Re-examining the global distribution of seamounts using neural network techniques

The distribution of seamounts and ocean islands, ranging from small hills just a few hundred meters in height to mountains such as Mauna Kea, offers key insights into the patterns of intraplate volcanism and their spatial and temporal variations. This may in turn provide information about the underlying lithosphere and asthenosphere. Mapping the global distribution of seamounts is complicated by the limitations of marine topography data and by the scale of the problem: 335 billion km² of ocean must be searched. Whilst seamounts are comparatively easy to identify by visual inspection, automating this identification is often difficult, since a mathematical description of the feature is required. Previous techniques include modelling seamounts using simple Gaussian or elliptical polynomial models (e.g. Kim & Wessel, 2011) or searching for sinks in inverted bathymetry data (e.g. Kitchingman & Lai, 2004). However, seamounts exhibit significant natural variation, and any particular model may not suit all examples. It is therefore difficult to develop an algorithm that detects the full range of seamount morphologies, yet excludes other features of different origin but similar general form, such as tectonic fabric or abyssal hills. One potential avenue lies in the use of neural networks, or other learning algorithms, inspired by the pattern-recognition ability of the brain. Rather than formulating an textit{a priori} description of the feature, a neural network assimilates characteristics from a set of hand-picked examples. Here, we use an autoencoder network to assess whether small patches of ocean floor display the representative characteristics learnt by our network. From this, we can build up a database of seamount locations, with early indications suggesting a reduction in false positives such as abyssal hills when compared with traditional methods, whilst results over large features are generally similar. Once the dataset is constructed, the distribution of the volcanism in space and time can then be analysed. We note that our underlying approach is quite general and could potentially be applied to a variety of geomorphological problems.