Chelsea Scott | arXiv Science

arXiv2022

Quantifying and analyzing rock trait distributions of rocky fault scarps using deep learning approach

We apply a deep learning model to segment and identify rock characteristics based on a Structure-from-Motion orthomap and digital elevation model of a rocky fault scarp in the Volcanic Tablelands, eastern California. By post-processing the deep learning results, we build a semantic rock map and analyze rock trait distributions. The resulting semantic map contains nearly 230,000 rocks with effective diameters ranging from 2 cm to 250 cm. Rock trait distributions provide a new perspective on rocky fault scarp development and extend past research on scarp geometry including slope, height, and length. Heatmaps indicate rock size spatial distributions on the fault scarp and surrounding topographic flats. Median grain size changes perpendicular to the fault scarp trace with the largest rocks exposed on and downslope from the scarp footwall. Correlation analyses of the segmented fault scarp illustrate the relationship between rock trait statistics and fault scarp geomorphology. Local fault scarp height correlates with median grain size (R2 of 0.6), the mean grain size of the largest rocks (R2 of 0.76), and the ratio of the number of small to large rocks (R2 of 0.40). The positive correlation (R2 of 0.81) between local fault scarp height and standard deviation of grain size suggests that rocks on a higher fault scarp are less well sorted. The correlation analysis between fault scarp height and rock orientation statistics supports a particle transportation model in which locally higher fault scarps have relatively more rocks with long axes parallel to fault scarp trace because rocks have a larger distance to roll and orient the long axes. Our work demonstrates a data-driven approach to geomorphology based on rock trait distributions, promising a greater understanding of fault scarp formation, as well as many other applications for which granulometry is an indicator of process.

physics.geo-ph↗

arXiv2019

Geomorphological Analysis Using Unpiloted Aircraft Systems, Structure from Motion, and Deep Learning

We present a pipeline for geomorphological analysis that uses structure from motion (SfM) and deep learning on close-range aerial imagery to estimate spatial distributions of rock traits (size, roundness, and orientation) along a tectonic fault scarp. The properties of the rocks on the fault scarp derive from the combination of initial volcanic fracturing and subsequent tectonic and geomorphic fracturing, and our pipeline allows scientists to leverage UAS-based imagery to gain a better understanding of such surface processes. We start by using SfM on aerial imagery to produce georeferenced orthomosaics and digital elevation models (DEM). A human expert then annotates rocks on a set of image tiles sampled from the orthomosaics, and these annotations are used to train a deep neural network to detect and segment individual rocks in the entire site. The extracted semantic information (rock masks) on large volumes of unlabeled, high-resolution SfM products allows subsequent structural analysis and shape descriptors to estimate rock size, roundness, and orientation. We present results of two experiments conducted along a fault scarp in the Volcanic Tablelands near Bishop, California. We conducted the first, proof-of-concept experiment with a DJI Phantom 4 Pro equipped with an RGB camera and inspected if elevation information assisted instance segmentation from RGB channels. Rock-trait histograms along and across the fault scarp were obtained with the neural network inference. In the second experiment, we deployed a hexrotor and a multispectral camera to produce a DEM and five spectral orthomosaics in red, green, blue, red edge, and near infrared. We focused on examining the effectiveness of different combinations of input channels in instance segmentation.

cs.RO↗