Polarimetric Synthetic Aperture Radar
Starting from early 2006 CPI has supported NRL in the development and implementation of algorithms for the analysis of Polarimetric Synthetic Aperture Radar (POLSAR) imagery. The latter provides information on changes in the polarization state of electromagnetic waves reflected from the earth surface that can be exploited to distinguish between man-made and natural objects. Research is primarily focused on extending the limits of information that can be extracted from POLSAR imagery acquired from radar systems on airborne and space-based platforms. This includes exploring what is capable using both full quadrupole polarimetry (quadpol) and dual polarimetry (dualpol) subsets. The focus of CPI's work in this area has been to advance development of POLSAR classification algorithms, implement polarimetric SAR processing capabilities to NRL's SAR Workbench, and develop improved algorithms for calibration of POLSAR imagery.
Polarimetric SAR Classificiation
POLSAR is sensitive to the orientation and character of objects from which the radar signal returns and has been extensively used for terrain classification. Target decomposition theory is one approach that has been widely used to extract information from POLSAR imagery by decomposing the radar returns in terms of different components such as surface or single-bounce scattering, double-bounce scattering, and volume scattering. Using such target decomposition concepts it is possible to classify objects within POLSAR imagery as being due to natural targets such as vegetation and water, and man-made targets such as buildings. CPI has developed and implemented several POLSAR classification algorithms for NRL and is active engaged in the leading edge of POLSAR terrain classification research.
CPI developed an algorithm for feature-driven classification in SAR images using the Intensity Driven Adaptive Neighborhood (IDAN) method for filtering and segmentation of polarimetric information. Both the filtering and the segmentation programs, originally designed to work with three-pole datasets, were adapted to work with a full six-pole dataset. The segmentation routine was found to work well with regions where there are clear intensity-based divisions (crops, roads, etc.), while regions with primarily polarization-based divisions (forests) are found to not be segmented effectively. The algorithm was enhanced to merge segments based on the Wishart distance between their clustering centers, and a function was incorporated to color and display the segmented image using the Freeman-Durden classification.
CPI also has extensive experience in target decomposition algorithms for terrain classification, having extended and implemented several algorithms within the NRL SAR Workbench: Cloude-Pottier, Freeman-Durden, Yamaguchi, and Wishart. Always striving to create the most useful software product for the end user, CPI devoted considerable attention to developing color rendering methods for each target decomposition algorithm to accentuate the differences between natural and manmade targets in the imagery. Examples of the color imagery are highlighted in the description of the NRL SAR Workbench.
NRL SAR Workbench
Over the course of several years the NRL SeaLab group has been developing a suite of software applications for analyzing SAR imagery referred to as the NRL SAR Workbench, and providing limited distribution of the software package to interested parties. This suite of tools are software applications with a graphical user interface (GUI) written in the Java programming language that provide visual displays of SAR images, interactive selection of areas of interest, and tools for processing and analysis of those selected regions. The image processing, visual color rendering, and analysis tools are written in the C programming language and are called from the Java source code as C compiled executables. Through work performed by NRL, its contractor CPI, and a MITRE consultant (Simplex, LLC), the NRL SAR Workbench was extended to include a polarimetric SAR processing capability. The POLSAR processing features were implemented via the addition of 3 panels organized into Image Display, Classification, and Individual Tools capabilities.
Image Display Panel
The Image Display panel provides for easy viewing of POLSAR imagery stored as scattering matrix (S-matrix) or covariance matrix data that are color rendered as red, green, and blue (RGB) images. Arbitrary selection of the polarimetric image components is available for the RGB channels, clipping and scaling of the imagery to enhance display, as well as transformations between polarimetric bases (circular, linear, and Pauli). Processing can be limited to any selected area of interest within the image.
Classification Panel
The classification panel provides an end-to-end sequence of processing steps that are commonly performed in classifying targets within POLSAR imagery: construction of covariance matrices, their initial spatial averaging (rebin), spatial filtering of the imagery, a subsequent spatial averaging, and then (finally) application of a target decomposition algorithm to identify scene elements within the imagery. Covariance matrices can be constructed for 2, 3, and 4 polarimetric components in linear, Pauli, and circular polarimetric bases. Spatial filtering was implemented for a Lee filter. Several target decomposition algorithms were implemented as classification options for the end user: Cloude-Pottier, Freeman-Durden, Yamaguchi, and Wishart. The Freeman-Durden and Yamaguchi classifiers include corrections for polarization orientation angle before applying the target decomposition. This correction removes look-dependent rotations of ground targets that can produce cross-pol contamination in the covariance matrix, thereby leading to incorrect estimates of double bounce, volume, and surface scattering contributions. Target decompositions of the POLSAR imagery can be limited to any selected area of interest within the image. An example Yamaguchi target decomposition of the Radar Satellite 2 (RADARSAT-2) image of Norfolk, VA is shown in Figure 1. Vegetation appears as green, bodies of water as blue, and buildings and bridges in red and magenta.
 |
| Figure 1. Example POLSAR classification produced by NRL SAR Workbench showing quadpol RADARSAT-2 image of Norfolk, VA in a Yamaguchi target decomposition. |
 |
| Figure 2. Covariance matrix of quadpol RADARSAT-2 image of Norfolk, VA in linear polarimetric basis as it appears as the first frame in the SDV animation window. |
Individual Tools Panel
The individual tools panel provides the user with the option of performing selected processing operations on the POLSAR imagery. Available tools are the construction of a covariance matrix from S-matrix data, eigenvalue/eigenvector decomposition of a covariance matrix, spatial averaging, spatial filtering, classification algorithms, and basis transformations between any one of a variety of polarimetric bases: linear, Pauli, and circular. Options are available to export the results as color rendered RGB images in one of two file formats: the native file format developed for use within SeaLab, and in Hierarchical Data Format (HDF).
