Synthetic aperture radar (SAR) is an essential sensor for military surface surveillance because of its unique ability to operate day or night through weather, smoke and dust. Of particular importance is the problem of automatic target recognition (ATR) which aims to identify targets of military significance within radar images. The Defence Evaluation and Research Agency (DERA) of the UK has a substantial programme of research into ATR algorithms for SAR battlefield surveillance. This covers both feature-based techniques which are the subject of this paper and model-based techniques. Feature-based ATR discriminates between target classes on the basis of the values taken by certain target features. The conventional approach is to select the best features for a particular task from a large set of features which have been pre-defined on the basis of physical intuition. A simple feature might be target area whilst a more sophisticated feature might be some measure of fractal dimension. ATR performance will be influenced by the type of features used and by the accuracy with which the statistical behaviour of these features has been characterised. This paper describes a technique which can be used to determine statistical feature behaviour despite limited examples of target realisations. It also addresses the problem of feature choice by introducing a method for adaptive feature design which automatically recognises the information not already contained in the existing feature set and develops a feature to represent this missing information. These ideas are illustrated by application to synthetic aperture radar (SAR) images of vehicles.