Ground-penetrating radar (GPR) is a versatile technology for subsurface sensing, and has shown promise in countermine applications when a target detection algorithm is employed. Because the soil environment is naturally heterogeneous and nonstationary,many detection algorithms have taken the form of adaptive filters operating on the real-aperture radar data. In particular, linear prediction techniques have received much attention for their ability to screen for anomalous signals that differ from the background. In this work,we demonstrate that linear predictionmay provide a low-dimensional feature set that is indicative of various soil properties. Experiments were performed with simulated and field-collected GPR data, and results provide greater understanding of how linear predictors might be useful in landmine detection over varying terrain. © Springer Science+Business Media New York 2014.