monitoring changes in vegetation type, stress, or vigor. Field observations detected zones of stressed vegetation accompanied by cracks during the drier months from October 2011 to February 2012. Field polygons on these stressed vegetation areas were used to classify speckle-filtered multi-temporal time series of HH, VV, and HH-VV polarization in supervised mode into 4 classes: stressed grass, green grass, dormant grass, and wet/moist soil. Comparison of the classification results showed that both HH and VV polarization were able to classify some common pixels of stressed grasses, and individually they classified some additional pixels. A GIS model was run to combine the individual outputs to produce a 3 class output showing the stressed grass class (Figure 10). Most of the known slide locations showed up as stressed grasses in the classification. Some of these areas unassociated with any known slide events were randomly checked in the field to reveal anomalous land cover dominated by Bermuda grass with cracks on the ground. These land cover vegetation characteristics were similar to those of the slide areas and are being considered as potentially vulnerable levee zones. CONCLUSIONS The various approaches of image analysis produced some meaningful methods of characterizing vulnerability of levee segments within the study area. SVM, RX anomaly detector and stressed vegetation maps provided useful methods of identifying areas that had previous slides, repaired slides and anomalies that could be the site of future failures. The lack of penetration depth by X-band SAR proved to be a limitation in mapping during full vegetation cover on the levee. This could be countered by collection of subsurface data through detailed soil investigation and geophysical tools in future studies. More intensive ground data to support image processing and interferometric SAR is also expected to provide better insight in future studies.

Back to list of proposals

© DLR 2004-2016