Geoacoustic inversion with generalized additive models

Geoacoustic parameter estimation is presented as a non-linear regression problem where prediction is performed using generalized additive models applied to features extracted from broadband acoustic time-series in a machine learning framework. Qualitatively, it can be seen that signals that have propagated in different environments have distinct structures: in some cases, a single mode is identified, in others, multiple modes can be seen; signals can also be distinguished by different energy levels. Features that are employed here comprise relative amplitudes of distinct peaks in the signals, signal kurtosis, signal strength, decay of the time-series with time, and time difference between distinct peaks of the received signals. Functions are sought that relate sediment sound speed and attenuation to these features. A multivariate generalized additive model is proposed using smoothing splines for the nonlinear regression problem of predicting geoacoustic properties using the features. The spline functions are estimated using noise-free training patterns from known environments. After this training step, the geoacoustic properties are predicted in an efficient manner using noisy testing patterns from a variety of different areas.