Comparison of seismic features extracted by digital signal processing techniques

During the past four years extensive effort has been made by this research group to digitally enhance the seismic data and to seek for the best mathematical features to discriminate between the natural earthquake and the nuclear explosion events. In this paper the recognition results based on different sets of seimic data base are reported. In particular, a critical comparison is made with the most recent seismic data base on the feature sets: autocovariance, power cepstrum, Alpha minus C energy estimate and entropy. The four feature sets are all effective but the autocovariance features provide the best performance with 89.32% correct recognition based on 16 features, 7 best learning samples from each class and the nearest neighbor classification rule. Although the theoretical comparison is not possible, computer results presented are reliable because of the reasonably large sample size used.