Discovery and analysis of topographic features using learning algorithms: A seamount case-study

In order to analyse the incidence and characteristics of geomorphological features, it is first necessary to identify and catalogue them. This process is typically reasonably straightforward "by eye", but difficult to automate---simple models fail to match the complexity of real structures. We demonstrate that a particular class of learning algorithm, the "autoencoder", is able to assimilate the characteristics of a small hand-picked catalogue. This can then be used for large-scale classification. In particular, we focus on the problem of identifying seamounts in bathymetric data.