Experiments on automatic classification of shallow water acoustic signal sources using two pattern recognition methods

The problem of classifying underwater acoustic signals has been approached from a pattern recognition point of view. The signals of 25 acoustic sources were recorded from shallow-water environments, including several disturbances. The classification was performed using two statistical methods: the learning subspace method and a method based on T. Kohonen´s (1981) self-organizing feature maps. In both methods the pattern memory was trained by several measurements of signals of these sources. The intention was automatic recognition of new recordings of the same sources using a separate class for each source. An overall accuracy of 80 to 90% was reached using signal samples that were present in the training process. The accuracy was about 40 to 50% using samples from entirely new recordings of the same signal sources, but varied significantly between individual classes