145 
SAR IMAGE REGISTRATION USING A NEW APPROACH BASED ON THE 
GENERALIZED HOUGH TRANSFORM 
C. Palmann*, S. Mavromatis, J. Sequeira. 
LSIS Laboratory - Equipe I&M (ESIL)Case 925, 163 avenue de Luminy 13288 Marseille cedex 9, France 
(palmann, mavromatis, sequeira)@uni vmed.fr 
Commission VII, WG VII/2 
KEY WORDS: Remote Sensing, Vision Sciences, Radar, Analysis, Registration, Algorithms, Matching, Transformation. 
ABSTRACT: 
Radar Imaging using SAR systems provides specific information that is very useful in the frame of “Digital Earth” applications (i.e. 
flood supervision, forestry or agriculture watch,). The main interest of such active systems is their capability to gather relevant data 
whatever the weather and the illumination conditions may be (cloudy, misty, during the night,). In addition, these systems give a 
useful “distance map” thanks to the wave coherence. Most applications require a follow-up of the situation during weeks or months. 
•Such aTollow-up can only be performed if we are able to register images captured at different times. This registration problem is a 
very classical one and has been widely studied in Remote Sensing, but the proposed solutions are often dedicated to specific contexts 
(sensors, type of scenes, known relevant elements).Many algorithms have been proposed to register SAR images, and we give, in this 
paper, a global overview of these methods depending on the chosen approach. They may use filtering or not prior to registration, and 
they may use landmarks or not; but, in all cases, there will be to take into account the speckle that reduces the efficiency of classical 
methods for extracting features (e.g. landmarks,) to be paired in both images. During the last years (since 2000), a new set of 
methods, related to the Hough Transform concept, have been proposed: the algorithm we introduce in this communication can be 
considered as being in this class of approaches. 
1. INTRODUCTION 
Using several images related to a given area improves the 
efficiency of Remote Sensing applications because it enables to 
integrate into a single model various information on this area. 
This integration process is directly dependent on the registration 
one that permits to geometrically superimpose two or more 
images. In this communication, we focus on a particular case of 
image registration process for Remote Sensing when images are 
SAR (Synthetic Aperture Radar) ones captured from a Spatial 
Platform. Only non-corrected images are studied because the 
registration process is not required when images have been 
corrected and thus are geocoded. 
When registering images at different times, we may be able to 
provide a follow up of area specific evolutions. For example, in 
the field of agriculture, it has been proved that there exits a 
linear correlation between the pixel values of SAR images and 
the height of cotton fields in the Ägrä region (Srivastava et al., 
2006), and thus we can use SAR images for controlling the 
agricultural process. We can also mention the use of SAR 
images for major risk management as, for example, in the case 
of flood (Stabel and Löffler, 2003) for disaster areas 
characterization. 
Various physical processes can be used to provide Radar images 
(Rees 2001) depending on sensor features (wavelength, 
polarization, viewing angle). This variety of images is 
interesting because of all the information they carry but it 
increases the difficulty of the registration process: geometry and 
radiometry of such images strongly depend on the acquisition 
process; in addition, all these images are modified by a noise 
called speckle. Finally, extracting information from such images 
in order to provide a registration with a subpixel precision, as it 
is often required (Eastman et al., 2007), is a very hard task to be 
performed automatically. 
This communication is structured as follows. Section 2 is a 
“State of the Art” on the registration process, especially related 
to Radar Imaging. In section 3, we introduce a new approach for 
registering SAR images that is based on the principle of the 
Hough Transform (Hough, 1962) when images have already 
been roughly registered. Results on the use of this approach are 
shown in section 4. 
2. STATE OF THE ART 
a) Registration 
Registration is a process that provides a geometrical 
correspondence between two images captured from different 
locations, or at different times, or using different sensors, or 
through different modalities. Usually, registration algorithms 
are sequenced as follows (Zitova and Flusser, 2003):