Adaptive Speckle Filtering in Radar Imagery [chapter]

Edmond Nezry
2014 Land Applications of Radar Remote Sensing  
whose coefficients S pq =|S pq |.exp(j.ϕ pq ) are the complex backscattered field amplitudes for ptransmit and q-received wave polarisations. They relate the backscattered wave vector Ed → to the incident wave vector E i → : Ed → = exp(j.k.r).S . Ei → . This complete radar measurement is said "fully polarimetric", often abbreviated in "polarimetric". Let consider the interaction of the radar wave with an extended random surface, i.e. a surface containing a sufficiently great number of
more » ... per resolution cell, with no preponderant scatterer with respect to the others. For whatever configuration pq of polarisation, the signal S pq received from such a surface by the antenna becomes, after quadratic detection in intensity I pq : This detected signal intensity I is proportional in average to the radar "backscattering coefficient" σ°. The backscattering coefficient σ°=4.π.|S pq | 2 is the average radar cross-section per surface unit [1] . σ°, expressed in m 2 /m 2 , is a dimensionless physical quantity. It is a physical property of the sensed surface, which depends principally on its roughness, its dielectric properties, its geometry, and the arrangement of its individual scatterers. Carrying the radiometric information with regard to the sensed target, σ° is a function of the frequency of the radar wave, of its angle of incidence upon the target, and of the configuration of polarisation. In terms of physical meaning, the radar backscattering coefficient is analogous to the bidirectional reflectance in the domain of optical wavelengths: σ° # 4 cos 2 θ, where is the incidence angle of illumination on the sensed target. Nevertheless, detected radar images look visually very noisy, exhibiting a very characteristic salt-and-pepper appearance with strong tonal variations from a pixel to the next. Indeed, since radar imaging systems are time-coherent, radar measurements over random rough surfaces are corrupted by "speckle" noise due to the random modulation of waves reflected by the elementary scatterers randomly located in the resolution cell. Then, coherent summation of the phases of elementary scatterers within the resolution cell results in a random phase of the complex pixel value. This speckle "noise" makes both photo-interpretation and the estimation of σ° extremely difficult. Actually, speckle is a physical phenomenon, which is inherent to all coherent imaging systems (radar, lidar, sonar, echography). In most remote sensing applications using radar/SAR imagery, speckle is generally considered a very strong noise that must be energically filtered to obtain an image on which classic and proven information extraction techniques could be further applied, in particular the techniques used for optical imagery acquired in the visible and near-infrared part of the electromagnetic spectrum. Therefore, speckle filtering and radar reflectivity restoration are among the main fields of interest in radar images processing for remote sensing. Speckle filtering is a pre-processing aiming at the restoration of σ° value in the first place. This pre-processing must account for both the particular properties of the speckle, and those of extended imaged targets (often called "clutter"). It must also account for the radar imaging system that has sensed the target and for Land Applications of Radar Remote Sensing 4 the processor that has generated the images. For stationary targets of infinite size, speckle filtering is equivalent to a simple smoothing using a moving processing (averaging) window. An ideal filter must nevertheless avoid image degradation through excessive smoothing of the signal. To this end, it must respect structural image information (road and water networks, etc.), and the contours of radiometric entities. In addition, it must also respect the particular texture of the clutter, in forested or urban environments for example. Last, it must also identify the strong local radiometric variations due to the presence of strong scatterers (often artificial in nature) from those due to spatially rapid speckle fluctuations. Therefore, an ideal speckle filter must satisfy to the following specifications: 1. Preserve accurately the local mean value of the radar reflectivity (i.e. the quantity actually measured by the radar, which is proportional to σ°) to enable, for example, the comparison of radar reflectivities in the framework of a multitemporal analysis of radar acquisition series. 2. Smooth as much as possible homogeneous image areas and therefore reduce the speckle to increase the Equivalent Number of Looks (ENL) of the radar image (cf. § 2.1.4). The minimum ENL depends on the desired radiometric accuracy. For example, a 1 dB accuracy with 90% confidence level (i.e. less than 25% variation of the radar intensity) requires an ENL value around 230, and a 2 dB accuracy with 90% confidence level (i.e. less than 60% variation of the radar intensity) needs an ENL value around 40. 3. Preserve texture as much as possible where it exists in the image (forests, non-homogeneous fields, etc.) to avoid confusions among radiometrically similar areas exhibiting different texture. Therefore, a speckle filter must be able to discriminate heterogeneity effects due to texture from those due to speckle. 4. Both preserve and denoise image structures (contours, lines) as well as the quasideterministic responses due to corner reflector effects within strongly textured areas such as urban environments. Indeed, the energy of artificial radar reflectors responses must be preserved to enable radiometric calibration, in particular when calibration targets are dispersed in the radar image. 5. Minimise, and whenever possible prevent loss in useful spatial resolution during the speckle filtering process. Statistical properties of speckle and texture in radar images In this section, the statistical properties of speckle in images produced by coherent imaging systems such as imaging radars, lidars or sonars, are exposed. Since a good speckle filter must restore the texture of the scene imaged by the radar, the statistical properties of texture in radar images are examined as well. This analysis intentionally restrains to the first order statistical properties, since only these are generally used by the estimation techniques involved in speckle reduction methods. Explicit use of second order statistical properties of both the speckle and the imaged scene in the filtering process is adressed in Section 4.
doi:10.5772/58593 fatcat:cs7ycqehsffsrkbtkylracfs3y