• 09:00 A new light on misclassificaiton results on the SOMA district in San Francisco, due to the difficulty to predict entropy

  Colin Koeniguer, Elise (1); Weissgeber, Flora (1,2); Trouvé, Nicolas (1) 1: Onera, France; 2: Telecom Paris Tech, France

  Show abstract

  Entropy is often used in classification algorithms based on the polarimetric information, by unsupervised Wishart classification in alpha entropy feature space for example. In this framework, entropy is supposed to be low for man-made targets. However, on most examples of classification results on San Francisco images, this parameter fails to well classify the SOMA district, one district containing a lot skyscrapers, and with a particular orientation.Even very recent studies fail to compensate the orientation effect on this area. Moreover, TerraSAR-X data show high entropy with poor contrast between natural and deterministic targets. Then, this paper investigates the reason of this issue. Several aspects are invsetigated: does entropy depends on noise ratio, wavelength, resolution size, orientation effect or complexity of the medium? Several assumptions have been proposed to answer these questions. They are discussed in this paper - by using EM simulation tools - by using statistitical simulations - by comparing six different polarimetric images of six different sensors on this site: AIRSAR, SIR-C, RADARSAT-2, ALOS, TerraSAR-X, and recent UAVSAR. This comprehensive study allows us to make the following conclusions: - Entropy is obviously affected by the thermal noise of the sensor. However, this influence is less compared to the following two influences: building orientation, and cell size resolution. - Disorientation in urban areas is not only accompanied by an increase in the cross pol signal, but also an increase in entropy. Disorientation lead to mix randomly the mechanisms and makes it impossible or at least very difficult to correct the effect of deorientation on the double bounce: even if we are able to highlight the presence of this effect, other mechanisms involved remain mixed in the resolution cell. This lead to the common misclassification results, even with deorientation effects. - High entropy in urban areas is strongly linked to improved cell resolution size. For a given wavelength, at L-band, entropy remains low when resolution is poor. Entropy increases with improved resolutions, probably because the spatial variability of polarimetric behavior is more important, and that the spatial averaging is not adapted in this case. For large resolution cells on the contrary, the behavior is more stable in a cell to another. Future recommandations are given on the use of an entropy computed using multitemporal images, especially when resolution improves.

  
   

• 10:00 EVALUATION OF MULTILOOK EFFECT IN ICA BASED ICTD FOR POLSAR DATA ANALYSIS

  Pralon, Leandro (1,2); Besic, Nikola (1); Vasile, Gabriel (1); Dalla Mura, Mauro (1); Chanussot, Jocelyn (1) 1: Grenoble Institute of Technology (Grenoble INP), France; 2: Brazilian Army Technological Center

