We extend a time-frequency discrimination algorithm, developed in an earlier article (Arrowsmith et al., 2006), for application to seismic-array data. Spectrograms evaluated at each component of an array are stacked and then converted into binary form for computation of discriminants. Because noise can bias the discriminants, we develop a procedure for removing the effect of noise on the discriminants. The binary spectrograms are randomized where the spectral amplitude of the signal is similar to the mean spectral amplitude of the pre-event noise at that frequency. The formulism of Arrowsmith et al. (2006) is further extended by modifying the objective function used to optimize the values of input parameters and by removing high-frequency and low-frequency spectral content. We apply the method to a dataset of regional recordings of earthquakes and delay-fired mine blasts recorded at the Pinedale seismic array in Wyoming. Our results show that the utilization of array data improves the success rate for source identification. Furthermore, we find that incorporating the noise-correction procedure increases the separation between earthquakes and cast overburden blasts (the largest type of delay-fired mine blasts). In total, the algorithm successfully identifies 97.4% of the events (74 of a total of 76 events, which comprise earthquakes and cast overburden blasts).