Machine learning for surficial geologic mapping

Surficial geologic maps contribute to decisions regarding natural hazard mitigation, land-use planning and infrastructure development. However, geologic maps may not adequately convey the uncertainty inherent in the information shown. In this study, we use machine learning and lidar elevation data to produce surficial geologic maps for parts of two quadrangles in Kentucky. We measured the performance of eight supervised machine learning methods by comparing the overall accuracy and F1 scores for each geologic unit. Surficial geologic units include residuum, colluvium, alluvial and lacustrine terraces, high-level alluvial deposits and modern alluvium. The importance of 41 moving-window geomorphic variables, including slope, roughness, residual topography, curvature, topographic wetness index, vertical distance to channel network and topographic flatness, was reduced to 12 variables by ranking the importance of each variable. The gradient-boosted trees model produced the classifier with the greatest overall accuracy, producing maps with overall accuracies of 87.4% to 90.7% in areas of simple geology and 80.7% to 81.6% in areas with more complex geology. The model produced high F1 scores of up to 96.2% for colluvium but was not as good at distinguishing between units found in the same geomorphic position, such as high-level alluvium and residuum, both of which are found on ridgelines. Probability values for each geologic unit at each cell are conveyed using gradations of colour and eliminate the need for drawn boundaries between units. Machine learning may be used to create accurate surficial geologic maps in areas of simple geology; in more complex areas, highlight that additional information obtained in the field is necessary to distinguish between units.