In this paper, we propose an automatic supervised classification of objects lying on the sea floor or buried in sediment layers. This pattern recognition provides a way to distinguish natural and manufactured objects and then should be helpful to detect mine, pipe-line, or wreckage. Proposed methods combine different techniques: pattern information extraction, relevant parameter search, and supervised classifier. Parameters are automatically selected using a principal component analysis to reduce misclassification rate and to simplify classifier structure. Performances of different parameters (two-dimensional and three-dimensional) are compared and discussed from testing on synthetic and real data bases.