Application of Polarimetric Features and Support Vector Machines for Classification of Improvised Explosive Devices

In this letter, we present a binary classification problem of improvised explosive devices, where true and false target features are taken from polarimetric ground-penetrating radar (GPR) signals. The radar measurements are carried out in laboratory and field scenarios by using dual-polarized Vivaldi antennas and ultrawideband multiple-input-multiple-output GPR (MIMO-GPR) systems in the frequency range from 0.8 to 5 GHz. Recursive algorithms, recursive least square, and linear predictive coding are used for clutter removal. To extract targets features of polarimetric measurements, eight data-processing methods are assembled, combining clutter removal algorithms, time-frequency transformations, and singular value decomposition. Moreover, for every method, 13 target feature vectors are constructed. Each vector is used to train and test a support vector machine algorithm for the classification problem. Classification results are validated by using leave-two-out cross-validation. Accuracy of 87.02% and a false positive rate of 10.53% in the best classifier were obtained.