ROCK TRAITS FROM MACHINE LEARNING: APPLICATIONS TO PRECARIOUSLY BALANCED ROCKS AND FAULT SCARPS IN SOUTHERN CALIFORNIA

Rock traits (size distribution, shapes, orientation, composition of pebbles, cobbles, clasts) distinguish many geologic features important in earthquake geology research. These include alluvial fans (fault activity), rock damage (fault zone rupture processes), precarious rocks (strong ground motion over kyr-time scales), and fault scarps (fault zone evolution, fault slip rates, earthquake timing). Machine learning (ML) has revolutionized data intensive computing problems for many scientists. We have begun to apply ML to rock traits for fault scarps and to develop a workflow for application to precariously balanced rocks (PBRs) and other fragile geologic features. Our preliminary work on fault scarps formed in the Bishop Tuff on the Volcanic Tablelands, California shows that deep neural networks trained on expert annotation of UAS-acquired rock imagery from a geological site leads to accurate and fast segmentation of rock surfaces. This capability of segmenting aerial imagery facilitates the estimation of distributions of rock traits such as angularity, size, and orientation. Rock size distributions reflect both the initial cooling joint fracture geometry, as well as faulting-induced fracturing. They vary with position as a function of strain magnitude and linkage characteristics. In addition, rock orientations indicate the degree of downslope transport along the fault scarps, enhancing our understanding of the fault scarp erosional processes. Large datasets of spatially explicit PBR fragility with detailed geomorphic and geologic context provide valuable assessment of sensitivity to ground motions. Instead of having a few fragilities, it should be possible to have a fragility spectrum for the fragile geologic features at a site. We are investigating safe and efficient imaging with UAS for mapping PBRs with particular emphasis on intelligent observation of basal contact relationships.