Kernel machines for object classification in high-resolution SAR data

The focus of this paper is the classification of military vehicles in high-resolution SAR images in an ATR framework. The usage of kernel machine classifiers is discussed. A new kernel machine, the relevance vector machine with integrated generator (RVMG) is introduced. Here, a single parameter controls the trade-off between speed and classification quality. Moreover classification heuristics and an adaptive feature extraction are used. These methods enable an improvement of the classification quality as well as a reduction of the computational effort. A parametrized reject criterion is presented to handle the classification of confusion objects. Therefore receiver operator characteristic (ROC) curves have been calculated. Tests have been performed using the MSTAR public target dataset and a fully polarimetric dataset from QinetiQ. An assessment of several polarimetric features has been performed.