Three-Dimensional Feature Models for Synthetic Aperture Radar and Experiments in Feature Extraction

This dissertation presents a new set of three-dimensional scattering feature models for synthetic aperture radar (SAR). We develop a set of parametric models of canonical shapes that capture aspect-dependent, high-frequency scattering for bistatic (and monostatic) 3D SAR phase history responses. The models are parameterized by the shape location, orientation, and size as well as the radar transmitter and receiver antenna aspects and frequency. We develop the models by combining physical optics (PO) and uniform theory of diffraction (UTD) planar scattering solutions to approximate 3D scattering responses of canonical shapes. We validate the models using scattering prediction software and show that the proposed models capture well the mainlobe responses of each shape. Thus, the proposed models may be used to accurately predict first-order scattering of scenes comprised of such shapes.

The second part of this dissertation focuses on the inverse problem of discerning the types of canonical shapes in a scene and estimating their corresponding model parameters from observed SAR phase history data. We present discrimination methods for classifying observed scattering into the geometric shape types. We compute the Cramer-Rao bounds for the models and characterize model parameter estimation accuracy for two estimation schemes. Finally, we present a feature extraction algorithm that classifies and estimates the canonical features from polarimetric phase history data. We use non-quadratic regularization to form sparsity-constrained 3D SAR images that are used to initialize the scatterer location, orientation, and size estimates. We test the feature extraction algorithm on simulated phase histories for densely-sampled and sparsely-sampled, monostatic and bistatic 3D SAR apertures. We show that even for sparsely-sampled apertures, the feature extraction algorithm is able to estimate geometric scattering features in the scene. Feature extraction for the proposed canonical shape models may be extended in future work for use in automatic target recognition.