Overview

Detecting primitives in a scene (see Chapter 8) is a major stage in automatic image processing. Classification, the purpose of which is to identify land cover, is another important operation. Unlike detection, which is often local in scope, classification is a regional, even global, approach to the scene involved. Its objective is to assign to every pixel in the scene a label that will bring together all pixels similar in nature and identify the class to which that pixel belongs.

Classification methods are the foundation on which shape recognition techniques are based [DUB 90]. They rely on a variety of tools provided by data analysis, discriminant analysis, probabilistic approach, decision theory and many other fields of applied mathematics. Not all of these techniques can be used in radar imaging. In addition, the remote sensing community has long been using optical field imaging tools. Therefore, it is naturally this type of methods that have predominantly been applied and extended to the radar domain. We will skip a general classification techniques presentation and instead describe only the methods that have been explored over the past few years and whose performances were good enough to have them rated as applicable to radar imaging. They mainly include Bayesian methods good at exploiting our a priori knowledge of the radar signal distribution laws. By contrast, other methods such as neural networks [HAR 94, CHE 96] or projection methods (main component analysis, factor analysis, etc.), albeit well developed in other shape recognition areas, have...