  Show abstract

  Polarimetric target decomposition is one of the most powerful and widespread tools for PolSAR image interpretation. The analysis of the interaction between the illuminated area and the transmitted waveform, to each polarimetric state of the latter, allows for a better prediction of the basic scattering mechanisms present on the scene, and to more efficiently propose classification, detection and geophysical parameter inversion algorithms. Many methods have been proposed in the literature to both decompose an image pixel into basic target vectors and to correctly retrieve quantitative information from them (parametrization). Concerning the latter, Cloude and Pottier's parameters (entropy, alpha and anisotropy) and Touzi's target scattering vector model are the most employed ones, whose usefulness have already been demonstrated by several authors. Regarding the decomposition, the algorithms are mainly classified in either coherent, if they are based on the scattering matrix analysis, or incoherent if their interest lies in the Hermitian, semidefinite positive coherency or covariance matrix. Among the incoherent target decompositions (ICTD), the eigenvector based one manages to decompose an image pixel into the three most dominant scatters from the averaged coherence matrix. Furthemore, it has an intrisic property that the derived scatters are orthogonal and uncorrelated, which for Gaussian clutters also means independence. The drawback of this kind of method emerge when the clutter is not Gaussian or not composed by orthogonal mechanisms, situations where the performance of the algorithm could be compromised. Besic et al. 2015 presented a new strategy to polarimetric target decomposition by incorporating the independent component analysis (ICA). The results proved it be a very promising area in polarimetry, mainly when non-Gaussian heterogeneous clutters (inherent to high resolution SAR systems) are under study. The theoretical potential in estimating similar entropy and first component, when compared to traditional eigenvector decomposition, but rather a second most dominant component independent with respect to the first one and unconstrained by the orthogonality introduces an alternative way of physically interpreting a polarimetric SAR image. The referred method is briefly summarised in three main steps: data selection, based on the statistical classification of the POLSAR image; estimation of independent components and parametrization of the derived target vectors. As stated by the authors, the principal drawback of the proposed method, is the size of the observation dataset, which has to be somewhat larger than the size of the sliding window used in the well established methods. This constraint lead them to use an unsupervised classification algorithm rather than relying on a very large sliding window, jeopardising the effectiveness of the method. The use of a classification algorithm limits the performance of the method in the sense that the image is segmented in a priori defined number of classes with variable sizes, what can lead to either over or under estimations of the target vectors parameters. Both consequences can compromise the correct interpretation of the scatters present in the area under study. The former implicates that a class contains more samples than it needs for a correct estimation of target`s parameters, meaning that spatial resolution, highly degraded with the use of this approach, is worse than it could be. The latter causes a bias in the parameters estimated, meaning that the values derived do not comply with ground truth. Within this context, this paper considers a Monte Carlo simulation approach to evaluate the optimal size of a sliding window for various medias, composed of Surface, Double Bounce and Volume returns. The simulation procedure is similar to the one presented by Lee et al. 2008 to evaluate the bias of multilook effect on Cloude and Pottier parameters in eigenvector based polarimetric SAR decomposition. The seed coherency matrix for each of the aforementioned type of scatters is extracted from real data, more precisely, in this paper a P-band airborne dataset acquired by the Office National d'Études et de Recherches Aérospatiales (ONERA) over the French Guiana in 2009 is taken into consideration. The main difference regarding the generation of the simulated data is that in Lee et al. approach only Gaussian variables were addressed and no texture was considered, meaning that the sampled coherency matrix has the complex Wishart distribution. In the present work the heterogeneous clutter is described by the Spherically Invariant Random Vectors (SIRV) model proposed by Vasile et al. 2010 and texture is characterised by a random variable. The final paper will be organised in five sections. Section II presents the ICA approach proposed by Besic et al 2015. as an ICTD method. Section III briefly describes the data simulation procedure, taken into account the SIRV model proposed by Vasile et al. 2010. In section IV Touzi's estimated parameters are presented for different window sizes and compared to the values obtained using the traditional eigenvector decomposition. Finally in Section V conclusions are drawn and future work possibilities are highlighted.

  
   

• 09:40 Classification Comparisons Between Compact Polarimetric and Quad-pol SAR Imagery

  Souissi, Boularbah (1); P. Doulgeris, Anthony (2); Eltoft, Torbjørn (2) 1: University of Science and Technology Houari Boumedienne (USTHB), Algeria; 2: Department of Physics and Technology, University of Tromsø, 9037 Tromsø, Norway

  Show abstract

  Recent interest in dual-pol SAR systems has lead to a novel approach, the so-called compact polarimetric imaging mode (CP) which is able to reproduce fully polarimetric information based on a few simple assumptions. The system transmits only one polarization, either π/4 or circular (right or left circular, RC or LC). The various dual-pol modes provide different information to the compact polarimetry model and thus generate different pseudo quad-pol imagery. In this work, The CP image is simulated from the full quad-pol (QP) image. We present here the initial comparison of polarimetric information content between QP and CP imaging modes. The analysis of multi-look polarimetric covariance matrix data uses an automated statistical clustering method based upon the expectation maximization (EM) algorithm for finite mixture modeling, using the complex Wishart probability density function. Our results showed that there are some different characteristics between the QP and CP modes which will be the subject of this paper. The classification is demonstrated using a E-SAR and Radarsat2 polarimetric SAR images acquired over DLR Oberpfaffenhofen in Germany and Algiers in Algeria respectively.

