BUILDING A COMMUNITY TOOL FOR AUTOMATED FACIES INTERPRETATION FROM SEDIMENTARY OUTCROP PHOTOS

Geological field data are essential for reconstructing historical conditions on Earth and Mars, finding and developing natural resources, and managing natural hazards. Interpreting field observations takes years to master and is among the most time-consuming aspects of geological research. Sedimentary geology relies on a particularly complicated set of visual cues and patterns in rock outcrops. We leveraged computer-science approaches to build a tool for labeling images and trained classification models using machine learning algorithms to identify key sedimentary structures useful for paleoenvironmental and sedimentary facies interpretation. OutcropSketch is a web-based interface that allows users to label outcrop images with geologic features. It provides users multiple tools for drawing and editing polygons on outcrop photos to label important sedimentary structures (cross strata, planar lamination, contorted bedding, graded bedding, and structureless beds), non-geologic features (e.g., vegetation, sky, people), and surfaces like fractures. User-generated image labels made in OutcropSketch are fed into a machine-learning algorithm to predict underlying structure types. Preliminary results indicate that identifying sedimentary structures in outcrop photos is feasible but will require many labeled images from a variety of settings spanning different weathering characteristics, lighting conditions, and rock colors. The process of labeling training images has helped identify more explicitly the scales and features expert sedimentary geologists rely on for outcrop interpretation. This insight will help adapt teaching approaches to quickly build observational skills in trainees. OutcropSketch also provides a mechanism of recording outcrop interpretations in a format that is flexible for a range of research analyses (e.g., bedform set mapping) which can be archived and shared in databases like StraboSpot. OutcropSketch offers the possibility of automating outcrop interpretation by pre-screening field photos for key sedimentary structures. This will save time, allowing experts to invest more resources into advanced analyses, select targeted sites for detailed observations, and will facilitate new, more accessible pathways for sharing sedimentary field observations.