Automated discrimination of fault scarps along an Arctic mid-ocean ridge using neural networks

Multibeam bathymetric data, acquired along mid-ocean ridges (MORs), provide critical information for the modeling of seabed terrains and the understanding of related geology. An automated detection of geological features, such as fault structures, helps to elucidate the structural characteristics of MORs and quantify e.g., the magnitude and spatial variability of geological phenomena such as faulting. For this purpose, this research presents a developed cross-sectional methodology where continuous elevation data are (1) collected across a MOR from individual transect lines at various spatial resolutions (50-150 m), and then (2) analyzed with a supervised learning algorithm to discriminate fault structures. An artificial neural network (ANN) is applied for the detection and classification of fault scarps which have either an east or west tilt orientation; the classification uses attributes of elevation data calculated from surface derivation, simulated relief shading and statistical analyses of transects. Results indicate an average detection accuracy of 92%, which is dependent on the data sampling resolution, the terrain complexity and the predictor variables considered. Both the variance and regression slope variables played a key role in the training phase for identifying and classifying the tectonic features. The cross-sectional learning method presented in this research finally evidences the possibility to achieve an automated quantification system for different landform types and emphasizes the need for complementary classification methods to deepen the interpretation of landform complexities and related geological processes at MORs.