  
   

• 09:20 A Physical Explanation of Texture for Polarimetric SAR data

  DENG, XINPING; LOPEZ-MARTINEZ, CARLOS Universitat Politecnica de Catalunya (UPC), Barcelona, SPAIN

  Show abstract

  In a polarimetric sythetic aperture radar (PolSAR) image, the response of a resolution cell of each polarimetric channel can be regarded as a sum of a number of complex phasors, each resulting from an individual scatterer or a separate subregion of a continuous scattering surface. The addition of the independent and randomly phased complex contributions gives rise to speckle. This procedure can be explained by a random walk in the complex plane, where the number of scatterers is treated as the step number and the amplitude as the step length. The classical model for homogeneous data assumes that the step number in each resolution cell is infinite, which results in Gaussian distributed speckle. However, many experiments have proved that the K-distribution, which can be obtained by letting the step number be negative binomial distributed in the random walk model, gives a more accurate representation for the data in areas like sea surface and forest. As shown by Ward, the K-distribution can be formulated into a product of two independent random variables, one is gamma distributed with a long correlation time, and the other follows a Gaussian distribution with a much shorter correlation time. The first component is referred to as texture and the second one as speckle by many researchers later. Since then, different distributions have been proposed to model the texture component, but most of them lack of an underlying physical explanation, such as the random walk model. In this paper, the random walk model is studied with the objective to obtain a physical explanation of the texture of PolSAR data. Different distributions for the step number and step length are tested with a simulator. Results show that the texture is mainly due to the distribution of the step number. The distribution of the step length has effect only when the number of scatterers in a resolution cell is very small, which appears in very high resolution data. Besides, it is found that the mixture of different targets can give non-Gaussian distributed statistics. For example, the mixture of point targets and distributed targets will lead to extremely heterogeneous appearance, which may be a clue to analyze the urban areas in PolSAR data. In the final paper, results based on simulated, as well as experimental data will be presented, analyzed and discussed.

  
   

• 10:20 Change detection in a time series of polarimetric SAR images

  Skriver, Henning (1); Nielsen, Allan Aasbjerg (2); Conradsen, Knut (2) 1: Technical University of Denmark, DTU Space, Denmark; 2: Technical University of Denmark, DTU Compute, Denmark

  Show abstract

  A test statistic for the equality of two or several variance-covariance matrices following the real (as opposed to the complex) Wishart distribution with an associated probability of finding a smaller value of the test statistic is described in the literature [1]. In 2003 we introduced a test statistic for the equality of two variance-covariance matrices following the complex Wishart distribution with an associated probability measure [2]. In that paper we also demonstrated the use of the test statistic to change detection over time in both fully polarimetric and azimuthal symmetric SAR data. To detect change in a series of k > 2 complex variance-covariance matrices the pairwise test described in [2] may be applied to either consecutive pairs or to all possible pairs. The former would lead to a lack of ability to detect weak trends over time, the latter to an increase in the probability of false positives (postulating a change when there actually is none) and/or false negatives (missing an actual change). Therefore we need to test for equality for all time points simultaneously. In this paper we demonstrate a new test statistic for the equality of several variance-covariance matrices from the real to the complex Wishart distribution and demonstrate its application to change detection in truly multi-temporal, polarimetric SAR data. Results will be shown that demonstrate the difference between applying to time series of polarimetric SAR images, pairwise comparisons or the new omnibus test statistic, where changes are clearly detected with the omnibus test, on the contrary to the pairwise comparisons, where no changes are detected. We also demonstrate how a factorization of the likelihood ratio statistic into a product of test statistics that each test simpler hypotheses of homogeneity up to a certain point can be used to detect at which points changes occur in the time series. [1] T. W. Anderson, An Introduction to Multivariate Statistical Analysis, John Wiley, New York, third edition, 2003. [2] K. Conradsen, A. A. Nielsen, J. Schou, and H. Skriver, “A test statistic in the complex Wishart distribution and its application to change detection in polarimetric SAR data,” IEEE Transactions on Geoscience and Remote Sensing, vol. 41, no. 1, pp. 4–19,

  
   

• 10:40 Round Table Discussion

  Show abstract

  Round Table Discussion

  
   

Highlights

In May 2013, the Biomass mission concept was selected to become the next in the series of satellites developed to further our understanding of Earth.

Released: 21/06/2013

REPLAY EARTH EXPLORER 7 USER CONSULTATION MEETING (PART 1)

Released: 11/03/2013

POLinSAR 2013 Workshop

Released: 22/02/2013