Probabilistic models of seafloor composition using multispectral acoustic backscatter: The benthic detectorists

We describe and compare two probabilistic models
for task-specific seafloor characterization based on multispectral
backscatter. We examine whether generative or discriminative
approaches to supervised seafloor characterization do better
at harnessing the greatly increased information about seafloor
substrate composition that is encoded in the backscattering
response across multiple frequencies. A Gaussian mixture model
(GMM) is proposed as a generative model, and a fully-connected
conditional random field (CRF) is proposed as a discriminative
model. Either model uses input data derived from monospectral
or multispectral backscatter without modification. The CRF
approach considers both the relative backscatter magnitudes of
different substrates as well as their relative proximity, and can
be optimized using parameters. The GMM model, in contrast,
includes no spatial information in its estimates, being based solely
on relative backscatter magnitudes. Both GMM and CRF modeling
approaches perform better with multispectral backscatter
compared to monospectral, significantly outperforming all three
monospectral frequencies. With multispectral backscatter inputs,
based on average classification accuracies alone, there was little
to choose between the two modeling approaches (classification
accuracy of 81% and 83% for GMM and CRF models, respectively,
evaluated using 50% of available bed observations to
train and 50% to test the models). However, a CRF model that
has been optimized with respect to its tunable parameters tends
to produce higher posterior probabilities (i.e. greater certainty)
for its classifications. Using monospectral backscatter inputs, the
CRF model significantly outperformed the GMM model in terms
of average classification accuracy. On balance, therefore, based
on the evidence presented here, the CRF is suggested to be
the superior approach for task-specific seafloor classification.
Although further work using additional data is required to
further examine this conclusion, the work presented here will
guide and focus subsequent research efforts as more areas of
the seafloor are mapped with the new technology. In order to
facilitate these efforts, the algorithms presented here are encoded
in a freely available python toolbox for Probabilistic acoustic
Sediment Mapping, called PriSM , that can be used for both
monospectral and multispectral backscatter. Finally, we show that
application of the CRF model to the outputs of a geoacoustical
model of seafloor scattering results in realistic substrate classification
boundaries. This hybrid CRF and physics-based approach
can predict the physical properties of the seafloor at a finer spatial
resolution than is possible using the geoacoustical model alone.