This proposal investigates spectral decomposition methods for identifying and classifying mine-related acoustic events. Minelaying operations produce a set of acoustic signature events ranging from the repeated splashes from mines dropped from an airplane, to the whirring of a motor unreeling a cable tethering a mine a set distance below the surface. Standard time-frequency (t-f) images, such as the spectrogram (Lofargram), are unable to resolve the transient structure of the images. Several advanced joint t-f algorithms have been studied to produce a better representation of the acoustic signal, but cannot yet reliably compute alias and artifact-free high-resolution representations. We propose to use statistical spectral decomposition methods to decompose an image into a set of basis functions for automated identification and classification. These methods have been widely used in image compression and medical imaging, but have only rarely been applied to transient acoustic analysis. We will investigate the use of the Karhunen-Loeve transform (KLT) and the discrete cosine transform (DCT) on sample waveforms to determine the efficiency of this method for classifying unknown signals. The image-dependent KLT has the highest theoretical efficiency but is computationally more intensive that the DCT, which is the most widely studied method.
