Learning Seabed Properties and Sensitivities from Ship Noise Spectrograms

In this work, deep neural networks are trained on synthetic spectrograms of transiting ships to find properties of the seafloor and uncertainty labels associated with those predictions. The spectrograms are labeled with the values of sediment layer thickness, sound speed, density, and attenuation; and a measure of the parameter information content or sensitivity for each parameter obtained from the Cramér-Rao bound (CRB). The CRB values for a given ocean environment and source-receiver geometry can be calculated and expressed as a relative uncertainty, for each sediment property. To obtain uncertainty labels for each spectrogram, the relative CRB is calculated for ten equally spaced frequencies for each sediment parameter of interest. These relative CRB values are divided into uncertainty classes, and the mode class across frequency is assigned as the uncertainty label for the spectrogram. The labeled synthetic spectrograms are used to train ResNet-18 networks, which can then tested on measured spectrograms. Comparisons are made between validation performance of networks that are trained to learn only the parameter value labels and those trained to learn both seabed parameter values and sensitivity class. The validation results indicate the potential to predict not only sediment properties but uncertainty in those predictions.