SAR Data Viewer
The SAR Data Viewer (SDV) is a companion stand-alone tool to NRL SAR Workbench that provides for interactive visual display of SAR imagery. SDV provides a host of capabilities, including brightness and contrast control, zoom, a line profile plot along a chosen path, histograms of areas of interest, and animations of sequences of SAR images stored as color rendered images in HDF. To fully exploit SDV's capabilities, CPI modified SAR Workbench to write the color rendered images of the input and output single look complex POLSAR data as a sequence of two images to the generated HDF files. This provided the capability to interactively switch between visual displays of the input and processed output POLSAR imagery within SDV to facilitate comparison of the results. An example is shown of a RADARSAT-2 image of Norfolk, VA that is 5384×3572 in azimuth and range for which the Cloude-Pottier target decomposition has been performed. The initial image displayed by SDV is the covariance matrix of the quadpol POLSAR data (Figure 2), while the second is the resulting target decomposition (Figure 3). Note that the Cloude-Pottier classification enhances the pixels of vegetation to a more prominent green color. It also more clearly delineates the bodies of water by blue, and provides for a much greater accentuation of the buildings shown in red.
Figure 3. Cloude-Pottier target decomposition of quadpol RADARSAT-2 image of Norfolk, VA in as it appears as the second frame in the SDV animation window.
Computational Efficiency, Supported Platforms, and File Formats
Particular attention has been paid in designing all POLSAR applications to handle processing of large images within the memory limitations of the computer platform. SAR Workbench is being supported by NRL for both Linux and Windows operating systems. SAR Workbench provides options for the import of a variety of image file formats, which include: Geographic Tagged Image File Format (GeoTIFF), RADARSAT Tagged Image File Format (TIFF), Sony Raw Image File (SRF), and TerraSAR-X (TSX).
Polarimetric SAR Calibration
The quality of POLSAR imagery and its polarimetric decompositions depends on the accuracy of polarimetric observations of the SAR system and its calibration. Polarization distortions on the polarimetric measurement can be incurred due to non-ideal system polarization quality and propagation factors, such as channel imbalance, cross-talk, and Faraday rotation at lower frequencies. All these distortions have varying impacts on different target types as well as different decomposition methods. CPI has assessed the polarization quality of the POLSAR system in the context of polarimetric imagery analysis and quantified the various effects of polarization distortions on polarization scattering decompositions.
A general evaluation method was developed to bridge the system polarization distortion and the polarimetric measurement quality with a generic metric based on the most general polarimetric system model. In this metric, the maximum normalized error (MNE) is evaluated as a single parameter,
where the superscript '*' stands for conjugate, λmax stands for the largest eigen-value of the enclosed matrix, i.e., A4*T(D-I)*T(D-I)A4, D is a 4×4 matrix representing the system distortion, and A4 is a 3×4 matrix
The new metric, MNE is adequately general to cover all polarization modes and all polarization distortions, reciprocal or non-reciprocal. It forms an accurate baseline for requirement analysis during radar system design and polarimetric data calibration. The metric preserves the full polarimetric information of the measurement and is basis-invariant. We used the term "maximum" in the sense that it caps the complete target space. This is very important but also practically meaningful because of the wide variation of the possible targets in POLSAR imagery. It enables convenient comparison of system merit across multiple platforms. Without this metric, it is not only difficult to quantify the overall distortion but also difficult to compare the polarization quality of different POLSAR systems, if not impossible. A real case study reveals it is a very tight upper bound for typical POLSAR imagery, which means the target subject to the maximum polarization error is not unrealistic in natural scenes and indicates the soundness of this new metric.
This new metric was employed in an investigation of the distortions due to channel imbalance, cross-talk, and Faraday rotation on several common polarimetric decomposition methods. The investigation demonstrated that the varying impact of these distortions depends on the specific decomposition method. Figure 4 shows the scatter plots of the resulting decomposition errors versus the MNE.
Figure 4. Scatter plots of the decomposition errors versus the MNE of the distorted observations for the canonical scattering targets: (a) The normalized decomposition error for the standard scattering targets using the Pauli decomposition; (b) The normalized decomposition error for the standard scattering targets using the Krogager decomposition; (c) The normalized decomposition error for the natural targets using the Freeman-Durden decomposition; (d) The bias on the entropy for the natural targets using the Cloude-Pottier decomposition; (e) The bias on the alpha angle for the natural targets using the Cloude-Pottier decomposition.
STAY CONNECTED 
