A new approach to the automated mapping of pockmarks in multi-beam bathymetry

Seabed pockmarks are of great interest to geologists and the marine geotechnical community. Identifying and mapping pockmarks rendered in multi-beam bathymetry data is an important but expensive manual process. In this paper, a new Machine Learning approach to automating the task is presented. Useful, low-dimensional feature vectors yielding very good classification accuracies are established. Overall process efficacy is subsequently evaluated by comparing counts of individual objects identified by the machine and a human analyst. Highest agreement (96.7%) occurs where there is a strong visual contrast between the pockmarks and the surrounding terrain. In low-contrast areas, our machine approach identifies several more objects than the human. Further, our process maps the boundaries of ≈ 2000 pockmarks in seconds - a task which would take days for a human to complete.