Data preprocessing for seabed characterization with machine learning

Data received when sound propagates in oceanic waveguides exhibit a behavior characterized by the physics of the propagation medium. These data can be presented as input to machine learning algorithms for environmental characterization, in our problem seabed identification. A key factor for the success of the task at hand is how the data are employed within the classification algorithms and whether raw signals or qualitative and quantitative features extracted from those are used at the input level. In this work, data are studied both in the frequency and time domains and their behavior is analyzed. It is observed that employing simple characteristic attributes extracted from recorded signals is equivalent to identifying seabed signatures, leading to effective and efficient sediment classification. [Work supported by ONR.]