

Supervised classification using neural networks based on polarimetric timefrequency signature
Mickaël Duquenoy^{(1)}, Jeanphilippe Ovarlez^{(1)}, Eric Pottier^{(2)}, Laurent FerroFamil^{(2)}, Christele Morisseau^{(1)} and Gilles Vieillard^{(1)}^{(1)} ONERA, ONERA Chemin de la hunière, 91761 Palaiseau cedex, France
^{(2)} IETR, Campus beaulieu, Rennes, France
Abstract
Conventional radar imaging assumes that all the scatterers are considered as bright points (isotropic for all observation angle and white in the frequency band). Recent studies based on multidimensional TimeFrequency Analysis, describe the angular and frequency behavior of scatterers and show that they are anisotropic and dispersive [1], [2], [3], [4]. Another information source in radar imaging is the polarimetry. Studies based on multidimensional wavelet and coherent decompositions allow to represent the angular and frequency polarimetric behavior and show the nonstationarity of this behavior [5], [6].
These representations called polarimetric hyperImages are 4dimension images. So, the HyperImage can be written I(x,y,f,theta) where (x;y) is the location of the scatterer, f the frequency emitted and theta the aspect angle. For each frequency f0 and each angle of radar illumination theta0, I(x; y; f0; theta0) represents a spatial repartition of reflectors which respond in energetic case or polarimetric case, at this frequency and this angle. Inversely, for each reflector located at r0 = (x0; y0), we can extract its feature I(x0; y0; f; theta) in frequency f and in angular theta. This is this aspect that we decided to point out in order to see if this quantity can be interpretable in terms of target characteristics.
This application drives to polarimetric timefequency signatures[7]. The choice of the multidimensional timefrequency distribution is the continuous wavelet because the work is oriented to anechoic chamber data. Indeed, in ONERA, there is an anechoic chamber which allows to, measure signatures of canonical targets as trihedral, dihedral, .... Continuous wavelet and polarimetric coherent decompositions allow to extract signatures of these targets.
The possibility to obtain a basis of signatures in anechoic chamber allows to made a supervised classification by neural networks applied on anechoic chamber data. So, these signatures are the learning basis.
The classifier is a multilayer perceptron (two layers) [8]. The number of neurons of the hidden layer is chosen in function of the number of output and the number of input. The input is the hyperimage I(x0; y0; f; theta).
This suprvised classification has been tested on a weapon named "Cyrano". A comparison is made between the results of the classification and the polarimetric coherent decomposition (Pauli, Krogager, Cameron).
[1] "J. Bertrand and P. Bertrand", The Concept of Hyperimage in WideBand Radar Imaging. Trans. IEEE Geoscience and Remote Sensing, vol 34,
number 5, p 11441150, september 1996.
[2] "J. P. Ovarlez and L. Vignaud and J. C. Castelli and M. Tria and M.
Benidir", Analysis of SAR images by multidimensional wavelet transform.
Trans. IEE Radar, Sonar and Navigation, vol 150, number 4, p 234241,
august 2003.
[3] "L. FerroFamil and P. Leducq and A. Reigber and E. Pottier", Extraction, of Information from TimeFrequency POLinSAR Response of Anisotropic Scatterers. Proc. IEEE International Geoscience and Remote Sensing Symposium (IGARSS’05), Seoul, South Corea, 2005.
[4] "T. Jin and Z. Zhou and W. Chang", Ultrawideband synthetic aperture radar timefrequency representation image formation. Trans. IEE Radar, Sonar and Navigation, vol 153, number 5, p 389395, december 2006.
[5] M. Duquenoy, J.P. Ovarlez, L. FerroFamil, L. Vignaud and E. Pottier, Study of Dispersive and Anisotropic Scatterers Behavior in Radar Imaging Using TimeFrequency Analysis and Polarimetric Coherent Decomposition, Proc. IEEE Radar International conference, 2427 april 2006, Verona, USA.
[6] M. Duquenoy, J.P. Ovarlez, L. FerroFamil, L. Vignaud and E. Pottier, Study of Dispersive and Anisotropic Scatterers Behavior in Radar Imaging Using TimeFrequency Analysis and Polarimetric Coherent Decompositions., Proc. EUSAR conference, 1618 may 2006, Dresden,
Germany.
[7] M. Duquenoy, J.P. Ovarlez, L. Vignaud, L. FerroFamil, E. Pottier,CLASSIFICATION BASED ON THE POLARIMETRIC DISPERSIVE AND ANISOTROPIC BEHAVIOR OF SCATTERERS, Proc. POLINSAR 2007, Frascati, Italy.
[8] B.D. Ripley, Pattern recognition and neural networks.

