We investigated the application of two state-of-the-art feature selection algorithms for subsurface target discrimination. One is called joint classification and feature optimization (JCFO), which imposes a sparse prior on the features, and optimizes the classifier and its predictors simultaneously via an expectation maximization (EM) algorithm. The other selects features by directly maximizing the hypothesis margin between targets and clutter. The results of feature selection and target discrimination are demonstrated using wideband electromagnetic induction data measured at data collected at the Aberdeen Proving Ground Standardized Test Site for UXO discrimination. It is shown that the classification performance is significantly improved by only including a compact set of relevant features.