| This thesis is concerned with developing new raw synthetic aperture radar (SAR) data compression schemes, as well as providing some processing techniques to extract features from the SAR images. The study involves neural networks, wavelets, wavelet packet, optimal wavelet, and multifractal analysis, with emphasis on quantization and bit allocation to compress such an SAR signal, and iterated function systems (IFS) to segment the SAR image.; A SAR system is a sophisticated remote sensing tool capable of producing high resolution images from a moving platform either in space or in the atmosphere. It offers the advantage over passive optical sensing of operating in either day or night and in all-weather. When such a radar is placed on board a satellite, compression of the raw SAR signal is necessary to reduce the large amount of collected data the downlink needs to transmit to a ground station within a restricted bandwidth. To this end, wavelets representations have proven to be very effective for signal compression, due to their time-frequency localization and their joint statistical regularities.; After a general presentation of SAR system and principles, this research investigates and implements methods for coding of raw SAR data which are based upon suitable transforms of the complex raw data. The Karhunen Loeve Transform (KLT) is implemented with a three layers back-propagation neural network. Then, a compression of the raw SAR signal using five kinds of wavelets is presented: Haar wavelet, the Battle-Lemarie wavelets (linear and quadratic), and Daubechies wavelets (D4 and D20). A generalization to wavelet packet is also proposed. Experimental results point out advantages and drawbacks of this approach. Due to noise-like characteristics of the raw SAR signal, standard wavelets are not very efficient in compacting energy in the transform domain. We finish by developing an optimal wavelet for raw SAR data compression. The optimality criterion is redundancy minimization in the transform domain. Experimental results show that the optimal wavelet outperforms the standard wavelets and the present reference technique used to compress raw SAR data: the block adaptive quantization (BAQ).; Also, a multifractal-based approach to the extraction of textural features from SAR images is proposed. First, the Holder exponents are estimated from the continuous wavelet transform of the image, then the singularity spectrum is computed. Each fractal component consisting of pixels having the same HOlder exponent can be an attractor of an IFS. The spectrum at each point is used as input for the K-means classifier. The theory and the algorithms for this segmentation approach are presented. Experimental results show that the approach is beneficial for SAR image segmentation, demonstrating better segmentation than those obtained by other techniques. |
