Adaptation and Comparison of Machine Learning Methods for Geomorphological Mapping and Terrace Prediction: A Case Study of the Buffalo River, Arkansas

Abstract

Working with a high-resolution LiDAR dataset, this research adapts existing techniques from machine learning to the problem of predicting general and specific geomorphic features in the complex terrain of the Buffalo River in Arkansas. This process is complicated by the fact that such geomorphic features are frequently not well-bounded and lack a one-to-one relationship between physical form and geomorphic class. These issues were addressed by analyzing terrain features within a spatial context using both local and regional land surface parameters. After selecting a horizontal resolution and degree of smoothing for creating our DEM that would maintain sufficient detail to distinguish the terraces and their boundaries while avoiding excessive noise, the study area was divided into three reaches based on general lithology and morphology. Using SAGA GIS, a free and open-source software, land surface parameters were calculated for each reach and five leaners—representing four distinct inductive biases—were tested. The results for each classifier were then validated using a dataset which combined field-mapped terraces and manually-delineated landform elements. It was found that Bayesian, Random Forest, and Support Vector Machine classifiers were the most accurate, while distance-based methods struggled to achieve acceptable accuracy. Support Vector Machines produced the smoothest class boundaries and mapped landforms in a way that was subjectively closest to manual methods; however, Bayesian and Random Forest approaches were more consistently accurate.