Detection and identification of objects in advanced Radar images

Synthetic Aperture Radars (SARs) are radars placed on a moving platform, typically a satellite or an aircraft for ground observation, which have large area coverage, as well as all-weather and day and night capabilities. Recent technological advances have made it possible to obtain high resolutions on the ground, of the order of magnitude of meters for spaceborne SAR and tens of centimeters for airborne SAR. These capabilities are of particular importance to find artificial objects such cars or boats at any time of the day or night, in any meteorological conditions, for time-critical applications such as organizing rescue operations. Likewise, this is particularly interesting for military applications, such as monitoring human activities and traffic in large areas and analyzing in real-time military movements and deployments. The data availability provided by SAR is also an undeniable advantage for urban development studying and planning. This thesis presents and illustrates different approaches to artificial object detection in fully polarimetric SAR images such as vehicles, buildings or ships, by making use of the specific scattering properties of artificial surfaces as opposed to natural scatterers. All examples provided in this thesis are high resolution fully polarimetric data sets, obtained either with the airborne F-SAR from the German Aerospace Center (DLR), or the PALSAR-2 aboard the ALOS-2 satellite from the Japanese Aerospace Exploration Agency.

The use of fully polarimetric SAR brings an increased amount of information about the scatterers reflecting the signal in the resolution cells regarding their shape, geometry, orientation, and dielectric properties. Different representations of the data, obtained through polarimetric decompositions and polarimetric parameters, provide a link between the measurements obtained and specific physical properties of the scatterers. An overview of the polarimetric decompositions and features used for target detection is presented, and their individual use for vehicle detection in natural background is thoroughly compared. The combined use of these parameters is thereafter studied with random forest classifiers, and optimal subsets of polarimetric features for vehicle detection are derived.

The received echo from a resolution cell corresponds to the coherent addition of the multiple echoes from many individual scatterers. The fundamental assumption made is that, for a homogeneous area and using the central limit theorem, the received signal should have a circular multivariate Gaussian distribution. However, this assumption does not always hold for high resolution radars, on artificial surfaces for which scatterers typically do not have random orientations. The application of several test statistics to test for the normality, zero mean and circularity of the measured single look complex data is presented, and its potential for artificial objects detection is illustrated.

The backscatter from natural areas is often assumed to be reflection symmetric, in which case the correlation between the co- and cross-polar channels is null. A new test statistic for reflection symmetry is proposed, and its application to high resolution SAR data shows its high potential for building detection. It is able to highlight and delineate isolated buildings in low density areas as well as densely built urban areas as opposed to parks and natural background. Its application to L-band data also showed potential for ship detection for coastal area monitoring. 

An alternative approach, already proposed in several publications, consists in studying the different point responses in the spectral domain using the so-called Time-Frequency decomposition. In this application, every point’s response is compared under several azimuth illumination angles and illumination frequencies. This type of study for target detection had already been proposed, and is applied and compared here to the other approaches proposed. The results found in this thesis constitute an investigation of the scattering properties which differentiate artificial scatterers from natural backgrounds. The potential of the different methods presented for artificial object detection are assessed and compared in different scenarios, and give a new insight on target detection in fully polarimetric SAR images.