Unsupervised Boundary Analysis of potential field data: a machine learning method

We propose a boundary analysis method, called Unsupervised Boundary Analysis, based on machine learning algorithms applied to potential fields. Its main purpose is to create a data-driven process yielding a good estimate of the source position and extension, which does not depend on choices or assumptions typically made by expert interpreters, such as a low-pass filtering or weights in the Enhanced Horizontal Derivative case. We first tested the simple synthetic case of two vertical faults, to understand the robustness of the method. We recognized three classes on the basis of their centroids and found that the sources edges could be detected at the transition between two of them. Subsequently, we applied the Unsupervised Boundary Analysis to the real magnetometric data of the archaeological site of Torre Galli (Calabria, Italy). We compared the results with those from two different boundary analysis techniques, the Enhanced Horizontal Derivative and the Tilt Derivative. The main sources were well recognized by our approach, in good agreement with the Enhanced Horizontal Derivative results, but the Unsupervised Boundary Analysis led us to have a more complete description of the lineaments and to retrieve further features of archaeologic interest in the area. Instead, the Tilt Derivative features were affected by noise, which made interpretation more complicated.