A new metric for morphologic variability using landform shape classification via supervised machine learning

The frequency and degree of change in landform-profile shape can provide insights into the evolution of complex landforms. Here, we present a workflow that can be used to quantify this aspect of morphologic variability along any landform, leveraging both the expertise of a geomorphologist and the efficiency of a machine-learning algorithm. As a case study, we tackle the problem of degradation of fault scarps in jointed bedrock. We made field observations of seven fault scarps in jointed bedrock from Hawai'i, California and Iceland and collected aerial imagery for Structure-from-Motion (SfM) photogrammetry. From these observations, we first manually classify fault-scarp profiles extracted from SfM-derived point clouds into six morphologic categories defined by a geomorphologist with a view towards geologic process. Then, we use principal component analysis with singular value decomposition to quantitatively distinguish morphologic classes. We follow this by employing the support vector machine (SVM) method to build a supervised classifier, using the principal-component coordinates of the classified profiles in principal component space as a training set. Classification performance was assessed using 5-fold cross validation (81% accuracy) and with independent test data (80% accuracy). Finally, we define a morphologic variability metric and calculate it by determining the number of classes represented and the standard deviation of their proportions in a moving window along a fault scarp. By analyzing the covariance between the morphologic variability metric and other geomorphic parameters, we can quantitatively determine the drivers of scarp form. We find that morphologic variability decreases with scarp maturity. Our results suggest that the morphologic variability metric is a promising tool to understand the evolution of complex landforms.